diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000021/group_0/index_tensor([6892, 6892])_group_0_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000021/group_0/index_tensor([6892, 6892])_group_0_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..dbc9628644b87637929a0f02b7dc87a6ac9291be --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000021/group_0/index_tensor([6892, 6892])_group_0_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "In an extreme slow-motion shot, the camera pans in a smooth and steady 360-degree arc, capturing, from a slightly upward angle, the decisive moment of a massive, crystal-clear sheet of heavy glass slowly tumbling and falling through the air. The entire scene is bathed in a chilling gray-blue hue, illuminated only by a cool, diffuse skylight cast from above, imbuing the scene with a quiet and melancholic atmosphere. The glass's edges are meticulously polished, refracting light as it falls, creating fleeting, tiny rainbows. When one of its sharp corners finally, almost silently and frozen in time, strikes the rough, cracked, and dust-covered concrete floor, a visible shockwave, like ripples on water, instantly spreads from the point of contact throughout the entire pane. Immediately, the glass's internal structure disintegrates, violently shattering in all directions into thousands of glittering fragments of varying sizes and shapes, stirring up a small circle of fine gray dust. Each fragment is like a carefully cut diamond, slowly spinning and flying in the air, its sharp facets constantly capturing and reflecting the cool light, creating a brief, brilliant dance of light full of destructive beauty. The whole process presents a poetic calmness that contrasts sharply with the violent impact.", + "seed": 11576, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000021/group_0/index_6892_seed_11576_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": -0.06839344752952456, + "vlm_logprob": -0.06839344752952456, + "vlm_yes_no_margin": 3.324999952316284, + "vlm_token_credit": -0.06839344752952456, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 in extreme slow-motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera pan in a smooth 360-degree arc from a slightly upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene bathed in a chilling gray-blue hue with cool diffuse skylight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a massive, crystal-clear sheet of heavy glass tumble and fall through the air before striking a rough, cracked, dust-covered concrete floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the glass shatter into thousands of glittering fragments that spin and fly while creating a visible shockwave ripple upon impact?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.14441950619220734 + ], + "opposite_logprob_tokens": [ + -2.1444194316864014 + ], + "logprob_reward": -0.14441950619220734, + "yes_no_margin_reward": 1.9999998807907104, + "token_credit_reward": -0.14441950619220734, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1448342204093933, + 0.14767944812774658, + 0.11880932003259659, + 0.09550121426582336, + 0.10552440583705902, + 0.13148316740989685, + 0.038308460265398026, + 0.10235045850276947, + 0.07534085214138031, + 0.04016845300793648 + ], + "credit_map": [ + 0.018815558403730392, + 0.019898775964975357, + 0.01952766254544258, + 0.014591051265597343, + 0.025714918971061707, + 0.018046151846647263, + 0.012866024859249592, + 0.015374084934592247, + 0.01178318914026022, + 0.017267510294914246, + 0.02153429388999939, + 0.012986461631953716, + 0.0226068627089262, + 0.02372409589588642, + 0.020349599421024323, + 0.017427437007427216, + 0.018807142972946167, + 0.016373250633478165, + 0.01502710860222578, + 0.012028906494379044, + 0.018802771344780922, + 0.01512052956968546, + 0.008531371131539345, + 0.01411823183298111, + 0.009496963582932949, + 0.010042868554592133, + 0.01717001385986805, + 0.013568203896284103, + 0.011622562073171139, + 0.010584784671664238, + 0.011472197249531746, + 0.011543624103069305, + 0.016311975196003914, + 0.012142420746386051, + 0.014438564889132977, + 0.010510940104722977, + 0.01429392397403717, + 0.01514813769608736, + 0.011347672902047634, + 0.011330779641866684, + 0.02075018361210823, + 0.022765453904867172, + 0.014704309403896332, + 0.01239415630698204, + 0.011822033673524857, + 0.015171949751675129, + 0.01179313100874424, + 0.02208193764090538, + 0.0050665466114878654, + 0.004663064610213041, + 0.00719464011490345, + 0.002308307448402047, + 0.0038019786588847637, + 0.004151946399360895, + 0.0032450559083372355, + 0.007876920513808727, + 0.007388602010905743, + 0.023049915209412575, + 0.013112070970237255, + 0.007503417786210775, + 0.011912438087165356, + 0.027375491335988045, + 0.005077880807220936, + 0.0069306460209190845, + 0.005217441823333502, + 0.007347970735281706, + 0.014312070794403553, + 0.010709294117987156, + 0.009039939381182194, + 0.007032179273664951, + 0.010574441403150558, + 0.011107510887086391, + 0.005208940710872412, + 0.008253894746303558, + 0.0037685290444642305, + 0.003545427229255438, + 0.006016023922711611, + 0.0054326835088431835, + 0.0033938090782612562, + 0.004549146629869938 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 5 + ], + "selected_reward": -0.14441950619220734 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 in extreme slow-motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera pan in a smooth 360-degree arc from a slightly upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene bathed in a chilling gray-blue hue with cool diffuse skylight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a massive, crystal-clear sheet of heavy glass tumble and fall through the air before striking a rough, cracked, dust-covered concrete floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the glass shatter into thousands of glittering fragments that spin and fly while creating a visible shockwave ripple upon impact?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.12708480656147003 + ], + "opposite_logprob_tokens": [ + -2.127084732055664 + ], + "logprob_reward": -0.12708480656147003, + "yes_no_margin_reward": 1.9999998807907104, + "token_credit_reward": -0.12708480656147003, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16936054825782776, + 0.17361950874328613, + 0.18622271716594696, + 0.0872313380241394, + 0.13228638470172882, + 0.04831749200820923, + 0.024662353098392487, + 0.07244404405355453, + 0.04039899632334709, + 0.06545661389827728 + ], + "credit_map": [ + 0.02290869504213333, + 0.01852748915553093, + 0.02841794118285179, + 0.01852826587855816, + 0.016450433060526848, + 0.020257947966456413, + 0.019367408007383347, + 0.024902373552322388, + 0.012942834757268429, + 0.01635492965579033, + 0.036390505731105804, + 0.01568426936864853, + 0.025753630325198174, + 0.02645176462829113, + 0.017759492620825768, + 0.0222820695489645, + 0.018029822036623955, + 0.01670672371983528, + 0.027292635291814804, + 0.019427992403507233, + 0.031483281403779984, + 0.030477730557322502, + 0.017044629901647568, + 0.025759896263480186, + 0.009073483757674694, + 0.009815963916480541, + 0.018039599061012268, + 0.010608229786157608, + 0.009928598999977112, + 0.009333563037216663, + 0.009332550689578056, + 0.011099351570010185, + 0.01869499869644642, + 0.03670625016093254, + 0.011842372827231884, + 0.011188250966370106, + 0.023495789617300034, + 0.014714223332703114, + 0.0064703915268182755, + 0.009174109436571598, + 0.006314065307378769, + 0.005977316293865442, + 0.006182088050991297, + 0.004071919247508049, + 0.004864518530666828, + 0.005188180133700371, + 0.004748563747853041, + 0.010970840230584145, + 0.003191394731402397, + 0.0032909817527979612, + 0.0035647808108478785, + 0.0021806773729622364, + 0.0033005194272845984, + 0.003017303068190813, + 0.0024833129718899727, + 0.0036333834286779165, + 0.006421474739909172, + 0.011995770037174225, + 0.009352927096188068, + 0.007802279200404882, + 0.007866175845265388, + 0.013761955313384533, + 0.005449399352073669, + 0.00979406014084816, + 0.003708573989570141, + 0.005353778600692749, + 0.006567682139575481, + 0.003913375549018383, + 0.005927765276283026, + 0.0036401874385774136, + 0.00433553010225296, + 0.006952103693038225, + 0.0061638145707547665, + 0.010689662769436836, + 0.006421948783099651, + 0.0049797347746789455, + 0.01134925801306963, + 0.009666655212640762, + 0.005784342531114817, + 0.010401197709143162 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 1, + 0 + ], + "selected_reward": -0.12708480656147003 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 in extreme slow-motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera pan in a smooth 360-degree arc from a slightly upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene bathed in a chilling gray-blue hue with cool diffuse skylight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a massive, crystal-clear sheet of heavy glass tumble and fall through the air before striking a rough, cracked, dust-covered concrete floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the glass shatter into thousands of glittering fragments that spin and fly while creating a visible shockwave ripple upon impact?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003623231779783964 + ], + "opposite_logprob_tokens": [ + -5.628623008728027 + ], + "logprob_reward": -0.003623231779783964, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.003623231779783964, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1756204515695572, + 0.17873714864253998, + 0.15358199179172516, + 0.06541768461465836, + 0.12985794246196747, + 0.04898440092802048, + 0.07715794444084167, + 0.10452058166265488, + 0.02398986928164959, + 0.04213191941380501 + ], + "credit_map": [ + 0.014466638676822186, + 0.024315517395734787, + 0.02470415271818638, + 0.017679333686828613, + 0.036114245653152466, + 0.022122278809547424, + 0.013288674876093864, + 0.02292962372303009, + 0.013432720676064491, + 0.017797524109482765, + 0.028666436672210693, + 0.015633955597877502, + 0.029954418540000916, + 0.029946258291602135, + 0.023800799623131752, + 0.019505053758621216, + 0.012954137288033962, + 0.013436386361718178, + 0.01898147165775299, + 0.012424112297594547, + 0.026521926745772362, + 0.03812661021947861, + 0.010240082629024982, + 0.020897276699543, + 0.005141925532370806, + 0.006008646916598082, + 0.015058094635605812, + 0.006305573042482138, + 0.009050902910530567, + 0.009605503641068935, + 0.004749520681798458, + 0.00949751865118742, + 0.023987989872694016, + 0.026723509654402733, + 0.012638803571462631, + 0.015988364815711975, + 0.016668546944856644, + 0.012736355885863304, + 0.011472455225884914, + 0.00964192021638155, + 0.007792401593178511, + 0.008227593265473843, + 0.00681759649887681, + 0.0036838275846093893, + 0.0037803957238793373, + 0.005268835928291082, + 0.004019110929220915, + 0.009394642896950245, + 0.009685441851615906, + 0.009375212714076042, + 0.01560295931994915, + 0.004462665878236294, + 0.00850777979940176, + 0.00865741167217493, + 0.0047603631392121315, + 0.016106121242046356, + 0.00814748089760542, + 0.020212063565850258, + 0.014257037080824375, + 0.0075272307731211185, + 0.010874403640627861, + 0.02555415965616703, + 0.005000330973416567, + 0.01294788345694542, + 0.0018927872879430652, + 0.003467425238341093, + 0.00509342597797513, + 0.002285531023517251, + 0.0034829843789339066, + 0.002767890691757202, + 0.0018706799019128084, + 0.0031291479244828224, + 0.0043927449733018875, + 0.006495455279946327, + 0.004142676014453173, + 0.003717048792168498, + 0.007096943445503712, + 0.00641527958214283, + 0.004782948177307844, + 0.005088824313133955 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 2 + ], + "selected_reward": -0.003623231779783964 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 in extreme slow-motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera pan in a smooth 360-degree arc from a slightly upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene bathed in a chilling gray-blue hue with cool diffuse skylight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a massive, crystal-clear sheet of heavy glass tumble and fall through the air before striking a rough, cracked, dust-covered concrete floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the glass shatter into thousands of glittering fragments that spin and fly while creating a visible shockwave ripple upon impact?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018190933391451836 + ], + "opposite_logprob_tokens": [ + -4.018190860748291 + ], + "logprob_reward": -0.018190933391451836, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018190933391451836, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1676245927810669, + 0.20807968080043793, + 0.1123875081539154, + 0.1849280744791031, + 0.07476600259542465, + 0.10368482023477554, + 0.04569825530052185, + 0.04388009011745453, + 0.029783587902784348, + 0.029167406260967255 + ], + "credit_map": [ + 0.022787073627114296, + 0.01965986005961895, + 0.02484779804944992, + 0.013416202738881111, + 0.02649860829114914, + 0.021753109991550446, + 0.016837282106280327, + 0.021824654191732407, + 0.016280192881822586, + 0.03648294135928154, + 0.03280818834900856, + 0.018787382170557976, + 0.02315574884414673, + 0.03275324031710625, + 0.021647943183779716, + 0.026164043694734573, + 0.009732786566019058, + 0.009651869535446167, + 0.01221455167979002, + 0.009845956228673458, + 0.020956173539161682, + 0.025239387527108192, + 0.009669648483395576, + 0.015077129006385803, + 0.015961818397045135, + 0.016792921349406242, + 0.04098345339298248, + 0.01845555752515793, + 0.022502215579152107, + 0.023797739297151566, + 0.018332690000534058, + 0.028101669624447823, + 0.0067992666736245155, + 0.01355099305510521, + 0.009574885480105877, + 0.007675327360630035, + 0.00729406950995326, + 0.014670629054307938, + 0.007725405972450972, + 0.00747542642056942, + 0.012788258492946625, + 0.010018464177846909, + 0.016795217990875244, + 0.005294754635542631, + 0.007951700128614902, + 0.0119142085313797, + 0.007071147672832012, + 0.03185106813907623, + 0.0058390614576637745, + 0.005417481996119022, + 0.008504115045070648, + 0.003281290177255869, + 0.005720408167690039, + 0.0052442955784499645, + 0.004248605575412512, + 0.007442997302860022, + 0.0038247706834226847, + 0.006979212164878845, + 0.008944381028413773, + 0.004102674778550863, + 0.006591332610696554, + 0.004445160739123821, + 0.0032957610674202442, + 0.0056967963464558125, + 0.001791996881365776, + 0.006433755159378052, + 0.003810975467786193, + 0.003087824210524559, + 0.005882681347429752, + 0.0026816362515091896, + 0.0029521079268306494, + 0.003142612287774682, + 0.002687990665435791, + 0.004969475790858269, + 0.004097312223166227, + 0.0029205798637121916, + 0.004267882090061903, + 0.003616519970819354, + 0.0032296322751790285, + 0.0033780147787183523 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 3, + 0 + ], + "selected_reward": -0.018190933391451836 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 in extreme slow-motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera pan in a smooth 360-degree arc from a slightly upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene bathed in a chilling gray-blue hue with cool diffuse skylight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a massive, crystal-clear sheet of heavy glass tumble and fall through the air before striking a rough, cracked, dust-covered concrete floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the glass shatter into thousands of glittering fragments that spin and fly while creating a visible shockwave ripple upon impact?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.048648759722709656 + ], + "opposite_logprob_tokens": [ + -3.0486488342285156 + ], + "logprob_reward": -0.048648759722709656, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.048648759722709656, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1288546323776245, + 0.047635626047849655, + 0.06603699177503586, + 0.03384658321738243, + 0.08546964079141617, + 0.0262666754424572, + 0.024660853669047356, + 0.5298593044281006, + 0.017452888190746307, + 0.039916832000017166 + ], + "credit_map": [ + 0.011398722417652607, + 0.024254178628325462, + 0.01868351548910141, + 0.010791398584842682, + 0.017745625227689743, + 0.016083506867289543, + 0.010866937227547169, + 0.019030753523111343, + 0.003326353384181857, + 0.009248106740415096, + 0.007497744634747505, + 0.006250301841646433, + 0.004183638375252485, + 0.007645843084901571, + 0.004212433472275734, + 0.005271206144243479, + 0.006572304293513298, + 0.008509672246873379, + 0.011570766568183899, + 0.007067008875310421, + 0.009010563604533672, + 0.010602756403386593, + 0.005658247042447329, + 0.007045675069093704, + 0.0038084376137703657, + 0.0035611570347100496, + 0.005013636779040098, + 0.0031786616891622543, + 0.003935062326490879, + 0.003571673994883895, + 0.0033193896524608135, + 0.007458564359694719, + 0.013745433650910854, + 0.021404104307293892, + 0.00823064986616373, + 0.008461964316666126, + 0.013158177025616169, + 0.00916715431958437, + 0.005857298616319895, + 0.005444861017167568, + 0.003657222492620349, + 0.0027592985425144434, + 0.003828273853287101, + 0.0027698592748492956, + 0.002383369952440262, + 0.003096991218626499, + 0.003477822756394744, + 0.004293839447200298, + 0.0032530499156564474, + 0.0029263074975460768, + 0.0035338865127414465, + 0.0030660582706332207, + 0.00282194628380239, + 0.0026469703298062086, + 0.0027375060599297285, + 0.003675126936286688, + 0.045103128999471664, + 0.08652301132678986, + 0.0838518738746643, + 0.037891361862421036, + 0.06011436879634857, + 0.10783182829618454, + 0.027066674083471298, + 0.08147703856229782, + 0.0015168795362114906, + 0.0034213236067444086, + 0.002044687746092677, + 0.0025426249485462904, + 0.0029673054814338684, + 0.0017830189317464828, + 0.0015884290914982557, + 0.001588617917150259, + 0.0038814449217170477, + 0.005908748600631952, + 0.0040272874757647514, + 0.0030805119313299656, + 0.007681123446673155, + 0.005955138709396124, + 0.004016000311821699, + 0.005366574507206678 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 4 + ], + "selected_reward": -0.048648759722709656 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In an extreme slow-motion shot, the camera pans in a smooth and steady 360-degree arc, capturing, from a slightly upward angle, the decisive moment of a massive, crystal-clear sheet of heavy glass slowly tumbling and falling through the air. The entire scene is bathed in a chilling gray-blue hue, illuminated only by a cool, diffuse skylight cast from above, imbuing the scene with a quiet and melancholic atmosphere. The glass's edges are meticulously polished, refracting light as it falls, creating fleeting, tiny rainbows. When one of its sharp corners finally, almost silently and frozen in time, strikes the rough, cracked, and dust-covered concrete floor, a visible shockwave, like ripples on water, instantly spreads from the point of contact throughout the entire pane. Immediately, the glass's internal structure disintegrates, violently shattering in all directions into thousands of glittering fragments of varying sizes and shapes, stirring up a small circle of fine gray dust. Each fragment is like a carefully cut diamond, slowly spinning and flying in the air, its sharp facets constantly capturing and reflecting the cool light, creating a brief, brilliant dance of light full of destructive beauty. The whole process presents a poetic calmness that contrasts sharply with the violent impact.", + "seed": 11577, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000021/group_0/index_6892_seed_11577_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": -0.057051450666040185, + "vlm_logprob": -0.057051450666040185, + "vlm_yes_no_margin": 3.975, + "vlm_token_credit": -0.057051450666040185, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 in extreme slow-motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera pan in a smooth 360-degree arc from a slightly upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene bathed in a chilling gray-blue hue with cool diffuse skylight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a massive, crystal-clear sheet of heavy glass tumble and fall through the air before striking a rough, cracked, dust-covered concrete floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the glass shatter into thousands of glittering fragments that spin and fly while creating a visible shockwave ripple upon impact?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03928142413496971 + ], + "opposite_logprob_tokens": [ + -3.7892813682556152 + ], + "logprob_reward": -0.03928142413496971, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.03928142413496971, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2097308486700058, + 0.21025802195072174, + 0.08012786507606506, + 0.18482370674610138, + 0.09972994774580002, + 0.037469830363988876, + 0.03674549609422684, + 0.08701994270086288, + 0.032568030059337616, + 0.021526360884308815 + ], + "credit_map": [ + 0.025216512382030487, + 0.04051179438829422, + 0.039769478142261505, + 0.024376271292567253, + 0.027770694345235825, + 0.01745656132698059, + 0.01918184943497181, + 0.015447662211954594, + 0.01856677047908306, + 0.030537229031324387, + 0.05271810665726662, + 0.02896486036479473, + 0.014650081284344196, + 0.022475233301520348, + 0.021018559113144875, + 0.021327156573534012, + 0.009166124276816845, + 0.01023042667657137, + 0.01443536952137947, + 0.008997890166938305, + 0.01227536890655756, + 0.010079871863126755, + 0.007995217107236385, + 0.006947587244212627, + 0.012450266629457474, + 0.019856149330735207, + 0.02791551873087883, + 0.02508017234504223, + 0.026979221031069756, + 0.02179115079343319, + 0.028197424486279488, + 0.022553790360689163, + 0.00837763212621212, + 0.012845824472606182, + 0.013534720055758953, + 0.022485418245196342, + 0.012399308383464813, + 0.008663772605359554, + 0.014295240864157677, + 0.007128025405108929, + 0.003315748879685998, + 0.00461776927113533, + 0.006375993601977825, + 0.005932834465056658, + 0.006100782658904791, + 0.0038618394173681736, + 0.0034264889545738697, + 0.003838371718302369, + 0.003336512017995119, + 0.003766455687582493, + 0.0072312988340854645, + 0.005611357279121876, + 0.004252939019352198, + 0.005049615167081356, + 0.0037313925568014383, + 0.003765922738239169, + 0.007650630548596382, + 0.009974688291549683, + 0.016190480440855026, + 0.013030503876507282, + 0.010141639038920403, + 0.012861405499279499, + 0.008206835947930813, + 0.008963756263256073, + 0.0029358388856053352, + 0.0037063283380120993, + 0.006074794568121433, + 0.00586328562349081, + 0.003988475538790226, + 0.0031134404707700014, + 0.0033938749693334103, + 0.0034919872414320707, + 0.0021287824492901564, + 0.002420972567051649, + 0.004045148380100727, + 0.003673495724797249, + 0.002565795788541436, + 0.001942208968102932, + 0.0026159179396927357, + 0.0021340367384254932 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 3 + ], + "selected_reward": -0.03928142413496971 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 in extreme slow-motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera pan in a smooth 360-degree arc from a slightly upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene bathed in a chilling gray-blue hue with cool diffuse skylight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a massive, crystal-clear sheet of heavy glass tumble and fall through the air before striking a rough, cracked, dust-covered concrete floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the glass shatter into thousands of glittering fragments that spin and fly while creating a visible shockwave ripple upon impact?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1270717829465866 + ], + "opposite_logprob_tokens": [ + -2.1270718574523926 + ], + "logprob_reward": -0.1270717829465866, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.1270717829465866, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10232183337211609, + 0.24915558099746704, + 0.1305413544178009, + 0.2585649788379669, + 0.12769418954849243, + 0.022563328966498375, + 0.02912795916199684, + 0.03980664908885956, + 0.02127678506076336, + 0.018947333097457886 + ], + "credit_map": [ + 0.01113413367420435, + 0.018395651131868362, + 0.016824347898364067, + 0.0167914517223835, + 0.013646324165165424, + 0.009919874370098114, + 0.009633596055209637, + 0.005976455751806498, + 0.020110271871089935, + 0.032902251929044724, + 0.05993657931685448, + 0.03223176300525665, + 0.01984294131398201, + 0.026585185900330544, + 0.023631738498806953, + 0.033914849162101746, + 0.009569738060235977, + 0.013935396447777748, + 0.019173897802829742, + 0.020142892375588417, + 0.0166928730905056, + 0.018786398693919182, + 0.018161876127123833, + 0.014078289270401001, + 0.023244915530085564, + 0.029576009139418602, + 0.029301006346940994, + 0.05158129706978798, + 0.036869656294584274, + 0.02744351513683796, + 0.03211529552936554, + 0.02843327820301056, + 0.010072892531752586, + 0.01505581196397543, + 0.022095711901783943, + 0.028489163145422935, + 0.015879139304161072, + 0.014122308231890202, + 0.014229193329811096, + 0.007749964948743582, + 0.0020346823148429394, + 0.0029591149650514126, + 0.0032927365973591805, + 0.003360931295901537, + 0.004156145732849836, + 0.002514451742172241, + 0.0026041772216558456, + 0.0016410874668508768, + 0.0025182708632200956, + 0.0033041329588741064, + 0.004211114253848791, + 0.003808346576988697, + 0.004306846763938665, + 0.004971067421138287, + 0.0030531533993780613, + 0.0029550264589488506, + 0.003463992616161704, + 0.004281798377633095, + 0.006924691144376993, + 0.006633346900343895, + 0.004328413400799036, + 0.005935585591942072, + 0.0038283010944724083, + 0.0044105201959609985, + 0.0018104652408510447, + 0.0023381474893540144, + 0.0032948029693216085, + 0.004047690890729427, + 0.0029439146164804697, + 0.0025001720059663057, + 0.0018229585839435458, + 0.0025186329148709774, + 0.0016031617997214198, + 0.0023600535932928324, + 0.0029554401990026236, + 0.0033452461939305067, + 0.0027285271789878607, + 0.0022372654639184475, + 0.0015627085231244564, + 0.0021549295634031296 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 1, + 2 + ], + "selected_reward": -0.1270717829465866 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 in extreme slow-motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera pan in a smooth 360-degree arc from a slightly upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene bathed in a chilling gray-blue hue with cool diffuse skylight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a massive, crystal-clear sheet of heavy glass tumble and fall through the air before striking a rough, cracked, dust-covered concrete floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the glass shatter into thousands of glittering fragments that spin and fly while creating a visible shockwave ripple upon impact?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0013390155509114265 + ], + "opposite_logprob_tokens": [ + -6.626338958740234 + ], + "logprob_reward": -0.0013390155509114265, + "yes_no_margin_reward": 6.625, + "token_credit_reward": -0.0013390155509114265, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15797719359397888, + 0.24496933817863464, + 0.15853393077850342, + 0.06951848417520523, + 0.04439130425453186, + 0.08094616234302521, + 0.04721401631832123, + 0.14372652769088745, + 0.024244436994194984, + 0.028478585183620453 + ], + "credit_map": [ + 0.02017544023692608, + 0.028068281710147858, + 0.024011798202991486, + 0.025883669033646584, + 0.018962953239679337, + 0.013886901549994946, + 0.012961308471858501, + 0.014026851393282413, + 0.02316170558333397, + 0.03715480864048004, + 0.06783349066972733, + 0.029675409197807312, + 0.01727522537112236, + 0.020969577133655548, + 0.021092846989631653, + 0.027806293219327927, + 0.020003970712423325, + 0.01827935129404068, + 0.026122618466615677, + 0.021118655800819397, + 0.018911760300397873, + 0.016944104805588722, + 0.01738552935421467, + 0.019767941907048225, + 0.008036903105676174, + 0.009472359903156757, + 0.01289432868361473, + 0.011615253984928131, + 0.008510963059961796, + 0.005931153893470764, + 0.007135894615203142, + 0.005921632517129183, + 0.005376175511628389, + 0.005201210733503103, + 0.006822030991315842, + 0.005731111858040094, + 0.005690861027687788, + 0.005272770766168833, + 0.00660689314827323, + 0.0036902506835758686, + 0.010422122664749622, + 0.01090964488685131, + 0.014903292991220951, + 0.009698689915239811, + 0.010601147077977657, + 0.013434900902211666, + 0.004933257587254047, + 0.0060431077145040035, + 0.0039717708714306355, + 0.005042413715273142, + 0.007580897770822048, + 0.005782166961580515, + 0.008924460038542747, + 0.008724742569029331, + 0.0032807933166623116, + 0.003906771540641785, + 0.015942441299557686, + 0.020441047847270966, + 0.021970896050333977, + 0.020935416221618652, + 0.02176015079021454, + 0.016804588958621025, + 0.013170967809855938, + 0.012701013125479221, + 0.001994403777644038, + 0.0029838550835847855, + 0.0036062400322407484, + 0.00530864717438817, + 0.002506854012608528, + 0.0022428748197853565, + 0.0029305412899702787, + 0.0026710203383117914, + 0.002595014637336135, + 0.0029361904598772526, + 0.006056747864931822, + 0.006033223122358322, + 0.0026295743882656097, + 0.0020271246321499348, + 0.0032390221022069454, + 0.002961687743663788 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 2, + 0 + ], + "selected_reward": -0.0013390155509114265 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 in extreme slow-motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera pan in a smooth 360-degree arc from a slightly upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene bathed in a chilling gray-blue hue with cool diffuse skylight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a massive, crystal-clear sheet of heavy glass tumble and fall through the air before striking a rough, cracked, dust-covered concrete floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the glass shatter into thousands of glittering fragments that spin and fly while creating a visible shockwave ripple upon impact?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0046377950347959995 + ], + "opposite_logprob_tokens": [ + -5.379637718200684 + ], + "logprob_reward": -0.0046377950347959995, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.0046377950347959995, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15186268091201782, + 0.24421802163124084, + 0.19688138365745544, + 0.11416438221931458, + 0.09937772154808044, + 0.03000367246568203, + 0.04953571781516075, + 0.0555637963116169, + 0.03252985700964928, + 0.025862718001008034 + ], + "credit_map": [ + 0.017283879220485687, + 0.02147781290113926, + 0.02651345543563366, + 0.03031584993004799, + 0.015817081555724144, + 0.012160899117588997, + 0.016606688499450684, + 0.011687013320624828, + 0.019068187102675438, + 0.03561585023999214, + 0.060422953218221664, + 0.027448734268546104, + 0.019890733063220978, + 0.031097158789634705, + 0.02663090080022812, + 0.024043522775173187, + 0.013730560429394245, + 0.019217416644096375, + 0.030043020844459534, + 0.023601561784744263, + 0.023868054151535034, + 0.031105872243642807, + 0.031505174934864044, + 0.023809734731912613, + 0.008245742879807949, + 0.013060711324214935, + 0.013871356844902039, + 0.022363608703017235, + 0.013737166300415993, + 0.01116873323917389, + 0.018940240144729614, + 0.012776828370988369, + 0.008243830874562263, + 0.011151459999382496, + 0.015077488496899605, + 0.021181168034672737, + 0.013425561599433422, + 0.012038353830575943, + 0.012116066180169582, + 0.006143799051642418, + 0.0024409168399870396, + 0.003417970845475793, + 0.005098801106214523, + 0.006091403774917126, + 0.00420468719676137, + 0.0027033693622797728, + 0.0034967113751918077, + 0.0025498138274997473, + 0.004018725827336311, + 0.004700516350567341, + 0.010191341862082481, + 0.009823581203818321, + 0.005494867917150259, + 0.005903309676796198, + 0.005409997887909412, + 0.003993377089500427, + 0.004264111164957285, + 0.0050467452965676785, + 0.010495313443243504, + 0.01108462829142809, + 0.006941583938896656, + 0.007304579485207796, + 0.005518474616110325, + 0.0049083661288022995, + 0.002886648289859295, + 0.0038758423179388046, + 0.006084142718464136, + 0.0059038191102445126, + 0.0041835857555270195, + 0.003587145823985338, + 0.0030723088420927525, + 0.0029363681096583605, + 0.0024967207573354244, + 0.0029763218481093645, + 0.005269245244562626, + 0.005465789698064327, + 0.002579947467893362, + 0.0020029752049595118, + 0.002659311518073082, + 0.002412407658994198 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 2, + 0 + ], + "selected_reward": -0.0046377950347959995 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 in extreme slow-motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera pan in a smooth 360-degree arc from a slightly upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene bathed in a chilling gray-blue hue with cool diffuse skylight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a massive, crystal-clear sheet of heavy glass tumble and fall through the air before striking a rough, cracked, dust-covered concrete floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the glass shatter into thousands of glittering fragments that spin and fly while creating a visible shockwave ripple upon impact?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11292723566293716 + ], + "opposite_logprob_tokens": [ + -2.237927198410034 + ], + "logprob_reward": -0.11292723566293716, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11292723566293716, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.30473098158836365, + 0.2188454419374466, + 0.07443063706159592, + 0.06045718491077423, + 0.08669060468673706, + 0.08006986230611801, + 0.03968695178627968, + 0.04611058905720711, + 0.04197762906551361, + 0.04700011387467384 + ], + "credit_map": [ + 0.03682328015565872, + 0.043778471648693085, + 0.06133180484175682, + 0.043040692806243896, + 0.0352790504693985, + 0.037355728447437286, + 0.02213144488632679, + 0.024990499019622803, + 0.01821637712419033, + 0.033429134637117386, + 0.05008545145392418, + 0.03150790557265282, + 0.017993291839957237, + 0.023400042206048965, + 0.020019838586449623, + 0.0241934172809124, + 0.007818158715963364, + 0.008565668016672134, + 0.010834640823304653, + 0.007865499705076218, + 0.00986004713922739, + 0.009903199970722198, + 0.010664445348083973, + 0.008918981067836285, + 0.007214131765067577, + 0.007409424055367708, + 0.009600362740457058, + 0.008152364753186703, + 0.006910175550729036, + 0.00798772368580103, + 0.00706485565751791, + 0.006118146702647209, + 0.007233705371618271, + 0.007578350603580475, + 0.011973974294960499, + 0.011684788390994072, + 0.010618459433317184, + 0.015618749894201756, + 0.011405427940189838, + 0.010577144101262093, + 0.006420592311769724, + 0.009789492934942245, + 0.012580504640936852, + 0.008946220390498638, + 0.016752423718571663, + 0.008576114661991596, + 0.009246105328202248, + 0.0077584064565598965, + 0.003339637303724885, + 0.004937959369271994, + 0.00662523927167058, + 0.006509119179099798, + 0.005972655490040779, + 0.005546771455556154, + 0.0034112853463739157, + 0.003344283439218998, + 0.003615324618294835, + 0.0042651137337088585, + 0.009797093458473682, + 0.00949268788099289, + 0.004518277011811733, + 0.00445935782045126, + 0.0045994785614311695, + 0.005363257136195898, + 0.003679528133943677, + 0.005174907855689526, + 0.007911049760878086, + 0.007566528394818306, + 0.00533377006649971, + 0.004537648055702448, + 0.003947036806493998, + 0.003827163251116872, + 0.0045714364387094975, + 0.0051511735655367374, + 0.008987349458038807, + 0.009381426498293877, + 0.005044478923082352, + 0.0043468656949698925, + 0.005278797820210457, + 0.004238590598106384 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 4 + ], + "selected_reward": -0.11292723566293716 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In an extreme slow-motion shot, the camera pans in a smooth and steady 360-degree arc, capturing, from a slightly upward angle, the decisive moment of a massive, crystal-clear sheet of heavy glass slowly tumbling and falling through the air. The entire scene is bathed in a chilling gray-blue hue, illuminated only by a cool, diffuse skylight cast from above, imbuing the scene with a quiet and melancholic atmosphere. The glass's edges are meticulously polished, refracting light as it falls, creating fleeting, tiny rainbows. When one of its sharp corners finally, almost silently and frozen in time, strikes the rough, cracked, and dust-covered concrete floor, a visible shockwave, like ripples on water, instantly spreads from the point of contact throughout the entire pane. Immediately, the glass's internal structure disintegrates, violently shattering in all directions into thousands of glittering fragments of varying sizes and shapes, stirring up a small circle of fine gray dust. Each fragment is like a carefully cut diamond, slowly spinning and flying in the air, its sharp facets constantly capturing and reflecting the cool light, creating a brief, brilliant dance of light full of destructive beauty. The whole process presents a poetic calmness that contrasts sharply with the violent impact.", + "seed": 11578, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000021/group_0/index_6892_seed_11578_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": -0.08412210890091956, + "vlm_logprob": -0.08412210890091956, + "vlm_yes_no_margin": 3.2249999761581423, + "vlm_token_credit": -0.08412210890091956, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 in extreme slow-motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera pan in a smooth 360-degree arc from a slightly upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene bathed in a chilling gray-blue hue with cool diffuse skylight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a massive, crystal-clear sheet of heavy glass tumble and fall through the air before striking a rough, cracked, dust-covered concrete floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the glass shatter into thousands of glittering fragments that spin and fly while creating a visible shockwave ripple upon impact?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.20142297446727753 + ], + "opposite_logprob_tokens": [ + -1.826422929763794 + ], + "logprob_reward": -0.20142297446727753, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.20142297446727753, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.25552332401275635, + 0.2298799604177475, + 0.12669755518436432, + 0.05369516462087631, + 0.04528966546058655, + 0.09319610893726349, + 0.028990760445594788, + 0.05590195208787918, + 0.0342383086681366, + 0.07658722251653671 + ], + "credit_map": [ + 0.02329111471772194, + 0.04833710193634033, + 0.03615310788154602, + 0.03138437867164612, + 0.03251584991812706, + 0.022591158747673035, + 0.024903371930122375, + 0.036347225308418274, + 0.020684391260147095, + 0.0382145531475544, + 0.04036865383386612, + 0.018591124564409256, + 0.04896888509392738, + 0.022876378148794174, + 0.018244881182909012, + 0.021931102499365807, + 0.010213104076683521, + 0.014789066277444363, + 0.013494766317307949, + 0.020990638062357903, + 0.017849860712885857, + 0.015366311185061932, + 0.022267717868089676, + 0.011726104654371738, + 0.005891794338822365, + 0.009574036113917828, + 0.007210993207991123, + 0.005588951054960489, + 0.00805790163576603, + 0.006832615006715059, + 0.004209279548376799, + 0.006329590920358896, + 0.004170910455286503, + 0.006414527539163828, + 0.007413564715534449, + 0.005751171614974737, + 0.006517298053950071, + 0.007073487155139446, + 0.004332476761192083, + 0.0036162289325147867, + 0.00579762551933527, + 0.0219479538500309, + 0.006312927696853876, + 0.004494518972933292, + 0.033577676862478256, + 0.01218167319893837, + 0.004108821507543325, + 0.004774908069521189, + 0.0027757741045206785, + 0.008623084984719753, + 0.0021272876765578985, + 0.0013569380389526486, + 0.0036419802345335484, + 0.006510632112622261, + 0.00214038765989244, + 0.0018146759830415249, + 0.005424537230283022, + 0.008204148150980473, + 0.0050180149264633656, + 0.003677504137158394, + 0.016543617472052574, + 0.009084831923246384, + 0.004212075844407082, + 0.0037372170481830835, + 0.003988670185208321, + 0.005524561740458012, + 0.0033496597316116095, + 0.0027205615770071745, + 0.006557785905897617, + 0.00555363018065691, + 0.0037486122455447912, + 0.0027948266360908747, + 0.003644478041678667, + 0.026297537609934807, + 0.003378651337698102, + 0.0032640057615935802, + 0.025920530781149864, + 0.006689330562949181, + 0.0036594003904610872, + 0.00373328966088593 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.20142297446727753 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 in extreme slow-motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera pan in a smooth 360-degree arc from a slightly upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene bathed in a chilling gray-blue hue with cool diffuse skylight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a massive, crystal-clear sheet of heavy glass tumble and fall through the air before striking a rough, cracked, dust-covered concrete floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the glass shatter into thousands of glittering fragments that spin and fly while creating a visible shockwave ripple upon impact?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.14285336434841156 + ], + "opposite_logprob_tokens": [ + -2.017853260040283 + ], + "logprob_reward": -0.14285336434841156, + "yes_no_margin_reward": 1.8749998807907104, + "token_credit_reward": -0.14285336434841156, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1525212675333023, + 0.08486819267272949, + 0.10982076078653336, + 0.1101161539554596, + 0.0551130473613739, + 0.06817284226417542, + 0.16546328365802765, + 0.11037226021289825, + 0.09638341516256332, + 0.0471687912940979 + ], + "credit_map": [ + 0.01665894128382206, + 0.03688397258520126, + 0.021852167323231697, + 0.018176916986703873, + 0.015368052758276463, + 0.013354799710214138, + 0.0143530098721385, + 0.01587340421974659, + 0.008325920440256596, + 0.009684991091489792, + 0.011221209540963173, + 0.007755178492516279, + 0.017192669212818146, + 0.011254465207457542, + 0.012207981199026108, + 0.0072257849387824535, + 0.010259747505187988, + 0.022095942869782448, + 0.008644766174256802, + 0.01869649812579155, + 0.015650661662220955, + 0.009668483398854733, + 0.015257840044796467, + 0.009546823799610138, + 0.008914312347769737, + 0.01999371498823166, + 0.012191125191748142, + 0.012575198896229267, + 0.01628100499510765, + 0.012038885615766048, + 0.01580071449279785, + 0.01232119183987379, + 0.005139941815286875, + 0.01075008511543274, + 0.00677754171192646, + 0.0055097369477152824, + 0.010324652306735516, + 0.005505421198904514, + 0.00502547062933445, + 0.006080200430005789, + 0.009145624004304409, + 0.010666796006262302, + 0.010253277607262135, + 0.007064632140100002, + 0.00956021063029766, + 0.00839094165712595, + 0.006788591854274273, + 0.006302766501903534, + 0.012884044088423252, + 0.05111004412174225, + 0.013492400757968426, + 0.009513718076050282, + 0.017675232142210007, + 0.044848646968603134, + 0.008359854109585285, + 0.007579340133816004, + 0.012822934426367283, + 0.01937011629343033, + 0.012400318868458271, + 0.009260721504688263, + 0.019845984876155853, + 0.018005825579166412, + 0.00981483981013298, + 0.008851507678627968, + 0.012769503518939018, + 0.017034323886036873, + 0.009106297045946121, + 0.011945610865950584, + 0.013844344764947891, + 0.012195507995784283, + 0.01139600295573473, + 0.008091830648481846, + 0.0021985291969031096, + 0.015755433589220047, + 0.0025773742236196995, + 0.002289151307195425, + 0.015633488073945045, + 0.004643318708986044, + 0.002014209283515811, + 0.002057286910712719 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 7 + ], + "selected_reward": -0.14285336434841156 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 in extreme slow-motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera pan in a smooth 360-degree arc from a slightly upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene bathed in a chilling gray-blue hue with cool diffuse skylight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a massive, crystal-clear sheet of heavy glass tumble and fall through the air before striking a rough, cracked, dust-covered concrete floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the glass shatter into thousands of glittering fragments that spin and fly while creating a visible shockwave ripple upon impact?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0031980343628674746 + ], + "opposite_logprob_tokens": [ + -5.753198146820068 + ], + "logprob_reward": -0.0031980343628674746, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.0031980343628674746, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07716329395771027, + 0.1250869184732437, + 0.3478720188140869, + 0.09369328618049622, + 0.053940389305353165, + 0.04577611759305, + 0.05277647450566292, + 0.047772541642189026, + 0.028861485421657562, + 0.1270574927330017 + ], + "credit_map": [ + 0.009292403236031532, + 0.01664242334663868, + 0.011115390807390213, + 0.011726860888302326, + 0.006850255653262138, + 0.008899746462702751, + 0.006321465130895376, + 0.006314751226454973, + 0.009074174799025059, + 0.02029242552816868, + 0.021693777292966843, + 0.012774212285876274, + 0.025549551472067833, + 0.00956976879388094, + 0.013641374185681343, + 0.012491638772189617, + 0.028112614527344704, + 0.06764587759971619, + 0.0305621474981308, + 0.0443793348968029, + 0.06754656881093979, + 0.0430368073284626, + 0.0373169369995594, + 0.029271749779582024, + 0.008584719151258469, + 0.015563368797302246, + 0.014363361522555351, + 0.013343257829546928, + 0.012889595702290535, + 0.009554819203913212, + 0.00920521467924118, + 0.01018894650042057, + 0.005253633018583059, + 0.008405677042901516, + 0.008781651966273785, + 0.005723803304135799, + 0.0073988353833556175, + 0.007677210029214621, + 0.005423584952950478, + 0.0052759950049221516, + 0.003049689345061779, + 0.009390074759721756, + 0.0044244262389838696, + 0.003317595226690173, + 0.014064796268939972, + 0.005037309601902962, + 0.0029808266554027796, + 0.0035114006604999304, + 0.005626876372843981, + 0.013500990346074104, + 0.0043569887056946754, + 0.002275667153298855, + 0.0137092974036932, + 0.007219606079161167, + 0.003111777128651738, + 0.0029752678237855434, + 0.005009994376450777, + 0.007034881506115198, + 0.005092451348900795, + 0.004045943263918161, + 0.012392349541187286, + 0.0073647224344313145, + 0.003861385863274336, + 0.002970813773572445, + 0.00356721761636436, + 0.004422674886882305, + 0.0039681727066636086, + 0.0030458017718046904, + 0.005577216390520334, + 0.003213366260752082, + 0.002348147565498948, + 0.002718888223171234, + 0.0064324140548706055, + 0.04071176052093506, + 0.007621751632541418, + 0.004542377311736345, + 0.03900688886642456, + 0.018003836274147034, + 0.0045441328547894955, + 0.006194334011524916 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 9, + 1 + ], + "selected_reward": -0.0031980343628674746 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 in extreme slow-motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera pan in a smooth 360-degree arc from a slightly upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene bathed in a chilling gray-blue hue with cool diffuse skylight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a massive, crystal-clear sheet of heavy glass tumble and fall through the air before striking a rough, cracked, dust-covered concrete floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the glass shatter into thousands of glittering fragments that spin and fly while creating a visible shockwave ripple upon impact?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018193157389760017 + ], + "opposite_logprob_tokens": [ + -4.018193244934082 + ], + "logprob_reward": -0.018193157389760017, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018193157389760017, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11725447326898575, + 0.1697685718536377, + 0.341079443693161, + 0.07100848853588104, + 0.05456014350056648, + 0.01714482717216015, + 0.04531913995742798, + 0.036516353487968445, + 0.0727190151810646, + 0.07462958991527557 + ], + "credit_map": [ + 0.01189118716865778, + 0.022026732563972473, + 0.016347136348485947, + 0.017219841480255127, + 0.012087710201740265, + 0.012457819655537605, + 0.013065381906926632, + 0.01215866208076477, + 0.014578092843294144, + 0.019063081592321396, + 0.036745335906744, + 0.01998113840818405, + 0.022908424958586693, + 0.015442842617630959, + 0.02142954058945179, + 0.019620127975940704, + 0.029528578743338585, + 0.06529136002063751, + 0.03393585979938507, + 0.04081615060567856, + 0.06582367420196533, + 0.04315546527504921, + 0.03197604790329933, + 0.03055230714380741, + 0.006085422355681658, + 0.010026849806308746, + 0.009189878590404987, + 0.008880152367055416, + 0.010126798413693905, + 0.008207087405025959, + 0.011317763477563858, + 0.007174536120146513, + 0.006120410282164812, + 0.009241240099072456, + 0.008624466136097908, + 0.005521794315427542, + 0.007397068198770285, + 0.007931149564683437, + 0.005294573958963156, + 0.0044294423423707485, + 0.0013633169000968337, + 0.00307634798809886, + 0.001502538681961596, + 0.001348867197521031, + 0.004879675805568695, + 0.0023225033655762672, + 0.00134802027605474, + 0.001303558237850666, + 0.0037419700529426336, + 0.015169653110206127, + 0.0028847239445894957, + 0.0027372867334634066, + 0.003913866356015205, + 0.012680497951805592, + 0.0023023802787065506, + 0.001888763508759439, + 0.004280528984963894, + 0.005159210413694382, + 0.003593460423871875, + 0.0027946345508098602, + 0.008890213444828987, + 0.006135717034339905, + 0.0032880865037441254, + 0.002374502830207348, + 0.007549222558736801, + 0.01137928944081068, + 0.006863776128739119, + 0.007992138154804707, + 0.010941548272967339, + 0.011783963069319725, + 0.009527319110929966, + 0.006681760307401419, + 0.005490160081535578, + 0.01793358474969864, + 0.006841254886239767, + 0.005589863285422325, + 0.01528860442340374, + 0.013091346248984337, + 0.004615353420376778, + 0.005779419559985399 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 1, + 0 + ], + "selected_reward": -0.018193157389760017 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 in extreme slow-motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera pan in a smooth 360-degree arc from a slightly upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene bathed in a chilling gray-blue hue with cool diffuse skylight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a massive, crystal-clear sheet of heavy glass tumble and fall through the air before striking a rough, cracked, dust-covered concrete floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the glass shatter into thousands of glittering fragments that spin and fly while creating a visible shockwave ripple upon impact?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.054943013936281204 + ], + "opposite_logprob_tokens": [ + -2.929943084716797 + ], + "logprob_reward": -0.054943013936281204, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.054943013936281204, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1871500462293625, + 0.11244577169418335, + 0.1555146425962448, + 0.07096469402313232, + 0.048893433064222336, + 0.03081461228430271, + 0.0665193572640419, + 0.07974149286746979, + 0.03423428535461426, + 0.21372169256210327 + ], + "credit_map": [ + 0.01713349111378193, + 0.03463909029960632, + 0.024425791576504707, + 0.020637616515159607, + 0.023109667003154755, + 0.016557933762669563, + 0.023848196491599083, + 0.026798265054821968, + 0.011698209680616856, + 0.020882563665509224, + 0.016196243464946747, + 0.00703927269205451, + 0.02680470608174801, + 0.011890536174178123, + 0.00795500073581934, + 0.009979241527616978, + 0.011589647270739079, + 0.015216046944260597, + 0.014492657035589218, + 0.031140992417931557, + 0.02128494158387184, + 0.017511142417788506, + 0.03107522241771221, + 0.013204002752900124, + 0.005932345986366272, + 0.015471058897674084, + 0.008895578794181347, + 0.00940097589045763, + 0.009531127288937569, + 0.007845966145396233, + 0.004769329447299242, + 0.009118311107158661, + 0.004411326255649328, + 0.007917727343738079, + 0.004972384311258793, + 0.004304802510887384, + 0.01363452710211277, + 0.005444605369120836, + 0.003320884658023715, + 0.0048871757462620735, + 0.002954116091132164, + 0.004811773542314768, + 0.004137815907597542, + 0.003025597659870982, + 0.005552324000746012, + 0.004754932131618261, + 0.0024867586325854063, + 0.003091292455792427, + 0.005745346657931805, + 0.023881081491708755, + 0.004562545567750931, + 0.0037858663126826286, + 0.004399020224809647, + 0.01862509921193123, + 0.0028340572025626898, + 0.0026863403618335724, + 0.00954418070614338, + 0.011638646945357323, + 0.010754606686532497, + 0.00794918555766344, + 0.014048978686332703, + 0.010844867676496506, + 0.008384313434362411, + 0.006576715502887964, + 0.003476038808003068, + 0.0056788865476846695, + 0.003108880715444684, + 0.003965181764215231, + 0.004841682966798544, + 0.00574647868052125, + 0.004231116268783808, + 0.003186020767316222, + 0.010489924810826778, + 0.07157491892576218, + 0.011122153140604496, + 0.008745876140892506, + 0.06872683763504028, + 0.023846127092838287, + 0.009601396508514881, + 0.009614458307623863 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 2 + ], + "selected_reward": -0.054943013936281204 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In an extreme slow-motion shot, the camera pans in a smooth and steady 360-degree arc, capturing, from a slightly upward angle, the decisive moment of a massive, crystal-clear sheet of heavy glass slowly tumbling and falling through the air. The entire scene is bathed in a chilling gray-blue hue, illuminated only by a cool, diffuse skylight cast from above, imbuing the scene with a quiet and melancholic atmosphere. The glass's edges are meticulously polished, refracting light as it falls, creating fleeting, tiny rainbows. When one of its sharp corners finally, almost silently and frozen in time, strikes the rough, cracked, and dust-covered concrete floor, a visible shockwave, like ripples on water, instantly spreads from the point of contact throughout the entire pane. Immediately, the glass's internal structure disintegrates, violently shattering in all directions into thousands of glittering fragments of varying sizes and shapes, stirring up a small circle of fine gray dust. Each fragment is like a carefully cut diamond, slowly spinning and flying in the air, its sharp facets constantly capturing and reflecting the cool light, creating a brief, brilliant dance of light full of destructive beauty. The whole process presents a poetic calmness that contrasts sharply with the violent impact.", + "seed": 11579, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000021/group_0/index_6892_seed_11579_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": -0.0870846651494503, + "vlm_logprob": -0.0870846651494503, + "vlm_yes_no_margin": 3.2, + "vlm_token_credit": -0.0870846651494503, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 in extreme slow-motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera pan in a smooth 360-degree arc from a slightly upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene bathed in a chilling gray-blue hue with cool diffuse skylight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a massive, crystal-clear sheet of heavy glass tumble and fall through the air before striking a rough, cracked, dust-covered concrete floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the glass shatter into thousands of glittering fragments that spin and fly while creating a visible shockwave ripple upon impact?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06450225412845612 + ], + "opposite_logprob_tokens": [ + -3.064502239227295 + ], + "logprob_reward": -0.06450225412845612, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.06450225412845612, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09161326289176941, + 0.05909290909767151, + 0.0768907442688942, + 0.05968962982296944, + 0.1911616027355194, + 0.050523541867733, + 0.09645101428031921, + 0.11493966728448868, + 0.21882419288158417, + 0.040813419967889786 + ], + "credit_map": [ + 0.007925600744783878, + 0.011133314110338688, + 0.013778848573565483, + 0.011493328958749771, + 0.012453493662178516, + 0.011146082542836666, + 0.015941012650728226, + 0.007741580251604319, + 0.008716423064470291, + 0.011843184009194374, + 0.011838849633932114, + 0.0037124445661902428, + 0.006317902356386185, + 0.005362542811781168, + 0.005152900703251362, + 0.0061486647464334965, + 0.011659287847578526, + 0.009810103103518486, + 0.012667846865952015, + 0.0069504654966294765, + 0.00845026783645153, + 0.01221423503011465, + 0.007191480603069067, + 0.007947059348225594, + 0.008784416131675243, + 0.008000795729458332, + 0.011351311579346657, + 0.006302367430180311, + 0.004789238795638084, + 0.009536700323224068, + 0.005491550080478191, + 0.005433251615613699, + 0.008975913748145103, + 0.06622187793254852, + 0.016499284654855728, + 0.008096491917967796, + 0.011593174189329147, + 0.0685347318649292, + 0.0058422088623046875, + 0.005397949833422899, + 0.005342455115169287, + 0.0068777091801166534, + 0.006479574367403984, + 0.006317481864243746, + 0.008512304164469242, + 0.006805608980357647, + 0.005429464858025312, + 0.004758941940963268, + 0.016745679080486298, + 0.021397842094302177, + 0.013161319307982922, + 0.012989810667932034, + 0.007797136437147856, + 0.013277419842779636, + 0.005073909647762775, + 0.0060078902170062065, + 0.012150273658335209, + 0.01229351107031107, + 0.015677543357014656, + 0.01939581148326397, + 0.012340489774942398, + 0.016562851145863533, + 0.01664871722459793, + 0.009870468638837337, + 0.0323832631111145, + 0.03337126225233078, + 0.02517849951982498, + 0.019501330330967903, + 0.03516309708356857, + 0.04635751247406006, + 0.014630880206823349, + 0.01223833579570055, + 0.00614288542419672, + 0.006319776177406311, + 0.0039017489179968834, + 0.0022860646713525057, + 0.0094737084582448, + 0.008551769889891148, + 0.0025218375958502293, + 0.0016156287165358663 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 4, + 7 + ], + "selected_reward": -0.06450225412845612 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 in extreme slow-motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera pan in a smooth 360-degree arc from a slightly upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene bathed in a chilling gray-blue hue with cool diffuse skylight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a massive, crystal-clear sheet of heavy glass tumble and fall through the air before striking a rough, cracked, dust-covered concrete floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the glass shatter into thousands of glittering fragments that spin and fly while creating a visible shockwave ripple upon impact?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.25211435556411743 + ], + "opposite_logprob_tokens": [ + -1.5021142959594727 + ], + "logprob_reward": -0.25211435556411743, + "yes_no_margin_reward": 1.25, + "token_credit_reward": -0.25211435556411743, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1163959875702858, + 0.17441138625144958, + 0.09359514713287354, + 0.0337948240339756, + 0.14995232224464417, + 0.18770408630371094, + 0.07634887099266052, + 0.03780791163444519, + 0.09782177209854126, + 0.032167643308639526 + ], + "credit_map": [ + 0.009997770190238953, + 0.011539201252162457, + 0.01497932430356741, + 0.020808415487408638, + 0.015838846564292908, + 0.013986124657094479, + 0.01970496028661728, + 0.009541352279484272, + 0.01993778720498085, + 0.022755011916160583, + 0.02712508849799633, + 0.019461078569293022, + 0.03081563115119934, + 0.021319054067134857, + 0.019534150138497353, + 0.013463608920574188, + 0.015687445178627968, + 0.010102959349751472, + 0.020787861198186874, + 0.00783043634146452, + 0.006852048449218273, + 0.01664138399064541, + 0.006195454858243465, + 0.009497568942606449, + 0.005677627865225077, + 0.004600265994668007, + 0.004486679099500179, + 0.0030943898018449545, + 0.0033619883470237255, + 0.005665648728609085, + 0.0032749008387327194, + 0.0036333247553557158, + 0.006545430980622768, + 0.061715591698884964, + 0.02251802198588848, + 0.005020052194595337, + 0.007405645679682493, + 0.040054623037576675, + 0.0029045536648482084, + 0.003788421396166086, + 0.01711234822869301, + 0.02224132977426052, + 0.02546769380569458, + 0.02702219784259796, + 0.03109964169561863, + 0.02422264777123928, + 0.01912780851125717, + 0.021410446614027023, + 0.011308525688946247, + 0.01119428500533104, + 0.011080359108746052, + 0.010540070943534374, + 0.007094447035342455, + 0.009312530979514122, + 0.0067605129443109035, + 0.009058145806193352, + 0.004218044690787792, + 0.004686844535171986, + 0.0043723853304982185, + 0.006008537486195564, + 0.004601154010742903, + 0.005164469126611948, + 0.005045776721090078, + 0.003710703691467643, + 0.010152840055525303, + 0.01335454173386097, + 0.01534014381468296, + 0.014920284040272236, + 0.013341301120817661, + 0.01623854972422123, + 0.006719129625707865, + 0.007754993159323931, + 0.0033126906491816044, + 0.00494436314329505, + 0.0025476280134171247, + 0.002268016804009676, + 0.009083939716219902, + 0.005055177491158247, + 0.0028500042390078306, + 0.0021058260463178158 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 1, + 4 + ], + "selected_reward": -0.25211435556411743 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 in extreme slow-motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera pan in a smooth 360-degree arc from a slightly upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene bathed in a chilling gray-blue hue with cool diffuse skylight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a massive, crystal-clear sheet of heavy glass tumble and fall through the air before striking a rough, cracked, dust-covered concrete floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the glass shatter into thousands of glittering fragments that spin and fly while creating a visible shockwave ripple upon impact?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0059540290385484695 + ], + "opposite_logprob_tokens": [ + -5.130954265594482 + ], + "logprob_reward": -0.0059540290385484695, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.0059540290385484695, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12197096645832062, + 0.12776300311088562, + 0.11241009831428528, + 0.06619390845298767, + 0.36436277627944946, + 0.04131427779793739, + 0.07867493480443954, + 0.01666353829205036, + 0.026816966012120247, + 0.043829482048749924 + ], + "credit_map": [ + 0.0077423653565347195, + 0.011381635442376137, + 0.013537696562707424, + 0.010421772487461567, + 0.027859585359692574, + 0.03461166471242905, + 0.00925306137651205, + 0.007163184229284525, + 0.01928822137415409, + 0.01947581022977829, + 0.02402186207473278, + 0.009689150378108025, + 0.01316444855183363, + 0.012312058359384537, + 0.014466725289821625, + 0.015344731509685516, + 0.021122511476278305, + 0.01546457502990961, + 0.01653389073908329, + 0.008269179612398148, + 0.008268680423498154, + 0.020225943997502327, + 0.007659705355763435, + 0.01486560981720686, + 0.00961360614746809, + 0.00708030303940177, + 0.01179613545536995, + 0.005005163606256247, + 0.006910673342645168, + 0.01600632630288601, + 0.004759234841912985, + 0.005022463388741016, + 0.019791696220636368, + 0.13694271445274353, + 0.03241041302680969, + 0.013720783405005932, + 0.01807897910475731, + 0.12434088438749313, + 0.008312960155308247, + 0.010764345526695251, + 0.004634354263544083, + 0.004780391696840525, + 0.008822396397590637, + 0.005243830382823944, + 0.0051521374844014645, + 0.004660139791667461, + 0.004228962119668722, + 0.0037920635659247637, + 0.011897115036845207, + 0.009805776178836823, + 0.011667197570204735, + 0.020269280299544334, + 0.005417515989392996, + 0.009592736139893532, + 0.005009866319596767, + 0.005015448201447725, + 0.002392514143139124, + 0.0019807915668934584, + 0.0019149825675413013, + 0.0022342081647366285, + 0.0021797246299684048, + 0.0019101339858025312, + 0.0023079842794686556, + 0.0017431980231776834, + 0.0032136805821210146, + 0.0033962256275117397, + 0.004110050853341818, + 0.00366679928265512, + 0.005311742424964905, + 0.003141779685392976, + 0.001796816592104733, + 0.002179871080443263, + 0.006568581331521273, + 0.006410520523786545, + 0.004016820341348648, + 0.0026515338104218245, + 0.010598045773804188, + 0.007555503398180008, + 0.00352838053368032, + 0.0025000981986522675 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 1, + 0 + ], + "selected_reward": -0.0059540290385484695 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 in extreme slow-motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera pan in a smooth 360-degree arc from a slightly upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene bathed in a chilling gray-blue hue with cool diffuse skylight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a massive, crystal-clear sheet of heavy glass tumble and fall through the air before striking a rough, cracked, dust-covered concrete floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the glass shatter into thousands of glittering fragments that spin and fly while creating a visible shockwave ripple upon impact?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012544417753815651 + ], + "opposite_logprob_tokens": [ + -4.387544631958008 + ], + "logprob_reward": -0.012544417753815651, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012544417753815651, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10089197009801865, + 0.0578208826482296, + 0.2853447496891022, + 0.03785797953605652, + 0.266047865152359, + 0.09466471523046494, + 0.05584998056292534, + 0.03224702179431915, + 0.0379127636551857, + 0.03136203810572624 + ], + "credit_map": [ + 0.0057315947487950325, + 0.010637818835675716, + 0.010215585120022297, + 0.013763022609055042, + 0.01775141805410385, + 0.018143996596336365, + 0.01746557094156742, + 0.007182958535850048, + 0.006240426562726498, + 0.007766274735331535, + 0.008385892026126385, + 0.006365353241562843, + 0.007227092981338501, + 0.006707728374749422, + 0.008074791170656681, + 0.007053324021399021, + 0.0492277555167675, + 0.032214708626270294, + 0.07310064882040024, + 0.017311589792370796, + 0.02089935727417469, + 0.0501352921128273, + 0.01426530722528696, + 0.028190089389681816, + 0.005472240038216114, + 0.005324109457433224, + 0.005599070340394974, + 0.0045362114906311035, + 0.0037417791318148375, + 0.005828311201184988, + 0.003547664498910308, + 0.00380859337747097, + 0.016579044982790947, + 0.07447841018438339, + 0.029879795387387276, + 0.01460779458284378, + 0.019865617156028748, + 0.08456209301948547, + 0.012262838892638683, + 0.013812270015478134, + 0.009062017314136028, + 0.013761903159320354, + 0.020206740126013756, + 0.010819390416145325, + 0.017214437946677208, + 0.012779650278389454, + 0.005803657695651054, + 0.0050169252790510654, + 0.007656970527023077, + 0.008216424845159054, + 0.012686493806540966, + 0.008873287588357925, + 0.004270206205546856, + 0.006192090455442667, + 0.003450485412031412, + 0.004504022654145956, + 0.0035853527951985598, + 0.004027533810585737, + 0.004765724763274193, + 0.004355339333415031, + 0.004287166520953178, + 0.003872690489515662, + 0.0033918481785804033, + 0.003961363807320595, + 0.004494378808885813, + 0.005445538554340601, + 0.004323996137827635, + 0.0046926140785217285, + 0.00577584421262145, + 0.007307186257094145, + 0.002730327658355236, + 0.003142877947539091, + 0.0031732069328427315, + 0.003781565697863698, + 0.004716127645224333, + 0.004616038873791695, + 0.005462885834276676, + 0.005094381980597973, + 0.002408955479040742, + 0.0021088754292577505 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 4, + 0 + ], + "selected_reward": -0.012544417753815651 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 in extreme slow-motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera pan in a smooth 360-degree arc from a slightly upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene bathed in a chilling gray-blue hue with cool diffuse skylight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a massive, crystal-clear sheet of heavy glass tumble and fall through the air before striking a rough, cracked, dust-covered concrete floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the glass shatter into thousands of glittering fragments that spin and fly while creating a visible shockwave ripple upon impact?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10030826926231384 + ], + "opposite_logprob_tokens": [ + -2.3503081798553467 + ], + "logprob_reward": -0.10030826926231384, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10030826926231384, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19796326756477356, + 0.08728425949811935, + 0.13124901056289673, + 0.021410344168543816, + 0.2367285192012787, + 0.06975670158863068, + 0.11542563140392303, + 0.04756719246506691, + 0.055182989686727524, + 0.03743208199739456 + ], + "credit_map": [ + 0.014614205807447433, + 0.02056911773979664, + 0.014126205816864967, + 0.01785592921078205, + 0.04763724282383919, + 0.053395263850688934, + 0.018467960879206657, + 0.011297355405986309, + 0.008882147260010242, + 0.011664538644254208, + 0.012206935323774815, + 0.009888840839266777, + 0.011288704350590706, + 0.008231326006352901, + 0.014083833433687687, + 0.011037942953407764, + 0.02159174717962742, + 0.016522763296961784, + 0.0347767136991024, + 0.007779501844197512, + 0.00882688257843256, + 0.022790564224123955, + 0.005842762533575296, + 0.013118070550262928, + 0.002936307108029723, + 0.0028129315469413996, + 0.0047488734126091, + 0.002545551396906376, + 0.0021474692039191723, + 0.002871630247682333, + 0.0018627186072990298, + 0.0014848660212010145, + 0.010914944112300873, + 0.06409195810556412, + 0.023534230887889862, + 0.012924975715577602, + 0.015511591918766499, + 0.08760470151901245, + 0.00936281867325306, + 0.012783308513462543, + 0.007638608105480671, + 0.010580706410109997, + 0.010313773527741432, + 0.009519709274172783, + 0.008693954907357693, + 0.008959394879639149, + 0.0067593976855278015, + 0.007291161455214024, + 0.020179152488708496, + 0.027545856311917305, + 0.015353425405919552, + 0.018108773976564407, + 0.012071709148585796, + 0.01189849991351366, + 0.004883809946477413, + 0.005384406074881554, + 0.0049813068471848965, + 0.005637107416987419, + 0.007318446412682533, + 0.005467013921588659, + 0.005521505139768124, + 0.005593433976173401, + 0.007902887649834156, + 0.005145492032170296, + 0.006374028045684099, + 0.0074057201854884624, + 0.006575811188668013, + 0.009397672489285469, + 0.009964993223547935, + 0.007690891623497009, + 0.00376518745906651, + 0.004008688032627106, + 0.003573888912796974, + 0.005288464482873678, + 0.0040496960282325745, + 0.0042950245551764965, + 0.009743081405758858, + 0.005422988906502724, + 0.0030877108220010996, + 0.001971227815374732 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 2 + ], + "selected_reward": -0.10030826926231384 + } + ], + "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/0000021/group_1/index_tensor([5218, 5218])_group_1_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000021/group_1/index_tensor([5218, 5218])_group_1_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..675a73d9d0fcecfe6af6b24ab229b8bb2986cc82 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000021/group_1/index_tensor([5218, 5218])_group_1_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "static, eye-level medium shot captures an intimate, portrait-style scene, centered on a young man with thoughtful, deep brown eyes and dark, damp hair styled into soft spikes. He turns toward a large window, the room's soft, warm light illuminating his pensive expression. The light sculpts the side of his face and captures the subtle moisture on his skin. He wears a simple, dark gray knit sweater, its soft texture adding to the cozy feel of the interior. Rain streaks vertically across the windowpane like a liquid curtain, distorting the blurred, rain-swept cityscape into a beautiful, cold-blue bokeh and twinkling, warm city lights. His gaze is distant and fixed, lost in quiet contemplation, his only movement a slow, deliberate blink. The warm, gentle light on the subject contrasts sharply with the cold, damp world outside, and combined with his profound stillness, creates an atmosphere of deep contemplation and tranquility.", + "seed": 9004, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000021/group_1/index_5218_seed_9004_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": -0.03856378719210625, + "vlm_logprob": -0.03856378719210625, + "vlm_yes_no_margin": 3.55, + "vlm_token_credit": -0.03856378719210625, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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, eye-level medium shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject a young man with dark, damp hair styled into soft spikes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the subject wearing a simple, dark gray knit sweater?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene show rain streaks on a windowpane with a blurred cityscape background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the subject perform only a slow, deliberate blink as his movement?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08494145423173904 + ], + "opposite_logprob_tokens": [ + -2.8349413871765137 + ], + "logprob_reward": -0.08494145423173904, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.08494145423173904, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2721821367740631, + 0.0946458950638771, + 0.07186046242713928, + 0.06473609805107117, + 0.05544266849756241, + 0.04628481715917587, + 0.1042621061205864, + 0.0810106173157692, + 0.08632458746433258, + 0.12325060367584229 + ], + "credit_map": [ + 0.03433844447135925, + 0.02375713735818863, + 0.030117692425847054, + 0.04429636895656586, + 0.025495294481515884, + 0.029076093807816505, + 0.037582725286483765, + 0.04751839116215706, + 0.018313176929950714, + 0.009327378123998642, + 0.012095771729946136, + 0.010399697348475456, + 0.011155711486935616, + 0.009872588329017162, + 0.013901123777031898, + 0.009580443613231182, + 0.008156382478773594, + 0.009178927168250084, + 0.0072603169828653336, + 0.009064125828444958, + 0.008655386976897717, + 0.007157531566917896, + 0.011064614169299603, + 0.0113231735303998, + 0.008356397971510887, + 0.006735533010214567, + 0.005950490944087505, + 0.009872323833405972, + 0.00851154513657093, + 0.006563506554812193, + 0.00990371871739626, + 0.008842580020427704, + 0.009513158351182938, + 0.007526449393481016, + 0.007351455744355917, + 0.006671941373497248, + 0.006517136003822088, + 0.004753210581839085, + 0.007934237830340862, + 0.005175077822059393, + 0.009531019255518913, + 0.004915822297334671, + 0.010111126117408276, + 0.003630514955148101, + 0.004218058194965124, + 0.004045093432068825, + 0.005230936221778393, + 0.004602246452122927, + 0.013552823103964329, + 0.013887842185795307, + 0.013353136368095875, + 0.01072856318205595, + 0.01486949436366558, + 0.014251554384827614, + 0.010166982188820839, + 0.013451711274683475, + 0.009916516952216625, + 0.013009520247578621, + 0.014872903935611248, + 0.008910007774829865, + 0.008342313580214977, + 0.008288505487143993, + 0.009503977373242378, + 0.008166875690221786, + 0.009143000468611717, + 0.014201650395989418, + 0.016952503472566605, + 0.009407814592123032, + 0.00878156628459692, + 0.009577821008861065, + 0.009043957106769085, + 0.009216275997459888, + 0.018875766545534134, + 0.01425141654908657, + 0.013046713545918465, + 0.01161646842956543, + 0.013532702811062336, + 0.017111200839281082, + 0.014278186485171318, + 0.02053814008831978 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 6 + ], + "selected_reward": -0.08494145423173904 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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, eye-level medium shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject a young man with dark, damp hair styled into soft spikes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the subject wearing a simple, dark gray knit sweater?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene show rain streaks on a windowpane with a blurred cityscape background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the subject perform only a slow, deliberate blink as his movement?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008654467761516571 + ], + "opposite_logprob_tokens": [ + -4.758654594421387 + ], + "logprob_reward": -0.008654467761516571, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008654467761516571, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.31585606932640076, + 0.09846297651529312, + 0.12882786989212036, + 0.08717534691095352, + 0.07322108745574951, + 0.07205245643854141, + 0.0757458508014679, + 0.054107870906591415, + 0.04754605144262314, + 0.04700436070561409 + ], + "credit_map": [ + 0.023547710850834846, + 0.03504912555217743, + 0.019569845870137215, + 0.03861046954989433, + 0.031223352998495102, + 0.046203117817640305, + 0.03038058429956436, + 0.09127189218997955, + 0.016662171110510826, + 0.010294659994542599, + 0.01623491197824478, + 0.008162898011505604, + 0.012527395971119404, + 0.00973263755440712, + 0.01352898683398962, + 0.011319329030811787, + 0.023782407864928246, + 0.017332449555397034, + 0.01577630452811718, + 0.013806297443807125, + 0.010889413766562939, + 0.01371103897690773, + 0.01435876078903675, + 0.019171223044395447, + 0.008481282740831375, + 0.013220428489148617, + 0.009197649545967579, + 0.012007726356387138, + 0.008594521321356297, + 0.009632926434278488, + 0.01023056823760271, + 0.015810249373316765, + 0.013033328577876091, + 0.008147460408508778, + 0.012587198987603188, + 0.006330381613224745, + 0.006105014588683844, + 0.007740223780274391, + 0.00904901884496212, + 0.010228468105196953, + 0.012933307327330112, + 0.009400635957717896, + 0.01175480242818594, + 0.006894200574606657, + 0.006646797060966492, + 0.006693877745419741, + 0.007335239555686712, + 0.010393600910902023, + 0.009210155345499516, + 0.010581363923847675, + 0.013429079204797745, + 0.008525238372385502, + 0.008389648050069809, + 0.00837089866399765, + 0.008156482130289078, + 0.00908299908041954, + 0.007212087977677584, + 0.005541777703911066, + 0.011584474705159664, + 0.004750577732920647, + 0.00552225299179554, + 0.007678473833948374, + 0.0051162331365048885, + 0.00670200027525425, + 0.005821185186505318, + 0.005390848498791456, + 0.007715635932981968, + 0.004466989543288946, + 0.00554285291582346, + 0.00692306412383914, + 0.005109534598886967, + 0.006575944367796183, + 0.005833703093230724, + 0.006151491776108742, + 0.005125875119119883, + 0.0037136413156986237, + 0.005170013289898634, + 0.0065535432659089565, + 0.005363190080970526, + 0.009092909283936024 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 1 + ], + "selected_reward": -0.008654467761516571 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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, eye-level medium shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject a young man with dark, damp hair styled into soft spikes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the subject wearing a simple, dark gray knit sweater?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene show rain streaks on a windowpane with a blurred cityscape background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the subject perform only a slow, deliberate blink as his movement?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029799478128552437 + ], + "opposite_logprob_tokens": [ + -3.529799461364746 + ], + "logprob_reward": -0.029799478128552437, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029799478128552437, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2277439832687378, + 0.23285315930843353, + 0.12993577122688293, + 0.04572037234902382, + 0.07938048988580704, + 0.12557078897953033, + 0.046406809240579605, + 0.05076606571674347, + 0.03088083676993847, + 0.030741697177290916 + ], + "credit_map": [ + 0.021153833717107773, + 0.026407824829220772, + 0.03012029640376568, + 0.044014908373355865, + 0.025851765647530556, + 0.021724320948123932, + 0.028107518330216408, + 0.030363531783223152, + 0.029846079647541046, + 0.02160758152604103, + 0.031475674360990524, + 0.03406471014022827, + 0.024093780666589737, + 0.023378686979413033, + 0.031103670597076416, + 0.03728296980261803, + 0.0231460053473711, + 0.014791539870202541, + 0.01721802167594433, + 0.017587531358003616, + 0.011982046999037266, + 0.011875570751726627, + 0.015031121671199799, + 0.018303941935300827, + 0.004406612366437912, + 0.003688110737130046, + 0.0052915094420313835, + 0.004427568521350622, + 0.004188207909464836, + 0.006940264254808426, + 0.004980280064046383, + 0.01179781835526228, + 0.01070055179297924, + 0.00780056556686759, + 0.012660917825996876, + 0.0077860550954937935, + 0.00937775056809187, + 0.008477662689983845, + 0.010035509243607521, + 0.012541478499770164, + 0.011195109225809574, + 0.01419759914278984, + 0.012640854343771935, + 0.011456714943051338, + 0.014651650562882423, + 0.016838684678077698, + 0.014444186352193356, + 0.03014599345624447, + 0.0053305719047784805, + 0.006371324881911278, + 0.006951135583221912, + 0.006065725814551115, + 0.006210715509951115, + 0.004429674241691828, + 0.005929726175963879, + 0.0051179323345422745, + 0.006431530229747295, + 0.005453381687402725, + 0.01037915050983429, + 0.005722550675272942, + 0.0047765374183654785, + 0.006236097775399685, + 0.005610709078609943, + 0.0061561111360788345, + 0.003577794414013624, + 0.0038640624843537807, + 0.0047786966897547245, + 0.0032760913018137217, + 0.0032380344346165657, + 0.0028593384195119143, + 0.004252873361110687, + 0.005033944733440876, + 0.0037411407101899385, + 0.004144333302974701, + 0.0027713284362107515, + 0.0027902356814593077, + 0.0033662300556898117, + 0.0032813872676342726, + 0.004347405396401882, + 0.006299634464085102 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 2 + ], + "selected_reward": -0.029799478128552437 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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, eye-level medium shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject a young man with dark, damp hair styled into soft spikes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the subject wearing a simple, dark gray knit sweater?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene show rain streaks on a windowpane with a blurred cityscape background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the subject perform only a slow, deliberate blink as his movement?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026360591873526573 + ], + "opposite_logprob_tokens": [ + -3.651360511779785 + ], + "logprob_reward": -0.026360591873526573, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026360591873526573, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14691630005836487, + 0.07865742594003677, + 0.2699415981769562, + 0.03108936920762062, + 0.20208929479122162, + 0.17867082357406616, + 0.027654508128762245, + 0.020978175103664398, + 0.016529591754078865, + 0.02747289277613163 + ], + "credit_map": [ + 0.020058730617165565, + 0.01306808553636074, + 0.016709383577108383, + 0.01322979386895895, + 0.012257586233317852, + 0.02070658840239048, + 0.012731182388961315, + 0.0381549596786499, + 0.013539710082113743, + 0.008378857746720314, + 0.011184655129909515, + 0.007104384247213602, + 0.008471298962831497, + 0.009428871795535088, + 0.00950254313647747, + 0.011047105304896832, + 0.05407090485095978, + 0.025712423026561737, + 0.042583923786878586, + 0.03123716451227665, + 0.024470599368214607, + 0.027431951835751534, + 0.025058289989829063, + 0.03937632963061333, + 0.005631173495203257, + 0.0036205053329467773, + 0.0037164618261158466, + 0.0043715136125683784, + 0.0034795161336660385, + 0.003443429246544838, + 0.0032728833612054586, + 0.003553885966539383, + 0.03567792847752571, + 0.025827981531620026, + 0.041257504373788834, + 0.02219805121421814, + 0.01994304358959198, + 0.01709982380270958, + 0.022139819338917732, + 0.017945146188139915, + 0.031224191188812256, + 0.01999320462346077, + 0.031924642622470856, + 0.01929948665201664, + 0.018541011959314346, + 0.015510395169258118, + 0.020287342369556427, + 0.021890537813305855, + 0.003453024197369814, + 0.0032741613686084747, + 0.004944424610584974, + 0.003047313541173935, + 0.002647069515660405, + 0.003461576532572508, + 0.0029585836455225945, + 0.003868355415761471, + 0.0027853455394506454, + 0.0024077650159597397, + 0.004299406427890062, + 0.002420635661110282, + 0.002344043692573905, + 0.0021945324260741472, + 0.0021464601159095764, + 0.0023799845948815346, + 0.002156962640583515, + 0.002452909480780363, + 0.0024582608602941036, + 0.0018829947803169489, + 0.0018243050435557961, + 0.0019166944548487663, + 0.0021945443004369736, + 0.00164292030967772, + 0.003822922008112073, + 0.0034736311063170433, + 0.0039635105058550835, + 0.0027355363126844168, + 0.002673997776582837, + 0.0036948940251022577, + 0.0029604455921798944, + 0.0041479552164673805 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 4, + 5 + ], + "selected_reward": -0.026360591873526573 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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, eye-level medium shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject a young man with dark, damp hair styled into soft spikes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the subject wearing a simple, dark gray knit sweater?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene show rain streaks on a windowpane with a blurred cityscape background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the subject perform only a slow, deliberate blink as his movement?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04306294396519661 + ], + "opposite_logprob_tokens": [ + -3.168062925338745 + ], + "logprob_reward": -0.04306294396519661, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04306294396519661, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3023383617401123, + 0.11339309066534042, + 0.0873074010014534, + 0.08673569560050964, + 0.06912041455507278, + 0.08325321972370148, + 0.06039230898022652, + 0.044048093259334564, + 0.06812455505132675, + 0.08528682589530945 + ], + "credit_map": [ + 0.04325750842690468, + 0.03216690570116043, + 0.0431516170501709, + 0.03426477685570717, + 0.03526225686073303, + 0.02881311997771263, + 0.040358006954193115, + 0.04506419226527214, + 0.014597931876778603, + 0.015550130046904087, + 0.012573419138789177, + 0.018903570249676704, + 0.014826594851911068, + 0.009855221956968307, + 0.015466215088963509, + 0.011620016768574715, + 0.01668570376932621, + 0.010168765671551228, + 0.008890504948794842, + 0.012350681237876415, + 0.007897313684225082, + 0.010037466883659363, + 0.009145202115178108, + 0.012131765484809875, + 0.008705309592187405, + 0.010572675615549088, + 0.0058693913742899895, + 0.00903546717017889, + 0.01147692371159792, + 0.00934691447764635, + 0.01642206497490406, + 0.015306956134736538, + 0.01097993552684784, + 0.008177208714187145, + 0.008410162292420864, + 0.008072592318058014, + 0.00887432973831892, + 0.006219227332621813, + 0.009670390747487545, + 0.008716579526662827, + 0.013240681029856205, + 0.012243302538990974, + 0.012551204301416874, + 0.00755834486335516, + 0.011209264397621155, + 0.008348057977855206, + 0.009029163978993893, + 0.009073211811482906, + 0.008284851908683777, + 0.007072650361806154, + 0.010304995812475681, + 0.006082920357584953, + 0.006820292212069035, + 0.007099388167262077, + 0.006787911057472229, + 0.007939303293824196, + 0.0053597092628479, + 0.0069096945226192474, + 0.007501136511564255, + 0.004556288477033377, + 0.00530160591006279, + 0.004006772302091122, + 0.004906217567622662, + 0.005506673827767372, + 0.008384964428842068, + 0.007120347581803799, + 0.015804780647158623, + 0.005984839983284473, + 0.007668201811611652, + 0.007742564659565687, + 0.007660737261176109, + 0.0077581265941262245, + 0.010425178334116936, + 0.013384331949055195, + 0.009598195552825928, + 0.010181691497564316, + 0.01173403114080429, + 0.008420387282967567, + 0.014172342605888844, + 0.007370670326054096 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.04306294396519661 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, eye-level medium shot captures an intimate, portrait-style scene, centered on a young man with thoughtful, deep brown eyes and dark, damp hair styled into soft spikes. He turns toward a large window, the room's soft, warm light illuminating his pensive expression. The light sculpts the side of his face and captures the subtle moisture on his skin. He wears a simple, dark gray knit sweater, its soft texture adding to the cozy feel of the interior. Rain streaks vertically across the windowpane like a liquid curtain, distorting the blurred, rain-swept cityscape into a beautiful, cold-blue bokeh and twinkling, warm city lights. His gaze is distant and fixed, lost in quiet contemplation, his only movement a slow, deliberate blink. The warm, gentle light on the subject contrasts sharply with the cold, damp world outside, and combined with his profound stillness, creates an atmosphere of deep contemplation and tranquility.", + "seed": 9005, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000021/group_1/index_5218_seed_9005_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": -0.07130904085934162, + "vlm_logprob": -0.07130904085934162, + "vlm_yes_no_margin": 3.075, + "vlm_token_credit": -0.07130904085934162, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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, eye-level medium shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject a young man with dark, damp hair styled into soft spikes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the subject wearing a simple, dark gray knit sweater?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene show rain streaks on a windowpane with a blurred cityscape background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the subject perform only a slow, deliberate blink as his movement?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08178271353244781 + ], + "opposite_logprob_tokens": [ + -3.456782817840576 + ], + "logprob_reward": -0.08178271353244781, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.08178271353244781, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2000749111175537, + 0.04895080253481865, + 0.04891880974173546, + 0.11035512387752533, + 0.2645586431026459, + 0.08908648788928986, + 0.07075250893831253, + 0.049664657562971115, + 0.0708882212638855, + 0.04674984887242317 + ], + "credit_map": [ + 0.021772025153040886, + 0.023079827427864075, + 0.03031875006854534, + 0.02304753288626671, + 0.014599548652768135, + 0.02100837044417858, + 0.017474422231316566, + 0.048774417489767075, + 0.007260262966156006, + 0.005848519969731569, + 0.006063993088901043, + 0.007628445513546467, + 0.006299962755292654, + 0.00535717373713851, + 0.0043092090636491776, + 0.0061832331120967865, + 0.00606969790533185, + 0.007179236970841885, + 0.006160857621580362, + 0.009000012651085854, + 0.0045957076363265514, + 0.0039808787405490875, + 0.003929777070879936, + 0.008002636954188347, + 0.010882883332669735, + 0.010162057355046272, + 0.011841867119073868, + 0.013669891282916069, + 0.007282385602593422, + 0.011104238219559193, + 0.018035970628261566, + 0.027375828474760056, + 0.049905888736248016, + 0.026375889778137207, + 0.05346785485744476, + 0.030767161399126053, + 0.017565378919243813, + 0.023248279467225075, + 0.022388234734535217, + 0.04083994776010513, + 0.007158797234296799, + 0.010529917664825916, + 0.009772064164280891, + 0.008598013781011105, + 0.011299076490104198, + 0.009502769447863102, + 0.01176035962998867, + 0.020465483888983727, + 0.012897208333015442, + 0.006597072351723909, + 0.0130052100867033, + 0.00921318493783474, + 0.004820889327675104, + 0.005625608377158642, + 0.009147316217422485, + 0.009446009993553162, + 0.0035867958795279264, + 0.006467535626143217, + 0.00434362655505538, + 0.005867194849997759, + 0.007081954274326563, + 0.004851414822041988, + 0.007993675768375397, + 0.009472455829381943, + 0.006124461069703102, + 0.013765247538685799, + 0.007214088924229145, + 0.010152316652238369, + 0.013944906182587147, + 0.005467040929943323, + 0.007779145147651434, + 0.006441006902605295, + 0.006038283929228783, + 0.004205403849482536, + 0.007355577312409878, + 0.005953563377261162, + 0.004645461216568947, + 0.00458518834784627, + 0.006187821738421917, + 0.00777854211628437 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 3 + ], + "selected_reward": -0.08178271353244781 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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, eye-level medium shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject a young man with dark, damp hair styled into soft spikes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the subject wearing a simple, dark gray knit sweater?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene show rain streaks on a windowpane with a blurred cityscape background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the subject perform only a slow, deliberate blink as his movement?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014217277988791466 + ], + "opposite_logprob_tokens": [ + -4.264217376708984 + ], + "logprob_reward": -0.014217277988791466, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014217277988791466, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5338855385780334, + 0.1130666732788086, + 0.06250637024641037, + 0.05074207857251167, + 0.07203682512044907, + 0.053890567272901535, + 0.0364765040576458, + 0.019812414422631264, + 0.023911088705062866, + 0.033671941608190536 + ], + "credit_map": [ + 0.05258248373866081, + 0.04717777669429779, + 0.0829479843378067, + 0.04764271900057793, + 0.04743604734539986, + 0.06885047256946564, + 0.03603248670697212, + 0.1512155681848526, + 0.013523988425731659, + 0.018965719267725945, + 0.008215031586587429, + 0.016053711995482445, + 0.012909921817481518, + 0.011144665069878101, + 0.01380448043346405, + 0.01844916306436062, + 0.005308022256940603, + 0.005636300891637802, + 0.0041503263637423515, + 0.007775267586112022, + 0.004498970229178667, + 0.008400964550673962, + 0.007035250775516033, + 0.019701266661286354, + 0.010740913450717926, + 0.00446219090372324, + 0.010624316520988941, + 0.005757727660238743, + 0.004333227872848511, + 0.00395079143345356, + 0.006361950654536486, + 0.004510962404310703, + 0.006556289736181498, + 0.0063231573440134525, + 0.00957863125950098, + 0.007988782599568367, + 0.009918374940752983, + 0.007278770674020052, + 0.010501105338335037, + 0.013891712762415409, + 0.006872245110571384, + 0.005382933653891087, + 0.007624604273587465, + 0.005152298137545586, + 0.004979745019227266, + 0.0059952386654913425, + 0.007351913955062628, + 0.010531592182815075, + 0.007250494323670864, + 0.0040157088078558445, + 0.006782082840800285, + 0.005980912130326033, + 0.0028321321588009596, + 0.0024825362488627434, + 0.004430324770510197, + 0.002702311845496297, + 0.0021372917108237743, + 0.002981728408485651, + 0.0018945157062262297, + 0.002718784147873521, + 0.0033705078531056643, + 0.001670581172220409, + 0.00324358232319355, + 0.0017954224022105336, + 0.002821615431457758, + 0.0033463938161730766, + 0.0029276448767632246, + 0.0032843947410583496, + 0.004118864890187979, + 0.00208073016256094, + 0.0026527701411396265, + 0.0026786753442138433, + 0.0026513116899877787, + 0.002835317049175501, + 0.0029238974675536156, + 0.004209422972053289, + 0.0030293972231447697, + 0.0043041701428592205, + 0.0041165901347994804, + 0.009601833298802376 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 4 + ], + "selected_reward": -0.014217277988791466 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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, eye-level medium shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject a young man with dark, damp hair styled into soft spikes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the subject wearing a simple, dark gray knit sweater?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene show rain streaks on a windowpane with a blurred cityscape background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the subject perform only a slow, deliberate blink as his movement?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.062042236328125 + ], + "opposite_logprob_tokens": [ + -2.812042236328125 + ], + "logprob_reward": -0.062042236328125, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.062042236328125, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12565511465072632, + 0.0887630432844162, + 0.04150997847318649, + 0.11139754205942154, + 0.28753700852394104, + 0.2545746862888336, + 0.016427433118224144, + 0.02585701271891594, + 0.03474796563386917, + 0.013530202209949493 + ], + "credit_map": [ + 0.013778873719274998, + 0.014250771142542362, + 0.02113637886941433, + 0.012816579081118107, + 0.012239163741469383, + 0.012668543495237827, + 0.011518855579197407, + 0.027245940640568733, + 0.008891184814274311, + 0.009060624055564404, + 0.006769238505512476, + 0.011706712655723095, + 0.0073471227660775185, + 0.009574733674526215, + 0.01564055308699608, + 0.019772877916693687, + 0.004128728993237019, + 0.0032542741391807795, + 0.0031474269926548004, + 0.004709395579993725, + 0.0033127067144960165, + 0.005027608014643192, + 0.007927087135612965, + 0.010002750903367996, + 0.01460585743188858, + 0.009607704356312752, + 0.013370418921113014, + 0.013767923228442669, + 0.010650058276951313, + 0.011084423400461674, + 0.017914921045303345, + 0.020396234467625618, + 0.02599179744720459, + 0.02557447925209999, + 0.03561006858944893, + 0.030927039682865143, + 0.03716963902115822, + 0.028907639905810356, + 0.04811767861247063, + 0.055238667875528336, + 0.02216442860662937, + 0.02265675738453865, + 0.032943032681941986, + 0.022765284404158592, + 0.022398125380277634, + 0.031545743346214294, + 0.03773006424307823, + 0.06237123906612396, + 0.003936876077204943, + 0.0012525964993983507, + 0.0035984429996460676, + 0.001579930423758924, + 0.001006747712381184, + 0.0016677960520610213, + 0.0009912805398926139, + 0.0023937635123729706, + 0.0023326240479946136, + 0.0033809072338044643, + 0.0028225649148225784, + 0.0031178926583379507, + 0.0027463368605822325, + 0.002605251269415021, + 0.004224283620715141, + 0.004627150949090719, + 0.003409851109609008, + 0.003908097743988037, + 0.005074925255030394, + 0.0037818425334990025, + 0.003509208559989929, + 0.0036488736514002085, + 0.004517716355621815, + 0.006897451356053352, + 0.0011331986170262098, + 0.0015231958823278546, + 0.0016114501049742103, + 0.0015754909254610538, + 0.001811363734304905, + 0.0012673453893512487, + 0.0019579448271542788, + 0.002650212962180376 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 5, + 0 + ], + "selected_reward": -0.062042236328125 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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, eye-level medium shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject a young man with dark, damp hair styled into soft spikes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the subject wearing a simple, dark gray knit sweater?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene show rain streaks on a windowpane with a blurred cityscape background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the subject perform only a slow, deliberate blink as his movement?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.16039592027664185 + ], + "opposite_logprob_tokens": [ + -1.910395860671997 + ], + "logprob_reward": -0.16039592027664185, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.16039592027664185, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3355981111526489, + 0.1437695026397705, + 0.19020411372184753, + 0.026145348325371742, + 0.03842570632696152, + 0.019371334463357925, + 0.07560615986585617, + 0.020715806633234024, + 0.017261328175663948, + 0.1329026073217392 + ], + "credit_map": [ + 0.03661220893263817, + 0.027708405628800392, + 0.053888946771621704, + 0.026431474834680557, + 0.02288641780614853, + 0.043392304331064224, + 0.02078775316476822, + 0.10389060527086258, + 0.016609234735369682, + 0.012635265477001667, + 0.017555339261889458, + 0.01374204270541668, + 0.011803112924098969, + 0.01664666086435318, + 0.018572267144918442, + 0.036205582320690155, + 0.015228244476020336, + 0.014408332295715809, + 0.01839357614517212, + 0.016952885314822197, + 0.012872792780399323, + 0.02608172781765461, + 0.01967189833521843, + 0.06659464538097382, + 0.00497416453436017, + 0.0028938709292560816, + 0.005532071925699711, + 0.003038843162357807, + 0.0018047834746539593, + 0.002263677306473255, + 0.0020535392686724663, + 0.0035843958612531424, + 0.005420840345323086, + 0.004630336072295904, + 0.006932270713150501, + 0.004807187709957361, + 0.0029850821010768414, + 0.0035171022173017263, + 0.004033297766000032, + 0.0060995896346867085, + 0.0027049765922129154, + 0.002413128735497594, + 0.0029430838767439127, + 0.0031050487887114286, + 0.0017521347617730498, + 0.0019961067009717226, + 0.002225469332188368, + 0.0022313871886581182, + 0.010763636790215969, + 0.005673521663993597, + 0.01007209625095129, + 0.006735576782375574, + 0.004820867441594601, + 0.009440453723073006, + 0.010439878329634666, + 0.017660127952694893, + 0.002190307481214404, + 0.0021406267769634724, + 0.002952049020677805, + 0.0017842288361862302, + 0.0013170352904126048, + 0.0026455277111381292, + 0.002042127773165703, + 0.005643903743475676, + 0.001718144048936665, + 0.002826808486133814, + 0.002171183004975319, + 0.002479282673448324, + 0.0024944727774709463, + 0.0014130899216979742, + 0.001855137525126338, + 0.0023032096214592457, + 0.011588122695684433, + 0.009744040668010712, + 0.012605163268744946, + 0.014919444918632507, + 0.010437586344778538, + 0.017689527943730354, + 0.019566727802157402, + 0.03635198622941971 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 1 + ], + "selected_reward": -0.16039592027664185 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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, eye-level medium shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject a young man with dark, damp hair styled into soft spikes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the subject wearing a simple, dark gray knit sweater?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene show rain streaks on a windowpane with a blurred cityscape background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the subject perform only a slow, deliberate blink as his movement?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03810705617070198 + ], + "opposite_logprob_tokens": [ + -3.28810715675354 + ], + "logprob_reward": -0.03810705617070198, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03810705617070198, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1722850352525711, + 0.1207752525806427, + 0.07017675042152405, + 0.027938177809119225, + 0.04363246262073517, + 0.06279973685741425, + 0.33650603890419006, + 0.05379082262516022, + 0.05244924873113632, + 0.059646494686603546 + ], + "credit_map": [ + 0.017981799319386482, + 0.03781189024448395, + 0.021264981478452682, + 0.027346163988113403, + 0.014966784045100212, + 0.013264640234410763, + 0.020990166813135147, + 0.018658610060811043, + 0.017136581242084503, + 0.018171196803450584, + 0.016529163345694542, + 0.011083177290856838, + 0.01320561207830906, + 0.009805644862353802, + 0.011551952920854092, + 0.023291923105716705, + 0.006711028516292572, + 0.007008712738752365, + 0.00928693450987339, + 0.007579715922474861, + 0.004410471301525831, + 0.008222343400120735, + 0.006191317457705736, + 0.02076622284948826, + 0.0034869713708758354, + 0.0032658781856298447, + 0.0028640124946832657, + 0.0033582004252821207, + 0.003587738610804081, + 0.0025564483366906643, + 0.0033728836569935083, + 0.005446044262498617, + 0.004565396346151829, + 0.003973347134888172, + 0.006172599270939827, + 0.005319662857800722, + 0.005358314607292414, + 0.004303315654397011, + 0.006498084869235754, + 0.00744174188002944, + 0.00824953243136406, + 0.007226577959954739, + 0.008402823470532894, + 0.006239931099116802, + 0.0064470525830984116, + 0.006671012379229069, + 0.007186626549810171, + 0.012376182712614536, + 0.0854129046201706, + 0.028544193133711815, + 0.07648781687021255, + 0.048825304955244064, + 0.02221585251390934, + 0.025990163907408714, + 0.0200578011572361, + 0.028971999883651733, + 0.006793583743274212, + 0.010014955885708332, + 0.004817171487957239, + 0.007517436519265175, + 0.0097221489995718, + 0.003996737767010927, + 0.006354769226163626, + 0.004574017133563757, + 0.005899116862565279, + 0.010174792259931564, + 0.005454191938042641, + 0.007393200881779194, + 0.0097197275608778, + 0.004022176843136549, + 0.004870383534580469, + 0.00491566164419055, + 0.00710071437060833, + 0.006141296587884426, + 0.007202169392257929, + 0.008877594955265522, + 0.008048472926020622, + 0.00642239348962903, + 0.008588017895817757, + 0.007265831343829632 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 1 + ], + "selected_reward": -0.03810705617070198 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, eye-level medium shot captures an intimate, portrait-style scene, centered on a young man with thoughtful, deep brown eyes and dark, damp hair styled into soft spikes. He turns toward a large window, the room's soft, warm light illuminating his pensive expression. The light sculpts the side of his face and captures the subtle moisture on his skin. He wears a simple, dark gray knit sweater, its soft texture adding to the cozy feel of the interior. Rain streaks vertically across the windowpane like a liquid curtain, distorting the blurred, rain-swept cityscape into a beautiful, cold-blue bokeh and twinkling, warm city lights. His gaze is distant and fixed, lost in quiet contemplation, his only movement a slow, deliberate blink. The warm, gentle light on the subject contrasts sharply with the cold, damp world outside, and combined with his profound stillness, creates an atmosphere of deep contemplation and tranquility.", + "seed": 9006, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000021/group_1/index_5218_seed_9006_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": -0.0261099960654974, + "vlm_logprob": -0.0261099960654974, + "vlm_yes_no_margin": 4.05, + "vlm_token_credit": -0.0261099960654974, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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, eye-level medium shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject a young man with dark, damp hair styled into soft spikes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the subject wearing a simple, dark gray knit sweater?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene show rain streaks on a windowpane with a blurred cityscape background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the subject perform only a slow, deliberate blink as his movement?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05616845190525055 + ], + "opposite_logprob_tokens": [ + -4.056168556213379 + ], + "logprob_reward": -0.05616845190525055, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.05616845190525055, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.8149113059043884, + 0.06713822484016418, + 0.018695633858442307, + 0.014611751772463322, + 0.010195696726441383, + 0.010296138934791088, + 0.013708557933568954, + 0.01749858260154724, + 0.012659519910812378, + 0.02028452791273594 + ], + "credit_map": [ + 0.1782340556383133, + 0.015927854925394058, + 0.3270778954029083, + 0.02564430981874466, + 0.0820486918091774, + 0.11103089898824692, + 0.040665462613105774, + 0.03428220748901367, + 0.012102135457098484, + 0.0020948133897036314, + 0.02358155883848667, + 0.0035429154522717, + 0.006547007709741592, + 0.011016550473868847, + 0.004145991988480091, + 0.004107252694666386, + 0.002496593166142702, + 0.0015247579431161284, + 0.002693169517442584, + 0.0018615747103467584, + 0.003454851917922497, + 0.0018745522247627378, + 0.002633675467222929, + 0.002156456932425499, + 0.0021339720115065575, + 0.0009938346920534968, + 0.002434865338727832, + 0.0015337209915742278, + 0.0027624655049294233, + 0.0009455482359044254, + 0.0027388380840420723, + 0.0010685065062716603, + 0.0013636212097480893, + 0.0007413371349684894, + 0.0010611129691824317, + 0.0010553185129538178, + 0.0024013826623559, + 0.0009109835955314338, + 0.001429099473170936, + 0.0012328403536230326, + 0.001355266198515892, + 0.0008387612760998309, + 0.0012342672562226653, + 0.0012134347343817353, + 0.0015260479412972927, + 0.0008957569370977581, + 0.0017921434482559562, + 0.0014404612593352795, + 0.002448422135785222, + 0.0009221776272170246, + 0.0027273192536085844, + 0.0012574620777741075, + 0.0023453913163393736, + 0.0010649641044437885, + 0.0013755845138803124, + 0.001567235914990306, + 0.002530706347897649, + 0.0011060505639761686, + 0.00378811196424067, + 0.002026139758527279, + 0.002750988118350506, + 0.00135546806268394, + 0.0026172734797000885, + 0.0013238429091870785, + 0.0016295085661113262, + 0.0010562109528109431, + 0.0016214505303651094, + 0.0016334481770172715, + 0.0023633348755538464, + 0.0011706209043040872, + 0.0015124910278245807, + 0.0016724542947486043, + 0.002937287325039506, + 0.0019347533816471696, + 0.0029475740157067776, + 0.0018048599595203996, + 0.00416593998670578, + 0.0018422332359477878, + 0.003223154926672578, + 0.0014287254307419062 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 9 + ], + "selected_reward": -0.05616845190525055 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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, eye-level medium shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject a young man with dark, damp hair styled into soft spikes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the subject wearing a simple, dark gray knit sweater?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene show rain streaks on a windowpane with a blurred cityscape background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the subject perform only a slow, deliberate blink as his movement?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008658958598971367 + ], + "opposite_logprob_tokens": [ + -4.7586588859558105 + ], + "logprob_reward": -0.008658958598971367, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008658958598971367, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.8666422963142395, + 0.05251259729266167, + 0.02241603285074234, + 0.0173143669962883, + 0.0074857990257442, + 0.00511134322732687, + 0.005678437650203705, + 0.010225176811218262, + 0.004295075312256813, + 0.008318876847624779 + ], + "credit_map": [ + 0.1893261969089508, + 0.01773705519735813, + 0.3459341526031494, + 0.02858303301036358, + 0.08694666624069214, + 0.11991393566131592, + 0.042469050735235214, + 0.035732124000787735, + 0.010936624370515347, + 0.0015017539262771606, + 0.019660070538520813, + 0.002312342869117856, + 0.005483119748532772, + 0.007143528666347265, + 0.003152610035613179, + 0.0023225389886647463, + 0.002652849769219756, + 0.0018632630817592144, + 0.0024765110574662685, + 0.0018284006509929895, + 0.006190502550452948, + 0.002138583455234766, + 0.0036149192601442337, + 0.0016509982524439692, + 0.001776610966771841, + 0.001499595702625811, + 0.0020962997805327177, + 0.0015697205672040582, + 0.004953118972480297, + 0.001243428559973836, + 0.002986806444823742, + 0.0011887841392308474, + 0.000930633454117924, + 0.0006097634322941303, + 0.0008659030427224934, + 0.000944696192163974, + 0.0013459831243380904, + 0.0007320362492464483, + 0.0010863711358979344, + 0.0009704112308099866, + 0.0005132066435180604, + 0.0004041314241476357, + 0.0005768255796283484, + 0.00047847680980339646, + 0.0012844217708334327, + 0.0004197743837721646, + 0.0008975142263807356, + 0.000536991807166487, + 0.0005120640271343291, + 0.000443664233898744, + 0.0006534691201522946, + 0.0004928219714201987, + 0.0015195368323475122, + 0.00048722518840804696, + 0.0009984393836930394, + 0.0005712161655537784, + 0.0013275975361466408, + 0.0007903644000180066, + 0.0018415139056742191, + 0.0010114265605807304, + 0.00220259977504611, + 0.0009183511137962341, + 0.001619097776710987, + 0.0005142249865457416, + 0.0005609869840554893, + 0.00030335644260048866, + 0.0006307113217189908, + 0.0005896333022974432, + 0.0005936644156463444, + 0.0004382025508675724, + 0.0005282546626403928, + 0.0006502649048343301, + 0.000995638663880527, + 0.0009262780658900738, + 0.0010115536861121655, + 0.0006708899163641036, + 0.00176912359893322, + 0.0007153001497499645, + 0.0014395283069461584, + 0.0007905639940872788 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.008658958598971367 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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, eye-level medium shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject a young man with dark, damp hair styled into soft spikes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the subject wearing a simple, dark gray knit sweater?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene show rain streaks on a windowpane with a blurred cityscape background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the subject perform only a slow, deliberate blink as his movement?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01607270911335945 + ], + "opposite_logprob_tokens": [ + -4.141072750091553 + ], + "logprob_reward": -0.01607270911335945, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.01607270911335945, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2914291024208069, + 0.11154482513666153, + 0.0941336378455162, + 0.0817224308848381, + 0.10098351538181305, + 0.061174824833869934, + 0.03406650573015213, + 0.08885233104228973, + 0.04193795844912529, + 0.09415485709905624 + ], + "credit_map": [ + 0.060596056282520294, + 0.012829813174903393, + 0.08954622596502304, + 0.018496204167604446, + 0.032617006450891495, + 0.022041238844394684, + 0.03086451068520546, + 0.024438057094812393, + 0.012647468596696854, + 0.006968238390982151, + 0.016053395345807076, + 0.011138794012367725, + 0.022404562681913376, + 0.010784835554659367, + 0.02168351225554943, + 0.009864016436040401, + 0.012526636943221092, + 0.009058239869773388, + 0.01086469180881977, + 0.010367249138653278, + 0.02206752821803093, + 0.00793370045721531, + 0.014597149565815926, + 0.006718446500599384, + 0.009453782811760902, + 0.007297447416931391, + 0.010970791801810265, + 0.007819251157343388, + 0.022056441754102707, + 0.00628282967954874, + 0.012295378372073174, + 0.005546506494283676, + 0.01080460473895073, + 0.006615551188588142, + 0.009700802154839039, + 0.015026802197098732, + 0.017910782247781754, + 0.01020592451095581, + 0.01722056046128273, + 0.013498494401574135, + 0.008300887420773506, + 0.0038862829096615314, + 0.007371900603175163, + 0.007188085466623306, + 0.012762237340211868, + 0.004939565900713205, + 0.008882525376975536, + 0.007843341678380966, + 0.0039043459109961987, + 0.0026916638016700745, + 0.0038093160837888718, + 0.004078988451510668, + 0.006833299528807402, + 0.003019335912540555, + 0.005733831319957972, + 0.003995722159743309, + 0.016604945063591003, + 0.006414292845875025, + 0.021813662722706795, + 0.011250464245676994, + 0.009841457940638065, + 0.007562709506601095, + 0.008281736634671688, + 0.007083062082529068, + 0.008377676829695702, + 0.003521651728078723, + 0.007439206819981337, + 0.004937184043228626, + 0.004865598864853382, + 0.003828390035778284, + 0.0044408561661839485, + 0.004527391400188208, + 0.018997423350811005, + 0.007014071103185415, + 0.01582450419664383, + 0.008302494883537292, + 0.012965344823896885, + 0.01306243147701025, + 0.009902402758598328, + 0.008086183108389378 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 4 + ], + "selected_reward": -0.01607270911335945 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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, eye-level medium shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject a young man with dark, damp hair styled into soft spikes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the subject wearing a simple, dark gray knit sweater?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene show rain streaks on a windowpane with a blurred cityscape background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the subject perform only a slow, deliberate blink as his movement?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02329878881573677 + ], + "opposite_logprob_tokens": [ + -3.773298740386963 + ], + "logprob_reward": -0.02329878881573677, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.02329878881573677, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.7934906482696533, + 0.012342802248895168, + 0.023224307224154472, + 0.052803535014390945, + 0.032029349356889725, + 0.020788943395018578, + 0.006430173758417368, + 0.028400516137480736, + 0.02124730497598648, + 0.009242500178515911 + ], + "credit_map": [ + 0.1744605451822281, + 0.016627077013254166, + 0.3165854215621948, + 0.027180897071957588, + 0.07905802875757217, + 0.10690408945083618, + 0.04026823118329048, + 0.03240622207522392, + 0.0018513401737436652, + 0.0007879705517552793, + 0.0019762206356972456, + 0.0018315217457711697, + 0.001789972186088562, + 0.0009497298742644489, + 0.0018670769641175866, + 0.0012889691861346364, + 0.0032571132760494947, + 0.0021062600426375866, + 0.003271406516432762, + 0.0036169840022921562, + 0.0032654646784067154, + 0.002118578879162669, + 0.003473106538876891, + 0.0021153916604816914, + 0.011969223618507385, + 0.003933926112949848, + 0.012001696042716503, + 0.006646374240517616, + 0.005631478037685156, + 0.0035793986171483994, + 0.0052360896952450275, + 0.003805344458669424, + 0.00791429728269577, + 0.002090698340907693, + 0.004752150736749172, + 0.003060665912926197, + 0.0045942808501422405, + 0.0027525536715984344, + 0.0032945263665169477, + 0.0035701743327081203, + 0.0038299388252198696, + 0.0013228405732661486, + 0.0023762870114296675, + 0.0014474651543423533, + 0.0040271892212331295, + 0.001891136053018272, + 0.002272883430123329, + 0.003621202427893877, + 0.0007271363283507526, + 0.000497671018820256, + 0.0009142435155808926, + 0.0007429402321577072, + 0.0011665301863104105, + 0.000569040363188833, + 0.0009816196979954839, + 0.0008309919503517449, + 0.004801443312317133, + 0.001798116136342287, + 0.006567489355802536, + 0.0031863513868302107, + 0.0035993284545838833, + 0.002285328460857272, + 0.0038752953987568617, + 0.0022871624678373337, + 0.00395711325109005, + 0.0014080919791013002, + 0.0039665913209319115, + 0.002753682667389512, + 0.002602128777652979, + 0.0020220610313117504, + 0.0022627408616244793, + 0.002274893457069993, + 0.001430296222679317, + 0.0008094831136986613, + 0.0013599375961348414, + 0.0008112229406833649, + 0.001464269938878715, + 0.0011878954246640205, + 0.0013166227145120502, + 0.0008627715287730098 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 4 + ], + "selected_reward": -0.02329878881573677 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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, eye-level medium shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject a young man with dark, damp hair styled into soft spikes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the subject wearing a simple, dark gray knit sweater?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene show rain streaks on a windowpane with a blurred cityscape background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the subject perform only a slow, deliberate blink as his movement?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026351071894168854 + ], + "opposite_logprob_tokens": [ + -3.651350975036621 + ], + "logprob_reward": -0.026351071894168854, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026351071894168854, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.7604201436042786, + 0.031188249588012695, + 0.019329745322465897, + 0.019823702052235603, + 0.05609773471951485, + 0.04707099124789238, + 0.026548780500888824, + 0.015567973256111145, + 0.011666106060147285, + 0.012286562472581863 + ], + "credit_map": [ + 0.17057055234909058, + 0.016441309824585915, + 0.30569103360176086, + 0.02515239641070366, + 0.07619993388652802, + 0.0987749695777893, + 0.038061946630477905, + 0.02952798642218113, + 0.0036615903954952955, + 0.0021103532053530216, + 0.004690695088356733, + 0.003114951541647315, + 0.006062531843781471, + 0.0039460319094359875, + 0.0048795719631016254, + 0.0027225245721638203, + 0.0019833152182400227, + 0.0014471843605861068, + 0.002196952933445573, + 0.0023276852443814278, + 0.002834449987858534, + 0.001936147571541369, + 0.00416957214474678, + 0.0024344390258193016, + 0.0037901608739048243, + 0.0011669351952150464, + 0.0034206577111035585, + 0.0020012555178254843, + 0.003217406338080764, + 0.0015816129744052887, + 0.0027098895516246557, + 0.0019357845885679126, + 0.006604909896850586, + 0.004241543356329203, + 0.005675753112882376, + 0.005616160109639168, + 0.014192928560078144, + 0.004714673385024071, + 0.009885529987514019, + 0.005166236776858568, + 0.0038982275873422623, + 0.004159666132181883, + 0.004408408887684345, + 0.0050371321849524975, + 0.011786060407757759, + 0.0043518380261957645, + 0.00863045547157526, + 0.004799200221896172, + 0.003466814523562789, + 0.0018345598364248872, + 0.004056289792060852, + 0.002542695729061961, + 0.004557169042527676, + 0.002411429537460208, + 0.0037424343172460794, + 0.003937385510653257, + 0.001921234535984695, + 0.0014529438922181726, + 0.001625223783776164, + 0.0015167482197284698, + 0.0030707966070622206, + 0.0014745164662599564, + 0.002822144189849496, + 0.0016843653284013271, + 0.0013573038158938289, + 0.0013511218130588531, + 0.0012288920115679502, + 0.001158029423095286, + 0.0021967338398098946, + 0.0011375299654901028, + 0.001665259711444378, + 0.0015712347812950611, + 0.0020342066418379545, + 0.0009685701807029545, + 0.0016086641699075699, + 0.0010068491101264954, + 0.002169159008190036, + 0.0014988778857514262, + 0.0018744685221463442, + 0.001125767477788031 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 5 + ], + "selected_reward": -0.026351071894168854 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, eye-level medium shot captures an intimate, portrait-style scene, centered on a young man with thoughtful, deep brown eyes and dark, damp hair styled into soft spikes. He turns toward a large window, the room's soft, warm light illuminating his pensive expression. The light sculpts the side of his face and captures the subtle moisture on his skin. He wears a simple, dark gray knit sweater, its soft texture adding to the cozy feel of the interior. Rain streaks vertically across the windowpane like a liquid curtain, distorting the blurred, rain-swept cityscape into a beautiful, cold-blue bokeh and twinkling, warm city lights. His gaze is distant and fixed, lost in quiet contemplation, his only movement a slow, deliberate blink. The warm, gentle light on the subject contrasts sharply with the cold, damp world outside, and combined with his profound stillness, creates an atmosphere of deep contemplation and tranquility.", + "seed": 9007, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000021/group_1/index_5218_seed_9007_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": -0.03712458834052086, + "vlm_logprob": -0.03712458834052086, + "vlm_yes_no_margin": 3.6, + "vlm_token_credit": -0.03712458834052086, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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, eye-level medium shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject a young man with dark, damp hair styled into soft spikes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the subject wearing a simple, dark gray knit sweater?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene show rain streaks on a windowpane with a blurred cityscape background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the subject perform only a slow, deliberate blink as his movement?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0633288323879242 + ], + "opposite_logprob_tokens": [ + -3.563328742980957 + ], + "logprob_reward": -0.0633288323879242, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.0633288323879242, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.359182208776474, + 0.14214858412742615, + 0.11507537961006165, + 0.05624169111251831, + 0.05378618463873863, + 0.0489027239382267, + 0.08311943709850311, + 0.03689626604318619, + 0.062287501990795135, + 0.04235999286174774 + ], + "credit_map": [ + 0.05501563102006912, + 0.01783471368253231, + 0.0886446088552475, + 0.06508410722017288, + 0.03589937463402748, + 0.024294234812259674, + 0.028452524915337563, + 0.04395704343914986, + 0.019954394549131393, + 0.013150518760085106, + 0.028793487697839737, + 0.017620041966438293, + 0.018376000225543976, + 0.010670899413526058, + 0.014788195490837097, + 0.018795043230056763, + 0.016916830092668533, + 0.012172780930995941, + 0.02196931280195713, + 0.01409807987511158, + 0.017189713194966316, + 0.007742604240775108, + 0.014387396164238453, + 0.01059865765273571, + 0.007530666887760162, + 0.00474349781870842, + 0.009671014733612537, + 0.006763846147805452, + 0.0063572539947927, + 0.005296683870255947, + 0.0068449946120381355, + 0.009033734910190105, + 0.008176169358193874, + 0.005031308624893427, + 0.009410173632204533, + 0.007382677868008614, + 0.005882479250431061, + 0.007435041479766369, + 0.005320850759744644, + 0.005147486459463835, + 0.007925833575427532, + 0.004957381635904312, + 0.008701704442501068, + 0.008118110708892345, + 0.005616522394120693, + 0.0026223917957395315, + 0.005792743060737848, + 0.005168038886040449, + 0.013806820847094059, + 0.007492430508136749, + 0.023985730484128, + 0.006361318286508322, + 0.007158743217587471, + 0.009558298625051975, + 0.008195520378649235, + 0.006560570560395718, + 0.005774995777755976, + 0.003348330734297633, + 0.008833864703774452, + 0.004418340511620045, + 0.003062742529436946, + 0.0022698116954416037, + 0.0037116510793566704, + 0.005476528313010931, + 0.009317628107964993, + 0.004998654592782259, + 0.010987242683768272, + 0.011246277950704098, + 0.006038584280759096, + 0.005505131557583809, + 0.008193232119083405, + 0.006000753492116928, + 0.006845921743661165, + 0.0037903408519923687, + 0.006864738650619984, + 0.0036192911211401224, + 0.00425752392038703, + 0.007724231109023094, + 0.004505809396505356, + 0.004752134904265404 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.0633288323879242 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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, eye-level medium shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject a young man with dark, damp hair styled into soft spikes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the subject wearing a simple, dark gray knit sweater?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene show rain streaks on a windowpane with a blurred cityscape background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the subject perform only a slow, deliberate blink as his movement?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00980265624821186 + ], + "opposite_logprob_tokens": [ + -4.63480281829834 + ], + "logprob_reward": -0.00980265624821186, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.00980265624821186, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3732270300388336, + 0.1871565580368042, + 0.12730032205581665, + 0.054928503930568695, + 0.03693762794137001, + 0.05619101598858833, + 0.05033012107014656, + 0.022210391238331795, + 0.05377482622861862, + 0.03794364258646965 + ], + "credit_map": [ + 0.0652223452925682, + 0.022897634655237198, + 0.12268135696649551, + 0.04297362640500069, + 0.03314074128866196, + 0.01744196191430092, + 0.03709687665104866, + 0.031772468239068985, + 0.028033427894115448, + 0.00962721649557352, + 0.0494297631084919, + 0.02121446654200554, + 0.01403411291539669, + 0.01656215637922287, + 0.020664500072598457, + 0.027590902522206306, + 0.015623334795236588, + 0.01308717392385006, + 0.02494112215936184, + 0.020629530772566795, + 0.01938997022807598, + 0.00952192023396492, + 0.014848090708255768, + 0.009259180165827274, + 0.007101677358150482, + 0.005936778150498867, + 0.01280464418232441, + 0.008861200883984566, + 0.00632873922586441, + 0.0026396182365715504, + 0.005687754135578871, + 0.0055680908262729645, + 0.0050925821997225285, + 0.003612934611737728, + 0.00888525228947401, + 0.00464092381298542, + 0.004125643987208605, + 0.002438024850562215, + 0.004370756447315216, + 0.003771508578211069, + 0.007446907460689545, + 0.004405310843139887, + 0.011224417015910149, + 0.01177273876965046, + 0.0052274493500590324, + 0.004257592372596264, + 0.006134106777608395, + 0.005722491070628166, + 0.007663874886929989, + 0.005747051443904638, + 0.011909368447959423, + 0.0037833387032151222, + 0.0045875059440732, + 0.006917380262166262, + 0.005793106276541948, + 0.003928496036678553, + 0.0032794680446386337, + 0.0024801702238619328, + 0.00515199638903141, + 0.0022672589402645826, + 0.002412691479548812, + 0.00240320828743279, + 0.0022690920159220695, + 0.0019465062068775296, + 0.007863971404731274, + 0.0049890135414898396, + 0.008629415184259415, + 0.0031993018928915262, + 0.005298646632581949, + 0.007861258462071419, + 0.006957091391086578, + 0.008976126089692116, + 0.005520670209079981, + 0.006030679680407047, + 0.006013236939907074, + 0.0030299928039312363, + 0.0044704913161695, + 0.004926224239170551, + 0.004512142390012741, + 0.0034402015153318644 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.00980265624821186 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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, eye-level medium shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject a young man with dark, damp hair styled into soft spikes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the subject wearing a simple, dark gray knit sweater?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene show rain streaks on a windowpane with a blurred cityscape background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the subject perform only a slow, deliberate blink as his movement?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.043060317635536194 + ], + "opposite_logprob_tokens": [ + -3.168060302734375 + ], + "logprob_reward": -0.043060317635536194, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.043060317635536194, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2185390442609787, + 0.31312796473503113, + 0.06540022790431976, + 0.027091009542346, + 0.07414533197879791, + 0.0994962602853775, + 0.062140341848134995, + 0.0391010083258152, + 0.05931563302874565, + 0.041643176227808 + ], + "credit_map": [ + 0.031250324100255966, + 0.01288626343011856, + 0.056121233850717545, + 0.02586822584271431, + 0.028308724984526634, + 0.015437901951372623, + 0.025391653180122375, + 0.023274710401892662, + 0.044344909489154816, + 0.017206331714987755, + 0.07212843745946884, + 0.038480617105960846, + 0.02544740028679371, + 0.027515387162566185, + 0.04725656658411026, + 0.04074830934405327, + 0.00842209905385971, + 0.0055352202616631985, + 0.011582726612687111, + 0.004273665137588978, + 0.009994735941290855, + 0.005457945633679628, + 0.009524204768240452, + 0.010609637014567852, + 0.004466243553906679, + 0.002458377508446574, + 0.006733921356499195, + 0.0032653426751494408, + 0.003027579979971051, + 0.0018890565261244774, + 0.0030572633258998394, + 0.002193224150687456, + 0.011359923519194126, + 0.0069819786585867405, + 0.021443165838718414, + 0.010851176455616951, + 0.005572612397372723, + 0.003322254167869687, + 0.006499702576547861, + 0.008114513009786606, + 0.014498729258775711, + 0.012189687229692936, + 0.0248640775680542, + 0.011205415241420269, + 0.0074384077452123165, + 0.008364367298781872, + 0.009743086993694305, + 0.011192485690116882, + 0.00891296099871397, + 0.007206103764474392, + 0.013182107359170914, + 0.0058914110995829105, + 0.004540002439171076, + 0.007432239595800638, + 0.005570245441049337, + 0.009405266493558884, + 0.006660898681730032, + 0.003963236231356859, + 0.010087595321238041, + 0.003566571045666933, + 0.0029371485579758883, + 0.0034514274448156357, + 0.0038617351092398167, + 0.004572398029267788, + 0.008587571792304516, + 0.004984028171747923, + 0.010017778724431992, + 0.002844357630237937, + 0.00566486781463027, + 0.008986299857497215, + 0.007161437068134546, + 0.011069294065237045, + 0.0058875479735434055, + 0.005618899129331112, + 0.01007277611643076, + 0.004731622524559498, + 0.002860505133867264, + 0.0044590323232114315, + 0.0036706668324768543, + 0.004342126660048962 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 5 + ], + "selected_reward": -0.043060317635536194 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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, eye-level medium shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject a young man with dark, damp hair styled into soft spikes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the subject wearing a simple, dark gray knit sweater?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene show rain streaks on a windowpane with a blurred cityscape background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the subject perform only a slow, deliberate blink as his movement?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0430804118514061 + ], + "opposite_logprob_tokens": [ + -3.1680803298950195 + ], + "logprob_reward": -0.0430804118514061, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.0430804118514061, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4800270199775696, + 0.14840106666088104, + 0.1394677609205246, + 0.054230690002441406, + 0.038897037506103516, + 0.03628958761692047, + 0.022773509845137596, + 0.019627772271633148, + 0.03661898151040077, + 0.023666607216000557 + ], + "credit_map": [ + 0.08379431813955307, + 0.028363842517137527, + 0.1668163686990738, + 0.06074719876050949, + 0.043583691120147705, + 0.017135856673121452, + 0.040760934352874756, + 0.038824792951345444, + 0.024024803191423416, + 0.008108978159725666, + 0.04112458974123001, + 0.022390956059098244, + 0.010200836695730686, + 0.011875302530825138, + 0.011806216090917587, + 0.018869372084736824, + 0.02379349060356617, + 0.00998967606574297, + 0.04102499410510063, + 0.0223244596272707, + 0.010517382062971592, + 0.0074354722164571285, + 0.011932185851037502, + 0.012450084090232849, + 0.008577053435146809, + 0.004163494799286127, + 0.01546090841293335, + 0.006901029963046312, + 0.00542167155072093, + 0.0033245987724512815, + 0.006291086785495281, + 0.00409083953127265, + 0.006434491835534573, + 0.0034296975936740637, + 0.009970562532544136, + 0.004349000286310911, + 0.00308254174888134, + 0.004042476881295443, + 0.0030821089167147875, + 0.004506154451519251, + 0.005264887586236, + 0.0030115144327282906, + 0.00951080210506916, + 0.004897385369986296, + 0.0032241614535450935, + 0.0036205947399139404, + 0.002880190499126911, + 0.0038800500333309174, + 0.0028983226511627436, + 0.0035826805979013443, + 0.0042271604761481285, + 0.001941878697834909, + 0.002250567777082324, + 0.002855285769328475, + 0.002690344350412488, + 0.002327266614884138, + 0.003390221856534481, + 0.0017648304346948862, + 0.005170488730072975, + 0.0015779964160174131, + 0.0017292217817157507, + 0.0027979332953691483, + 0.0014791813446208835, + 0.0017178973648697138, + 0.005200113169848919, + 0.0032343738712370396, + 0.0074087511748075485, + 0.006146556232124567, + 0.0029428552370518446, + 0.004518464673310518, + 0.0037785933818668127, + 0.0033892712090164423, + 0.003587489714846015, + 0.002218534704297781, + 0.005944453179836273, + 0.0031081109773367643, + 0.0023458560463041067, + 0.002086036605760455, + 0.002493577543646097, + 0.0018825478618964553 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.0430804118514061 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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, eye-level medium shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject a young man with dark, damp hair styled into soft spikes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the subject wearing a simple, dark gray knit sweater?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene show rain streaks on a windowpane with a blurred cityscape background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the subject perform only a slow, deliberate blink as his movement?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026350723579525948 + ], + "opposite_logprob_tokens": [ + -3.651350736618042 + ], + "logprob_reward": -0.026350723579525948, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026350723579525948, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16525277495384216, + 0.08791874349117279, + 0.11320165544748306, + 0.06683021783828735, + 0.13022570312023163, + 0.19726450741291046, + 0.05495359003543854, + 0.09208965301513672, + 0.04107479378581047, + 0.051188334822654724 + ], + "credit_map": [ + 0.0229052621871233, + 0.011764783412218094, + 0.03858571499586105, + 0.018358202651143074, + 0.027797428891062737, + 0.01069097314029932, + 0.019047899171710014, + 0.016102522611618042, + 0.012459360994398594, + 0.009006681852042675, + 0.014435243792831898, + 0.009619692340493202, + 0.01392449252307415, + 0.009720771573483944, + 0.010013340972363949, + 0.008739163167774677, + 0.01715167798101902, + 0.006606684532016516, + 0.02594519592821598, + 0.010958626866340637, + 0.016334466636180878, + 0.005921663250774145, + 0.01458507589995861, + 0.01569826528429985, + 0.009615940973162651, + 0.005949352867901325, + 0.011460685171186924, + 0.00637391023337841, + 0.01103392243385315, + 0.004374340176582336, + 0.012405834160745144, + 0.005616237409412861, + 0.02051054872572422, + 0.01218490768224001, + 0.02306528016924858, + 0.020908350124955177, + 0.012122385203838348, + 0.017044376581907272, + 0.014126750640571117, + 0.010263114236295223, + 0.024889906868338585, + 0.01708441786468029, + 0.034517284482717514, + 0.0478200763463974, + 0.012891221791505814, + 0.017320500686764717, + 0.019553961232304573, + 0.023187167942523956, + 0.007645105943083763, + 0.006806227378547192, + 0.008862419053912163, + 0.0037037942092865705, + 0.006340942345559597, + 0.008201749995350838, + 0.007746030110865831, + 0.0056473203003406525, + 0.014917464926838875, + 0.007409806363284588, + 0.017325757071375847, + 0.006257213652133942, + 0.008695323020219803, + 0.016286088153719902, + 0.011153900995850563, + 0.010044111870229244, + 0.0054539223201572895, + 0.0033173055853694677, + 0.006107003893703222, + 0.0039600045420229435, + 0.004500485025346279, + 0.004237332846969366, + 0.0052651953883469105, + 0.00823354721069336, + 0.007419705856591463, + 0.005880266893655062, + 0.010034948587417603, + 0.008143654093146324, + 0.00605106120929122, + 0.0026672922540456057, + 0.006584524177014828, + 0.004406887572258711 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 4 + ], + "selected_reward": -0.026350723579525948 + } + ], + "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/0000021/group_10/index_tensor([6990, 6990])_group_10_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000021/group_10/index_tensor([6990, 6990])_group_10_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..0015bd52538a870a6583402ff9c50caa2ad8ad38 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000021/group_10/index_tensor([6990, 6990])_group_10_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "This static, low-angle, extreme close-up image, rendered in ultra-detailed 4K resolution, focuses on a lone Coprinus comatus mushroom, sprouting from a dark, jagged crack in the damp sidewalk of a vast cityscape. The environment is bathed in a stunning cyberpunk aesthetic, with harsh, pulsating light from scorched burgundy and cyan neon signs reflecting off the damp asphalt, casting long, dramatic shadows. The mushroom's originally pristine white, furry cap, each delicate scale clearly visible, begins to rapidly decay. As the structure withers and its lower edge curls upward, its color meticulously shifts from white to a deep black. The cap dissolves into thick, viscous streams of black ink, shimmering beneath the vibrant city lights. These inky streams steadily flow downward, eventually forming heavy droplets that break off and fall, crashing onto a discarded, gleaming chrome plate below. Every impact is captured with crystal clarity, black liquid splattering into beautiful, chaotic patterns that starkly reflect the dust-soaked futuristic metropolis, creating a mesmerizing and melancholic atmosphere of organic decay within a synthetic world.", + "seed": 32012, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000021/group_10/index_6990_seed_32012_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": -0.0629024950787425, + "vlm_logprob": -0.0629024950787425, + "vlm_yes_no_margin": 3.525, + "vlm_token_credit": -0.0629024950787425, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 lone Coprinus comatus mushroom sprouting from a crack in a sidewalk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a vast cityscape with a cyberpunk aesthetic featuring burgundy and cyan neon lights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the mushroom's cap visibly shift in color from white to deep black as it decays?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the decaying mushroom cap dissolve into thick, viscous streams of black ink that flow downward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the black ink droplets fall and splash onto a discarded chrome plate below?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.18712174892425537 + ], + "opposite_logprob_tokens": [ + -1.8121217489242554 + ], + "logprob_reward": -0.18712174892425537, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.18712174892425537, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1584603637456894, + 0.09075361490249634, + 0.06174379214644432, + 0.07043031603097916, + 0.15181906521320343, + 0.1788746863603592, + 0.1820523887872696, + 0.03232024610042572, + 0.055208541452884674, + 0.01833704113960266 + ], + "credit_map": [ + 0.011802876368165016, + 0.017973897978663445, + 0.01893812045454979, + 0.017531191930174828, + 0.03428402543067932, + 0.0209200382232666, + 0.01692025549709797, + 0.02008995972573757, + 0.00999907124787569, + 0.008625969290733337, + 0.009996069595217705, + 0.011193573474884033, + 0.013735219836235046, + 0.011288844980299473, + 0.012155718170106411, + 0.013759148307144642, + 0.0120899248868227, + 0.004928173031657934, + 0.007809251546859741, + 0.0057273320853710175, + 0.005874330177903175, + 0.008993186987936497, + 0.0057270037941634655, + 0.010594590567052364, + 0.015135252848267555, + 0.005655054468661547, + 0.00819436926394701, + 0.007733497768640518, + 0.006496603600680828, + 0.008256232365965843, + 0.009334344416856766, + 0.009624965488910675, + 0.029094958677887917, + 0.0116947703063488, + 0.01580902561545372, + 0.012635121122002602, + 0.015842344611883163, + 0.029805228114128113, + 0.012285027652978897, + 0.02465258166193962, + 0.013321857899427414, + 0.014674820005893707, + 0.0250996220856905, + 0.014477535150945187, + 0.027146361768245697, + 0.03173375502228737, + 0.024998581036925316, + 0.02742215245962143, + 0.0187044870108366, + 0.01714448258280754, + 0.022146744653582573, + 0.017417561262845993, + 0.022034553810954094, + 0.03416639193892479, + 0.021262075752019882, + 0.02917609177529812, + 0.0030007746536284685, + 0.003440071130171418, + 0.0026214916724711657, + 0.00400145910680294, + 0.004890893120318651, + 0.003940146882086992, + 0.004905025940388441, + 0.005520387087017298, + 0.006567981094121933, + 0.0032764167990535498, + 0.003780300496146083, + 0.0020161038264632225, + 0.0028137790504842997, + 0.022709105163812637, + 0.0027265120297670364, + 0.011318345554172993, + 0.0019960403442382812, + 0.0022375821135938168, + 0.001987190218642354, + 0.0018655641470104456, + 0.003239196026697755, + 0.002529850462451577, + 0.0022260411642491817, + 0.0022555754985660315 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 5, + 0 + ], + "selected_reward": -0.18712174892425537 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 lone Coprinus comatus mushroom sprouting from a crack in a sidewalk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a vast cityscape with a cyberpunk aesthetic featuring burgundy and cyan neon lights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the mushroom's cap visibly shift in color from white to deep black as it decays?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the decaying mushroom cap dissolve into thick, viscous streams of black ink that flow downward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the black ink droplets fall and splash onto a discarded chrome plate below?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009801475331187248 + ], + "opposite_logprob_tokens": [ + -4.634801387786865 + ], + "logprob_reward": -0.009801475331187248, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009801475331187248, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22001828253269196, + 0.16538983583450317, + 0.0914594978094101, + 0.06017601117491722, + 0.13086912035942078, + 0.10283122956752777, + 0.09913088381290436, + 0.04532014578580856, + 0.05220940709114075, + 0.032595615833997726 + ], + "credit_map": [ + 0.01971548981964588, + 0.021073997020721436, + 0.0310494564473629, + 0.018820999190211296, + 0.029429106041789055, + 0.04074236378073692, + 0.025140559300780296, + 0.03404631093144417, + 0.018454521894454956, + 0.013837780803442001, + 0.018422601744532585, + 0.02191573940217495, + 0.023080911487340927, + 0.0268753319978714, + 0.016673123463988304, + 0.0261298306286335, + 0.01893618144094944, + 0.006466612219810486, + 0.010502668097615242, + 0.007601411547511816, + 0.006151974201202393, + 0.01833348721265793, + 0.006774449720978737, + 0.016692711040377617, + 0.0070770676247775555, + 0.004276806954294443, + 0.007336620707064867, + 0.006607020273804665, + 0.004914010409265757, + 0.013784532435238361, + 0.006056952755898237, + 0.010122999548912048, + 0.02650049701333046, + 0.011847825720906258, + 0.013090278953313828, + 0.014295211993157864, + 0.012259673327207565, + 0.020806102082133293, + 0.012752421200275421, + 0.019317101687192917, + 0.015097626484930515, + 0.008926083333790302, + 0.011184883303940296, + 0.008747287094593048, + 0.01252223551273346, + 0.017020171508193016, + 0.013472642749547958, + 0.01586030051112175, + 0.011561308056116104, + 0.009865397587418556, + 0.012313967570662498, + 0.009789509698748589, + 0.012155327945947647, + 0.01673753559589386, + 0.011974839493632317, + 0.014732995070517063, + 0.008594867773354053, + 0.004154342692345381, + 0.005529386922717094, + 0.003472432028502226, + 0.004329860210418701, + 0.008045804686844349, + 0.003579120384529233, + 0.00761432945728302, + 0.00991244800388813, + 0.005436106584966183, + 0.008460811339318752, + 0.004263362847268581, + 0.005123103968799114, + 0.008754868060350418, + 0.0034513860009610653, + 0.006807323079556227, + 0.0047278632409870625, + 0.003185655688866973, + 0.003642439842224121, + 0.0034779931884258986, + 0.0033141253516077995, + 0.00518434401601553, + 0.002784013282507658, + 0.006279179360717535 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 4 + ], + "selected_reward": -0.009801475331187248 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 lone Coprinus comatus mushroom sprouting from a crack in a sidewalk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a vast cityscape with a cyberpunk aesthetic featuring burgundy and cyan neon lights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the mushroom's cap visibly shift in color from white to deep black as it decays?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the decaying mushroom cap dissolve into thick, viscous streams of black ink that flow downward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the black ink droplets fall and splash onto a discarded chrome plate below?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02330787293612957 + ], + "opposite_logprob_tokens": [ + -3.7733078002929688 + ], + "logprob_reward": -0.02330787293612957, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.02330787293612957, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20655696094036102, + 0.07769530266523361, + 0.058217938989400864, + 0.07789444178342819, + 0.09154906123876572, + 0.1653565764427185, + 0.15827438235282898, + 0.05643168464303017, + 0.02620711922645569, + 0.08181647211313248 + ], + "credit_map": [ + 0.014243264682590961, + 0.02862163633108139, + 0.024910081177949905, + 0.018540898337960243, + 0.03934001177549362, + 0.03419248387217522, + 0.018264872953295708, + 0.028443722054362297, + 0.007016289513558149, + 0.010250749066472054, + 0.007556885480880737, + 0.009132442064583302, + 0.016506696119904518, + 0.007326383143663406, + 0.010494895279407501, + 0.009410963393747807, + 0.007760730106383562, + 0.006284942850470543, + 0.006482521537691355, + 0.0060366615653038025, + 0.010373403318226337, + 0.006825434044003487, + 0.006530381739139557, + 0.007923861965537071, + 0.007048368453979492, + 0.007295485120266676, + 0.006785576231777668, + 0.011762399226427078, + 0.011094015091657639, + 0.013675665482878685, + 0.010676244273781776, + 0.009556681849062443, + 0.016383083537220955, + 0.007887891493737698, + 0.00847502052783966, + 0.008728232234716415, + 0.010595673695206642, + 0.015937160700559616, + 0.011116995476186275, + 0.012425004504621029, + 0.010802584700286388, + 0.012576773762702942, + 0.021756760776042938, + 0.013042597100138664, + 0.026029672473669052, + 0.030816424638032913, + 0.02408115565776825, + 0.02625061199069023, + 0.019449623301625252, + 0.015784405171871185, + 0.018007509410381317, + 0.01605520024895668, + 0.020219434052705765, + 0.021711712703108788, + 0.020818591117858887, + 0.02622789703309536, + 0.004551596939563751, + 0.005128895398229361, + 0.0074979267083108425, + 0.006344887427985668, + 0.007132810074836016, + 0.00890337210148573, + 0.007315043825656176, + 0.009557154029607773, + 0.004612012766301632, + 0.002654851647093892, + 0.0026826688554137945, + 0.002748889150097966, + 0.0032446179538965225, + 0.004334126133471727, + 0.0022009722888469696, + 0.0037289790343493223, + 0.013007976114749908, + 0.009831265546381474, + 0.008067192509770393, + 0.008957276120781898, + 0.009874098002910614, + 0.01335894875228405, + 0.005606654100120068, + 0.01311306469142437 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 6 + ], + "selected_reward": -0.02330787293612957 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 lone Coprinus comatus mushroom sprouting from a crack in a sidewalk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a vast cityscape with a cyberpunk aesthetic featuring burgundy and cyan neon lights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the mushroom's cap visibly shift in color from white to deep black as it decays?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the decaying mushroom cap dissolve into thick, viscous streams of black ink that flow downward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the black ink droplets fall and splash onto a discarded chrome plate below?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0052552176639437675 + ], + "opposite_logprob_tokens": [ + -5.255255222320557 + ], + "logprob_reward": -0.0052552176639437675, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.0052552176639437675, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2517867088317871, + 0.18119412660598755, + 0.0681791752576828, + 0.05403565987944603, + 0.10780711472034454, + 0.09133223444223404, + 0.10311295092105865, + 0.036007460206747055, + 0.0655616968870163, + 0.04098288342356682 + ], + "credit_map": [ + 0.027663402259349823, + 0.02550158090889454, + 0.030192574486136436, + 0.01970798149704933, + 0.043903328478336334, + 0.04810316488146782, + 0.0202637892216444, + 0.03645087406039238, + 0.02546258457005024, + 0.014329697005450726, + 0.015750424936413765, + 0.014747717417776585, + 0.019873296841979027, + 0.05064136162400246, + 0.014439939521253109, + 0.02594910003244877, + 0.006506992504000664, + 0.004886938259005547, + 0.009049545042216778, + 0.007919369265437126, + 0.009010658599436283, + 0.011488598771393299, + 0.010819390416145325, + 0.008497687056660652, + 0.005104224197566509, + 0.004204897675663233, + 0.005586350802332163, + 0.005636748392134905, + 0.007477770559489727, + 0.011515704914927483, + 0.007156647741794586, + 0.007353317458182573, + 0.012394647113978863, + 0.008471331559121609, + 0.012986698187887669, + 0.010779169388115406, + 0.014996851794421673, + 0.0171371977776289, + 0.014149628579616547, + 0.016891587525606155, + 0.008782532997429371, + 0.0069443597458302975, + 0.018284762278199196, + 0.011205377988517284, + 0.010951393283903599, + 0.008469663560390472, + 0.012786218896508217, + 0.013907922431826591, + 0.018698623403906822, + 0.011330736801028252, + 0.01064370758831501, + 0.008959028869867325, + 0.01242391299456358, + 0.01388544775545597, + 0.011723935604095459, + 0.015447564423084259, + 0.002936073811724782, + 0.003095104591920972, + 0.0030272763688117266, + 0.004188748076558113, + 0.003790575545281172, + 0.009204580448567867, + 0.003966763149946928, + 0.005798337049782276, + 0.008147085085511208, + 0.005834877956658602, + 0.005382218863815069, + 0.005803472362458706, + 0.008540459908545017, + 0.01606100983917713, + 0.004358269274234772, + 0.011434304527938366, + 0.006949549075216055, + 0.00415174663066864, + 0.0034484874922782183, + 0.004177446011453867, + 0.004372264724224806, + 0.008884629234671593, + 0.0032561812549829483, + 0.005742580164223909 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 4 + ], + "selected_reward": -0.0052552176639437675 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 lone Coprinus comatus mushroom sprouting from a crack in a sidewalk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a vast cityscape with a cyberpunk aesthetic featuring burgundy and cyan neon lights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the mushroom's cap visibly shift in color from white to deep black as it decays?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the decaying mushroom cap dissolve into thick, viscous streams of black ink that flow downward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the black ink droplets fall and splash onto a discarded chrome plate below?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08902616053819656 + ], + "opposite_logprob_tokens": [ + -2.4640262126922607 + ], + "logprob_reward": -0.08902616053819656, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08902616053819656, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19166775047779083, + 0.1128554493188858, + 0.22678810358047485, + 0.0755797028541565, + 0.10108661651611328, + 0.08147315680980682, + 0.06420397758483887, + 0.05028451606631279, + 0.031366001814603806, + 0.06469467282295227 + ], + "credit_map": [ + 0.016065094619989395, + 0.02452687919139862, + 0.02160795032978058, + 0.02349843457341194, + 0.03233952820301056, + 0.02565373107790947, + 0.023675333708524704, + 0.02430080808699131, + 0.017453204840421677, + 0.009822411462664604, + 0.010885873809456825, + 0.01009813416749239, + 0.021710101515054703, + 0.013456623069941998, + 0.01375584863126278, + 0.015673251822590828, + 0.024770913645625114, + 0.014621400274336338, + 0.03507496789097786, + 0.023714778944849968, + 0.018579328432679176, + 0.04099059849977493, + 0.03332308679819107, + 0.03571305051445961, + 0.017001213505864143, + 0.006576461251825094, + 0.007715992629528046, + 0.009314336813986301, + 0.007512348238378763, + 0.007638819050043821, + 0.009364353492856026, + 0.010456186719238758, + 0.014199323952198029, + 0.008118011988699436, + 0.012155517935752869, + 0.007856527343392372, + 0.012179073877632618, + 0.02040662057697773, + 0.011540799401700497, + 0.014630754478275776, + 0.008551178500056267, + 0.006439079996198416, + 0.016185227781534195, + 0.009049876593053341, + 0.00992605835199356, + 0.00877316389232874, + 0.011219360865652561, + 0.011329220607876778, + 0.004823601339012384, + 0.005783087573945522, + 0.004979041405022144, + 0.006804200354963541, + 0.011053158901631832, + 0.01236051321029663, + 0.007383118849247694, + 0.011017263866961002, + 0.006433070171624422, + 0.0050333114340901375, + 0.0049892328679561615, + 0.006656005047261715, + 0.007116398774087429, + 0.006463430356234312, + 0.006067138630896807, + 0.007525928318500519, + 0.0035753531847149134, + 0.002841950161382556, + 0.003987214062362909, + 0.0027239799965173006, + 0.004300220869481564, + 0.006039069965481758, + 0.0034737687092274427, + 0.004424448125064373, + 0.014045480638742447, + 0.006409483030438423, + 0.009165563620626926, + 0.004862225614488125, + 0.006364053580909967, + 0.010721355676651001, + 0.004414663650095463, + 0.008711855858564377 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 1 + ], + "selected_reward": -0.08902616053819656 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "This static, low-angle, extreme close-up image, rendered in ultra-detailed 4K resolution, focuses on a lone Coprinus comatus mushroom, sprouting from a dark, jagged crack in the damp sidewalk of a vast cityscape. The environment is bathed in a stunning cyberpunk aesthetic, with harsh, pulsating light from scorched burgundy and cyan neon signs reflecting off the damp asphalt, casting long, dramatic shadows. The mushroom's originally pristine white, furry cap, each delicate scale clearly visible, begins to rapidly decay. As the structure withers and its lower edge curls upward, its color meticulously shifts from white to a deep black. The cap dissolves into thick, viscous streams of black ink, shimmering beneath the vibrant city lights. These inky streams steadily flow downward, eventually forming heavy droplets that break off and fall, crashing onto a discarded, gleaming chrome plate below. Every impact is captured with crystal clarity, black liquid splattering into beautiful, chaotic patterns that starkly reflect the dust-soaked futuristic metropolis, creating a mesmerizing and melancholic atmosphere of organic decay within a synthetic world.", + "seed": 32013, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000021/group_10/index_6990_seed_32013_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": -0.14442673455923796, + "vlm_logprob": -0.14442673455923796, + "vlm_yes_no_margin": 3.05, + "vlm_token_credit": -0.14442673455923796, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 lone Coprinus comatus mushroom sprouting from a crack in a sidewalk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a vast cityscape with a cyberpunk aesthetic featuring burgundy and cyan neon lights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the mushroom's cap visibly shift in color from white to deep black as it decays?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the decaying mushroom cap dissolve into thick, viscous streams of black ink that flow downward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the black ink droplets fall and splash onto a discarded chrome plate below?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.5838863253593445 + ], + "opposite_logprob_tokens": [ + -0.8338863253593445 + ], + "logprob_reward": -0.5838863253593445, + "yes_no_margin_reward": 0.25, + "token_credit_reward": -0.5838863253593445, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5076428651809692, + 0.10537223517894745, + 0.05687678977847099, + 0.10356945544481277, + 0.06519201397895813, + 0.04823259636759758, + 0.04252913221716881, + 0.024337373673915863, + 0.018177993595600128, + 0.02806960605084896 + ], + "credit_map": [ + 0.045749031007289886, + 0.07021038234233856, + 0.041513681411743164, + 0.07429211586713791, + 0.0567810945212841, + 0.07928818464279175, + 0.07826057076454163, + 0.06154778599739075, + 0.013245201669633389, + 0.011814902536571026, + 0.009595169685781002, + 0.013250024057924747, + 0.012044306844472885, + 0.02070562355220318, + 0.010914548300206661, + 0.013802451081573963, + 0.006814209278672934, + 0.00574774295091629, + 0.005507211666554213, + 0.007134019862860441, + 0.00762179447337985, + 0.010085961781442165, + 0.006698952056467533, + 0.0072668916545808315, + 0.01780117116868496, + 0.009707299061119556, + 0.008132927119731903, + 0.012044848874211311, + 0.017702888697385788, + 0.014729529619216919, + 0.012729021720588207, + 0.010721752420067787, + 0.006733274087309837, + 0.006297883577644825, + 0.0072327121160924435, + 0.007721115369349718, + 0.01188638061285019, + 0.008055616170167923, + 0.007779229898005724, + 0.00948579702526331, + 0.004673206713050604, + 0.0034453412517905235, + 0.0036919577978551388, + 0.004578443709760904, + 0.004284643102437258, + 0.012735043652355671, + 0.005477570928633213, + 0.00934638362377882, + 0.007812398951500654, + 0.0034379486460238695, + 0.004422292113304138, + 0.0036388859152793884, + 0.004681229125708342, + 0.006448561325669289, + 0.004803509451448917, + 0.0072843036614358425, + 0.0027910417411476374, + 0.0018043483141809702, + 0.002775362925603986, + 0.0026112610939890146, + 0.0023739063180983067, + 0.00441809231415391, + 0.0041922302916646, + 0.003371128113940358, + 0.002584171947091818, + 0.0016820282908156514, + 0.0015889984788373113, + 0.003304115030914545, + 0.0014315638691186905, + 0.002806937089189887, + 0.002616794779896736, + 0.0021633836440742016, + 0.003783406689763069, + 0.002455176319926977, + 0.003046443685889244, + 0.005106127820909023, + 0.0025853237602859735, + 0.003708950709551573, + 0.004235462285578251, + 0.003148710820823908 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 3 + ], + "selected_reward": -0.5838863253593445 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 lone Coprinus comatus mushroom sprouting from a crack in a sidewalk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a vast cityscape with a cyberpunk aesthetic featuring burgundy and cyan neon lights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the mushroom's cap visibly shift in color from white to deep black as it decays?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the decaying mushroom cap dissolve into thick, viscous streams of black ink that flow downward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the black ink droplets fall and splash onto a discarded chrome plate below?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01421516202390194 + ], + "opposite_logprob_tokens": [ + -4.264214992523193 + ], + "logprob_reward": -0.01421516202390194, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.01421516202390194, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4139184355735779, + 0.19675898551940918, + 0.10143325477838516, + 0.03942360728979111, + 0.03270791098475456, + 0.05606287345290184, + 0.08372670412063599, + 0.036445118486881256, + 0.01792248710989952, + 0.02160060778260231 + ], + "credit_map": [ + 0.04484214261174202, + 0.045879822224378586, + 0.023169251158833504, + 0.03551211953163147, + 0.05752895772457123, + 0.11538366228342056, + 0.03203711286187172, + 0.05956535041332245, + 0.013649479486048222, + 0.011309562250971794, + 0.012404090724885464, + 0.014730154536664486, + 0.029113490134477615, + 0.07013870030641556, + 0.02038097381591797, + 0.025032535195350647, + 0.009300964884459972, + 0.01200336217880249, + 0.008563232608139515, + 0.017619939520955086, + 0.012497926130890846, + 0.01565997302532196, + 0.015371560119092464, + 0.010416300036013126, + 0.0037436732091009617, + 0.002890523988753557, + 0.0023861583322286606, + 0.006650450173765421, + 0.007639592047780752, + 0.005811231676489115, + 0.0056958976201713085, + 0.004606080241501331, + 0.0043488419614732265, + 0.003090913873165846, + 0.002998822834342718, + 0.003019553143531084, + 0.008177324198186398, + 0.0033433381468057632, + 0.0036456985399127007, + 0.004083418752998114, + 0.007590234279632568, + 0.005385258700698614, + 0.004417291842401028, + 0.004411390516906977, + 0.010638747364282608, + 0.00827146228402853, + 0.005900555290281773, + 0.009447930380702019, + 0.006743521895259619, + 0.0030137246940284967, + 0.003955894149839878, + 0.0037002130411565304, + 0.0041784183122217655, + 0.02917463146150112, + 0.0045309425331652164, + 0.02842935360968113, + 0.005645205266773701, + 0.0034447552170604467, + 0.003423028625547886, + 0.00412391871213913, + 0.004780581686645746, + 0.00526751484721899, + 0.005914117209613323, + 0.0038459990173578262, + 0.0034033660776913166, + 0.0015156621811911464, + 0.0019358029821887612, + 0.002102408092468977, + 0.0016078659100458026, + 0.0026875799521803856, + 0.0020258789882063866, + 0.0026439232751727104, + 0.0036097948905080557, + 0.0016736777033656836, + 0.0023109838366508484, + 0.0022803544998168945, + 0.0023691831156611443, + 0.003408282995223999, + 0.0029777144081890583, + 0.00297061656601727 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.01421516202390194 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 lone Coprinus comatus mushroom sprouting from a crack in a sidewalk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a vast cityscape with a cyberpunk aesthetic featuring burgundy and cyan neon lights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the mushroom's cap visibly shift in color from white to deep black as it decays?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the decaying mushroom cap dissolve into thick, viscous streams of black ink that flow downward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the black ink droplets fall and splash onto a discarded chrome plate below?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02636604942381382 + ], + "opposite_logprob_tokens": [ + -3.6513659954071045 + ], + "logprob_reward": -0.02636604942381382, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.02636604942381382, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24367180466651917, + 0.1929912120103836, + 0.28018754720687866, + 0.04338710755109787, + 0.028977693989872932, + 0.08175917714834213, + 0.07530799508094788, + 0.02730400115251541, + 0.011479155160486698, + 0.01493431068956852 + ], + "credit_map": [ + 0.017517367377877235, + 0.02651977725327015, + 0.01275662425905466, + 0.04163020849227905, + 0.03388393297791481, + 0.03866405412554741, + 0.04187948256731033, + 0.03082035481929779, + 0.014562122523784637, + 0.017528770491480827, + 0.01866086944937706, + 0.020414991304278374, + 0.031012360006570816, + 0.04740098491311073, + 0.018001753836870193, + 0.025409340858459473, + 0.02051115408539772, + 0.026200473308563232, + 0.030392998829483986, + 0.055009517818689346, + 0.026835892349481583, + 0.040663834661245346, + 0.04869603365659714, + 0.03187762200832367, + 0.006497918628156185, + 0.0031734751537442207, + 0.004972604103386402, + 0.006314935628324747, + 0.006122876890003681, + 0.006063543725758791, + 0.005391167476773262, + 0.004850582219660282, + 0.0026900372467935085, + 0.002062136773020029, + 0.003260168479755521, + 0.0052891680970788, + 0.0032711199019104242, + 0.004156905226409435, + 0.004026582930237055, + 0.004221574869006872, + 0.00954410433769226, + 0.006559576373547316, + 0.006548114586621523, + 0.008108601905405521, + 0.016575369983911514, + 0.010049319826066494, + 0.011159555986523628, + 0.01321452297270298, + 0.005125551950186491, + 0.005009409040212631, + 0.005181632004678249, + 0.003879701951518655, + 0.005883336998522282, + 0.02190985344350338, + 0.006133515387773514, + 0.022184984758496284, + 0.0032330588437616825, + 0.0024918175768107176, + 0.0021478277631103992, + 0.003180135041475296, + 0.004186784382909536, + 0.003862264333292842, + 0.004959443584084511, + 0.003242668230086565, + 0.0017771146958693862, + 0.0008695821743458509, + 0.0011254233540967107, + 0.0016090943245217204, + 0.0012136633740738034, + 0.0019297839608043432, + 0.0016383228357881308, + 0.0013161697424948215, + 0.002031800104305148, + 0.0011902019614353776, + 0.001433775294572115, + 0.0023024382535368204, + 0.0018490240909159184, + 0.002418854972347617, + 0.0021709471475332975, + 0.0015372679335996509 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 1 + ], + "selected_reward": -0.02636604942381382 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a lone Coprinus comatus mushroom sprouting from a crack in a sidewalk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a vast cityscape with a cyberpunk aesthetic featuring burgundy and cyan neon lights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the mushroom's cap visibly shift in color from white to deep black as it decays?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the decaying mushroom cap dissolve into thick, viscous streams of black ink that flow downward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the black ink droplets fall and splash onto a discarded chrome plate below?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00863910373300314 + ], + "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": [ + 0.3373282551765442, + 0.15063633024692535, + 0.082891084253788, + 0.06076699495315552, + 0.03324044495820999, + 0.10299330949783325, + 0.13011643290519714, + 0.05350933596491814, + 0.022699635475873947, + 0.02581820823252201 + ], + "credit_map": [ + 0.021548015996813774, + 0.034130919724702835, + 0.02450701594352722, + 0.08082515001296997, + 0.035134121775627136, + 0.03540753200650215, + 0.07083077728748322, + 0.034944698214530945, + 0.010903498157858849, + 0.007909530773758888, + 0.009151641279459, + 0.011390539817512035, + 0.018710331991314888, + 0.052658289670944214, + 0.016807498410344124, + 0.023104969412088394, + 0.012934607453644276, + 0.008124588057398796, + 0.007896182127296925, + 0.005541418679058552, + 0.018386123701930046, + 0.011906780302524567, + 0.009608889929950237, + 0.008492486551404, + 0.011723203584551811, + 0.005059572868049145, + 0.005565133411437273, + 0.0065846829675138, + 0.007673706393688917, + 0.009466174989938736, + 0.007333794608712196, + 0.007360721938312054, + 0.004299262072890997, + 0.0026580779813230038, + 0.0034903804771602154, + 0.005145695060491562, + 0.003630533115938306, + 0.00523046450689435, + 0.004054222255945206, + 0.00473180552944541, + 0.00964956171810627, + 0.024198148399591446, + 0.005461398512125015, + 0.013339245691895485, + 0.02329779602587223, + 0.009123899973928928, + 0.008112896233797073, + 0.009810354560613632, + 0.011595695279538631, + 0.009047445841133595, + 0.009402865543961525, + 0.00992924626916647, + 0.011473173275589943, + 0.031216708943247795, + 0.014222864992916584, + 0.03322840854525566, + 0.006479153875261545, + 0.0038127212319523096, + 0.004864464979618788, + 0.004570464603602886, + 0.005737609695643187, + 0.009955833666026592, + 0.00941260252147913, + 0.008676479570567608, + 0.0033201684709638357, + 0.0017479516100138426, + 0.002324472414329648, + 0.0025135481264442205, + 0.003064854070544243, + 0.003314285073429346, + 0.003276221686974168, + 0.0031381307635456324, + 0.004446790087968111, + 0.0019641455728560686, + 0.0028383769094944, + 0.00237640761770308, + 0.003231586655601859, + 0.003936811815947294, + 0.0033150718081742525, + 0.0037090135738253593 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 6 + ], + "selected_reward": -0.00863910373300314 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 lone Coprinus comatus mushroom sprouting from a crack in a sidewalk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a vast cityscape with a cyberpunk aesthetic featuring burgundy and cyan neon lights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the mushroom's cap visibly shift in color from white to deep black as it decays?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the decaying mushroom cap dissolve into thick, viscous streams of black ink that flow downward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the black ink droplets fall and splash onto a discarded chrome plate below?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0890270322561264 + ], + "opposite_logprob_tokens": [ + -2.464026927947998 + ], + "logprob_reward": -0.0890270322561264, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.0890270322561264, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.27648454904556274, + 0.15602806210517883, + 0.07572568953037262, + 0.046401213854551315, + 0.04187268763780594, + 0.11619915068149567, + 0.11754917353391647, + 0.1103956401348114, + 0.028393113985657692, + 0.030950676649808884 + ], + "credit_map": [ + 0.019281890243291855, + 0.02866315469145775, + 0.018984118476510048, + 0.04035047069191933, + 0.039274122565984726, + 0.04212456941604614, + 0.05182402580976486, + 0.03598221018910408, + 0.010941457003355026, + 0.009808245114982128, + 0.009189482778310776, + 0.013408022932708263, + 0.017261790111660957, + 0.0537886768579483, + 0.020002072677016258, + 0.021628331393003464, + 0.010957751423120499, + 0.010808772407472134, + 0.00549160223454237, + 0.006363480817526579, + 0.01189984567463398, + 0.013621322810649872, + 0.006605712231248617, + 0.009977200999855995, + 0.007522590458393097, + 0.004059404134750366, + 0.005475119221955538, + 0.004106253385543823, + 0.00785971898585558, + 0.005402667447924614, + 0.005330507177859545, + 0.0066449581645429134, + 0.006517881993204355, + 0.0036018744576722383, + 0.00341977016068995, + 0.005183110013604164, + 0.008268516510725021, + 0.005101157817989588, + 0.004978201352059841, + 0.004802179988473654, + 0.010463193990290165, + 0.01143795158714056, + 0.006315212231129408, + 0.00872865505516529, + 0.02371816337108612, + 0.022550638765096664, + 0.009655434638261795, + 0.023329908028244972, + 0.011640776880085468, + 0.009997733868658543, + 0.009743685834109783, + 0.012032723985612392, + 0.014435934834182262, + 0.02145099639892578, + 0.013206308707594872, + 0.025041019544005394, + 0.01186833530664444, + 0.008917208760976791, + 0.009365399368107319, + 0.010281595401465893, + 0.011507091112434864, + 0.01856381818652153, + 0.023311583325266838, + 0.01658061146736145, + 0.003174638142809272, + 0.0022005222272127867, + 0.003488389076665044, + 0.0029548141174018383, + 0.0035081258974969387, + 0.0040779379196465015, + 0.004213049076497555, + 0.0047756400890648365, + 0.004761831369251013, + 0.002833671635016799, + 0.003339822869747877, + 0.0036031680647283792, + 0.0031943318899720907, + 0.0038341106846928596, + 0.004690510220825672, + 0.004693232011049986 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 6 + ], + "selected_reward": -0.0890270322561264 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "This static, low-angle, extreme close-up image, rendered in ultra-detailed 4K resolution, focuses on a lone Coprinus comatus mushroom, sprouting from a dark, jagged crack in the damp sidewalk of a vast cityscape. The environment is bathed in a stunning cyberpunk aesthetic, with harsh, pulsating light from scorched burgundy and cyan neon signs reflecting off the damp asphalt, casting long, dramatic shadows. The mushroom's originally pristine white, furry cap, each delicate scale clearly visible, begins to rapidly decay. As the structure withers and its lower edge curls upward, its color meticulously shifts from white to a deep black. The cap dissolves into thick, viscous streams of black ink, shimmering beneath the vibrant city lights. These inky streams steadily flow downward, eventually forming heavy droplets that break off and fall, crashing onto a discarded, gleaming chrome plate below. Every impact is captured with crystal clarity, black liquid splattering into beautiful, chaotic patterns that starkly reflect the dust-soaked futuristic metropolis, creating a mesmerizing and melancholic atmosphere of organic decay within a synthetic world.", + "seed": 32014, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000021/group_10/index_6990_seed_32014_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": -0.05156862586736679, + "vlm_logprob": -0.05156862586736679, + "vlm_yes_no_margin": 3.425, + "vlm_token_credit": -0.05156862586736679, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 lone Coprinus comatus mushroom sprouting from a crack in a sidewalk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a vast cityscape with a cyberpunk aesthetic featuring burgundy and cyan neon lights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the mushroom's cap visibly shift in color from white to deep black as it decays?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the decaying mushroom cap dissolve into thick, viscous streams of black ink that flow downward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the black ink droplets fall and splash onto a discarded chrome plate below?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05557086318731308 + ], + "opposite_logprob_tokens": [ + -3.1805708408355713 + ], + "logprob_reward": -0.05557086318731308, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.05557086318731308, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10472548007965088, + 0.033479560166597366, + 0.05420544371008873, + 0.13175098598003387, + 0.055751826614141464, + 0.123900406062603, + 0.06798513978719711, + 0.2557799220085144, + 0.11643625795841217, + 0.055984996259212494 + ], + "credit_map": [ + 0.008361141197383404, + 0.009807928465306759, + 0.01240328885614872, + 0.01693566143512726, + 0.015039673075079918, + 0.007952692918479443, + 0.024454042315483093, + 0.009771055541932583, + 0.0030635129660367966, + 0.006028680596500635, + 0.004032876342535019, + 0.0034378296695649624, + 0.003504151711240411, + 0.00471047218888998, + 0.003321352880448103, + 0.005380684044212103, + 0.008755132555961609, + 0.004744917619973421, + 0.007851842790842056, + 0.00541877094656229, + 0.008261744864284992, + 0.006986317224800587, + 0.006243038922548294, + 0.00594367878511548, + 0.031341973692178726, + 0.00931436475366354, + 0.013821841217577457, + 0.0075798481702804565, + 0.007112792227417231, + 0.016134653240442276, + 0.010657007806003094, + 0.03578851372003555, + 0.010630461387336254, + 0.0050807069055736065, + 0.005637808702886105, + 0.004492478910833597, + 0.003976429346948862, + 0.009399531409144402, + 0.005762186367064714, + 0.0107722207903862, + 0.012401129119098186, + 0.014353073202073574, + 0.013672789558768272, + 0.013590933755040169, + 0.019006580114364624, + 0.014598086476325989, + 0.018874824047088623, + 0.017403002828359604, + 0.006982908118516207, + 0.00721068074926734, + 0.00834927149116993, + 0.008552192710340023, + 0.013106566853821278, + 0.007440680637955666, + 0.008292211219668388, + 0.008050627075135708, + 0.03461068868637085, + 0.031058555468916893, + 0.040982600301504135, + 0.02844751812517643, + 0.04050146043300629, + 0.029659047722816467, + 0.03292340040206909, + 0.017596658319234848, + 0.020877383649349213, + 0.012038610875606537, + 0.014326442964375019, + 0.014362461864948273, + 0.016621047630906105, + 0.012654770165681839, + 0.01741647906601429, + 0.008139058016240597, + 0.0059761470183730125, + 0.005940568167716265, + 0.008831583894789219, + 0.01039336621761322, + 0.006409184541553259, + 0.005793436896055937, + 0.005196098703891039, + 0.007444609422236681 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 3, + 5 + ], + "selected_reward": -0.05557086318731308 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 lone Coprinus comatus mushroom sprouting from a crack in a sidewalk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a vast cityscape with a cyberpunk aesthetic featuring burgundy and cyan neon lights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the mushroom's cap visibly shift in color from white to deep black as it decays?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the decaying mushroom cap dissolve into thick, viscous streams of black ink that flow downward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the black ink droplets fall and splash onto a discarded chrome plate below?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020603444427251816 + ], + "opposite_logprob_tokens": [ + -3.8956034183502197 + ], + "logprob_reward": -0.020603444427251816, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020603444427251816, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16924414038658142, + 0.03712724149227142, + 0.05514562502503395, + 0.26427939534187317, + 0.04225468263030052, + 0.06263449788093567, + 0.04910463094711304, + 0.07709567993879318, + 0.057278506457805634, + 0.18583565950393677 + ], + "credit_map": [ + 0.024196114391088486, + 0.018919218331575394, + 0.015412183478474617, + 0.02242010459303856, + 0.019683631137013435, + 0.02777238003909588, + 0.022912422195076942, + 0.017928065732121468, + 0.003071522805839777, + 0.0042562056332826614, + 0.003839021548628807, + 0.004345407709479332, + 0.006423612590879202, + 0.0053041791543364525, + 0.005457090679556131, + 0.004430197644978762, + 0.008291587233543396, + 0.0032612462528049946, + 0.008504727855324745, + 0.007118096109479666, + 0.004693813156336546, + 0.010349924676120281, + 0.006131461821496487, + 0.006794759538024664, + 0.05655594915151596, + 0.013860436156392097, + 0.02216595783829689, + 0.017314482480287552, + 0.01230559777468443, + 0.03531326353549957, + 0.027872983366250992, + 0.07889071106910706, + 0.006277948152273893, + 0.0034863525070250034, + 0.004401111509650946, + 0.0045369816944003105, + 0.0039867679588496685, + 0.009781762026250362, + 0.0042936233803629875, + 0.005490130744874477, + 0.00733719440177083, + 0.005281704943627119, + 0.010884965769946575, + 0.006550780031830072, + 0.007429248653352261, + 0.007139438763260841, + 0.00986381247639656, + 0.008147348649799824, + 0.005502353422343731, + 0.005566435866057873, + 0.0062267412431538105, + 0.006908062379807234, + 0.005773310083895922, + 0.005180943291634321, + 0.00705728679895401, + 0.006889491342008114, + 0.009165809489786625, + 0.010012808255851269, + 0.013629384338855743, + 0.009738195687532425, + 0.011161601170897484, + 0.00826654490083456, + 0.009069515392184258, + 0.006051815114915371, + 0.009550472721457481, + 0.005290982313454151, + 0.006427221931517124, + 0.008612764067947865, + 0.008575452491641045, + 0.005479475948959589, + 0.007166822906583548, + 0.006175308022648096, + 0.0225294791162014, + 0.021983662620186806, + 0.028350766748189926, + 0.031229281798005104, + 0.022104280069470406, + 0.020494624972343445, + 0.017845474183559418, + 0.02129807323217392 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 9, + 0 + ], + "selected_reward": -0.020603444427251816 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 lone Coprinus comatus mushroom sprouting from a crack in a sidewalk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a vast cityscape with a cyberpunk aesthetic featuring burgundy and cyan neon lights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the mushroom's cap visibly shift in color from white to deep black as it decays?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the decaying mushroom cap dissolve into thick, viscous streams of black ink that flow downward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the black ink droplets fall and splash onto a discarded chrome plate below?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04867442324757576 + ], + "opposite_logprob_tokens": [ + -3.0486743450164795 + ], + "logprob_reward": -0.04867442324757576, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04867442324757576, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13369914889335632, + 0.051796965301036835, + 0.10014939308166504, + 0.10701017081737518, + 0.10539612919092178, + 0.16229534149169922, + 0.15563081204891205, + 0.11106168478727341, + 0.03510596603155136, + 0.037854406982660294 + ], + "credit_map": [ + 0.010097164660692215, + 0.012482338584959507, + 0.01444148924201727, + 0.01923675835132599, + 0.027041995897889137, + 0.009124797768890858, + 0.027243247255682945, + 0.014031353406608105, + 0.006597690749913454, + 0.004976988770067692, + 0.005228855647146702, + 0.006152322981506586, + 0.006428809836506844, + 0.00574371637776494, + 0.0070793465711176395, + 0.00958923902362585, + 0.010600699111819267, + 0.009448183700442314, + 0.013448907993733883, + 0.009683765470981598, + 0.01490664854645729, + 0.012167282402515411, + 0.012583184987306595, + 0.017310718074440956, + 0.01668129302561283, + 0.006291639059782028, + 0.012079358100891113, + 0.013450728729367256, + 0.008979056030511856, + 0.011779545806348324, + 0.017033075913786888, + 0.020715465769171715, + 0.015922144055366516, + 0.010891243815422058, + 0.012902812100946903, + 0.012502636760473251, + 0.008499141782522202, + 0.020058125257492065, + 0.00995541363954544, + 0.01466460432857275, + 0.016858309507369995, + 0.017102060839533806, + 0.017443738877773285, + 0.018692845478653908, + 0.02432083897292614, + 0.014484087005257607, + 0.03252660855650902, + 0.02086685784161091, + 0.021954135969281197, + 0.017623048275709152, + 0.020182792097330093, + 0.027971550822257996, + 0.02246949076652527, + 0.012662327848374844, + 0.021074941381812096, + 0.011692526750266552, + 0.015232783742249012, + 0.009806170128285885, + 0.018856177106499672, + 0.013019165024161339, + 0.018805885687470436, + 0.010531089268624783, + 0.01194065436720848, + 0.01286975760012865, + 0.0052064755000174046, + 0.003559883451089263, + 0.005356406792998314, + 0.004523769486695528, + 0.005664183292537928, + 0.0030839447863399982, + 0.00426614610478282, + 0.003445157315582037, + 0.00397541094571352, + 0.005389457102864981, + 0.00538634043186903, + 0.006365401670336723, + 0.005357240326702595, + 0.004341096617281437, + 0.004233182407915592, + 0.002806280739605427 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 6, + 0 + ], + "selected_reward": -0.04867442324757576 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 lone Coprinus comatus mushroom sprouting from a crack in a sidewalk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a vast cityscape with a cyberpunk aesthetic featuring burgundy and cyan neon lights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the mushroom's cap visibly shift in color from white to deep black as it decays?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the decaying mushroom cap dissolve into thick, viscous streams of black ink that flow downward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the black ink droplets fall and splash onto a discarded chrome plate below?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005949288606643677 + ], + "opposite_logprob_tokens": [ + -5.1309494972229 + ], + "logprob_reward": -0.005949288606643677, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005949288606643677, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05290345102548599, + 0.04405416548252106, + 0.04839690029621124, + 0.09360509365797043, + 0.06825380772352219, + 0.09996643662452698, + 0.26876693964004517, + 0.11029596626758575, + 0.03634339198470116, + 0.17741383612155914 + ], + "credit_map": [ + 0.009354853071272373, + 0.005405801814049482, + 0.006286470219492912, + 0.007407957687973976, + 0.007153591141104698, + 0.006065604276955128, + 0.007380286231637001, + 0.0038488840218633413, + 0.004417744465172291, + 0.004863800015300512, + 0.004716416820883751, + 0.0044111027382314205, + 0.00819836650043726, + 0.005607565399259329, + 0.004702566657215357, + 0.007136601954698563, + 0.007637382950633764, + 0.0028130647260695696, + 0.007174914702773094, + 0.005620577372610569, + 0.005688684061169624, + 0.006761973723769188, + 0.006455748341977596, + 0.006244551856070757, + 0.014393776655197144, + 0.005166159477084875, + 0.009531726129353046, + 0.00652516633272171, + 0.007219038903713226, + 0.013456085696816444, + 0.01539838407188654, + 0.021914750337600708, + 0.01296590082347393, + 0.004700413439422846, + 0.007255277130752802, + 0.006441932171583176, + 0.005246918182820082, + 0.012235400266945362, + 0.007480158004909754, + 0.01192780677229166, + 0.012972318567335606, + 0.008860744535923004, + 0.013261154294013977, + 0.010126293636858463, + 0.01251912023872137, + 0.007637527771294117, + 0.02222001552581787, + 0.01236925832927227, + 0.03619185462594032, + 0.02567790262401104, + 0.02870367094874382, + 0.053672634065151215, + 0.036585189402103424, + 0.03125547990202904, + 0.0309736467897892, + 0.025706559419631958, + 0.01268148422241211, + 0.016397444531321526, + 0.012390686199069023, + 0.010356918908655643, + 0.021800516173243523, + 0.013778853230178356, + 0.014246679842472076, + 0.0086433757096529, + 0.00530072720721364, + 0.0039409068413078785, + 0.005581132601946592, + 0.004900321364402771, + 0.003884423291310668, + 0.003983117640018463, + 0.005049684084951878, + 0.0037030766252428293, + 0.019930336624383926, + 0.02046879567205906, + 0.027629824355244637, + 0.03305176645517349, + 0.01987040601670742, + 0.020788006484508514, + 0.0175374336540699, + 0.018137270584702492 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 9, + 7 + ], + "selected_reward": -0.005949288606643677 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 lone Coprinus comatus mushroom sprouting from a crack in a sidewalk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a vast cityscape with a cyberpunk aesthetic featuring burgundy and cyan neon lights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the mushroom's cap visibly shift in color from white to deep black as it decays?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the decaying mushroom cap dissolve into thick, viscous streams of black ink that flow downward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the black ink droplets fall and splash onto a discarded chrome plate below?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.12704510986804962 + ], + "opposite_logprob_tokens": [ + -2.127045154571533 + ], + "logprob_reward": -0.12704510986804962, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.12704510986804962, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05045406520366669, + 0.08001438528299332, + 0.02320966310799122, + 0.05453117936849594, + 0.050440095365047455, + 0.11585183441638947, + 0.07007928937673569, + 0.12203095108270645, + 0.06152749061584473, + 0.371861070394516 + ], + "credit_map": [ + 0.004965577740222216, + 0.004590165335685015, + 0.006913402117788792, + 0.005930495914071798, + 0.007975428365170956, + 0.004802808165550232, + 0.010586874559521675, + 0.004689312539994717, + 0.007733243051916361, + 0.008812054991722107, + 0.0091779213398695, + 0.008341523818671703, + 0.015365853905677795, + 0.008706540800631046, + 0.009850256145000458, + 0.012026996351778507, + 0.0023969959001988173, + 0.001968854106962681, + 0.0033768671564757824, + 0.002299066400155425, + 0.0032154882792383432, + 0.0026293674018234015, + 0.0048218946903944016, + 0.002501127775758505, + 0.01113833300769329, + 0.003234629752114415, + 0.006251398008316755, + 0.003598774317651987, + 0.004056851379573345, + 0.005865500774234533, + 0.007676661480218172, + 0.012709027156233788, + 0.010597809217870235, + 0.0037318251561373472, + 0.004948061890900135, + 0.004680152516812086, + 0.0037154245655983686, + 0.007719902787357569, + 0.005129446275532246, + 0.009917473420500755, + 0.011795098893344402, + 0.013873533345758915, + 0.012825416401028633, + 0.013866265304386616, + 0.016901783645153046, + 0.013890902511775494, + 0.01715925708413124, + 0.015539578162133694, + 0.008624337613582611, + 0.007294397801160812, + 0.007944626733660698, + 0.008225226774811745, + 0.010808592662215233, + 0.00842638686299324, + 0.010519227012991905, + 0.008236492983996868, + 0.018623869866132736, + 0.013324272818863392, + 0.017249010503292084, + 0.012513609603047371, + 0.024752216413617134, + 0.012149744667112827, + 0.01336432434618473, + 0.010053903795778751, + 0.008851008489727974, + 0.00600793119519949, + 0.00946991890668869, + 0.00947667472064495, + 0.007652949076145887, + 0.006289258133620024, + 0.007987594231963158, + 0.005792154930531979, + 0.042526405304670334, + 0.04395256191492081, + 0.057773198932409286, + 0.06621868908405304, + 0.043171126395463943, + 0.04347224906086922, + 0.034302689135074615, + 0.04044413939118385 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 7, + 5 + ], + "selected_reward": -0.12704510986804962 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "This static, low-angle, extreme close-up image, rendered in ultra-detailed 4K resolution, focuses on a lone Coprinus comatus mushroom, sprouting from a dark, jagged crack in the damp sidewalk of a vast cityscape. The environment is bathed in a stunning cyberpunk aesthetic, with harsh, pulsating light from scorched burgundy and cyan neon signs reflecting off the damp asphalt, casting long, dramatic shadows. The mushroom's originally pristine white, furry cap, each delicate scale clearly visible, begins to rapidly decay. As the structure withers and its lower edge curls upward, its color meticulously shifts from white to a deep black. The cap dissolves into thick, viscous streams of black ink, shimmering beneath the vibrant city lights. These inky streams steadily flow downward, eventually forming heavy droplets that break off and fall, crashing onto a discarded, gleaming chrome plate below. Every impact is captured with crystal clarity, black liquid splattering into beautiful, chaotic patterns that starkly reflect the dust-soaked futuristic metropolis, creating a mesmerizing and melancholic atmosphere of organic decay within a synthetic world.", + "seed": 32015, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000021/group_10/index_6990_seed_32015_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": -0.23760966751724483, + "vlm_logprob": -0.23760966751724483, + "vlm_yes_no_margin": 2.999999988079071, + "vlm_token_credit": -0.23760966751724483, + "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 of a lone Coprinus comatus mushroom sprouting from a crack in a sidewalk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a vast cityscape with a cyberpunk aesthetic featuring burgundy and cyan neon lights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the mushroom's cap visibly shift in color from white to deep black as it decays?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the decaying mushroom cap dissolve into thick, viscous streams of black ink that flow downward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the black ink droplets fall and splash onto a discarded chrome plate below?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.0582821369171143 + ], + "opposite_logprob_tokens": [ + -0.4332820773124695 + ], + "logprob_reward": -1.0582821369171143, + "yes_no_margin_reward": -0.6250000596046448, + "token_credit_reward": -1.0582821369171143, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.08020106703042984, + 0.4351319372653961, + 0.033702552318573, + 0.08175215870141983, + 0.10612455010414124, + 0.013122190721333027, + 0.02804134041070938, + 0.16709284484386444, + 0.030008483678102493, + 0.024822870269417763 + ], + "credit_map": [ + 0.007075140718370676, + 0.010624749585986137, + 0.02081252820789814, + 0.008902071043848991, + 0.009677884168922901, + 0.007152029778808355, + 0.005510066635906696, + 0.010446601547300816, + 0.05149826407432556, + 0.07560466229915619, + 0.0654577687382698, + 0.03799983114004135, + 0.07882564514875412, + 0.04070072993636131, + 0.034317780286073685, + 0.0507272444665432, + 0.0025067180395126343, + 0.0034366720356047153, + 0.009081725031137466, + 0.0030093977693468332, + 0.003170893294736743, + 0.002802270231768489, + 0.00566111458465457, + 0.004033761564642191, + 0.00732269836589694, + 0.010641250759363174, + 0.008188340812921524, + 0.010937890969216824, + 0.009135729633271694, + 0.006132860202342272, + 0.014236001297831535, + 0.015157385729253292, + 0.02333475463092327, + 0.010628693737089634, + 0.013263501226902008, + 0.008923890069127083, + 0.009108534082770348, + 0.018387697637081146, + 0.009109953418374062, + 0.013367523439228535, + 0.0016839518211781979, + 0.0018969913944602013, + 0.0012789492029696703, + 0.0019691104535013437, + 0.0017943276325240731, + 0.0010740229627117515, + 0.0016449630493298173, + 0.0017798749031499028, + 0.0020890424493700266, + 0.0035563367418944836, + 0.00358747155405581, + 0.004863890819251537, + 0.0036998193245381117, + 0.003369950456544757, + 0.003033821005374193, + 0.0038410071283578873, + 0.03608548641204834, + 0.01346630696207285, + 0.02292824164032936, + 0.010702943429350853, + 0.016109393909573555, + 0.032437555491924286, + 0.011254195123910904, + 0.024108711630105972, + 0.00256805750541389, + 0.003436920465901494, + 0.004692986607551575, + 0.0027930336073040962, + 0.005102123599499464, + 0.0037128927651792765, + 0.0032619209960103035, + 0.004440547898411751, + 0.0015607003588229418, + 0.0033098976127803326, + 0.003022077027708292, + 0.002542744157835841, + 0.004322955384850502, + 0.002457862254232168, + 0.0025063229259103537, + 0.0051003098487854 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 7, + 4 + ], + "selected_reward": -1.0582821369171143 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 lone Coprinus comatus mushroom sprouting from a crack in a sidewalk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a vast cityscape with a cyberpunk aesthetic featuring burgundy and cyan neon lights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the mushroom's cap visibly shift in color from white to deep black as it decays?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the decaying mushroom cap dissolve into thick, viscous streams of black ink that flow downward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the black ink droplets fall and splash onto a discarded chrome plate below?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009801830165088177 + ], + "opposite_logprob_tokens": [ + -4.634801864624023 + ], + "logprob_reward": -0.009801830165088177, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009801830165088177, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.176352396607399, + 0.15098530054092407, + 0.07554160058498383, + 0.09801452606916428, + 0.03155764564871788, + 0.031374044716358185, + 0.037907674908638, + 0.35680118203163147, + 0.020623333752155304, + 0.02084229327738285 + ], + "credit_map": [ + 0.017965728417038918, + 0.024216335266828537, + 0.025360964238643646, + 0.030651969835162163, + 0.025893086567521095, + 0.015228060074150562, + 0.01808115653693676, + 0.018955102190375328, + 0.017697256058454514, + 0.022753482684493065, + 0.03106859140098095, + 0.013074049726128578, + 0.022115403786301613, + 0.011752358637750149, + 0.011665758676826954, + 0.020858393982052803, + 0.007906455546617508, + 0.006885108537971973, + 0.013688827864825726, + 0.005259516183286905, + 0.006689354311674833, + 0.010793301276862621, + 0.015081355348229408, + 0.009237682446837425, + 0.008306263014674187, + 0.012589247897267342, + 0.012570574879646301, + 0.015400426462292671, + 0.010959699749946594, + 0.011041085235774517, + 0.013420822098851204, + 0.013726405799388885, + 0.0034201836679130793, + 0.004013495519757271, + 0.003687883261591196, + 0.003798599587753415, + 0.004089157097041607, + 0.003920297604054213, + 0.004378515295684338, + 0.004249512683600187, + 0.004555842373520136, + 0.003428040072321892, + 0.0028068532701581717, + 0.003310748841613531, + 0.003908226266503334, + 0.005746165290474892, + 0.0034668438602238894, + 0.004151326138526201, + 0.005280671641230583, + 0.003970012068748474, + 0.004697325639426708, + 0.005171690136194229, + 0.004090680740773678, + 0.0058250706642866135, + 0.0037391202058643103, + 0.005133104044944048, + 0.07974139600992203, + 0.028005050495266914, + 0.049765024334192276, + 0.023079462349414825, + 0.033716484904289246, + 0.06872224807739258, + 0.024038827046751976, + 0.04973268881440163, + 0.0029513284098356962, + 0.002556375227868557, + 0.0027237385511398315, + 0.0019026114605367184, + 0.0026773083955049515, + 0.002086867345497012, + 0.0026455523911863565, + 0.0030795522034168243, + 0.0021389417815953493, + 0.002102479338645935, + 0.002555910963565111, + 0.0017457744106650352, + 0.0026783302891999483, + 0.0021314008627086878, + 0.002344652311876416, + 0.005144804250448942 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 1 + ], + "selected_reward": -0.009801830165088177 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a lone Coprinus comatus mushroom sprouting from a crack in a sidewalk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a vast cityscape with a cyberpunk aesthetic featuring burgundy and cyan neon lights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the mushroom's cap visibly shift in color from white to deep black as it decays?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the decaying mushroom cap dissolve into thick, viscous streams of black ink that flow downward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the black ink droplets fall and splash onto a discarded chrome plate below?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023302050307393074 + ], + "opposite_logprob_tokens": [ + -3.7733020782470703 + ], + "logprob_reward": -0.023302050307393074, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023302050307393074, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06674075126647949, + 0.33834123611450195, + 0.04170311242341995, + 0.06777611374855042, + 0.0503133125603199, + 0.03804833069443703, + 0.02596607245504856, + 0.3080788254737854, + 0.04598435014486313, + 0.01704786717891693 + ], + "credit_map": [ + 0.009406105615198612, + 0.008345827460289001, + 0.01040889136493206, + 0.007864311337471008, + 0.007764390204101801, + 0.009135941974818707, + 0.005619415547698736, + 0.008195869624614716, + 0.04031028226017952, + 0.06398465484380722, + 0.04083527624607086, + 0.030394582077860832, + 0.06708730757236481, + 0.03464435413479805, + 0.027541393414139748, + 0.033543385565280914, + 0.0036843104753643274, + 0.006002188194543123, + 0.005755475722253323, + 0.0038110220339149237, + 0.0036230392288416624, + 0.005423830822110176, + 0.007145212963223457, + 0.006258026696741581, + 0.004212069325149059, + 0.009940996766090393, + 0.004955232609063387, + 0.01053551770746708, + 0.01031810138374567, + 0.006098231300711632, + 0.011222586035728455, + 0.010493370704352856, + 0.010635649785399437, + 0.00522974319756031, + 0.007191210053861141, + 0.004548749420791864, + 0.004376617725938559, + 0.007653550710529089, + 0.004224680829793215, + 0.006453110370784998, + 0.004770082887262106, + 0.004274522885680199, + 0.0054911780171096325, + 0.00539062637835741, + 0.003805270418524742, + 0.003883844008669257, + 0.003374414285644889, + 0.007058385759592056, + 0.003576982533559203, + 0.0030440466944128275, + 0.0031279143877327442, + 0.0029118398670107126, + 0.0028503567446023226, + 0.004058797378093004, + 0.0026482720859348774, + 0.0037478613667190075, + 0.06781008094549179, + 0.024477481842041016, + 0.04184861481189728, + 0.018799157813191414, + 0.03082164376974106, + 0.06085857003927231, + 0.020789405331015587, + 0.04267387092113495, + 0.0044761658646166325, + 0.004626653157174587, + 0.005872062407433987, + 0.0043038842268288136, + 0.009999135509133339, + 0.005709667690098286, + 0.005344110075384378, + 0.005652667488902807, + 0.001739415223710239, + 0.0020544580183923244, + 0.0021548327058553696, + 0.0015307243447750807, + 0.0030715588945895433, + 0.0016770868096500635, + 0.0020048508886247873, + 0.002814937848597765 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 7, + 3 + ], + "selected_reward": -0.023302050307393074 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 lone Coprinus comatus mushroom sprouting from a crack in a sidewalk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a vast cityscape with a cyberpunk aesthetic featuring burgundy and cyan neon lights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the mushroom's cap visibly shift in color from white to deep black as it decays?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the decaying mushroom cap dissolve into thick, viscous streams of black ink that flow downward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the black ink droplets fall and splash onto a discarded chrome plate below?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007629185914993286 + ], + "opposite_logprob_tokens": [ + -4.88262939453125 + ], + "logprob_reward": -0.007629185914993286, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007629185914993286, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09949910640716553, + 0.2426193356513977, + 0.02405896596610546, + 0.05457627400755882, + 0.2542695701122284, + 0.016543027013540268, + 0.01722842827439308, + 0.24424023926258087, + 0.02889110893011093, + 0.018073929473757744 + ], + "credit_map": [ + 0.009829393588006496, + 0.011261647567152977, + 0.0196559876203537, + 0.015244591049849987, + 0.01146006491035223, + 0.01151102501899004, + 0.009235376492142677, + 0.011301016435027122, + 0.02161462791264057, + 0.02866290509700775, + 0.05412320792675018, + 0.018272148445248604, + 0.030820945277810097, + 0.017449617385864258, + 0.017610251903533936, + 0.05406561493873596, + 0.0020008368883281946, + 0.004136850591748953, + 0.0037828877102583647, + 0.002325810259208083, + 0.0028174410108476877, + 0.002688597422093153, + 0.002812805352732539, + 0.0034937357995659113, + 0.006517407018691301, + 0.008002325892448425, + 0.004458510782569647, + 0.008541998453438282, + 0.006600405089557171, + 0.005047873593866825, + 0.008557173423469067, + 0.006850580219179392, + 0.05732345953583717, + 0.027820324525237083, + 0.03296220675110817, + 0.022705398499965668, + 0.02076452597975731, + 0.04263400286436081, + 0.019935034215450287, + 0.030124634504318237, + 0.0025107366964221, + 0.0021957566495984793, + 0.002278649713844061, + 0.0018808564636856318, + 0.0017410877626389265, + 0.001960906432941556, + 0.0016350022051483393, + 0.002340030623599887, + 0.002405736828222871, + 0.0018166614463552833, + 0.0023760925978422165, + 0.0024842924904078245, + 0.0015533191617578268, + 0.0024110255762934685, + 0.001772149815224111, + 0.002409151289612055, + 0.055171020328998566, + 0.018785299733281136, + 0.03374491631984711, + 0.015087856911122799, + 0.02296692691743374, + 0.04813618212938309, + 0.016470205038785934, + 0.03387785330414772, + 0.0023421519435942173, + 0.0036200121976435184, + 0.0031944240909069777, + 0.0032651936635375023, + 0.005796067416667938, + 0.0030667195096611977, + 0.0036109518259763718, + 0.003995588980615139, + 0.0013668769970536232, + 0.0024638394825160503, + 0.002220120979472995, + 0.0019470301922410727, + 0.002759290626272559, + 0.0017243832116946578, + 0.0018130299868062139, + 0.0037793563678860664 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 7, + 1 + ], + "selected_reward": -0.007629185914993286 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 lone Coprinus comatus mushroom sprouting from a crack in a sidewalk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a vast cityscape with a cyberpunk aesthetic featuring burgundy and cyan neon lights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the mushroom's cap visibly shift in color from white to deep black as it decays?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the decaying mushroom cap dissolve into thick, viscous streams of black ink that flow downward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the black ink droplets fall and splash onto a discarded chrome plate below?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08903313428163528 + ], + "opposite_logprob_tokens": [ + -2.4640331268310547 + ], + "logprob_reward": -0.08903313428163528, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08903313428163528, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07886531203985214, + 0.27714431285858154, + 0.045807551592588425, + 0.06759367883205414, + 0.22172406315803528, + 0.015986472368240356, + 0.03153355047106743, + 0.22090813517570496, + 0.01641499251127243, + 0.024021942168474197 + ], + "credit_map": [ + 0.009572876617312431, + 0.010589484125375748, + 0.009903778322041035, + 0.01583455130457878, + 0.010814685374498367, + 0.007053061854094267, + 0.007830236107110977, + 0.007266638800501823, + 0.03286932781338692, + 0.052125148475170135, + 0.03128264844417572, + 0.023748811334371567, + 0.05432652682065964, + 0.028895746916532516, + 0.02167980931699276, + 0.03221631795167923, + 0.00444902665913105, + 0.005627151113003492, + 0.009067460894584656, + 0.0029435008764266968, + 0.005555956624448299, + 0.005915082059800625, + 0.007486187852919102, + 0.0047631822526454926, + 0.007370661944150925, + 0.009991667233407497, + 0.008655041456222534, + 0.007381000556051731, + 0.007512221112847328, + 0.007861408405005932, + 0.008443538099527359, + 0.01037813350558281, + 0.04637467488646507, + 0.02640017867088318, + 0.031903110444545746, + 0.02226279117166996, + 0.01843332126736641, + 0.034043971449136734, + 0.01785970665514469, + 0.02444630116224289, + 0.0014205769402906299, + 0.0018923599272966385, + 0.002713122172281146, + 0.0023366541136056185, + 0.0017736322479322553, + 0.0019460864132270217, + 0.0018512089736759663, + 0.0020528328604996204, + 0.002889305353164673, + 0.0032299079466611147, + 0.004942771513015032, + 0.004703608341515064, + 0.0033760485239326954, + 0.004801159258931875, + 0.003235942218452692, + 0.004354808945208788, + 0.04870091751217842, + 0.017366627231240273, + 0.033798426389694214, + 0.013421410694718361, + 0.019372694194316864, + 0.0426139160990715, + 0.014995867386460304, + 0.030638262629508972, + 0.0029176052194088697, + 0.0017888305010274053, + 0.0024555823765695095, + 0.0013945354148745537, + 0.002435344271361828, + 0.0019832307007163763, + 0.0015716450288891792, + 0.0018682185327634215, + 0.0020326084922999144, + 0.003163550514727831, + 0.0027525057084858418, + 0.0033266106620430946, + 0.003461372572928667, + 0.0025308048352599144, + 0.0033634991850703955, + 0.0033909897319972515 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 4, + 7 + ], + "selected_reward": -0.08903313428163528 + } + ], + "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/0000021/group_11/index_tensor([856, 856])_group_11_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000021/group_11/index_tensor([856, 856])_group_11_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..4af77351da2f0d99836c845225380c3569831e8d --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000021/group_11/index_tensor([856, 856])_group_11_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "high-angle, sweeping wide shot slowly descends and advances across a vast, nearly deserted public square paved with weathered cobblestones, taking on a grand and majestic scale. The soft, late afternoon sunlight casts long, dramatic shadows from the surrounding classical stone buildings, bathing the scene in a serene golden hue. In the center of the square stands a massive equestrian statue depicting a heroic figure in armor, its ancient bronze surface riddled with green veins. As the camera continues its deliberate, steady approach, transitioning to a medium shot, a deeply unsettling detail becomes apparent: the entire statue is trembling, accompanied by subtle, low-frequency vibrations. The bronze's dull metallic luster is disturbed by these constant, minute vibrations, causing golden light to shimmer and ripple across its surface as if it were liquid. A layer of fine dust on the thick granite base trembles and dances with the unsettling vibrations that seem to emanate from the metal's very core, creating an tense and surreal atmosphere that contrasts starkly with the serene beauty of the surroundings.", + "seed": 10283, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000021/group_11/index_856_seed_10283_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": -0.10501961489208042, + "vlm_logprob": -0.10501961489208042, + "vlm_yes_no_margin": 3.549999988079071, + "vlm_token_credit": -0.10501961489208042, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 vast public square paved with weathered cobblestones surrounded by classical stone buildings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive equestrian statue of a heroic figure in armor visible in the center of the square?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the statue have an ancient bronze surface with green veins and a dull metallic luster?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera starting with a high-angle sweeping wide shot that descends and advances before transitioning to a medium shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the entire statue visibly trembling with low-frequency vibrations that cause the bronze surface to shimmer and ripple?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018691591918468475 + ], + "opposite_logprob_tokens": [ + -4.518691539764404 + ], + "logprob_reward": -0.018691591918468475, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.018691591918468475, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09978608042001724, + 0.12404775619506836, + 0.0537397675216198, + 0.0734633207321167, + 0.09141528606414795, + 0.14163461327552795, + 0.12126728892326355, + 0.07759267836809158, + 0.08904970437288284, + 0.12800349295139313 + ], + "credit_map": [ + 0.010857210494577885, + 0.012708090245723724, + 0.01306137628853321, + 0.010813389904797077, + 0.016975242644548416, + 0.013182192109525204, + 0.011053730733692646, + 0.011134848929941654, + 0.01164035964757204, + 0.008997289463877678, + 0.02849521115422249, + 0.009215394966304302, + 0.006963765248656273, + 0.017993168905377388, + 0.007754210848361254, + 0.03298835828900337, + 0.007241822313517332, + 0.006324261426925659, + 0.01052248477935791, + 0.003638667520135641, + 0.005658049136400223, + 0.007547345012426376, + 0.0037478813901543617, + 0.009059257805347443, + 0.012302027083933353, + 0.008604377508163452, + 0.01458700094372034, + 0.005600460804998875, + 0.00560236070305109, + 0.011145457625389099, + 0.004843337461352348, + 0.010778296738862991, + 0.014805546030402184, + 0.006507477257400751, + 0.019846811890602112, + 0.007168247364461422, + 0.007273468188941479, + 0.009942613542079926, + 0.010037004016339779, + 0.01583411730825901, + 0.019810384139418602, + 0.011108417063951492, + 0.022214461117982864, + 0.015033131465315819, + 0.01319892518222332, + 0.01628095470368862, + 0.02362951822578907, + 0.020358815789222717, + 0.01218113861978054, + 0.008667794056236744, + 0.021266860887408257, + 0.006891882978379726, + 0.009822414256632328, + 0.018423335626721382, + 0.007387496996670961, + 0.03662636876106262, + 0.016309093683958054, + 0.006404206156730652, + 0.011754190549254417, + 0.004656218457967043, + 0.005993008613586426, + 0.012555938214063644, + 0.005323807243257761, + 0.014596213586628437, + 0.010148541070520878, + 0.008835469372570515, + 0.011326576583087444, + 0.012051074765622616, + 0.008833890780806541, + 0.015808144584298134, + 0.00858825072646141, + 0.01345774345099926, + 0.009781244210898876, + 0.009284790605306625, + 0.018198497593402863, + 0.00891220010817051, + 0.008692210540175438, + 0.04260074347257614, + 0.00996122881770134, + 0.020572585985064507 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 9, + 1 + ], + "selected_reward": -0.018691591918468475 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 vast public square paved with weathered cobblestones surrounded by classical stone buildings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive equestrian statue of a heroic figure in armor visible in the center of the square?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the statue have an ancient bronze surface with green veins and a dull metallic luster?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera starting with a high-angle sweeping wide shot that descends and advances before transitioning to a medium shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the entire statue visibly trembling with low-frequency vibrations that cause the bronze surface to shimmer and ripple?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002508709440007806 + ], + "opposite_logprob_tokens": [ + -6.002508640289307 + ], + "logprob_reward": -0.002508709440007806, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.002508709440007806, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1122557669878006, + 0.10490565001964569, + 0.10097750276327133, + 0.07186456769704819, + 0.06617844104766846, + 0.10973458737134933, + 0.09915080666542053, + 0.10832657665014267, + 0.11720313131809235, + 0.10940296947956085 + ], + "credit_map": [ + 0.01231923047453165, + 0.010424243286252022, + 0.013194163329899311, + 0.020029494538903236, + 0.01606377214193344, + 0.00830396730452776, + 0.020639335736632347, + 0.011281559243798256, + 0.01016763225197792, + 0.012724125757813454, + 0.02172933705151081, + 0.012103759683668613, + 0.011918766424059868, + 0.0071794274263083935, + 0.010105759836733341, + 0.0189768448472023, + 0.007043900899589062, + 0.013796873390674591, + 0.020768946036696434, + 0.01604590192437172, + 0.006711204536259174, + 0.006882160436362028, + 0.010911923833191395, + 0.018816588446497917, + 0.006382741499692202, + 0.00752074783667922, + 0.016388339921832085, + 0.009109430015087128, + 0.005007910076528788, + 0.008518239483237267, + 0.005025690421462059, + 0.013911469839513302, + 0.007098563946783543, + 0.005360888782888651, + 0.016219643875956535, + 0.004544323310256004, + 0.007157692685723305, + 0.0061336359940469265, + 0.00397915206849575, + 0.015684539452195168, + 0.009553490206599236, + 0.015434286557137966, + 0.013398168608546257, + 0.022544359788298607, + 0.009675376117229462, + 0.012932252138853073, + 0.012669273652136326, + 0.013527381233870983, + 0.01406826265156269, + 0.006361916661262512, + 0.011354806832969189, + 0.0069718011654913425, + 0.008587492629885674, + 0.014464040286839008, + 0.006389235612004995, + 0.030953241512179375, + 0.011132815852761269, + 0.007023296784609556, + 0.027476074174046516, + 0.008129516616463661, + 0.007865125313401222, + 0.017846720293164253, + 0.009842971339821815, + 0.019010059535503387, + 0.011554821394383907, + 0.00824830960482359, + 0.014353645034134388, + 0.012790560722351074, + 0.012117614038288593, + 0.030635809525847435, + 0.007966467179358006, + 0.019535908475518227, + 0.008075165562331676, + 0.007553064730018377, + 0.03640438988804817, + 0.0039968411438167095, + 0.0036809251178056, + 0.02275848761200905, + 0.0030331932939589024, + 0.02390090376138687 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 5 + ], + "selected_reward": -0.002508709440007806 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 vast public square paved with weathered cobblestones surrounded by classical stone buildings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive equestrian statue of a heroic figure in armor visible in the center of the square?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the statue have an ancient bronze surface with green veins and a dull metallic luster?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera starting with a high-angle sweeping wide shot that descends and advances before transitioning to a medium shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the entire statue visibly trembling with low-frequency vibrations that cause the bronze surface to shimmer and ripple?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026352116838097572 + ], + "opposite_logprob_tokens": [ + -3.6513521671295166 + ], + "logprob_reward": -0.026352116838097572, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026352116838097572, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10567013919353485, + 0.15315507352352142, + 0.21276673674583435, + 0.03621331602334976, + 0.06308002024888992, + 0.14478978514671326, + 0.10496731847524643, + 0.06614518165588379, + 0.06817691773176193, + 0.04503554478287697 + ], + "credit_map": [ + 0.015572466887533665, + 0.0128213781863451, + 0.010943424887955189, + 0.010671768337488174, + 0.02209905907511711, + 0.013079113326966763, + 0.010501806624233723, + 0.00998111441731453, + 0.02036435529589653, + 0.016021469607949257, + 0.03119651786983013, + 0.01610906608402729, + 0.011888107284903526, + 0.018507279455661774, + 0.011373338289558887, + 0.02769492194056511, + 0.015234723687171936, + 0.022984579205513, + 0.047163691371679306, + 0.01749705709517002, + 0.025493133813142776, + 0.013902327977120876, + 0.017950942739844322, + 0.05254025757312775, + 0.003307062666863203, + 0.004472574684768915, + 0.006963465362787247, + 0.004492933861911297, + 0.0031921803019940853, + 0.0030384582933038473, + 0.0043142326176166534, + 0.006432404741644859, + 0.004945286083966494, + 0.006536933593451977, + 0.014888878911733627, + 0.004918741527944803, + 0.007496600039303303, + 0.005084995646029711, + 0.006705441977828741, + 0.012503137812018394, + 0.017203854396939278, + 0.013901501893997192, + 0.016903232783079147, + 0.023316966369748116, + 0.012313627637922764, + 0.02318229153752327, + 0.015217973850667477, + 0.022750334814190865, + 0.010779154486954212, + 0.010611386969685555, + 0.01415707916021347, + 0.007390834856778383, + 0.012476802803575993, + 0.015627441927790642, + 0.008640395477414131, + 0.02528422325849533, + 0.008263450115919113, + 0.0035548810847103596, + 0.015562089160084724, + 0.004737271461635828, + 0.005015852861106396, + 0.010812071152031422, + 0.0057954927906394005, + 0.012404073029756546, + 0.007336584851145744, + 0.006874276325106621, + 0.009448944590985775, + 0.007882801815867424, + 0.008335092104971409, + 0.01341736689209938, + 0.005412188358604908, + 0.0094696544110775, + 0.004151713103055954, + 0.004132598172873259, + 0.010851466096937656, + 0.004103412386029959, + 0.0040267352014780045, + 0.007514151744544506, + 0.0031872487161308527, + 0.00706821633502841 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 1, + 5 + ], + "selected_reward": -0.026352116838097572 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 vast public square paved with weathered cobblestones surrounded by classical stone buildings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive equestrian statue of a heroic figure in armor visible in the center of the square?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the statue have an ancient bronze surface with green veins and a dull metallic luster?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera starting with a high-angle sweeping wide shot that descends and advances before transitioning to a medium shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the entire statue visibly trembling with low-frequency vibrations that cause the bronze surface to shimmer and ripple?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04865012317895889 + ], + "opposite_logprob_tokens": [ + -3.048650026321411 + ], + "logprob_reward": -0.04865012317895889, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04865012317895889, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09695913642644882, + 0.13242600858211517, + 0.08693940937519073, + 0.14752201735973358, + 0.16842377185821533, + 0.09826619178056717, + 0.06424363702535629, + 0.08679603785276413, + 0.09380745142698288, + 0.02461639791727066 + ], + "credit_map": [ + 0.015717968344688416, + 0.010598394088447094, + 0.010441455990076065, + 0.011963472701609135, + 0.0161447674036026, + 0.01091818232089281, + 0.01117308996617794, + 0.010001800023019314, + 0.016193076968193054, + 0.011578014120459557, + 0.030672505497932434, + 0.01095147430896759, + 0.008107153698801994, + 0.01681813970208168, + 0.009512413293123245, + 0.028593221679329872, + 0.009277199395000935, + 0.011973023414611816, + 0.015033960342407227, + 0.008312798105180264, + 0.011371354572474957, + 0.008127763867378235, + 0.007610292639583349, + 0.015233014710247517, + 0.012790996581315994, + 0.021426385268568993, + 0.028073575347661972, + 0.017768722027540207, + 0.013645204715430737, + 0.014949513599276543, + 0.018581995740532875, + 0.020285623148083687, + 0.012965047731995583, + 0.02400035969913006, + 0.0260753370821476, + 0.013755668886005878, + 0.029457176104187965, + 0.017031649127602577, + 0.018114833161234856, + 0.027023697271943092, + 0.011495768092572689, + 0.015603240579366684, + 0.011942163109779358, + 0.011719047091901302, + 0.010517681017518044, + 0.014687660150229931, + 0.010125243104994297, + 0.012175386771559715, + 0.0073633319698274136, + 0.007622164208441973, + 0.010398699901998043, + 0.005824206862598658, + 0.00913372915238142, + 0.007307646330446005, + 0.006569613702595234, + 0.010024243034422398, + 0.014551611617207527, + 0.00547434575855732, + 0.016095008701086044, + 0.00507564190775156, + 0.007184205576777458, + 0.01460372842848301, + 0.006355587858706713, + 0.01745590567588806, + 0.008271617814898491, + 0.008213126100599766, + 0.008660389110445976, + 0.013562628999352455, + 0.00842115469276905, + 0.024065516889095306, + 0.006347597576677799, + 0.016265418380498886, + 0.0024100092705339193, + 0.0020686795469373465, + 0.002482056152075529, + 0.0029850956052541733, + 0.0022414170671254396, + 0.005523550324141979, + 0.0027222533244639635, + 0.004183335695415735 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 3, + 1 + ], + "selected_reward": -0.04865012317895889 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 vast public square paved with weathered cobblestones surrounded by classical stone buildings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive equestrian statue of a heroic figure in armor visible in the center of the square?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the statue have an ancient bronze surface with green veins and a dull metallic luster?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera starting with a high-angle sweeping wide shot that descends and advances before transitioning to a medium shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the entire statue visibly trembling with low-frequency vibrations that cause the bronze surface to shimmer and ripple?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.4288955330848694 + ], + "opposite_logprob_tokens": [ + -1.0538954734802246 + ], + "logprob_reward": -0.4288955330848694, + "yes_no_margin_reward": 0.6249999403953552, + "token_credit_reward": -0.4288955330848694, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12168291956186295, + 0.06572916358709335, + 0.04356439411640167, + 0.0525796115398407, + 0.11131525784730911, + 0.23693996667861938, + 0.1044217050075531, + 0.1100626215338707, + 0.05079851672053337, + 0.10290585458278656 + ], + "credit_map": [ + 0.020582472905516624, + 0.010158400982618332, + 0.013131050392985344, + 0.018806884065270424, + 0.014579271897673607, + 0.010736620053648949, + 0.01614142581820488, + 0.017546799033880234, + 0.01086188480257988, + 0.009272615425288677, + 0.011856002733111382, + 0.0058427355252206326, + 0.00559787405654788, + 0.007034700829535723, + 0.0059502399526536465, + 0.009313113056123257, + 0.005933496169745922, + 0.003966440446674824, + 0.006430082954466343, + 0.004595641512423754, + 0.005677259061485529, + 0.004735421855002642, + 0.004968137014657259, + 0.0072579151019454, + 0.007969856262207031, + 0.00738738477230072, + 0.008513856679201126, + 0.004262164235115051, + 0.004120596218854189, + 0.008627805858850479, + 0.004920592065900564, + 0.006777356844395399, + 0.009808661416172981, + 0.007198844105005264, + 0.03489089757204056, + 0.006025388371199369, + 0.007377085275948048, + 0.007587153930217028, + 0.00705693569034338, + 0.03137030079960823, + 0.02830839902162552, + 0.014563395641744137, + 0.03372709080576897, + 0.03248359262943268, + 0.017438452690839767, + 0.035958290100097656, + 0.0349465012550354, + 0.03951423987746239, + 0.012377272360026836, + 0.013486148789525032, + 0.015411295928061008, + 0.008864698931574821, + 0.015926113352179527, + 0.014020217582583427, + 0.009192842058837414, + 0.01514312345534563, + 0.013101981952786446, + 0.015118371695280075, + 0.016611779108643532, + 0.0074634249322116375, + 0.01154474075883627, + 0.0176040381193161, + 0.008789249695837498, + 0.019829044118523598, + 0.00473316153511405, + 0.005296058487147093, + 0.006232161074876785, + 0.008655941113829613, + 0.0048691933043301105, + 0.00812872126698494, + 0.004036423284560442, + 0.008846859447658062, + 0.0077059222385287285, + 0.008082976564764977, + 0.028910698369145393, + 0.005508649628609419, + 0.0058901370503008366, + 0.01877620443701744, + 0.004799327347427607, + 0.02323193848133087 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 4 + ], + "selected_reward": -0.4288955330848694 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "high-angle, sweeping wide shot slowly descends and advances across a vast, nearly deserted public square paved with weathered cobblestones, taking on a grand and majestic scale. The soft, late afternoon sunlight casts long, dramatic shadows from the surrounding classical stone buildings, bathing the scene in a serene golden hue. In the center of the square stands a massive equestrian statue depicting a heroic figure in armor, its ancient bronze surface riddled with green veins. As the camera continues its deliberate, steady approach, transitioning to a medium shot, a deeply unsettling detail becomes apparent: the entire statue is trembling, accompanied by subtle, low-frequency vibrations. The bronze's dull metallic luster is disturbed by these constant, minute vibrations, causing golden light to shimmer and ripple across its surface as if it were liquid. A layer of fine dust on the thick granite base trembles and dances with the unsettling vibrations that seem to emanate from the metal's very core, creating an tense and surreal atmosphere that contrasts starkly with the serene beauty of the surroundings.", + "seed": 10284, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000021/group_11/index_856_seed_10284_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": -0.22846194133162498, + "vlm_logprob": -0.22846194133162498, + "vlm_yes_no_margin": 2.074999988079071, + "vlm_token_credit": -0.22846194133162498, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 vast public square paved with weathered cobblestones surrounded by classical stone buildings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive equestrian statue of a heroic figure in armor visible in the center of the square?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the statue have an ancient bronze surface with green veins and a dull metallic luster?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera starting with a high-angle sweeping wide shot that descends and advances before transitioning to a medium shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the entire statue visibly trembling with low-frequency vibrations that cause the bronze surface to shimmer and ripple?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.4001642167568207 + ], + "opposite_logprob_tokens": [ + -1.150164246559143 + ], + "logprob_reward": -0.4001642167568207, + "yes_no_margin_reward": 0.75, + "token_credit_reward": -0.4001642167568207, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.46186479926109314, + 0.2060670554637909, + 0.049121737480163574, + 0.03393019363284111, + 0.01221494097262621, + 0.05976884067058563, + 0.050529565662145615, + 0.03544372320175171, + 0.061376627534627914, + 0.0296824648976326 + ], + "credit_map": [ + 0.07723260670900345, + 0.05004670098423958, + 0.05998601019382477, + 0.03365902602672577, + 0.06562929600477219, + 0.06556402146816254, + 0.04827617481350899, + 0.06147097796201706, + 0.01973530650138855, + 0.009201761335134506, + 0.06077934429049492, + 0.015258532017469406, + 0.010222387500107288, + 0.01984785497188568, + 0.010898618027567863, + 0.0601232685148716, + 0.011229956522583961, + 0.003366763237863779, + 0.0072079612873494625, + 0.00418611615896225, + 0.003272316185757518, + 0.010340221226215363, + 0.0026041909586638212, + 0.006914208177477121, + 0.004734978545457125, + 0.003856642171740532, + 0.004750507418066263, + 0.003949272911995649, + 0.0035949493758380413, + 0.005213534459471703, + 0.0029159123077988625, + 0.004914396908134222, + 0.001540748286060989, + 0.001191916991956532, + 0.0016329013742506504, + 0.0018086213385686278, + 0.0015082551399245858, + 0.0017142462311312556, + 0.0011351013090461493, + 0.001683149836026132, + 0.010487064719200134, + 0.008043782785534859, + 0.009289187379181385, + 0.006332149729132652, + 0.0066053858026862144, + 0.005729012656956911, + 0.0076743233948946, + 0.005607936531305313, + 0.012584058567881584, + 0.0042359367944300175, + 0.008150270208716393, + 0.00451701832935214, + 0.003720973851159215, + 0.005668829660862684, + 0.008069915696978569, + 0.003582565812394023, + 0.005628566723316908, + 0.0039724325761199, + 0.004247805569320917, + 0.002214346779510379, + 0.00497864605858922, + 0.005622229538857937, + 0.006036716513335705, + 0.0027429780457168818, + 0.005357249639928341, + 0.003364951815456152, + 0.021075356751680374, + 0.004504870623350143, + 0.0028850603848695755, + 0.0030797957442700863, + 0.01751747541129589, + 0.0035918685607612133, + 0.003485057968646288, + 0.003922092262655497, + 0.0038993549533188343, + 0.003077287459746003, + 0.003712612437084317, + 0.003292578272521496, + 0.005453584250062704, + 0.002839895896613598 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 8 + ], + "selected_reward": -0.4001642167568207 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 vast public square paved with weathered cobblestones surrounded by classical stone buildings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive equestrian statue of a heroic figure in armor visible in the center of the square?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the statue have an ancient bronze surface with green veins and a dull metallic luster?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera starting with a high-angle sweeping wide shot that descends and advances before transitioning to a medium shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the entire statue visibly trembling with low-frequency vibrations that cause the bronze surface to shimmer and ripple?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01110893115401268 + ], + "opposite_logprob_tokens": [ + -4.511108875274658 + ], + "logprob_reward": -0.01110893115401268, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.01110893115401268, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24204376339912415, + 0.3619731366634369, + 0.06408649682998657, + 0.08992982655763626, + 0.020824695006012917, + 0.07844968885183334, + 0.027518754824995995, + 0.049257904291152954, + 0.041693855077028275, + 0.02422194369137287 + ], + "credit_map": [ + 0.03840094432234764, + 0.024739664047956467, + 0.03242936730384827, + 0.021742619574069977, + 0.03342919051647186, + 0.03261948749423027, + 0.026544565334916115, + 0.032137881964445114, + 0.03595660999417305, + 0.016822539269924164, + 0.10240345448255539, + 0.028863904997706413, + 0.02134297601878643, + 0.033074088394641876, + 0.02267010323703289, + 0.10083941370248795, + 0.012493492104113102, + 0.008518950082361698, + 0.008308833464980125, + 0.007516221143305302, + 0.006651739589869976, + 0.0077139888890087605, + 0.006442895159125328, + 0.006440365687012672, + 0.012035285122692585, + 0.011407701298594475, + 0.011707221157848835, + 0.013439132831990719, + 0.010555143468081951, + 0.009867277927696705, + 0.010227184742689133, + 0.010690868832170963, + 0.0026331290137022734, + 0.0025247400626540184, + 0.0024783448316156864, + 0.0028155334293842316, + 0.003202515421435237, + 0.0022122471127659082, + 0.002870688447728753, + 0.0020874920301139355, + 0.014462980441749096, + 0.01180475763976574, + 0.012324674054980278, + 0.00855197012424469, + 0.00790436565876007, + 0.007829517126083374, + 0.008498220704495907, + 0.007073195651173592, + 0.006900832522660494, + 0.0028588552959263325, + 0.003789928974583745, + 0.0024657880421727896, + 0.0024266543332487345, + 0.0033915583044290543, + 0.0030970678199082613, + 0.002588064642623067, + 0.007424325682222843, + 0.005548014771193266, + 0.0061411443166434765, + 0.0035680318251252174, + 0.00673983059823513, + 0.008515767753124237, + 0.0079738674685359, + 0.0033469125628471375, + 0.006323083303868771, + 0.003549146233126521, + 0.006187213584780693, + 0.00405485276132822, + 0.0035724169574677944, + 0.004246470518410206, + 0.0061920518055558205, + 0.0075686099007725716, + 0.0029178184922784567, + 0.002697551157325506, + 0.0030856116209179163, + 0.0032212946098297834, + 0.0028828252106904984, + 0.003139892825856805, + 0.003418168518692255, + 0.0028587779961526394 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 3 + ], + "selected_reward": -0.01110893115401268 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 vast public square paved with weathered cobblestones surrounded by classical stone buildings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive equestrian statue of a heroic figure in armor visible in the center of the square?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the statue have an ancient bronze surface with green veins and a dull metallic luster?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera starting with a high-angle sweeping wide shot that descends and advances before transitioning to a medium shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the entire statue visibly trembling with low-frequency vibrations that cause the bronze surface to shimmer and ripple?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.033724263310432434 + ], + "opposite_logprob_tokens": [ + -3.408724308013916 + ], + "logprob_reward": -0.033724263310432434, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.033724263310432434, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.46283817291259766, + 0.10914348810911179, + 0.03926147148013115, + 0.03699348866939545, + 0.029794706031680107, + 0.07545871287584305, + 0.041956137865781784, + 0.07094325125217438, + 0.07807343453168869, + 0.05553722009062767 + ], + "credit_map": [ + 0.07418269664049149, + 0.05303644388914108, + 0.051944732666015625, + 0.028951482847332954, + 0.07060320675373077, + 0.07070925831794739, + 0.050849657505750656, + 0.06256067752838135, + 0.012109166011214256, + 0.009381582960486412, + 0.022575199604034424, + 0.013054972514510155, + 0.00958045944571495, + 0.0093691386282444, + 0.012889226898550987, + 0.020183730870485306, + 0.005142281763255596, + 0.007975815795361996, + 0.004472873639315367, + 0.0051191505044698715, + 0.004916641861200333, + 0.0028572119772434235, + 0.006106714718043804, + 0.002670777030289173, + 0.005016354378312826, + 0.005126604810357094, + 0.004910556133836508, + 0.003606280777603388, + 0.005759524181485176, + 0.003037603572010994, + 0.004375041462481022, + 0.00516152149066329, + 0.0030088392086327076, + 0.0030715009197592735, + 0.0036839277017861605, + 0.004713915754109621, + 0.0043733692727983, + 0.0039901225827634335, + 0.0037700477987527847, + 0.0031829781364649534, + 0.009048253297805786, + 0.012565272860229015, + 0.010562567971646786, + 0.011486598290503025, + 0.009364490397274494, + 0.009015986695885658, + 0.006650688126683235, + 0.0067648496478796005, + 0.006596709601581097, + 0.006025693845003843, + 0.005474770441651344, + 0.0031304783187806606, + 0.005729553755372763, + 0.007664498407393694, + 0.004210267215967178, + 0.003124161157757044, + 0.009193262085318565, + 0.00808921828866005, + 0.009683878161013126, + 0.0047982907854020596, + 0.008485643193125725, + 0.010923163965344429, + 0.014188242144882679, + 0.0055815428495407104, + 0.011368999257683754, + 0.008003883063793182, + 0.015733584761619568, + 0.006106515880674124, + 0.005734303966164589, + 0.006576623301953077, + 0.013293457217514515, + 0.011256064288318157, + 0.005396348424255848, + 0.007476686965674162, + 0.00872692558914423, + 0.006829071324318647, + 0.00644518481567502, + 0.00545489601790905, + 0.009527302347123623, + 0.0056807976216077805 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 8 + ], + "selected_reward": -0.033724263310432434 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 vast public square paved with weathered cobblestones surrounded by classical stone buildings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive equestrian statue of a heroic figure in armor visible in the center of the square?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the statue have an ancient bronze surface with green veins and a dull metallic luster?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera starting with a high-angle sweeping wide shot that descends and advances before transitioning to a medium shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the entire statue visibly trembling with low-frequency vibrations that cause the bronze surface to shimmer and ripple?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.3486517667770386 + ], + "opposite_logprob_tokens": [ + -1.2236517667770386 + ], + "logprob_reward": -0.3486517667770386, + "yes_no_margin_reward": 0.875, + "token_credit_reward": -0.3486517667770386, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4151667356491089, + 0.352794885635376, + 0.02375497855246067, + 0.03814546391367912, + 0.015928972512483597, + 0.029253289103507996, + 0.013675478287041187, + 0.017450379207730293, + 0.06572895497083664, + 0.02810082398355007 + ], + "credit_map": [ + 0.06189144775271416, + 0.05346072465181351, + 0.04461675509810448, + 0.026788249611854553, + 0.05753752216696739, + 0.06235097348690033, + 0.04568715766072273, + 0.06283389776945114, + 0.031201668083667755, + 0.01835997775197029, + 0.10014566779136658, + 0.03165208920836449, + 0.02308390475809574, + 0.03031359240412712, + 0.020976414903998375, + 0.09706158190965652, + 0.0037965322844684124, + 0.003554942784830928, + 0.002681389218196273, + 0.0030297019984573126, + 0.002908496418967843, + 0.0025798380374908447, + 0.0030708243139088154, + 0.00213325466029346, + 0.006426407489925623, + 0.005785203538835049, + 0.004996578674763441, + 0.00407096603885293, + 0.004030429758131504, + 0.004036430269479752, + 0.0038135042414069176, + 0.004985940642654896, + 0.0019222411792725325, + 0.001687851850874722, + 0.001967690186575055, + 0.0028118265327066183, + 0.0019787519704550505, + 0.0015639864141121507, + 0.0018691103905439377, + 0.0021275137551128864, + 0.003495902754366398, + 0.0036828494630753994, + 0.004085929598659277, + 0.003364103613421321, + 0.003848091932013631, + 0.004395590163767338, + 0.002595550613477826, + 0.003785270731896162, + 0.0018518506549298763, + 0.0015743110561743379, + 0.001948711578734219, + 0.001069903140887618, + 0.0014576497487723827, + 0.002196329180151224, + 0.001852695713751018, + 0.0017240269808098674, + 0.0023297974839806557, + 0.002750196261331439, + 0.0023491359315812588, + 0.000839544809423387, + 0.002351271454244852, + 0.0021698232740163803, + 0.003199470927938819, + 0.0014611377846449614, + 0.006484683137387037, + 0.005807545036077499, + 0.02052159234881401, + 0.003979329951107502, + 0.0046407850459218025, + 0.004135603550821543, + 0.015681743621826172, + 0.004477672278881073, + 0.003458403516560793, + 0.004091027192771435, + 0.003318127943202853, + 0.0036815679632127285, + 0.004165117163211107, + 0.002745198318734765, + 0.00431494414806366, + 0.0023264368064701557 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 8 + ], + "selected_reward": -0.3486517667770386 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 vast public square paved with weathered cobblestones surrounded by classical stone buildings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive equestrian statue of a heroic figure in armor visible in the center of the square?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the statue have an ancient bronze surface with green veins and a dull metallic luster?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera starting with a high-angle sweeping wide shot that descends and advances before transitioning to a medium shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the entire statue visibly trembling with low-frequency vibrations that cause the bronze surface to shimmer and ripple?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.34866052865982056 + ], + "opposite_logprob_tokens": [ + -1.2236604690551758 + ], + "logprob_reward": -0.34866052865982056, + "yes_no_margin_reward": 0.8749999403953552, + "token_credit_reward": -0.34866052865982056, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4152242839336395, + 0.14415539801120758, + 0.03806724771857262, + 0.0649389997124672, + 0.023040523752570152, + 0.08976716548204422, + 0.023848960176110268, + 0.05450554937124252, + 0.08304809033870697, + 0.06340381503105164 + ], + "credit_map": [ + 0.06263411790132523, + 0.050560206174850464, + 0.04460331052541733, + 0.025982173159718513, + 0.06334593892097473, + 0.0623319149017334, + 0.0476221889257431, + 0.05814444646239281, + 0.016100972890853882, + 0.009549037553369999, + 0.03159742429852486, + 0.013910681009292603, + 0.010683037340641022, + 0.016139866784214973, + 0.01336030662059784, + 0.03281407058238983, + 0.005041787400841713, + 0.0056961639784276485, + 0.0042990827932953835, + 0.005220584571361542, + 0.004757083486765623, + 0.0035404022783041, + 0.006437151692807674, + 0.003074991051107645, + 0.009138455614447594, + 0.008922476321458817, + 0.008434928022325039, + 0.009169300086796284, + 0.007845762185752392, + 0.006646518129855394, + 0.008661475032567978, + 0.006120085250586271, + 0.003516026074066758, + 0.0029868807177990675, + 0.0028042097110301256, + 0.003297315677627921, + 0.0031839178409427404, + 0.0024926296900957823, + 0.0022545920219272375, + 0.002504951786249876, + 0.010956540703773499, + 0.015285667963325977, + 0.011576244607567787, + 0.010323649272322655, + 0.012678083032369614, + 0.015483242459595203, + 0.00537429703399539, + 0.00808944646269083, + 0.004915987607091665, + 0.0031590028665959835, + 0.0036272683646529913, + 0.002319777850061655, + 0.0020228768698871136, + 0.0026970100589096546, + 0.0029402298387140036, + 0.0021668055560439825, + 0.009174293838441372, + 0.00642790412530303, + 0.0070126657374203205, + 0.0035141287371516228, + 0.00664732838049531, + 0.00810256414115429, + 0.009420270100235939, + 0.004206395708024502, + 0.01327511202543974, + 0.008939811959862709, + 0.018401699140667915, + 0.007383799180388451, + 0.007046043407171965, + 0.005601586773991585, + 0.016017943620681763, + 0.006382100284099579, + 0.005517409183084965, + 0.009137128479778767, + 0.007695105858147144, + 0.006616669241338968, + 0.007976522669196129, + 0.010118435136973858, + 0.009691421873867512, + 0.006651120726019144 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 5 + ], + "selected_reward": -0.34866052865982056 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "high-angle, sweeping wide shot slowly descends and advances across a vast, nearly deserted public square paved with weathered cobblestones, taking on a grand and majestic scale. The soft, late afternoon sunlight casts long, dramatic shadows from the surrounding classical stone buildings, bathing the scene in a serene golden hue. In the center of the square stands a massive equestrian statue depicting a heroic figure in armor, its ancient bronze surface riddled with green veins. As the camera continues its deliberate, steady approach, transitioning to a medium shot, a deeply unsettling detail becomes apparent: the entire statue is trembling, accompanied by subtle, low-frequency vibrations. The bronze's dull metallic luster is disturbed by these constant, minute vibrations, causing golden light to shimmer and ripple across its surface as if it were liquid. A layer of fine dust on the thick granite base trembles and dances with the unsettling vibrations that seem to emanate from the metal's very core, creating an tense and surreal atmosphere that contrasts starkly with the serene beauty of the surroundings.", + "seed": 10285, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000021/group_11/index_856_seed_10285_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": -0.1736407306045294, + "vlm_logprob": -0.1736407306045294, + "vlm_yes_no_margin": 2.35, + "vlm_token_credit": -0.1736407306045294, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 vast public square paved with weathered cobblestones surrounded by classical stone buildings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive equestrian statue of a heroic figure in armor visible in the center of the square?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the statue have an ancient bronze surface with green veins and a dull metallic luster?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera starting with a high-angle sweeping wide shot that descends and advances before transitioning to a medium shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the entire statue visibly trembling with low-frequency vibrations that cause the bronze surface to shimmer and ripple?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.14424388110637665 + ], + "opposite_logprob_tokens": [ + -2.1442439556121826 + ], + "logprob_reward": -0.14424388110637665, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.14424388110637665, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.522209107875824, + 0.1928318589925766, + 0.03515481576323509, + 0.0994022861123085, + 0.025028850883245468, + 0.009267977438867092, + 0.02066776528954506, + 0.030759038403630257, + 0.04535543546080589, + 0.019322911277413368 + ], + "credit_map": [ + 0.05843076482415199, + 0.06724946200847626, + 0.07438396662473679, + 0.04136662557721138, + 0.0657723918557167, + 0.0719982162117958, + 0.046798352152109146, + 0.09620930254459381, + 0.026777509599924088, + 0.01585821807384491, + 0.041272491216659546, + 0.01958819292485714, + 0.014426402747631073, + 0.019691085442900658, + 0.018047111108899117, + 0.03717084228992462, + 0.0033932975493371487, + 0.0029524543788284063, + 0.010366871021687984, + 0.003197294194251299, + 0.002396452007815242, + 0.0032838317565619946, + 0.002829653210937977, + 0.006734958849847317, + 0.011950886808335781, + 0.0067496392875909805, + 0.027493055909872055, + 0.014545531943440437, + 0.004311438649892807, + 0.010065080597996712, + 0.0066383914090693, + 0.017648249864578247, + 0.0027533124666661024, + 0.002170613734051585, + 0.005949386395514011, + 0.00172961107455194, + 0.0015189106343314052, + 0.004046454094350338, + 0.0013917883625254035, + 0.005468773189932108, + 0.0012162326602265239, + 0.001139564672484994, + 0.0014894242631271482, + 0.001134285586886108, + 0.0010192834306508303, + 0.0007789934752508998, + 0.0010739878052845597, + 0.0014162055449560285, + 0.003330148523673415, + 0.0041606188751757145, + 0.0021530534140765667, + 0.002324778586626053, + 0.004052901640534401, + 0.0017020232044160366, + 0.0013520271750167012, + 0.0015922138700261712, + 0.0026835440658032894, + 0.004167383536696434, + 0.0033788266591727734, + 0.004682039842009544, + 0.004340199287980795, + 0.0033266793470829725, + 0.005209808237850666, + 0.0029705550987273455, + 0.0034150006249547005, + 0.005404611583799124, + 0.011518415063619614, + 0.007252637762576342, + 0.005213210359215736, + 0.0035278366412967443, + 0.005109126679599285, + 0.003914592321962118, + 0.001713853213004768, + 0.004222908988595009, + 0.0019359355792403221, + 0.0014920500107109547, + 0.0019605904817581177, + 0.0023943327832967043, + 0.003134892089292407, + 0.0024683440569788218 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 3 + ], + "selected_reward": -0.14424388110637665 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 vast public square paved with weathered cobblestones surrounded by classical stone buildings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive equestrian statue of a heroic figure in armor visible in the center of the square?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the statue have an ancient bronze surface with green veins and a dull metallic luster?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera starting with a high-angle sweeping wide shot that descends and advances before transitioning to a medium shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the entire statue visibly trembling with low-frequency vibrations that cause the bronze surface to shimmer and ripple?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020611969754099846 + ], + "opposite_logprob_tokens": [ + -3.8956120014190674 + ], + "logprob_reward": -0.020611969754099846, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020611969754099846, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5782428979873657, + 0.08017919957637787, + 0.025486549362540245, + 0.03376717120409012, + 0.05651526525616646, + 0.007505254354327917, + 0.028955653309822083, + 0.04400838539004326, + 0.056358855217695236, + 0.08898069709539413 + ], + "credit_map": [ + 0.06181486323475838, + 0.053569890558719635, + 0.09477560222148895, + 0.05444129556417465, + 0.07983668148517609, + 0.06832310557365417, + 0.06814426928758621, + 0.09733718633651733, + 0.009276369586586952, + 0.008783881552517414, + 0.015687130391597748, + 0.007966822013258934, + 0.006961185950785875, + 0.008787589147686958, + 0.007988172583281994, + 0.01472805067896843, + 0.0019864938221871853, + 0.002764703007414937, + 0.005638619884848595, + 0.0026588416658341885, + 0.002310331678017974, + 0.002287955256178975, + 0.0038286743219941854, + 0.004010929726064205, + 0.004142195917665958, + 0.0023807547986507416, + 0.010280556976795197, + 0.003443369874730706, + 0.0022576111368834972, + 0.0034283322747796774, + 0.0016372452955693007, + 0.006197101902216673, + 0.004456411115825176, + 0.0054658809676766396, + 0.011799881234765053, + 0.005209131632000208, + 0.005562801379710436, + 0.005478405859321356, + 0.005811392329633236, + 0.012731354683637619, + 0.0009122684132307768, + 0.0010412728879600763, + 0.0010094130411744118, + 0.0009827213361859322, + 0.0008242320036515594, + 0.000811986334156245, + 0.0008301771013066173, + 0.0010931836441159248, + 0.005347320344299078, + 0.006369998212903738, + 0.0030661553610116243, + 0.0036305589601397514, + 0.003721626242622733, + 0.002592627424746752, + 0.001683238078840077, + 0.002544128568843007, + 0.0036758105270564556, + 0.0062099844217300415, + 0.004376541823148727, + 0.006500807125121355, + 0.0066034793853759766, + 0.004570394288748503, + 0.007762144785374403, + 0.0043092225678265095, + 0.003452154342085123, + 0.006990205962210894, + 0.01049777865409851, + 0.006702763959765434, + 0.006920501124113798, + 0.009506182745099068, + 0.006288896314799786, + 0.006000373046845198, + 0.00893071573227644, + 0.018423201516270638, + 0.010381869040429592, + 0.007795579265803099, + 0.010636929422616959, + 0.011237001977860928, + 0.011879941448569298, + 0.009695456363260746 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 1 + ], + "selected_reward": -0.020611969754099846 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 vast public square paved with weathered cobblestones surrounded by classical stone buildings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive equestrian statue of a heroic figure in armor visible in the center of the square?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the statue have an ancient bronze surface with green veins and a dull metallic luster?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera starting with a high-angle sweeping wide shot that descends and advances before transitioning to a medium shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the entire statue visibly trembling with low-frequency vibrations that cause the bronze surface to shimmer and ripple?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.038124725222587585 + ], + "opposite_logprob_tokens": [ + -3.2881247997283936 + ], + "logprob_reward": -0.038124725222587585, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.038124725222587585, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.6533180475234985, + 0.02877013199031353, + 0.024218633770942688, + 0.0566067174077034, + 0.018861379474401474, + 0.010026749223470688, + 0.11713182181119919, + 0.016516074538230896, + 0.03519846126437187, + 0.03935198113322258 + ], + "credit_map": [ + 0.07300758361816406, + 0.06917965412139893, + 0.10062506049871445, + 0.061465535312891006, + 0.08086711168289185, + 0.0883619412779808, + 0.06587490439414978, + 0.11393624544143677, + 0.003319264156743884, + 0.002960099373012781, + 0.005406045820564032, + 0.003017356153577566, + 0.002142244018614292, + 0.003559600096195936, + 0.003300558775663376, + 0.005064961966127157, + 0.0022922141943126917, + 0.0023682103492319584, + 0.00635848194360733, + 0.0022065998055040836, + 0.0020487364381551743, + 0.0024447643663734198, + 0.001942675095051527, + 0.004556951578706503, + 0.006467613857239485, + 0.004562983289361, + 0.014066610485315323, + 0.007965528406202793, + 0.003366942750290036, + 0.006808992475271225, + 0.0036830371245741844, + 0.009685012511909008, + 0.0012746585998684168, + 0.002899959683418274, + 0.00319065828807652, + 0.0020273500122129917, + 0.002279838314279914, + 0.001413842779584229, + 0.002169877290725708, + 0.0036051920615136623, + 0.0010265125893056393, + 0.0011354463640600443, + 0.0014140033163130283, + 0.0014988875482231379, + 0.001121313776820898, + 0.0005084334407001734, + 0.0024158775340765715, + 0.0009062742465175688, + 0.019699346274137497, + 0.01609059050679207, + 0.012468723580241203, + 0.013609771616756916, + 0.02357233129441738, + 0.013271219097077847, + 0.009880129247903824, + 0.008539709262549877, + 0.0015305393608286977, + 0.0026628493797034025, + 0.0020114174112677574, + 0.002000844106078148, + 0.0033131507225334644, + 0.002088263165205717, + 0.001618415699340403, + 0.001290595275349915, + 0.0033923019655048847, + 0.007280988153070211, + 0.002609926974400878, + 0.0032176789827644825, + 0.0061250729486346245, + 0.006421900354325771, + 0.002879983512684703, + 0.003270606277510524, + 0.004395083524286747, + 0.007321603130549192, + 0.005181387532502413, + 0.002393081085756421, + 0.004320562817156315, + 0.00524291954934597, + 0.004594543948769569, + 0.005902799312025309 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 3 + ], + "selected_reward": -0.038124725222587585 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 vast public square paved with weathered cobblestones surrounded by classical stone buildings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive equestrian statue of a heroic figure in armor visible in the center of the square?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the statue have an ancient bronze surface with green veins and a dull metallic luster?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera starting with a high-angle sweeping wide shot that descends and advances before transitioning to a medium shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the entire statue visibly trembling with low-frequency vibrations that cause the bronze surface to shimmer and ripple?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08904458582401276 + ], + "opposite_logprob_tokens": [ + -2.4640445709228516 + ], + "logprob_reward": -0.08904458582401276, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08904458582401276, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5164157748222351, + 0.07319153100252151, + 0.02887178771197796, + 0.08736536651849747, + 0.06359098106622696, + 0.041397999972105026, + 0.031002910807728767, + 0.05247771367430687, + 0.05418257787823677, + 0.05150337889790535 + ], + "credit_map": [ + 0.05115329474210739, + 0.06411281228065491, + 0.07251232862472534, + 0.046418771147727966, + 0.06332681328058243, + 0.06067531183362007, + 0.06566477566957474, + 0.09255167096853256, + 0.007408151403069496, + 0.006128598470240831, + 0.016946658492088318, + 0.00975023303180933, + 0.004756964277476072, + 0.007291530724614859, + 0.006821052636951208, + 0.014088350348174572, + 0.0028184556867927313, + 0.002475777640938759, + 0.0072576431557536125, + 0.0029647420160472393, + 0.0021540275774896145, + 0.0033701409120112658, + 0.002597783226519823, + 0.005233218893408775, + 0.008535053580999374, + 0.00759351160377264, + 0.020141683518886566, + 0.017622480168938637, + 0.004731147550046444, + 0.007693218532949686, + 0.006616219878196716, + 0.014432059600949287, + 0.00749229546636343, + 0.005838689859956503, + 0.014200222678482533, + 0.0054655675776302814, + 0.00489871297031641, + 0.010266565717756748, + 0.0045552062802016735, + 0.01087372750043869, + 0.006307546515017748, + 0.004717106930911541, + 0.0070306118577718735, + 0.004471668042242527, + 0.0038802477065473795, + 0.0044829146936535835, + 0.003823678707703948, + 0.006684229709208012, + 0.00647638039663434, + 0.003718449268490076, + 0.004186710808426142, + 0.004087962210178375, + 0.0044357324950397015, + 0.003651835024356842, + 0.001933498540893197, + 0.0025123455561697483, + 0.005633510183542967, + 0.007227135356515646, + 0.006219119299203157, + 0.005971816834062338, + 0.007587865926325321, + 0.008780653588473797, + 0.006087683606892824, + 0.004969934467226267, + 0.0028890317771583796, + 0.007741106208413839, + 0.008698776364326477, + 0.00675016175955534, + 0.006383021827787161, + 0.009782287292182446, + 0.006253410130739212, + 0.005684786941856146, + 0.0051569207571446896, + 0.009234976023435593, + 0.009216643869876862, + 0.005083261989057064, + 0.004656459670513868, + 0.007409280631691217, + 0.005565833766013384, + 0.005180011037737131 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 1 + ], + "selected_reward": -0.08904458582401276 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 vast public square paved with weathered cobblestones surrounded by classical stone buildings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive equestrian statue of a heroic figure in armor visible in the center of the square?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the statue have an ancient bronze surface with green veins and a dull metallic luster?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera starting with a high-angle sweeping wide shot that descends and advances before transitioning to a medium shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the entire statue visibly trembling with low-frequency vibrations that cause the bronze surface to shimmer and ripple?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.5761784911155701 + ], + "opposite_logprob_tokens": [ + -0.8261784911155701 + ], + "logprob_reward": -0.5761784911155701, + "yes_no_margin_reward": 0.25, + "token_credit_reward": -0.5761784911155701, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.6537967324256897, + 0.04882455989718437, + 0.026184165850281715, + 0.023484690114855766, + 0.015280257910490036, + 0.008712071925401688, + 0.07408417761325836, + 0.06478434056043625, + 0.0270346999168396, + 0.05781431496143341 + ], + "credit_map": [ + 0.07417835295200348, + 0.07694675773382187, + 0.09971478581428528, + 0.05303175747394562, + 0.08694391697645187, + 0.09151291102170944, + 0.05458883196115494, + 0.1168794259428978, + 0.0063914889469742775, + 0.0030862127896398306, + 0.0113774873316288, + 0.004883227404206991, + 0.0031368313357234, + 0.0045689851976931095, + 0.0053055165335536, + 0.010074811987578869, + 0.0021385392174124718, + 0.0029932830948382616, + 0.006074054166674614, + 0.0028445490170270205, + 0.002003774279728532, + 0.0023710280656814575, + 0.0030892477370798588, + 0.004669690504670143, + 0.0024489318020641804, + 0.001909436658024788, + 0.005838923156261444, + 0.003128323471173644, + 0.0015073640970513225, + 0.0026242914609611034, + 0.001993519254028797, + 0.004033902194350958, + 0.000983265694230795, + 0.0014477978693321347, + 0.0034906507935374975, + 0.0015631288988515735, + 0.001348048448562622, + 0.0009514327975921333, + 0.0013501527719199657, + 0.004145781043916941, + 0.0008000726229511201, + 0.0012183436192572117, + 0.0013252794742584229, + 0.0010690375929698348, + 0.0014263542834669352, + 0.0006467290222644806, + 0.0012153678108006716, + 0.0010108876740559936, + 0.012210302986204624, + 0.009111377410590649, + 0.007957122288644314, + 0.008423144929111004, + 0.0147086875513196, + 0.00937483087182045, + 0.007121029309928417, + 0.005177681799978018, + 0.007071180269122124, + 0.009235450066626072, + 0.008606120012700558, + 0.0063170637004077435, + 0.008091744966804981, + 0.0096689872443676, + 0.008982128463685513, + 0.0068116686306893826, + 0.0020916436333209276, + 0.004379726015031338, + 0.0029199346899986267, + 0.003106118179857731, + 0.0034865722991526127, + 0.00514418538659811, + 0.0028134400490671396, + 0.003093081060796976, + 0.005442108493298292, + 0.012152398005127907, + 0.005533522926270962, + 0.004518003668636084, + 0.007794859353452921, + 0.006865689996629953, + 0.006772318854928017, + 0.0087354127317667 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 7 + ], + "selected_reward": -0.5761784911155701 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "high-angle, sweeping wide shot slowly descends and advances across a vast, nearly deserted public square paved with weathered cobblestones, taking on a grand and majestic scale. The soft, late afternoon sunlight casts long, dramatic shadows from the surrounding classical stone buildings, bathing the scene in a serene golden hue. In the center of the square stands a massive equestrian statue depicting a heroic figure in armor, its ancient bronze surface riddled with green veins. As the camera continues its deliberate, steady approach, transitioning to a medium shot, a deeply unsettling detail becomes apparent: the entire statue is trembling, accompanied by subtle, low-frequency vibrations. The bronze's dull metallic luster is disturbed by these constant, minute vibrations, causing golden light to shimmer and ripple across its surface as if it were liquid. A layer of fine dust on the thick granite base trembles and dances with the unsettling vibrations that seem to emanate from the metal's very core, creating an tense and surreal atmosphere that contrasts starkly with the serene beauty of the surroundings.", + "seed": 10286, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000021/group_11/index_856_seed_10286_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": -0.23293669000267983, + "vlm_logprob": -0.23293669000267983, + "vlm_yes_no_margin": 1.95, + "vlm_token_credit": -0.23293669000267983, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 vast public square paved with weathered cobblestones surrounded by classical stone buildings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive equestrian statue of a heroic figure in armor visible in the center of the square?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the statue have an ancient bronze surface with green veins and a dull metallic luster?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera starting with a high-angle sweeping wide shot that descends and advances before transitioning to a medium shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the entire statue visibly trembling with low-frequency vibrations that cause the bronze surface to shimmer and ripple?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1263575255870819 + ], + "opposite_logprob_tokens": [ + -2.2513575553894043 + ], + "logprob_reward": -0.1263575255870819, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.1263575255870819, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12243437767028809, + 0.05545792356133461, + 0.023384975269436836, + 0.04536481201648712, + 0.06345640122890472, + 0.15030813217163086, + 0.06516421586275101, + 0.24430830776691437, + 0.17475423216819763, + 0.0553666353225708 + ], + "credit_map": [ + 0.010970397852361202, + 0.013906578533351421, + 0.01986229233443737, + 0.013193991966545582, + 0.00884826760739088, + 0.01803070679306984, + 0.015166237019002438, + 0.02245590090751648, + 0.008722053840756416, + 0.004076223820447922, + 0.010050295852124691, + 0.004990574903786182, + 0.007316591218113899, + 0.007044446654617786, + 0.0055182152427732944, + 0.007739521563053131, + 0.003801564686000347, + 0.0018871971406042576, + 0.003957970067858696, + 0.0021437248215079308, + 0.0021012781653553247, + 0.00390422809869051, + 0.0031789096537977457, + 0.0024101033341139555, + 0.0035660581197589636, + 0.00660286471247673, + 0.006492013111710548, + 0.005141507834196091, + 0.005141356959939003, + 0.0056892866268754005, + 0.005376092158257961, + 0.007355631794780493, + 0.009578988887369633, + 0.00754373986274004, + 0.012881052680313587, + 0.005615470465272665, + 0.005060490220785141, + 0.0081287557259202, + 0.004036843776702881, + 0.01061105914413929, + 0.017154820263385773, + 0.011540932580828667, + 0.02206309512257576, + 0.012249676510691643, + 0.011747865937650204, + 0.04810618236660957, + 0.012761091813445091, + 0.014684455469250679, + 0.007430181372910738, + 0.005950788967311382, + 0.009001318365335464, + 0.006602162029594183, + 0.004952979739755392, + 0.01716114580631256, + 0.006266764365136623, + 0.007798876613378525, + 0.01908107101917267, + 0.0201979111880064, + 0.033776964992284775, + 0.0192258283495903, + 0.026533348485827446, + 0.05790994316339493, + 0.02517196163535118, + 0.04241127893328667, + 0.02262016385793686, + 0.014923996292054653, + 0.024936990812420845, + 0.0180580485612154, + 0.01622704043984413, + 0.03362704813480377, + 0.021240470930933952, + 0.02312047965824604, + 0.005738870706409216, + 0.003954605665057898, + 0.009899639524519444, + 0.007623558398336172, + 0.006486865226179361, + 0.004987609107047319, + 0.008540193550288677, + 0.008135288953781128 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 8, + 5 + ], + "selected_reward": -0.1263575255870819 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 vast public square paved with weathered cobblestones surrounded by classical stone buildings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive equestrian statue of a heroic figure in armor visible in the center of the square?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the statue have an ancient bronze surface with green veins and a dull metallic luster?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera starting with a high-angle sweeping wide shot that descends and advances before transitioning to a medium shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the entire statue visibly trembling with low-frequency vibrations that cause the bronze surface to shimmer and ripple?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02637963369488716 + ], + "opposite_logprob_tokens": [ + -3.6513795852661133 + ], + "logprob_reward": -0.02637963369488716, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.02637963369488716, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14824503660202026, + 0.1610909104347229, + 0.09922479838132858, + 0.27790406346321106, + 0.057904522866010666, + 0.08257577568292618, + 0.039313096553087234, + 0.046428829431533813, + 0.05832044407725334, + 0.02899252623319626 + ], + "credit_map": [ + 0.01582670770585537, + 0.022687822580337524, + 0.022196929901838303, + 0.0182842705398798, + 0.015961993485689163, + 0.011444858275353909, + 0.018483353778719902, + 0.02335907518863678, + 0.012308331206440926, + 0.010444124229252338, + 0.03449847549200058, + 0.017357712611556053, + 0.02561110630631447, + 0.013750485144555569, + 0.02065463364124298, + 0.02646603062748909, + 0.005990229081362486, + 0.01478977408260107, + 0.01835738867521286, + 0.011059600859880447, + 0.012205548584461212, + 0.00788863655179739, + 0.015089718624949455, + 0.013843893073499203, + 0.033378973603248596, + 0.041550468653440475, + 0.04323292896151543, + 0.030766556039452553, + 0.023419931530952454, + 0.03643670678138733, + 0.026799900457262993, + 0.04231859743595123, + 0.004378821235150099, + 0.00902036763727665, + 0.011306818574666977, + 0.007054531015455723, + 0.005859397351741791, + 0.004823839291930199, + 0.0065292357467114925, + 0.008931507356464863, + 0.005269906017929316, + 0.009151825681328773, + 0.022130798548460007, + 0.008573916740715504, + 0.006913505494594574, + 0.009212708100676537, + 0.007620106916874647, + 0.013703005388379097, + 0.0054651652462780476, + 0.0034654231276363134, + 0.005627009551972151, + 0.004047649912536144, + 0.0030726108234375715, + 0.009707411751151085, + 0.00409119576215744, + 0.0038366271182894707, + 0.004655409604310989, + 0.004751523490995169, + 0.006751690991222858, + 0.005434110760688782, + 0.005317806266248226, + 0.006729831453412771, + 0.006141701713204384, + 0.006646754685789347, + 0.004502341151237488, + 0.005540027283132076, + 0.008067620918154716, + 0.005454414524137974, + 0.004386866930872202, + 0.007879660464823246, + 0.014949432574212551, + 0.007540075108408928, + 0.003206438384950161, + 0.0018286991398781538, + 0.004436914809048176, + 0.0056730820797383785, + 0.002393227769061923, + 0.0026974682696163654, + 0.0043248385190963745, + 0.00443185493350029 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 1, + 0 + ], + "selected_reward": -0.02637963369488716 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 vast public square paved with weathered cobblestones surrounded by classical stone buildings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive equestrian statue of a heroic figure in armor visible in the center of the square?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the statue have an ancient bronze surface with green veins and a dull metallic luster?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera starting with a high-angle sweeping wide shot that descends and advances before transitioning to a medium shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the entire statue visibly trembling with low-frequency vibrations that cause the bronze surface to shimmer and ripple?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04867658019065857 + ], + "opposite_logprob_tokens": [ + -3.0486764907836914 + ], + "logprob_reward": -0.04867658019065857, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04867658019065857, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09403100609779358, + 0.36366012692451477, + 0.04194305092096329, + 0.07415689527988434, + 0.040818266570568085, + 0.06970414519309998, + 0.02620464563369751, + 0.09942271560430527, + 0.13765159249305725, + 0.05240752920508385 + ], + "credit_map": [ + 0.01192930806428194, + 0.006357290782034397, + 0.015807801857590675, + 0.008846726268529892, + 0.010431957431137562, + 0.015308056958019733, + 0.010739144869148731, + 0.014610731042921543, + 0.022597819566726685, + 0.026210017502307892, + 0.06961547583341599, + 0.04785240814089775, + 0.06259850412607193, + 0.024679938331246376, + 0.04921882972121239, + 0.06088715419173241, + 0.004088267683982849, + 0.009817533195018768, + 0.005764031782746315, + 0.0044781058095395565, + 0.005348686594516039, + 0.004293979145586491, + 0.004209939390420914, + 0.003942510113120079, + 0.005807946436107159, + 0.015250246971845627, + 0.01123900432139635, + 0.010984944179654121, + 0.005907073151320219, + 0.005744125694036484, + 0.008457757532596588, + 0.010765803046524525, + 0.009774823673069477, + 0.002642502309754491, + 0.007492927368730307, + 0.00380580872297287, + 0.002911235671490431, + 0.005988216493278742, + 0.0022844106424599886, + 0.005918344482779503, + 0.00887330062687397, + 0.005436589941382408, + 0.014392314478754997, + 0.005065664649009705, + 0.005087122321128845, + 0.013658074662089348, + 0.004697232972830534, + 0.012493852525949478, + 0.0026068217121064663, + 0.0026467335410416126, + 0.003712708130478859, + 0.003377871122211218, + 0.002373218536376953, + 0.00517693068832159, + 0.0032535826321691275, + 0.0030567795038223267, + 0.007681929040700197, + 0.008586038835346699, + 0.02083948254585266, + 0.009608360007405281, + 0.0075057330541312695, + 0.020011894404888153, + 0.011166607961058617, + 0.01402267999947071, + 0.019107673317193985, + 0.010644068010151386, + 0.019994881004095078, + 0.012994278222322464, + 0.014258324168622494, + 0.026540303602814674, + 0.01611536368727684, + 0.017996711656451225, + 0.006638215389102697, + 0.004810435231775045, + 0.006888024974614382, + 0.006319226697087288, + 0.004923093132674694, + 0.007480056956410408, + 0.007008030079305172, + 0.008340450003743172 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 8, + 7 + ], + "selected_reward": -0.04867658019065857 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 vast public square paved with weathered cobblestones surrounded by classical stone buildings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive equestrian statue of a heroic figure in armor visible in the center of the square?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the statue have an ancient bronze surface with green veins and a dull metallic luster?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera starting with a high-angle sweeping wide shot that descends and advances before transitioning to a medium shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the entire statue visibly trembling with low-frequency vibrations that cause the bronze surface to shimmer and ripple?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.576155424118042 + ], + "opposite_logprob_tokens": [ + -0.826155424118042 + ], + "logprob_reward": -0.576155424118042, + "yes_no_margin_reward": 0.25, + "token_credit_reward": -0.576155424118042, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0847407728433609, + 0.0778164342045784, + 0.04335186257958412, + 0.36906856298446655, + 0.05433496832847595, + 0.1260138601064682, + 0.042308978736400604, + 0.05321306735277176, + 0.05128868669271469, + 0.09786279499530792 + ], + "credit_map": [ + 0.008435525000095367, + 0.005479094106703997, + 0.0172712579369545, + 0.010415850207209587, + 0.009938759729266167, + 0.0063079786486923695, + 0.010129876434803009, + 0.016762420535087585, + 0.005269782152026892, + 0.005516624543815851, + 0.017453284934163094, + 0.009098346345126629, + 0.00997253879904747, + 0.007141694892197847, + 0.009269076399505138, + 0.014095085673034191, + 0.0044418275356292725, + 0.005049513187259436, + 0.007542418781667948, + 0.00429672934114933, + 0.006248774938285351, + 0.0032267868518829346, + 0.005824990104883909, + 0.006720819044858217, + 0.04063088074326515, + 0.04675602167844772, + 0.062458280473947525, + 0.04540964215993881, + 0.033832889050245285, + 0.046244338154792786, + 0.036526668816804886, + 0.05720982328057289, + 0.0059894537553191185, + 0.006147271953523159, + 0.009959014132618904, + 0.006489875260740519, + 0.004513431806117296, + 0.005257915705442429, + 0.007425825111567974, + 0.00855217780917883, + 0.020894352346658707, + 0.011281564831733704, + 0.02576332539319992, + 0.008690571412444115, + 0.010187491774559021, + 0.016192397102713585, + 0.009002764709293842, + 0.024001378566026688, + 0.005376683082431555, + 0.0045599364675581455, + 0.005652951542288065, + 0.0050357189029455185, + 0.003892979584634304, + 0.007593894377350807, + 0.00515084620565176, + 0.005045964848250151, + 0.004399845842272043, + 0.0044589051976799965, + 0.015049696899950504, + 0.004728697706013918, + 0.003606525482609868, + 0.006406181957572699, + 0.005004481878131628, + 0.009558723308146, + 0.0032607109751552343, + 0.003763205371797085, + 0.009419836103916168, + 0.009608090855181217, + 0.004628028720617294, + 0.004138940013945103, + 0.00841781497001648, + 0.008052052929997444, + 0.010325536131858826, + 0.007384434342384338, + 0.013126782141625881, + 0.010061954148113728, + 0.010364469140768051, + 0.010860300622880459, + 0.019118905067443848, + 0.016620399430394173 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 5, + 9 + ], + "selected_reward": -0.576155424118042 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 vast public square paved with weathered cobblestones surrounded by classical stone buildings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive equestrian statue of a heroic figure in armor visible in the center of the square?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the statue have an ancient bronze surface with green veins and a dull metallic luster?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera starting with a high-angle sweeping wide shot that descends and advances before transitioning to a medium shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the entire statue visibly trembling with low-frequency vibrations that cause the bronze surface to shimmer and ripple?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.3871142864227295 + ], + "opposite_logprob_tokens": [ + -1.1371142864227295 + ], + "logprob_reward": -0.3871142864227295, + "yes_no_margin_reward": 0.75, + "token_credit_reward": -0.3871142864227295, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08638334274291992, + 0.13388802111148834, + 0.04514536261558533, + 0.28398197889328003, + 0.04502822458744049, + 0.10170120745897293, + 0.026379141956567764, + 0.155443012714386, + 0.06349421292543411, + 0.05855552479624748 + ], + "credit_map": [ + 0.011161329224705696, + 0.009565928019583225, + 0.012752042151987553, + 0.010064560920000076, + 0.010613037273287773, + 0.00998520478606224, + 0.007229461334645748, + 0.015011770650744438, + 0.008084606379270554, + 0.009607311338186264, + 0.026891939342021942, + 0.018515441566705704, + 0.022773537784814835, + 0.008687804453074932, + 0.016755899414420128, + 0.022571465000510216, + 0.004720564465969801, + 0.006226432975381613, + 0.0065086293034255505, + 0.006174340844154358, + 0.006057938560843468, + 0.005428553093224764, + 0.0033837445080280304, + 0.006645153276622295, + 0.032524269074201584, + 0.03835956007242203, + 0.047694262117147446, + 0.03487492725253105, + 0.024406176060438156, + 0.035666726529598236, + 0.026785606518387794, + 0.043670423328876495, + 0.00643149483948946, + 0.005332956090569496, + 0.00675128772854805, + 0.005252750124782324, + 0.004124643746763468, + 0.005106851924210787, + 0.005027018021792173, + 0.007001216523349285, + 0.014171476475894451, + 0.007961525581777096, + 0.022089414298534393, + 0.008738242089748383, + 0.007370660547167063, + 0.01467923540621996, + 0.007889628410339355, + 0.018801016733050346, + 0.0029669899959117174, + 0.0032961717806756496, + 0.002893987577408552, + 0.003768275259062648, + 0.0037289808969944715, + 0.002812967635691166, + 0.0034381994046270847, + 0.0034735656809061766, + 0.010714397765696049, + 0.016650864854454994, + 0.03745763748884201, + 0.019651878625154495, + 0.011934501118957996, + 0.012065358459949493, + 0.01528108585625887, + 0.03168727457523346, + 0.008740896359086037, + 0.006390294060111046, + 0.008659101091325283, + 0.008375796489417553, + 0.00786543171852827, + 0.007149644661694765, + 0.007210917770862579, + 0.009102119132876396, + 0.004336696583777666, + 0.004919921513646841, + 0.00892679113894701, + 0.013125423341989517, + 0.004744973499327898, + 0.005662545096129179, + 0.009728342294692993, + 0.007110824342817068 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 7, + 1 + ], + "selected_reward": -0.3871142864227295 + } + ], + "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/0000021/group_12/index_tensor([3731, 3731])_group_12_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000021/group_12/index_tensor([3731, 3731])_group_12_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..7337bee275227d2ee5155c41f1085b7a3c05e94d --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000021/group_12/index_tensor([3731, 3731])_group_12_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "tight, close-up shot from a slightly upward angle focuses on a Tang Dynasty poet. He is a thin, middle-aged man with a neatly trimmed beard. His long, jet-black hair is tied high with an ancient jade hairpin, a few strands of hair blowing gently against his temples in the night breeze. He wears a round-necked, dark blue silk robe embroidered with elegant cloud patterns, which gleams softly in the moonlight. The cool, bright moonlight pours down like mercury, serving as the primary source of light, not only outlining the three-dimensional contours of his face but also casting sharp shadows on his high nose and cheekbones. At this moment, his cheeks are slightly tipsy, his eyes narrowed, a glint of intoxication and inspiration mingling. A faint smile plays at the corners of his mouth, and his brows are filled with a carefree and relaxed expression. He slowly raised a simple white porcelain wine glass to eye level, as if greeting the bright moon. His lips moved slightly, as if he were whispering a line of poetry that had just come to mind. The background was the carved railings of a pavilion and the silhouettes of a few distant bamboo clumps swaying gently in the night breeze. The entire scene, dominated by deep blue and inky black, was punctuated by the silvery moonlight, creating a tranquil, intoxicating, and poetic atmosphere filled with creative passion.", + "seed": 9948, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000021/group_12/index_3731_seed_9948_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": -0.13203499019145964, + "vlm_logprob": -0.13203499019145964, + "vlm_yes_no_margin": 2.25, + "vlm_token_credit": -0.13203499019145964, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a thin, middle-aged man with a neatly trimmed beard?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a round-necked, dark blue silk robe with cloud patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene lit primarily by cool, bright moonlight casting sharp shadows on the face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the man holding a simple white porcelain wine glass raised to eye level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background composed of carved pavilion railings and distant swaying bamboo silhouettes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.09050320088863373 + ], + "opposite_logprob_tokens": [ + -2.465503215789795 + ], + "logprob_reward": -0.09050320088863373, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.09050320088863373, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3564831614494324, + 0.09701903164386749, + 0.04609673097729683, + 0.08510903269052505, + 0.11489379405975342, + 0.10595155507326126, + 0.04593484103679657, + 0.09356251358985901, + 0.03056265413761139, + 0.0243866965174675 + ], + "credit_map": [ + 0.062346700578927994, + 0.027171926572918892, + 0.054466016590595245, + 0.02382841147482395, + 0.02606307342648506, + 0.07570585608482361, + 0.0346698984503746, + 0.05223128944635391, + 0.01605468988418579, + 0.01062499824911356, + 0.012420892715454102, + 0.007226685527712107, + 0.01061288546770811, + 0.017092175781726837, + 0.010315376333892345, + 0.012671319767832756, + 0.005123570561408997, + 0.003358413465321064, + 0.006801567040383816, + 0.004048678558319807, + 0.005911857821047306, + 0.005635106936097145, + 0.005250164773315191, + 0.009967371821403503, + 0.006641873624175787, + 0.007340278476476669, + 0.012383311986923218, + 0.0065895067527890205, + 0.012549220584332943, + 0.010919460095465183, + 0.011798558756709099, + 0.016886822879314423, + 0.009032508358359337, + 0.008268608711659908, + 0.01027736347168684, + 0.011831142008304596, + 0.02127382531762123, + 0.01545930840075016, + 0.016418753191828728, + 0.02233227714896202, + 0.008513657376170158, + 0.01269044540822506, + 0.013762439601123333, + 0.015970291569828987, + 0.01006434764713049, + 0.013029202818870544, + 0.01552655640989542, + 0.016394605860114098, + 0.005602909252047539, + 0.004549932200461626, + 0.006098294164985418, + 0.003478727303445339, + 0.007483682595193386, + 0.005593066103756428, + 0.0076285176910459995, + 0.00549970893189311, + 0.009103621356189251, + 0.010202118195593357, + 0.009967544116079807, + 0.009041178040206432, + 0.01236567273736, + 0.011179324239492416, + 0.014722565189003944, + 0.016980476677417755, + 0.0034350333735346794, + 0.004160743672400713, + 0.0036549903452396393, + 0.002859038533642888, + 0.004844577983021736, + 0.0039062276482582092, + 0.003784329164773226, + 0.003917710855603218, + 0.0030464613810181618, + 0.0025995851028710604, + 0.003499042708426714, + 0.001638838555663824, + 0.004069410730153322, + 0.004125142470002174, + 0.002461389871314168, + 0.0029468252323567867 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 5 + ], + "selected_reward": -0.09050320088863373 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a thin, middle-aged man with a neatly trimmed beard?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a round-necked, dark blue silk robe with cloud patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene lit primarily by cool, bright moonlight casting sharp shadows on the face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the man holding a simple white porcelain wine glass raised to eye level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background composed of carved pavilion railings and distant swaying bamboo silhouettes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.048694971948862076 + ], + "opposite_logprob_tokens": [ + -3.0486950874328613 + ], + "logprob_reward": -0.048694971948862076, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.048694971948862076, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.29180845618247986, + 0.06802650541067123, + 0.10985169559717178, + 0.09825173020362854, + 0.1325845867395401, + 0.05055486410856247, + 0.07639595866203308, + 0.060488756746053696, + 0.07721364498138428, + 0.034823764115571976 + ], + "credit_map": [ + 0.033061642199754715, + 0.030486786738038063, + 0.039556775242090225, + 0.021814025938510895, + 0.032318275421857834, + 0.04580581560730934, + 0.04302292689681053, + 0.04574223607778549, + 0.009676922112703323, + 0.007233036682009697, + 0.010309161618351936, + 0.005450481083244085, + 0.00900882389396429, + 0.010676776990294456, + 0.00768628902733326, + 0.007985008880496025, + 0.012295059859752655, + 0.006918883882462978, + 0.01822848618030548, + 0.008123008534312248, + 0.010138907469809055, + 0.03478008136153221, + 0.007812211290001869, + 0.011555054225027561, + 0.007753339130431414, + 0.009366455487906933, + 0.013928505592048168, + 0.00727849779650569, + 0.01456541195511818, + 0.012072106823325157, + 0.014452891424298286, + 0.018834520131349564, + 0.014636032283306122, + 0.020832251757383347, + 0.021534182131290436, + 0.012905553914606571, + 0.016222404316067696, + 0.016010422259569168, + 0.016725091263651848, + 0.013718655332922935, + 0.005902217701077461, + 0.008299898356199265, + 0.006979083176702261, + 0.00311103044077754, + 0.007021612022072077, + 0.006721651181578636, + 0.006272525060921907, + 0.0062468452379107475, + 0.009687880985438824, + 0.010162229649722576, + 0.009731496684253216, + 0.007653920445591211, + 0.009899614378809929, + 0.009693164378404617, + 0.008021382614970207, + 0.011546271853148937, + 0.005503026302903891, + 0.005807892885059118, + 0.007237457670271397, + 0.008232380263507366, + 0.012011241167783737, + 0.007191146723926067, + 0.007496435195207596, + 0.0070091732777655125, + 0.009324582293629646, + 0.013348371721804142, + 0.011713783256709576, + 0.007666362915188074, + 0.009555523283779621, + 0.009647211991250515, + 0.007766653783619404, + 0.008191149681806564, + 0.004086520057171583, + 0.005602490622550249, + 0.005736910738050938, + 0.0029551598709076643, + 0.004156067967414856, + 0.00529963243752718, + 0.0032011503353714943, + 0.0037858306895941496 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 2 + ], + "selected_reward": -0.048694971948862076 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a thin, middle-aged man with a neatly trimmed beard?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a round-necked, dark blue silk robe with cloud patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene lit primarily by cool, bright moonlight casting sharp shadows on the face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the man holding a simple white porcelain wine glass raised to eye level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background composed of carved pavilion railings and distant swaying bamboo silhouettes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.28133493661880493 + ], + "opposite_logprob_tokens": [ + -1.4063348770141602 + ], + "logprob_reward": -0.28133493661880493, + "yes_no_margin_reward": 1.125, + "token_credit_reward": -0.28133493661880493, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3421897888183594, + 0.12084949761629105, + 0.056985966861248016, + 0.0615164078772068, + 0.08013055473566055, + 0.10449270159006119, + 0.06676728278398514, + 0.10555440187454224, + 0.046566128730773926, + 0.01494729146361351 + ], + "credit_map": [ + 0.059280894696712494, + 0.04091687127947807, + 0.04647642746567726, + 0.026057200506329536, + 0.03915074095129967, + 0.04604141786694527, + 0.042166758328676224, + 0.042099494487047195, + 0.018172748386859894, + 0.01623212918639183, + 0.015103844925761223, + 0.009140127338469028, + 0.01641269400715828, + 0.014028127305209637, + 0.017898565158247948, + 0.013861266896128654, + 0.006285089999437332, + 0.0070403506979346275, + 0.007127250079065561, + 0.006995375733822584, + 0.0070984335616230965, + 0.006521602626889944, + 0.008857488632202148, + 0.007060378789901733, + 0.0066643934696912766, + 0.004457621369510889, + 0.009316028095781803, + 0.009256658144295216, + 0.007128539960831404, + 0.007194221019744873, + 0.008648466318845749, + 0.008850477635860443, + 0.009921499527990818, + 0.01096199732273817, + 0.010031507350504398, + 0.008088337257504463, + 0.011631936766207218, + 0.010315257124602795, + 0.009575210511684418, + 0.00960481259971857, + 0.01002497412264347, + 0.016551274806261063, + 0.01630101352930069, + 0.01432703249156475, + 0.010788108222186565, + 0.013342408463358879, + 0.013973894529044628, + 0.009183995425701141, + 0.009952863678336143, + 0.009852844290435314, + 0.010167831555008888, + 0.0038118830416351557, + 0.009157839231193066, + 0.008714796975255013, + 0.008200296200811863, + 0.006908927112817764, + 0.017833955585956573, + 0.017841940745711327, + 0.014968539588153362, + 0.009278200566768646, + 0.011867809109389782, + 0.014050520956516266, + 0.010147630237042904, + 0.00956580601632595, + 0.007037840783596039, + 0.006107608787715435, + 0.0074876719154417515, + 0.003826296655461192, + 0.006429713219404221, + 0.005733151454478502, + 0.004946053959429264, + 0.004997792653739452, + 0.0021950420923531055, + 0.0017096729716286063, + 0.0023862754460424185, + 0.0017174602253362536, + 0.001805146224796772, + 0.001739990315400064, + 0.0015856096288189292, + 0.0018080944428220391 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 7 + ], + "selected_reward": -0.28133493661880493 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a thin, middle-aged man with a neatly trimmed beard?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a round-necked, dark blue silk robe with cloud patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene lit primarily by cool, bright moonlight casting sharp shadows on the face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the man holding a simple white porcelain wine glass raised to eye level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background composed of carved pavilion railings and distant swaying bamboo silhouettes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.20154500007629395 + ], + "opposite_logprob_tokens": [ + -1.701545000076294 + ], + "logprob_reward": -0.20154500007629395, + "yes_no_margin_reward": 1.5, + "token_credit_reward": -0.20154500007629395, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22923101484775543, + 0.1900511234998703, + 0.1325353980064392, + 0.12017953395843506, + 0.060381192713975906, + 0.05721569433808327, + 0.05240286514163017, + 0.08354754000902176, + 0.043397460132837296, + 0.031058160588145256 + ], + "credit_map": [ + 0.022230371832847595, + 0.015336326323449612, + 0.037396468222141266, + 0.020952772349119186, + 0.022634878754615784, + 0.04079470410943031, + 0.0336509570479393, + 0.03623455762863159, + 0.01832352951169014, + 0.01339240837842226, + 0.025981958955526352, + 0.024214968085289, + 0.026833198964595795, + 0.03021448850631714, + 0.023357652127742767, + 0.027732914313673973, + 0.013289893046021461, + 0.011307070963084698, + 0.019286654889583588, + 0.01967497542500496, + 0.01435078214854002, + 0.01642574369907379, + 0.015776384621858597, + 0.022423891350626945, + 0.010401763021945953, + 0.01864958181977272, + 0.015755750238895416, + 0.013717212714254856, + 0.013768717646598816, + 0.01585857756435871, + 0.013287115842103958, + 0.018740814179182053, + 0.0055559538304805756, + 0.006828125100582838, + 0.007349974010139704, + 0.005290285684168339, + 0.012638852931559086, + 0.006970491260290146, + 0.007703239098191261, + 0.008044270798563957, + 0.005147307179868221, + 0.004692545626312494, + 0.006443304009735584, + 0.007846557535231113, + 0.008846556767821312, + 0.007590034045279026, + 0.009194114245474339, + 0.007455279119312763, + 0.006724334321916103, + 0.006589723750948906, + 0.007459391374140978, + 0.0056595164351165295, + 0.0060798656195402145, + 0.0072988346219062805, + 0.005406157113611698, + 0.007185039576143026, + 0.0073255496099591255, + 0.009129849262535572, + 0.010020246729254723, + 0.014597505331039429, + 0.00774178933352232, + 0.010916049592196941, + 0.010109787806868553, + 0.013706761412322521, + 0.004847856238484383, + 0.007315963506698608, + 0.005771998316049576, + 0.0053209904581308365, + 0.006019475404173136, + 0.005123666487634182, + 0.00501205213367939, + 0.0039854575879871845, + 0.003364546922966838, + 0.002855211729183793, + 0.00355966342613101, + 0.005042545031756163, + 0.00428807083517313, + 0.0036251505371183157, + 0.004362341482192278, + 0.00396063132211566 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.20154500007629395 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a thin, middle-aged man with a neatly trimmed beard?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a round-necked, dark blue silk robe with cloud patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene lit primarily by cool, bright moonlight casting sharp shadows on the face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the man holding a simple white porcelain wine glass raised to eye level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background composed of carved pavilion railings and distant swaying bamboo silhouettes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0380968414247036 + ], + "opposite_logprob_tokens": [ + -3.2880969047546387 + ], + "logprob_reward": -0.0380968414247036, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.0380968414247036, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.315200537443161, + 0.046241942793130875, + 0.036259185522794724, + 0.0441594123840332, + 0.10225842148065567, + 0.13907083868980408, + 0.050784919410943985, + 0.04102275148034096, + 0.0707990899682045, + 0.1542029231786728 + ], + "credit_map": [ + 0.05164952203631401, + 0.020818885415792465, + 0.04577792435884476, + 0.022033745422959328, + 0.02553238719701767, + 0.05283751338720322, + 0.039065245538949966, + 0.05748530849814415, + 0.0063347951509058475, + 0.0044405097141861916, + 0.006020208355039358, + 0.004157816991209984, + 0.005963492207229137, + 0.0056426082737743855, + 0.005457751918584108, + 0.008224756456911564, + 0.004950468894094229, + 0.003805858315899968, + 0.005425584502518177, + 0.003362370654940605, + 0.004326620139181614, + 0.006734492257237434, + 0.0037165889516472816, + 0.003937202971428633, + 0.004914938472211361, + 0.006113761570304632, + 0.005417026113718748, + 0.005686563439667225, + 0.004691454581916332, + 0.005820797756314278, + 0.00546139944344759, + 0.006053467746824026, + 0.011819224804639816, + 0.019512875005602837, + 0.014289340935647488, + 0.009347343817353249, + 0.011462575756013393, + 0.012644154950976372, + 0.011369595304131508, + 0.01181331928819418, + 0.012471314519643784, + 0.0172720979899168, + 0.02130000665783882, + 0.02272072806954384, + 0.012946952134370804, + 0.01718904823064804, + 0.018025485798716545, + 0.017145192250609398, + 0.0053158532828092575, + 0.006536348722875118, + 0.00729727977886796, + 0.004042442422360182, + 0.007326518651098013, + 0.0064161308109760284, + 0.005581580568104982, + 0.008268763311207294, + 0.0047807879745960236, + 0.007283298298716545, + 0.006111472845077515, + 0.0028034455608576536, + 0.005422266665846109, + 0.004866296425461769, + 0.004060971084982157, + 0.005694212391972542, + 0.00977842602878809, + 0.009930230677127838, + 0.011446956545114517, + 0.00701738428324461, + 0.008237123489379883, + 0.008951607160270214, + 0.006200307514518499, + 0.00923705380409956, + 0.019942941144108772, + 0.0250827856361866, + 0.024998698383569717, + 0.016839055344462395, + 0.01564568094909191, + 0.02560311369597912, + 0.012807830236852169, + 0.013282799161970615 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 5 + ], + "selected_reward": -0.0380968414247036 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "tight, close-up shot from a slightly upward angle focuses on a Tang Dynasty poet. He is a thin, middle-aged man with a neatly trimmed beard. His long, jet-black hair is tied high with an ancient jade hairpin, a few strands of hair blowing gently against his temples in the night breeze. He wears a round-necked, dark blue silk robe embroidered with elegant cloud patterns, which gleams softly in the moonlight. The cool, bright moonlight pours down like mercury, serving as the primary source of light, not only outlining the three-dimensional contours of his face but also casting sharp shadows on his high nose and cheekbones. At this moment, his cheeks are slightly tipsy, his eyes narrowed, a glint of intoxication and inspiration mingling. A faint smile plays at the corners of his mouth, and his brows are filled with a carefree and relaxed expression. He slowly raised a simple white porcelain wine glass to eye level, as if greeting the bright moon. His lips moved slightly, as if he were whispering a line of poetry that had just come to mind. The background was the carved railings of a pavilion and the silhouettes of a few distant bamboo clumps swaying gently in the night breeze. The entire scene, dominated by deep blue and inky black, was punctuated by the silvery moonlight, creating a tranquil, intoxicating, and poetic atmosphere filled with creative passion.", + "seed": 9949, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000021/group_12/index_3731_seed_9949_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": -0.10570017285645009, + "vlm_logprob": -0.10570017285645009, + "vlm_yes_no_margin": 2.499999976158142, + "vlm_token_credit": -0.10570017285645009, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a thin, middle-aged man with a neatly trimmed beard?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a round-necked, dark blue silk robe with cloud patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene lit primarily by cool, bright moonlight casting sharp shadows on the face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the man holding a simple white porcelain wine glass raised to eye level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background composed of carved pavilion railings and distant swaying bamboo silhouettes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07988493144512177 + ], + "opposite_logprob_tokens": [ + -2.5798850059509277 + ], + "logprob_reward": -0.07988493144512177, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07988493144512177, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20484119653701782, + 0.09736145287752151, + 0.042647041380405426, + 0.06173916906118393, + 0.10663044452667236, + 0.11639206111431122, + 0.05249899625778198, + 0.04530685395002365, + 0.12270532548427582, + 0.1498773843050003 + ], + "credit_map": [ + 0.030233336612582207, + 0.026447324082255363, + 0.02427375130355358, + 0.011006440035998821, + 0.02749515324831009, + 0.025100497528910637, + 0.03260945901274681, + 0.027675267308950424, + 0.01266724057495594, + 0.013500894419848919, + 0.012053357437252998, + 0.0062663317658007145, + 0.012088807299733162, + 0.013852331787347794, + 0.011661218479275703, + 0.015271280892193317, + 0.004706254694610834, + 0.004786405246704817, + 0.005337456241250038, + 0.005389419849961996, + 0.0052739884704351425, + 0.0059069120325148106, + 0.006475544068962336, + 0.004771064501255751, + 0.00726351048797369, + 0.007660640869289637, + 0.00881034042686224, + 0.004999179858714342, + 0.006266321055591106, + 0.009569196961820126, + 0.010676654055714607, + 0.006493328604847193, + 0.01273112092167139, + 0.012625287286937237, + 0.015200113877654076, + 0.008343570865690708, + 0.015441937372088432, + 0.01708175614476204, + 0.013805648311972618, + 0.011401020921766758, + 0.014880655333399773, + 0.014776797033846378, + 0.016043920069932938, + 0.007778319995850325, + 0.01845107413828373, + 0.01930464431643486, + 0.013911244459450245, + 0.011245406232774258, + 0.011958691291511059, + 0.004924869630485773, + 0.011761985719203949, + 0.0027145729400217533, + 0.007081580813974142, + 0.005342832766473293, + 0.004573841113597155, + 0.004140628036111593, + 0.0064262947998940945, + 0.005502322688698769, + 0.006236718501895666, + 0.003186103655025363, + 0.0057435911148786545, + 0.007302100770175457, + 0.005547407083213329, + 0.005362317897379398, + 0.012420636601746082, + 0.0196387879550457, + 0.016243409365415573, + 0.0117372190579772, + 0.01399148628115654, + 0.011617806740105152, + 0.026210427284240723, + 0.010845555923879147, + 0.012512341141700745, + 0.024328958243131638, + 0.020054766908288002, + 0.010180842131376266, + 0.014948690310120583, + 0.0232253335416317, + 0.022395381703972816, + 0.022231079638004303 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 8 + ], + "selected_reward": -0.07988493144512177 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a thin, middle-aged man with a neatly trimmed beard?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a round-necked, dark blue silk robe with cloud patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene lit primarily by cool, bright moonlight casting sharp shadows on the face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the man holding a simple white porcelain wine glass raised to eye level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background composed of carved pavilion railings and distant swaying bamboo silhouettes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.043107133358716965 + ], + "opposite_logprob_tokens": [ + -3.168107032775879 + ], + "logprob_reward": -0.043107133358716965, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.043107133358716965, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15975795686244965, + 0.17186738550662994, + 0.11356070637702942, + 0.06516841799020767, + 0.04146227613091469, + 0.10863642394542694, + 0.052704714238643646, + 0.07756099104881287, + 0.05351172760128975, + 0.15576942265033722 + ], + "credit_map": [ + 0.026545802131295204, + 0.02306950092315674, + 0.021843837574124336, + 0.01668546348810196, + 0.02020370401442051, + 0.0196763277053833, + 0.016349976882338524, + 0.015383319929242134, + 0.020769882947206497, + 0.02256675809621811, + 0.019607821479439735, + 0.02849763073027134, + 0.017882026731967926, + 0.02416233904659748, + 0.016253160312771797, + 0.022127751260995865, + 0.012950021773576736, + 0.014261550270020962, + 0.015891162678599358, + 0.01019533071666956, + 0.01312381960451603, + 0.02257484570145607, + 0.011324030347168446, + 0.013239927589893341, + 0.007172402460128069, + 0.008670401759445667, + 0.009587510488927364, + 0.004236237611621618, + 0.00842994637787342, + 0.01010164339095354, + 0.009517064318060875, + 0.007453201804310083, + 0.006563973147422075, + 0.004975747317075729, + 0.00806038361042738, + 0.003023055847734213, + 0.004795928485691547, + 0.004719767719507217, + 0.004349617753177881, + 0.004973796661943197, + 0.029722662642598152, + 0.00838368572294712, + 0.029196450486779213, + 0.0044986167922616005, + 0.00945224892348051, + 0.011908020824193954, + 0.006479422561824322, + 0.008995306678116322, + 0.011107097379863262, + 0.004824759904295206, + 0.010265604592859745, + 0.0037556777242571115, + 0.004395931959152222, + 0.006817317102104425, + 0.005768369417637587, + 0.005769954528659582, + 0.009808779694139957, + 0.01248896773904562, + 0.009180554188787937, + 0.006105930078774691, + 0.008236988447606564, + 0.010593673214316368, + 0.010035114362835884, + 0.011110974475741386, + 0.006011691875755787, + 0.006733358837664127, + 0.009772993624210358, + 0.00439060851931572, + 0.004923544824123383, + 0.01069596502929926, + 0.00487028993666172, + 0.006113269366323948, + 0.016530776396393776, + 0.022418128326535225, + 0.027683265507221222, + 0.01802249252796173, + 0.01491689682006836, + 0.027096468955278397, + 0.01363950502127409, + 0.015461878851056099 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 9 + ], + "selected_reward": -0.043107133358716965 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a thin, middle-aged man with a neatly trimmed beard?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a round-necked, dark blue silk robe with cloud patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene lit primarily by cool, bright moonlight casting sharp shadows on the face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the man holding a simple white porcelain wine glass raised to eye level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background composed of carved pavilion railings and distant swaying bamboo silhouettes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.25213611125946045 + ], + "opposite_logprob_tokens": [ + -1.5021361112594604 + ], + "logprob_reward": -0.25213611125946045, + "yes_no_margin_reward": 1.25, + "token_credit_reward": -0.25213611125946045, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13687573373317719, + 0.057626135647296906, + 0.07909366488456726, + 0.11701766401529312, + 0.26186323165893555, + 0.08193095773458481, + 0.07627834379673004, + 0.05547052621841431, + 0.071524977684021, + 0.06231878325343132 + ], + "credit_map": [ + 0.014172211289405823, + 0.017361927777528763, + 0.016179025173187256, + 0.019927313551306725, + 0.022289477288722992, + 0.01937669888138771, + 0.01327060628682375, + 0.014298481866717339, + 0.007457369472831488, + 0.008406084962189198, + 0.00810174085199833, + 0.006757060997188091, + 0.008055275306105614, + 0.007700550369918346, + 0.00498130451887846, + 0.006166749633848667, + 0.007875875569880009, + 0.01098982896655798, + 0.009944532997906208, + 0.01158599928021431, + 0.00886025745421648, + 0.013612247072160244, + 0.00773094454780221, + 0.008493981324136257, + 0.01303871814161539, + 0.01552835013717413, + 0.030336206778883934, + 0.008179457858204842, + 0.011115609668195248, + 0.01343537401407957, + 0.013330555520951748, + 0.012053394690155983, + 0.05964341387152672, + 0.028523635119199753, + 0.07014361023902893, + 0.01807723008096218, + 0.020590245723724365, + 0.02563086524605751, + 0.019020292907953262, + 0.02023390866816044, + 0.013632082380354404, + 0.008531919680535793, + 0.013162411749362946, + 0.004760394804179668, + 0.011451617814600468, + 0.011973151005804539, + 0.009626430459320545, + 0.008792947977781296, + 0.012572681531310081, + 0.009533751755952835, + 0.010492702946066856, + 0.005579493939876556, + 0.009696651250123978, + 0.01357925683259964, + 0.007988328114151955, + 0.006835481151938438, + 0.010282795876264572, + 0.008354291319847107, + 0.01217286940664053, + 0.003741412190720439, + 0.005308247171342373, + 0.006243784911930561, + 0.0058018798008561134, + 0.003565241815522313, + 0.010373431257903576, + 0.011666984297335148, + 0.01149960607290268, + 0.006213484797626734, + 0.0073921955190598965, + 0.00784976501017809, + 0.009534797631204128, + 0.006994714494794607, + 0.007893506437540054, + 0.008624269627034664, + 0.00751475989818573, + 0.0056845382787287235, + 0.008266258053481579, + 0.0092769805341959, + 0.008042301051318645, + 0.0070161703042685986 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 3 + ], + "selected_reward": -0.25213611125946045 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a thin, middle-aged man with a neatly trimmed beard?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a round-necked, dark blue silk robe with cloud patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene lit primarily by cool, bright moonlight casting sharp shadows on the face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the man holding a simple white porcelain wine glass raised to eye level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background composed of carved pavilion railings and distant swaying bamboo silhouettes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.12703020870685577 + ], + "opposite_logprob_tokens": [ + -2.12703013420105 + ], + "logprob_reward": -0.12703020870685577, + "yes_no_margin_reward": 1.9999998807907104, + "token_credit_reward": -0.12703020870685577, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11301474273204803, + 0.404621958732605, + 0.0839269682765007, + 0.0651504397392273, + 0.04381886124610901, + 0.03875381126999855, + 0.056941449642181396, + 0.10128992050886154, + 0.040148116648197174, + 0.05233367905020714 + ], + "credit_map": [ + 0.013498487882316113, + 0.014907319098711014, + 0.010525166057050228, + 0.026907728984951973, + 0.010714417323470116, + 0.01331839058548212, + 0.010786443017423153, + 0.012356805615127087, + 0.06104019284248352, + 0.06371897459030151, + 0.05771690607070923, + 0.08404728770256042, + 0.02834845520555973, + 0.03919519856572151, + 0.028619928285479546, + 0.041935067623853683, + 0.008489779196679592, + 0.011615145951509476, + 0.011662141419947147, + 0.01682826690375805, + 0.007454948965460062, + 0.009736466221511364, + 0.008912363089621067, + 0.009227865375578403, + 0.0074224285781383514, + 0.008503642864525318, + 0.01486550085246563, + 0.005129679571837187, + 0.005823349114507437, + 0.005813432391732931, + 0.01205269992351532, + 0.0055397129617631435, + 0.008637622930109501, + 0.006571539212018251, + 0.0056096636690199375, + 0.0035562431439757347, + 0.005133805796504021, + 0.003851289628073573, + 0.004572093952447176, + 0.005886607337743044, + 0.006244315765798092, + 0.0031413473188877106, + 0.005755800288170576, + 0.0032087264116853476, + 0.005239960737526417, + 0.005136897321790457, + 0.005343506578356028, + 0.004683257546275854, + 0.012687228620052338, + 0.005238467361778021, + 0.011708253063261509, + 0.0035746817011386156, + 0.006188100669533014, + 0.007745000533759594, + 0.005056197755038738, + 0.00474352715536952, + 0.014400728046894073, + 0.01471067313104868, + 0.01382905151695013, + 0.00760603928938508, + 0.012453224509954453, + 0.015514211729168892, + 0.01127285324037075, + 0.011503146030008793, + 0.004973600618541241, + 0.005342161748558283, + 0.008306038565933704, + 0.00337063311599195, + 0.004074915312230587, + 0.004679809790104628, + 0.004236381035298109, + 0.0051645818166434765, + 0.006183408200740814, + 0.006812694948166609, + 0.010165796615183353, + 0.0049824840389192104, + 0.005652454681694508, + 0.005842682905495167, + 0.004968863911926746, + 0.007725302129983902 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 7 + ], + "selected_reward": -0.12703020870685577 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a thin, middle-aged man with a neatly trimmed beard?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a round-necked, dark blue silk robe with cloud patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene lit primarily by cool, bright moonlight casting sharp shadows on the face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the man holding a simple white porcelain wine glass raised to eye level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background composed of carved pavilion railings and distant swaying bamboo silhouettes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02634247951209545 + ], + "opposite_logprob_tokens": [ + -3.6513423919677734 + ], + "logprob_reward": -0.02634247951209545, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.02634247951209545, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1096532940864563, + 0.03490334376692772, + 0.10638688504695892, + 0.06480591744184494, + 0.042805902659893036, + 0.030638674274086952, + 0.07609907537698746, + 0.07481887936592102, + 0.2412765473127365, + 0.21861150860786438 + ], + "credit_map": [ + 0.022336680442094803, + 0.018864944577217102, + 0.01905767060816288, + 0.008002270013093948, + 0.009393609128892422, + 0.014807934872806072, + 0.007863487116992474, + 0.0093266936019063, + 0.004536642227321863, + 0.0050659324042499065, + 0.0043393527157604694, + 0.005308253690600395, + 0.004574054852128029, + 0.0036591121461242437, + 0.004120988305658102, + 0.0032990097533911467, + 0.010968001559376717, + 0.014397278428077698, + 0.014183226972818375, + 0.01685798168182373, + 0.010399316437542439, + 0.016973111778497696, + 0.010360070504248142, + 0.012247900478541851, + 0.010426433756947517, + 0.010284962132573128, + 0.016649678349494934, + 0.0055352672934532166, + 0.00467654038220644, + 0.005410206038504839, + 0.005120591260492802, + 0.006702235899865627, + 0.008232071064412594, + 0.006196877919137478, + 0.005657975561916828, + 0.0033429013565182686, + 0.0051461318507790565, + 0.004760868847370148, + 0.00568907568231225, + 0.0037799987476319075, + 0.004361052066087723, + 0.0032358861062675714, + 0.0034636566415429115, + 0.0024839979596436024, + 0.0032622432336211205, + 0.004855156876146793, + 0.004430001135915518, + 0.004546681419014931, + 0.016318587586283684, + 0.007651807740330696, + 0.015087737701833248, + 0.005543322768062353, + 0.006799809634685516, + 0.009625893086194992, + 0.005728744901716709, + 0.009343169629573822, + 0.01165354810655117, + 0.009741448797285557, + 0.009977513924241066, + 0.006094498559832573, + 0.010012785904109478, + 0.00947333499789238, + 0.007358063943684101, + 0.010507683269679546, + 0.043666768819093704, + 0.04231750965118408, + 0.047410838305950165, + 0.018817562609910965, + 0.01867152564227581, + 0.03298081085085869, + 0.018750883638858795, + 0.018660636618733406, + 0.029482174664735794, + 0.029504816979169846, + 0.036841124296188354, + 0.033433180302381516, + 0.018131988123059273, + 0.029012732207775116, + 0.02216610498726368, + 0.020039385184645653 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 0 + ], + "selected_reward": -0.02634247951209545 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "tight, close-up shot from a slightly upward angle focuses on a Tang Dynasty poet. He is a thin, middle-aged man with a neatly trimmed beard. His long, jet-black hair is tied high with an ancient jade hairpin, a few strands of hair blowing gently against his temples in the night breeze. He wears a round-necked, dark blue silk robe embroidered with elegant cloud patterns, which gleams softly in the moonlight. The cool, bright moonlight pours down like mercury, serving as the primary source of light, not only outlining the three-dimensional contours of his face but also casting sharp shadows on his high nose and cheekbones. At this moment, his cheeks are slightly tipsy, his eyes narrowed, a glint of intoxication and inspiration mingling. A faint smile plays at the corners of his mouth, and his brows are filled with a carefree and relaxed expression. He slowly raised a simple white porcelain wine glass to eye level, as if greeting the bright moon. His lips moved slightly, as if he were whispering a line of poetry that had just come to mind. The background was the carved railings of a pavilion and the silhouettes of a few distant bamboo clumps swaying gently in the night breeze. The entire scene, dominated by deep blue and inky black, was punctuated by the silvery moonlight, creating a tranquil, intoxicating, and poetic atmosphere filled with creative passion.", + "seed": 9950, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000021/group_12/index_3731_seed_9950_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": -0.11334843281656504, + "vlm_logprob": -0.11334843281656504, + "vlm_yes_no_margin": 2.6, + "vlm_token_credit": -0.11334843281656504, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a thin, middle-aged man with a neatly trimmed beard?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a round-necked, dark blue silk robe with cloud patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene lit primarily by cool, bright moonlight casting sharp shadows on the face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the man holding a simple white porcelain wine glass raised to eye level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background composed of carved pavilion railings and distant swaying bamboo silhouettes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05577415972948074 + ], + "opposite_logprob_tokens": [ + -2.930774211883545 + ], + "logprob_reward": -0.05577415972948074, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05577415972948074, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08739591389894485, + 0.043039463460445404, + 0.07886472344398499, + 0.18350563943386078, + 0.2281673550605774, + 0.18752025067806244, + 0.023472167551517487, + 0.03652403503656387, + 0.028246421366930008, + 0.10326404869556427 + ], + "credit_map": [ + 0.013948087580502033, + 0.01404778752475977, + 0.009271008893847466, + 0.00835349503904581, + 0.0074069323018193245, + 0.009682024829089642, + 0.012457659468054771, + 0.01222891453653574, + 0.0071549322456121445, + 0.008563324809074402, + 0.004533689469099045, + 0.0038959993980824947, + 0.004226604010909796, + 0.005180262494832277, + 0.004064379725605249, + 0.005420268978923559, + 0.026036672294139862, + 0.0232985932379961, + 0.0069306823424994946, + 0.00411650724709034, + 0.0032896436750888824, + 0.004456193186342716, + 0.003993177320808172, + 0.006743252743035555, + 0.016388706862926483, + 0.0630519762635231, + 0.011146108619868755, + 0.0038029449060559273, + 0.008762347511947155, + 0.007082977797836065, + 0.006896074395626783, + 0.06637448817491531, + 0.06644986569881439, + 0.026626259088516235, + 0.01572991907596588, + 0.007218676619231701, + 0.0074472310952842236, + 0.012922576628625393, + 0.00613043736666441, + 0.08564239740371704, + 0.03416217491030693, + 0.024919072166085243, + 0.017093710601329803, + 0.009695920161902905, + 0.02444274164736271, + 0.02059534192085266, + 0.021767936646938324, + 0.03484337031841278, + 0.0033526949118822813, + 0.0034687118604779243, + 0.00281511596404016, + 0.002384948544204235, + 0.002480240073055029, + 0.0031030080281198025, + 0.002021081978455186, + 0.0038463661912828684, + 0.006047265138477087, + 0.006305328570306301, + 0.0037861245218664408, + 0.0030907760374248028, + 0.004958445206284523, + 0.0048283776268363, + 0.0031415701378136873, + 0.004366149660199881, + 0.005155864637345076, + 0.0060225846245884895, + 0.0029917960055172443, + 0.003263586200773716, + 0.0027289418503642082, + 0.0024811194743961096, + 0.0025161614175885916, + 0.003086367156356573, + 0.020668305456638336, + 0.01676921360194683, + 0.009592561982572079, + 0.004978789482265711, + 0.019825546070933342, + 0.014541634358465672, + 0.0064541371539235115, + 0.010433858260512352 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 5, + 3 + ], + "selected_reward": -0.05577415972948074 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a thin, middle-aged man with a neatly trimmed beard?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a round-necked, dark blue silk robe with cloud patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene lit primarily by cool, bright moonlight casting sharp shadows on the face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the man holding a simple white porcelain wine glass raised to eye level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background composed of carved pavilion railings and distant swaying bamboo silhouettes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04309263080358505 + ], + "opposite_logprob_tokens": [ + -3.168092727661133 + ], + "logprob_reward": -0.04309263080358505, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04309263080358505, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19248807430267334, + 0.11129453778266907, + 0.07087546586990356, + 0.2782647907733917, + 0.0602496974170208, + 0.04770753160119057, + 0.03230082988739014, + 0.04316065460443497, + 0.03848530724644661, + 0.12517307698726654 + ], + "credit_map": [ + 0.02839852124452591, + 0.026423361152410507, + 0.029485056176781654, + 0.027385754510760307, + 0.01867259480059147, + 0.0152998436242342, + 0.02136850170791149, + 0.025454457849264145, + 0.02010667510330677, + 0.01784970611333847, + 0.01826733909547329, + 0.01843693107366562, + 0.007355307694524527, + 0.011647361330688, + 0.008269139565527439, + 0.009362082928419113, + 0.015386504121124744, + 0.01950654946267605, + 0.007308875676244497, + 0.005248186178505421, + 0.0050529432483017445, + 0.006187885999679565, + 0.0056124539114534855, + 0.006572069134563208, + 0.04963850975036621, + 0.10960539430379868, + 0.016025420278310776, + 0.006149309687316418, + 0.011702282354235649, + 0.00784615334123373, + 0.006972134578973055, + 0.07032559812068939, + 0.015661245211958885, + 0.017477678135037422, + 0.00559692457318306, + 0.004564215429127216, + 0.004013321362435818, + 0.004074336029589176, + 0.0030795487109571695, + 0.005782432854175568, + 0.007619925774633884, + 0.007930499501526356, + 0.006055578589439392, + 0.004944098647683859, + 0.005460965912789106, + 0.005834316369146109, + 0.003611920168623328, + 0.006250231061130762, + 0.004820309113711119, + 0.00415033521130681, + 0.0039192186668515205, + 0.002614022931084037, + 0.004372036084532738, + 0.003598588053137064, + 0.0030011306516826153, + 0.0058251917362213135, + 0.006149062421172857, + 0.0057234251871705055, + 0.0047141206450760365, + 0.0050721061415970325, + 0.005613774061203003, + 0.00526910787448287, + 0.0038223485462367535, + 0.006796712521463633, + 0.0068945735692977905, + 0.005880053620785475, + 0.004538342356681824, + 0.005978603847324848, + 0.002956916345283389, + 0.0032285433262586594, + 0.003243149956688285, + 0.005765127018094063, + 0.041699595749378204, + 0.03550370782613754, + 0.009794800542294979, + 0.005641137715429068, + 0.010832768864929676, + 0.009888558648526669, + 0.004179118201136589, + 0.00763340387493372 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 9 + ], + "selected_reward": -0.04309263080358505 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a thin, middle-aged man with a neatly trimmed beard?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a round-necked, dark blue silk robe with cloud patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene lit primarily by cool, bright moonlight casting sharp shadows on the face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the man holding a simple white porcelain wine glass raised to eye level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background composed of carved pavilion railings and distant swaying bamboo silhouettes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2521255612373352 + ], + "opposite_logprob_tokens": [ + -1.5021255016326904 + ], + "logprob_reward": -0.2521255612373352, + "yes_no_margin_reward": 1.25, + "token_credit_reward": -0.2521255612373352, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18904365599155426, + 0.10667107999324799, + 0.10595323145389557, + 0.06770048290491104, + 0.1429506540298462, + 0.059299372136592865, + 0.11432826519012451, + 0.031248027458786964, + 0.08503405749797821, + 0.09777110815048218 + ], + "credit_map": [ + 0.0357898585498333, + 0.03517453745007515, + 0.0279097817838192, + 0.023745452985167503, + 0.014610163867473602, + 0.017095431685447693, + 0.017049362882971764, + 0.01766907423734665, + 0.020926572382450104, + 0.023259826004505157, + 0.013407541438937187, + 0.01063846331089735, + 0.0077371359802782536, + 0.010105517692863941, + 0.009367691352963448, + 0.011228346265852451, + 0.02764594741165638, + 0.03381088748574257, + 0.008916917257010937, + 0.0046544671058654785, + 0.0069238427095115185, + 0.006703549064695835, + 0.008727588690817356, + 0.008570040576159954, + 0.01899406500160694, + 0.022789424285292625, + 0.005611265078186989, + 0.0036466061137616634, + 0.0036254730075597763, + 0.002355479635298252, + 0.0033047397155314684, + 0.007373438216745853, + 0.03769766166806221, + 0.018023574724793434, + 0.011176724918186665, + 0.00788317434489727, + 0.004903323017060757, + 0.006474408321082592, + 0.005779214669018984, + 0.05101258307695389, + 0.01397637091577053, + 0.008838987909257412, + 0.006240843329578638, + 0.006150529254227877, + 0.005503401160240173, + 0.004743973258882761, + 0.004607704468071461, + 0.009237567894160748, + 0.02030755952000618, + 0.027153614908456802, + 0.014356672763824463, + 0.008333631791174412, + 0.008992955088615417, + 0.009737811051309109, + 0.005501989740878344, + 0.019944043830037117, + 0.004639832768589258, + 0.0057909125462174416, + 0.003473562654107809, + 0.0032567817252129316, + 0.002826334908604622, + 0.0037787759210914373, + 0.0034911169204860926, + 0.00399071304127574, + 0.02906409464776516, + 0.02499292977154255, + 0.006256901193410158, + 0.004077457822859287, + 0.004645892884582281, + 0.0030660522170364857, + 0.004578740801662207, + 0.008351998403668404, + 0.022242166101932526, + 0.019547374919056892, + 0.011608581058681011, + 0.008651578798890114, + 0.010524412617087364, + 0.010914741083979607, + 0.006100831553339958, + 0.008181426674127579 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 6 + ], + "selected_reward": -0.2521255612373352 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a thin, middle-aged man with a neatly trimmed beard?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a round-necked, dark blue silk robe with cloud patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene lit primarily by cool, bright moonlight casting sharp shadows on the face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the man holding a simple white porcelain wine glass raised to eye level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background composed of carved pavilion railings and distant swaying bamboo silhouettes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2015569806098938 + ], + "opposite_logprob_tokens": [ + -1.701556921005249 + ], + "logprob_reward": -0.2015569806098938, + "yes_no_margin_reward": 1.5, + "token_credit_reward": -0.2015569806098938, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1725408136844635, + 0.11290058493614197, + 0.0454404279589653, + 0.04906085506081581, + 0.327481746673584, + 0.03436104208230972, + 0.04801533371210098, + 0.1177564486861229, + 0.05648702383041382, + 0.0359557569026947 + ], + "credit_map": [ + 0.028429463505744934, + 0.036519553512334824, + 0.022589514032006264, + 0.01755348965525627, + 0.016113486140966415, + 0.012614336796104908, + 0.016236476600170135, + 0.022484512999653816, + 0.018528763204813004, + 0.02078094892203808, + 0.012150131165981293, + 0.007290900219231844, + 0.010923266410827637, + 0.013032671995460987, + 0.013969537802040577, + 0.016224361956119537, + 0.007766406983137131, + 0.008060460910201073, + 0.004087183624505997, + 0.002588598057627678, + 0.005162571091204882, + 0.006257835775613785, + 0.004922308027744293, + 0.006595062557607889, + 0.00799112394452095, + 0.0146404430270195, + 0.004583158530294895, + 0.00328338285908103, + 0.003280142555013299, + 0.003092917613685131, + 0.0030313972383737564, + 0.00915828999131918, + 0.121178038418293, + 0.019405120983719826, + 0.026190537959337234, + 0.013130173087120056, + 0.007603057660162449, + 0.014069258235394955, + 0.0070786383002996445, + 0.11882689595222473, + 0.006069052964448929, + 0.003995983861386776, + 0.005345535464584827, + 0.0036242101341485977, + 0.0027667914982885122, + 0.002195685636252165, + 0.0026329089887440205, + 0.007730873301625252, + 0.0068158418871462345, + 0.010085069574415684, + 0.0058014546521008015, + 0.005526434164494276, + 0.004030351527035236, + 0.004528365097939968, + 0.0035546866711229086, + 0.007673132233321667, + 0.027393227443099022, + 0.024381594732403755, + 0.009870299138128757, + 0.015793513506650925, + 0.009006025269627571, + 0.0095642926171422, + 0.009325416758656502, + 0.012422081083059311, + 0.00788117479532957, + 0.010713182389736176, + 0.0066528026945889, + 0.007568578235805035, + 0.0046112388372421265, + 0.004816221073269844, + 0.005203738808631897, + 0.009040084667503834, + 0.005743509158492088, + 0.005915993824601173, + 0.003951665014028549, + 0.0027315744664520025, + 0.005775138735771179, + 0.004242525435984135, + 0.0026756569277495146, + 0.004919694270938635 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 7 + ], + "selected_reward": -0.2015569806098938 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a thin, middle-aged man with a neatly trimmed beard?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a round-necked, dark blue silk robe with cloud patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene lit primarily by cool, bright moonlight casting sharp shadows on the face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the man holding a simple white porcelain wine glass raised to eye level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background composed of carved pavilion railings and distant swaying bamboo silhouettes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"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": [ + 0.1319267600774765, + 0.08375824242830276, + 0.07423274964094162, + 0.05541609227657318, + 0.14352236688137054, + 0.11335041373968124, + 0.059954047203063965, + 0.03486014902591705, + 0.02699972689151764, + 0.2759794294834137 + ], + "credit_map": [ + 0.0243516992777586, + 0.021530387923121452, + 0.019047824665904045, + 0.022798867896199226, + 0.009677227586507797, + 0.010891285724937916, + 0.011428732424974442, + 0.012200742028653622, + 0.016108883544802666, + 0.012667338363826275, + 0.012226194143295288, + 0.015223536640405655, + 0.0051653762347996235, + 0.007367163430899382, + 0.007403295952826738, + 0.007596451323479414, + 0.021977247670292854, + 0.02073539048433304, + 0.006893807556480169, + 0.006201983895152807, + 0.0032702942844480276, + 0.004482877440750599, + 0.004125385079532862, + 0.00654577137902379, + 0.00904581043869257, + 0.019900264218449593, + 0.0035475054755806923, + 0.0019423342309892178, + 0.0027259550988674164, + 0.0018347775330767035, + 0.0026256833225488663, + 0.013793770223855972, + 0.057713303714990616, + 0.03423064202070236, + 0.011185661889612675, + 0.006291833706200123, + 0.004173208028078079, + 0.005217938218265772, + 0.0033520390279591084, + 0.02135774865746498, + 0.040191374719142914, + 0.025733737275004387, + 0.008921343833208084, + 0.007848028093576431, + 0.0036158470902591944, + 0.0037641741801053286, + 0.004441923927515745, + 0.018833985552191734, + 0.011803739704191685, + 0.009052815847098827, + 0.008525107055902481, + 0.007134758401662111, + 0.004051013384014368, + 0.005866724997758865, + 0.00487653911113739, + 0.008643352426588535, + 0.005292393732815981, + 0.005939972586929798, + 0.0050799003802239895, + 0.006173478905111551, + 0.0026933259796351194, + 0.0028008408844470978, + 0.0028343142475932837, + 0.004045926034450531, + 0.00437224330380559, + 0.004884932655841112, + 0.003405624534934759, + 0.0026691004168242216, + 0.0027632517740130424, + 0.0034370601642876863, + 0.002081142505630851, + 0.003386375028640032, + 0.0657920241355896, + 0.08391834050416946, + 0.02879512496292591, + 0.02812661975622177, + 0.022787857800722122, + 0.020571105182170868, + 0.010357736609876156, + 0.01563061773777008 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 4, + 0 + ], + "selected_reward": -0.014192831702530384 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "tight, close-up shot from a slightly upward angle focuses on a Tang Dynasty poet. He is a thin, middle-aged man with a neatly trimmed beard. His long, jet-black hair is tied high with an ancient jade hairpin, a few strands of hair blowing gently against his temples in the night breeze. He wears a round-necked, dark blue silk robe embroidered with elegant cloud patterns, which gleams softly in the moonlight. The cool, bright moonlight pours down like mercury, serving as the primary source of light, not only outlining the three-dimensional contours of his face but also casting sharp shadows on his high nose and cheekbones. At this moment, his cheeks are slightly tipsy, his eyes narrowed, a glint of intoxication and inspiration mingling. A faint smile plays at the corners of his mouth, and his brows are filled with a carefree and relaxed expression. He slowly raised a simple white porcelain wine glass to eye level, as if greeting the bright moon. His lips moved slightly, as if he were whispering a line of poetry that had just come to mind. The background was the carved railings of a pavilion and the silhouettes of a few distant bamboo clumps swaying gently in the night breeze. The entire scene, dominated by deep blue and inky black, was punctuated by the silvery moonlight, creating a tranquil, intoxicating, and poetic atmosphere filled with creative passion.", + "seed": 9951, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000021/group_12/index_3731_seed_9951_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": -0.10797260664403438, + "vlm_logprob": -0.10797260664403438, + "vlm_yes_no_margin": 2.525, + "vlm_token_credit": -0.10797260664403438, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a thin, middle-aged man with a neatly trimmed beard?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a round-necked, dark blue silk robe with cloud patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene lit primarily by cool, bright moonlight casting sharp shadows on the face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the man holding a simple white porcelain wine glass raised to eye level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background composed of carved pavilion railings and distant swaying bamboo silhouettes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06517516821622849 + ], + "opposite_logprob_tokens": [ + -2.8151750564575195 + ], + "logprob_reward": -0.06517516821622849, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06517516821622849, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2932127118110657, + 0.08592117577791214, + 0.03710692375898361, + 0.04260358586907387, + 0.07810398191213608, + 0.08343248069286346, + 0.07917400449514389, + 0.24240735173225403, + 0.028173254802823067, + 0.02986452542245388 + ], + "credit_map": [ + 0.054681431502103806, + 0.019143184646964073, + 0.0531085841357708, + 0.02691335417330265, + 0.021197617053985596, + 0.046885885298252106, + 0.03403551131486893, + 0.037247125059366226, + 0.011774774640798569, + 0.006549373269081116, + 0.011828786693513393, + 0.006455505732446909, + 0.007838231511414051, + 0.014681805856525898, + 0.012824987061321735, + 0.013967717066407204, + 0.004446149338036776, + 0.004251830279827118, + 0.0050488184206187725, + 0.003938887733966112, + 0.0036405990831553936, + 0.004783423617482185, + 0.005865262355655432, + 0.005131954792886972, + 0.005200119689106941, + 0.0033061327412724495, + 0.005866338964551687, + 0.003737048013135791, + 0.0039054632652550936, + 0.00615562591701746, + 0.0068433345295488834, + 0.007589525077491999, + 0.011048060841858387, + 0.006293287966400385, + 0.01051158644258976, + 0.0060350727289915085, + 0.010080928914248943, + 0.009295034222304821, + 0.013184024021029472, + 0.011655987240374088, + 0.005953507497906685, + 0.004883558489382267, + 0.010388895869255066, + 0.007162858732044697, + 0.010492686182260513, + 0.010894370265305042, + 0.022891337051987648, + 0.01076527126133442, + 0.00806974247097969, + 0.004658036399632692, + 0.00912402756512165, + 0.0039842878468334675, + 0.009208347648382187, + 0.014147326350212097, + 0.011764035560190678, + 0.01821819692850113, + 0.017405422404408455, + 0.0168477613478899, + 0.02855457179248333, + 0.018029438331723213, + 0.03579892963171005, + 0.03746272251009941, + 0.04984397813677788, + 0.03846452757716179, + 0.002423609606921673, + 0.002062711399048567, + 0.00408187136054039, + 0.0025060423649847507, + 0.002624716842547059, + 0.005156895145773888, + 0.004675030708312988, + 0.004642377607524395, + 0.003749822499230504, + 0.0024580650497227907, + 0.0050847711972892284, + 0.0019955544266849756, + 0.0018417277606204152, + 0.005404125899076462, + 0.003241786966100335, + 0.006088672671467066 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 1 + ], + "selected_reward": -0.06517516821622849 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a thin, middle-aged man with a neatly trimmed beard?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a round-necked, dark blue silk robe with cloud patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene lit primarily by cool, bright moonlight casting sharp shadows on the face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the man holding a simple white porcelain wine glass raised to eye level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background composed of carved pavilion railings and distant swaying bamboo silhouettes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04869156703352928 + ], + "opposite_logprob_tokens": [ + -3.048691511154175 + ], + "logprob_reward": -0.04869156703352928, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04869156703352928, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20539380609989166, + 0.09941092878580093, + 0.11165523529052734, + 0.05973371863365173, + 0.08760856837034225, + 0.09879288822412491, + 0.08799852430820465, + 0.12674753367900848, + 0.037126701325178146, + 0.08553209155797958 + ], + "credit_map": [ + 0.019263189285993576, + 0.021878426894545555, + 0.021762417629361153, + 0.019513973966240883, + 0.019546600058674812, + 0.029549162834882736, + 0.039989955723285675, + 0.03389009088277817, + 0.010762274265289307, + 0.011415045708417892, + 0.011331714689731598, + 0.011400447227060795, + 0.010161536745727062, + 0.015585770830512047, + 0.015540335327386856, + 0.013213802129030228, + 0.0196788739413023, + 0.010296620428562164, + 0.01661004312336445, + 0.008304393850266933, + 0.009329442866146564, + 0.01482907310128212, + 0.015882940962910652, + 0.016723841428756714, + 0.00542509974911809, + 0.004543545190244913, + 0.008264596574008465, + 0.005436723120510578, + 0.005801670718938112, + 0.007148640230298042, + 0.010730646550655365, + 0.012382794171571732, + 0.005997523665428162, + 0.007094442844390869, + 0.013151477091014385, + 0.006191413849592209, + 0.012410893104970455, + 0.010173233225941658, + 0.021198784932494164, + 0.011390794068574905, + 0.01130308024585247, + 0.0060752443969249725, + 0.01188827957957983, + 0.007972784340381622, + 0.00937830749899149, + 0.013540409505367279, + 0.018814342096447945, + 0.019820434972643852, + 0.011544839479029179, + 0.006721668876707554, + 0.013018984347581863, + 0.006600877270102501, + 0.005784286186099052, + 0.015997696667909622, + 0.009450571611523628, + 0.018879594281315804, + 0.008341805078089237, + 0.008250640705227852, + 0.015697605907917023, + 0.013797505758702755, + 0.019620047882199287, + 0.01889822445809841, + 0.027987198904156685, + 0.014154507778584957, + 0.0036828904412686825, + 0.004921368323266506, + 0.004918928258121014, + 0.0030317765194922686, + 0.003528706729412079, + 0.0052625262178480625, + 0.006827103439718485, + 0.004953398369252682, + 0.014051820151507854, + 0.004656821954995394, + 0.015782328322529793, + 0.005018739495426416, + 0.004537345841526985, + 0.016122708097100258, + 0.007855714298784733, + 0.017506616190075874 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 2 + ], + "selected_reward": -0.04869156703352928 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a thin, middle-aged man with a neatly trimmed beard?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a round-necked, dark blue silk robe with cloud patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene lit primarily by cool, bright moonlight casting sharp shadows on the face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the man holding a simple white porcelain wine glass raised to eye level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background composed of carved pavilion railings and distant swaying bamboo silhouettes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2255670428276062 + ], + "opposite_logprob_tokens": [ + -1.600567102432251 + ], + "logprob_reward": -0.2255670428276062, + "yes_no_margin_reward": 1.375, + "token_credit_reward": -0.2255670428276062, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.41425952315330505, + 0.10231903195381165, + 0.08724014461040497, + 0.0499725416302681, + 0.0928785428404808, + 0.02699384093284607, + 0.06316711008548737, + 0.05927605554461479, + 0.04362867772579193, + 0.06026458367705345 + ], + "credit_map": [ + 0.06303941458463669, + 0.04802005738019943, + 0.06757117062807083, + 0.03508702293038368, + 0.027096997946500778, + 0.06843609362840652, + 0.0355655811727047, + 0.06944317370653152, + 0.011907928623259068, + 0.015217586420476437, + 0.01133382972329855, + 0.014736912213265896, + 0.005771780852228403, + 0.012690695002675056, + 0.014481917954981327, + 0.016178373247385025, + 0.013655919581651688, + 0.007084303069859743, + 0.012400311417877674, + 0.011289595626294613, + 0.008038509637117386, + 0.010442322120070457, + 0.010242260992527008, + 0.01408692542463541, + 0.0038129554595798254, + 0.004831478465348482, + 0.006229042075574398, + 0.005126446019858122, + 0.006705166306346655, + 0.00687817158177495, + 0.007059355732053518, + 0.009329925291240215, + 0.0062620132230222225, + 0.005818193778395653, + 0.012214678339660168, + 0.007277703378349543, + 0.011640314012765884, + 0.01338090281933546, + 0.014688012190163136, + 0.021596716716885567, + 0.0027124902699142694, + 0.0029960377141833305, + 0.0033957764972001314, + 0.0018270554719492793, + 0.003351248800754547, + 0.0032413352746516466, + 0.005630834028124809, + 0.003839062759652734, + 0.012440184131264687, + 0.006868474651128054, + 0.009883308783173561, + 0.003985780756920576, + 0.004476028494536877, + 0.010211783461272717, + 0.006306577939540148, + 0.00899496953934431, + 0.006398103199899197, + 0.004767211154103279, + 0.0096243591979146, + 0.004996098577976227, + 0.0067793503403663635, + 0.007170819211751223, + 0.010184056125581264, + 0.009356056340038776, + 0.004676796495914459, + 0.003861888311803341, + 0.005392955616116524, + 0.004799395799636841, + 0.003975193947553635, + 0.0054635158739984035, + 0.0056206053122878075, + 0.009838324040174484, + 0.009997847490012646, + 0.0036974840331822634, + 0.010451128706336021, + 0.0039527202025055885, + 0.004869057331234217, + 0.00993106048554182, + 0.005731854122132063, + 0.0116334343329072 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 4 + ], + "selected_reward": -0.2255670428276062 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a thin, middle-aged man with a neatly trimmed beard?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a round-necked, dark blue silk robe with cloud patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene lit primarily by cool, bright moonlight casting sharp shadows on the face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the man holding a simple white porcelain wine glass raised to eye level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background composed of carved pavilion railings and distant swaying bamboo silhouettes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1798529028892517 + ], + "opposite_logprob_tokens": [ + -1.8048529624938965 + ], + "logprob_reward": -0.1798529028892517, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.1798529028892517, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3067745566368103, + 0.1201997846364975, + 0.0949481949210167, + 0.07584257423877716, + 0.06841591000556946, + 0.05147447809576988, + 0.042998407036066055, + 0.07898224890232086, + 0.07186266779899597, + 0.08850117772817612 + ], + "credit_map": [ + 0.057705506682395935, + 0.020324774086475372, + 0.04820976406335831, + 0.01993384212255478, + 0.02511826902627945, + 0.05573008209466934, + 0.03354647010564804, + 0.04620585963129997, + 0.013816875405609608, + 0.010571826249361038, + 0.017018146812915802, + 0.00831652618944645, + 0.012466097250580788, + 0.022032976150512695, + 0.020075974985957146, + 0.015901358798146248, + 0.00841081328690052, + 0.009071962907910347, + 0.011932538822293282, + 0.010913864709436893, + 0.010153873823583126, + 0.015845179557800293, + 0.011343623511493206, + 0.017276328057050705, + 0.006567216943949461, + 0.006752098444849253, + 0.009950609877705574, + 0.005496187135577202, + 0.0106864208355546, + 0.009160039015114307, + 0.014877010136842728, + 0.012352993711829185, + 0.00521687138825655, + 0.006238680332899094, + 0.00978278461843729, + 0.00628001568838954, + 0.00787560734897852, + 0.008196641691029072, + 0.01421381812542677, + 0.010611485689878464, + 0.003278101095929742, + 0.003434415441006422, + 0.0060393656603991985, + 0.002989343600347638, + 0.00773630803450942, + 0.00846881978213787, + 0.011282739229500294, + 0.008245385251939297, + 0.007024661637842655, + 0.0036079566925764084, + 0.007244115229696035, + 0.0028190528973937035, + 0.0042142667807638645, + 0.00625265296548605, + 0.005271250382065773, + 0.006564452312886715, + 0.008123157545924187, + 0.0048849210143089294, + 0.010590852238237858, + 0.005021593999117613, + 0.010626308619976044, + 0.01229024212807417, + 0.01334045734256506, + 0.014104718342423439, + 0.006707926746457815, + 0.008646545000374317, + 0.010097757913172245, + 0.008289304561913013, + 0.003624578472226858, + 0.010981976985931396, + 0.005433822050690651, + 0.018080754205584526, + 0.008810246363282204, + 0.005632498767226934, + 0.011087704449892044, + 0.007285833824425936, + 0.013200027868151665, + 0.01141977496445179, + 0.013124680146574974, + 0.01794040948152542 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.1798529028892517 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a thin, middle-aged man with a neatly trimmed beard?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a round-necked, dark blue silk robe with cloud patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene lit primarily by cool, bright moonlight casting sharp shadows on the face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the man holding a simple white porcelain wine glass raised to eye level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background composed of carved pavilion railings and distant swaying bamboo silhouettes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02057635225355625 + ], + "opposite_logprob_tokens": [ + -3.895576238632202 + ], + "logprob_reward": -0.02057635225355625, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.02057635225355625, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.426602840423584, + 0.19060541689395905, + 0.09159399569034576, + 0.04054708778858185, + 0.04935404658317566, + 0.03094634972512722, + 0.014145788736641407, + 0.04071692004799843, + 0.09290902316570282, + 0.022578516975045204 + ], + "credit_map": [ + 0.07426412403583527, + 0.03523148223757744, + 0.07930566370487213, + 0.03505022078752518, + 0.02498750016093254, + 0.07973404228687286, + 0.03426769748330116, + 0.06376209110021591, + 0.03905797749757767, + 0.016201771795749664, + 0.03826310485601425, + 0.015365354716777802, + 0.011250183917582035, + 0.035497914999723434, + 0.012394510209560394, + 0.02257460728287697, + 0.0171981044113636, + 0.00873317290097475, + 0.012806170620024204, + 0.007651427295058966, + 0.0074111949652433395, + 0.012598471716046333, + 0.010939659550786018, + 0.014255797490477562, + 0.005422702990472317, + 0.004999401047825813, + 0.005258041433990002, + 0.004704800900071859, + 0.0024067421909421682, + 0.006330638658255339, + 0.004281866364181042, + 0.007142896298319101, + 0.009054761379957199, + 0.004287455230951309, + 0.008927738294005394, + 0.005292780697345734, + 0.003334669629111886, + 0.006709542125463486, + 0.0037878563161939383, + 0.007959242910146713, + 0.0031170162837952375, + 0.0038132246118038893, + 0.004050303250551224, + 0.002071988768875599, + 0.0021068546921014786, + 0.005344914272427559, + 0.002708189655095339, + 0.007733855862170458, + 0.0021687515545636415, + 0.0014574740780517459, + 0.002118155360221863, + 0.0009055695845745504, + 0.001681615482084453, + 0.0018212664872407913, + 0.0023074366617947817, + 0.001685519004240632, + 0.003758111037313938, + 0.00492634205147624, + 0.005188910290598869, + 0.004368005786091089, + 0.0033321629744023085, + 0.006418643984943628, + 0.003572966204956174, + 0.009151777252554893, + 0.010060456581413746, + 0.009309119544923306, + 0.013663897290825844, + 0.00938440766185522, + 0.005082345567643642, + 0.016549380496144295, + 0.0063248807564377785, + 0.022534536197781563, + 0.004234823398292065, + 0.0015660310164093971, + 0.004064151085913181, + 0.0013422429328784347, + 0.0011848908616229892, + 0.004510348662734032, + 0.0022289890330284834, + 0.003447040216997266 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 8 + ], + "selected_reward": -0.02057635225355625 + } + ], + "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/0000021/group_13/index_tensor([3040, 3040])_group_13_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000021/group_13/index_tensor([3040, 3040])_group_13_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..422967209c4cfd1d1262418eb7d1db2dee0a4490 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000021/group_13/index_tensor([3040, 3040])_group_13_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "slow, eye-level wide-angle shot gracefully pans from left to right, capturing a group of four joyful children, aged between six and nine, celebrating in a sun-drenched, historic cobblestone town square. The children are dressed in festive attire in the distinctive colors of the Polish flag. A young girl with blonde braids tied with a red ribbon, dressed in a plain white short-sleeved shirt and a lively red pleated skirt, sways with laughter. Beside her, a boy in a bright red polo shirt and crisp white shorts enthusiastically waves a small paper Polish flag on a thin wooden stick. Another girl wears a simple, elegant dress, split horizontally: a white top and a red bottom. The warm, golden light of the afternoon sun casts long, soft shadows on the stone, highlighting the cheerful, irreverent expressions on their faces. In the background, the charming, colorful facades of old European buildings appear softly out of focus, creating a beautiful bokeh effect that enhances the scene's depth. The overall atmosphere is one of pure, innocent patriotism and exuberant happiness, with the red and white of the vivid children’s clothing becoming the powerful, emotional centerpiece of the visual narrative.", + "seed": 28883, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000021/group_13/index_3040_seed_28883_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": -0.09170901579782367, + "vlm_logprob": -0.09170901579782367, + "vlm_yes_no_margin": 3.275, + "vlm_token_credit": -0.09170901579782367, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 eye-level wide-angle view that pans smoothly from left to right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are there exactly four children visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a sun-drenched historic cobblestone town square with old European buildings in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is one of the children a girl with blonde braids wearing a white shirt and a red pleated skirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is one of the children a boy in a red polo shirt and white shorts waving a small paper Polish flag on a stick?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1729726493358612 + ], + "opposite_logprob_tokens": [ + -2.2979726791381836 + ], + "logprob_reward": -0.1729726493358612, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.1729726493358612, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1112896203994751, + 0.303213894367218, + 0.08606202900409698, + 0.05461154505610466, + 0.11693815886974335, + 0.037818290293216705, + 0.07994363456964493, + 0.07377742975950241, + 0.08699125796556473, + 0.04935410991311073 + ], + "credit_map": [ + 0.018128011375665665, + 0.012180688790977001, + 0.015570572577416897, + 0.013400839641690254, + 0.014952221885323524, + 0.01647385023534298, + 0.00934155285358429, + 0.011241881176829338, + 0.0330590195953846, + 0.027699965983629227, + 0.042575426399707794, + 0.03900602087378502, + 0.03583087772130966, + 0.022758539766073227, + 0.06139606982469559, + 0.0408879779279232, + 0.010052582249045372, + 0.01111958920955658, + 0.01057436317205429, + 0.006715217139571905, + 0.011221880093216896, + 0.011454997584223747, + 0.006595015060156584, + 0.018328381702303886, + 0.006966584827750921, + 0.009984982199966908, + 0.004519032780081034, + 0.003998642321676016, + 0.006371039431542158, + 0.007016582880169153, + 0.007185048423707485, + 0.008569630794227123, + 0.01122039183974266, + 0.02123347297310829, + 0.011862609535455704, + 0.010314729064702988, + 0.014859642833471298, + 0.027703797444701195, + 0.00861409306526184, + 0.011129424907267094, + 0.004673144780099392, + 0.00955238938331604, + 0.0031125827226787806, + 0.0025509006809443235, + 0.007306643761694431, + 0.005326281767338514, + 0.002921168226748705, + 0.0023751782719045877, + 0.011282754130661488, + 0.009906145744025707, + 0.009584575891494751, + 0.01777729019522667, + 0.011807809583842754, + 0.008488466031849384, + 0.006664209067821503, + 0.004432385321706533, + 0.0071044797077775, + 0.00971780251711607, + 0.008504982106387615, + 0.007336293812841177, + 0.011530404910445213, + 0.012533226981759071, + 0.010781117714941502, + 0.006269123870879412, + 0.008473941124975681, + 0.012068833224475384, + 0.00962417759001255, + 0.012210257351398468, + 0.009649083949625492, + 0.01161743700504303, + 0.011369386687874794, + 0.01197813916951418, + 0.006830156315118074, + 0.008756129071116447, + 0.006672777701169252, + 0.00540842255577445, + 0.00656867166981101, + 0.006163828074932098, + 0.004882629960775375, + 0.004071493167430162 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 4, + 0 + ], + "selected_reward": -0.1729726493358612 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 eye-level wide-angle view that pans smoothly from left to right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are there exactly four children visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a sun-drenched historic cobblestone town square with old European buildings in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is one of the children a girl with blonde braids wearing a white shirt and a red pleated skirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is one of the children a boy in a red polo shirt and white shorts waving a small paper Polish flag on a stick?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2256619930267334 + ], + "opposite_logprob_tokens": [ + -1.6006619930267334 + ], + "logprob_reward": -0.2256619930267334, + "yes_no_margin_reward": 1.375, + "token_credit_reward": -0.2256619930267334, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12301087379455566, + 0.0673213079571724, + 0.3637585937976837, + 0.039505064487457275, + 0.05639934539794922, + 0.023918364197015762, + 0.12839920818805695, + 0.051496103405952454, + 0.07788659632205963, + 0.06830456107854843 + ], + "credit_map": [ + 0.017907308414578438, + 0.01331675797700882, + 0.015799159184098244, + 0.015179356560111046, + 0.013964153826236725, + 0.009891479276120663, + 0.02373368851840496, + 0.013218976557254791, + 0.0076378765515983105, + 0.00935481209307909, + 0.012081746943295002, + 0.007325357291847467, + 0.007960441522300243, + 0.01024699304252863, + 0.004942175000905991, + 0.007771908771246672, + 0.033788204193115234, + 0.05664501339197159, + 0.032485123723745346, + 0.03431868180632591, + 0.02910301461815834, + 0.0738367810845375, + 0.018038708716630936, + 0.08554301410913467, + 0.004658675752580166, + 0.008203347213566303, + 0.0036500589922070503, + 0.0033817817457020283, + 0.00600856076925993, + 0.006328742019832134, + 0.003109378507360816, + 0.004164517857134342, + 0.0065972828306257725, + 0.008276920765638351, + 0.005384825170040131, + 0.005279665347188711, + 0.008158736862242222, + 0.012110093608498573, + 0.004821732174605131, + 0.005770084448158741, + 0.0030235429294407368, + 0.004819836933165789, + 0.002359559992328286, + 0.0023881932720541954, + 0.003921296447515488, + 0.0036991548258811235, + 0.0016515945317223668, + 0.002055185381323099, + 0.015148979611694813, + 0.012897043488919735, + 0.012759329751133919, + 0.033156558871269226, + 0.016691263765096664, + 0.01434986013919115, + 0.015140704810619354, + 0.008255470544099808, + 0.0045901075936853886, + 0.0069109005853533745, + 0.006174618378281593, + 0.011702795512974262, + 0.006542467046529055, + 0.00599678372964263, + 0.005940230563282967, + 0.0036382018588483334, + 0.015683917328715324, + 0.00902788806706667, + 0.009130831807851791, + 0.008359819650650024, + 0.01254637073725462, + 0.008910028263926506, + 0.0074014184065163136, + 0.006826321594417095, + 0.006278275512158871, + 0.009847831912338734, + 0.00869195256382227, + 0.007384580560028553, + 0.012576096691191196, + 0.010030732490122318, + 0.00784965232014656, + 0.005645436700433493 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 6, + 0 + ], + "selected_reward": -0.2256619930267334 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 eye-level wide-angle view that pans smoothly from left to right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are there exactly four children visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a sun-drenched historic cobblestone town square with old European buildings in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is one of the children a girl with blonde braids wearing a white shirt and a red pleated skirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is one of the children a boy in a red polo shirt and white shorts waving a small paper Polish flag on a stick?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03811129927635193 + ], + "opposite_logprob_tokens": [ + -3.2881112098693848 + ], + "logprob_reward": -0.03811129927635193, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03811129927635193, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0911884531378746, + 0.4422300457954407, + 0.09338051825761795, + 0.09602408856153488, + 0.027094194665551186, + 0.11358045041561127, + 0.051992569118738174, + 0.02108108438551426, + 0.024539383128285408, + 0.038889236748218536 + ], + "credit_map": [ + 0.013269342482089996, + 0.00958306435495615, + 0.01230805553495884, + 0.008705704472959042, + 0.009582570753991604, + 0.021844591945409775, + 0.006689996458590031, + 0.009205126203596592, + 0.0489550344645977, + 0.041591059416532516, + 0.06057238206267357, + 0.04747045785188675, + 0.05514822155237198, + 0.04205195978283882, + 0.08928167074918747, + 0.05715925619006157, + 0.010322156362235546, + 0.011583942919969559, + 0.005872961133718491, + 0.004829194396734238, + 0.008781321346759796, + 0.023663340136408806, + 0.004617400001734495, + 0.02371019683778286, + 0.01195001695305109, + 0.021508680656552315, + 0.007162090390920639, + 0.006908407434821129, + 0.01792042702436447, + 0.013338077813386917, + 0.006081480998545885, + 0.011154904030263424, + 0.002218521200120449, + 0.003007159335538745, + 0.0025105574168264866, + 0.0029800692573189735, + 0.005757275968790054, + 0.004068808164447546, + 0.004126199055463076, + 0.00242560263723135, + 0.013615008443593979, + 0.0284161027520895, + 0.0113162687048316, + 0.010799926705658436, + 0.02249118685722351, + 0.012328713200986385, + 0.006695680320262909, + 0.007917555049061775, + 0.004121845588088036, + 0.005019524600356817, + 0.0073380665853619576, + 0.013895625248551369, + 0.0072220079600811005, + 0.004010955337435007, + 0.006216634064912796, + 0.004167903680354357, + 0.00229283282533288, + 0.0024463923182338476, + 0.002569806296378374, + 0.002833469770848751, + 0.003811604343354702, + 0.0027200570330023766, + 0.0024105163756757975, + 0.001996404957026243, + 0.003588553052395582, + 0.004205899778753519, + 0.0024290578439831734, + 0.0025042237248271704, + 0.003649201476946473, + 0.0025542702060192823, + 0.0030871296767145395, + 0.0025210471358150244, + 0.007550391834229231, + 0.007511792238801718, + 0.005925115197896957, + 0.00280062947422266, + 0.004964353982359171, + 0.00466120382770896, + 0.003095046617090702, + 0.0023806975223124027 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 5, + 3 + ], + "selected_reward": -0.03811129927635193 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 eye-level wide-angle view that pans smoothly from left to right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are there exactly four children visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a sun-drenched historic cobblestone town square with old European buildings in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is one of the children a girl with blonde braids wearing a white shirt and a red pleated skirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is one of the children a boy in a red polo shirt and white shorts waving a small paper Polish flag on a stick?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"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": [ + 0.1990475058555603, + 0.21726730465888977, + 0.21298712491989136, + 0.03525774925947189, + 0.03129268437623978, + 0.02059674635529518, + 0.09326289594173431, + 0.05517715588212013, + 0.09067361801862717, + 0.0444372184574604 + ], + "credit_map": [ + 0.037575915455818176, + 0.024814598262310028, + 0.023166770115494728, + 0.022135397419333458, + 0.01864466816186905, + 0.02582252398133278, + 0.025511983782052994, + 0.021375667303800583, + 0.021765554323792458, + 0.024028653278946877, + 0.03048987314105034, + 0.02744375355541706, + 0.02487117238342762, + 0.022716496139764786, + 0.04009612649679184, + 0.025855684652924538, + 0.02257140353322029, + 0.04779236763715744, + 0.027852391824126244, + 0.022389566525816917, + 0.024424247443675995, + 0.02448424883186817, + 0.019792035222053528, + 0.02368086576461792, + 0.004347093403339386, + 0.005767499096691608, + 0.003599561285227537, + 0.0030976193957030773, + 0.005596828181296587, + 0.005855647847056389, + 0.003225184977054596, + 0.0037683160044252872, + 0.003887598402798176, + 0.004100078251212835, + 0.0026972356718033552, + 0.0027278659399598837, + 0.005970282480120659, + 0.006010118871927261, + 0.0030575233977288008, + 0.0028419797308743, + 0.00271973735652864, + 0.005028360523283482, + 0.0016757291741669178, + 0.0019962326623499393, + 0.0031045505311340094, + 0.003156412160024047, + 0.0016299956478178501, + 0.0012857269030064344, + 0.008196625858545303, + 0.009008098393678665, + 0.011361165903508663, + 0.027214176952838898, + 0.009108608588576317, + 0.010210422798991203, + 0.008633273653686047, + 0.009530522860586643, + 0.005095839500427246, + 0.007307438179850578, + 0.005808369722217321, + 0.012356306426227093, + 0.006277532782405615, + 0.006106514949351549, + 0.005717795807868242, + 0.006507360376417637, + 0.00612994097173214, + 0.010640759021043777, + 0.011227188631892204, + 0.01220371387898922, + 0.01054074801504612, + 0.011810468509793282, + 0.0151364179328084, + 0.012984382919967175, + 0.004403367172926664, + 0.005103251896798611, + 0.007587152067571878, + 0.006643518339842558, + 0.005999178625643253, + 0.004141108598560095, + 0.008173714391887188, + 0.0023859282955527306 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 2, + 0 + ], + "selected_reward": -0.0036201435141265392 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 eye-level wide-angle view that pans smoothly from left to right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are there exactly four children visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a sun-drenched historic cobblestone town square with old European buildings in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is one of the children a girl with blonde braids wearing a white shirt and a red pleated skirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is one of the children a boy in a red polo shirt and white shorts waving a small paper Polish flag on a stick?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018178993836045265 + ], + "opposite_logprob_tokens": [ + -4.018178939819336 + ], + "logprob_reward": -0.018178993836045265, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018178993836045265, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17347532510757446, + 0.139073446393013, + 0.16542434692382812, + 0.0234816987067461, + 0.11015766113996506, + 0.021023498848080635, + 0.22492554783821106, + 0.04759082570672035, + 0.04823489487171173, + 0.046612780541181564 + ], + "credit_map": [ + 0.026955032721161842, + 0.016514167189598083, + 0.022927364334464073, + 0.018137099221348763, + 0.02212442271411419, + 0.029886964708566666, + 0.018460392951965332, + 0.01846987009048462, + 0.015570783987641335, + 0.012925050221383572, + 0.019276786595582962, + 0.01603948511183262, + 0.01641024462878704, + 0.012403969652950764, + 0.028464822098612785, + 0.01798231154680252, + 0.02114887349307537, + 0.027066977694630623, + 0.027456263080239296, + 0.019912313669919968, + 0.02111753821372986, + 0.01568719558417797, + 0.016682712361216545, + 0.016352469101548195, + 0.002627173438668251, + 0.0031018683221191168, + 0.001873220200650394, + 0.00216209189966321, + 0.0044929273426532745, + 0.0029553023632615805, + 0.0024665945675224066, + 0.003802519990131259, + 0.011867530643939972, + 0.014034817926585674, + 0.011118991300463676, + 0.0118919238448143, + 0.023529211059212685, + 0.01894880272448063, + 0.00986387673765421, + 0.008902499452233315, + 0.002558917971327901, + 0.0050719003193080425, + 0.0016902253264561296, + 0.0028573828749358654, + 0.002814145525917411, + 0.0033183542545884848, + 0.001318443100899458, + 0.0013941291254013777, + 0.01825924590229988, + 0.02146216295659542, + 0.029527820646762848, + 0.06226063892245293, + 0.027896054089069366, + 0.020743150264024734, + 0.028260577470064163, + 0.016515888273715973, + 0.006678987294435501, + 0.006517038680613041, + 0.005548701621592045, + 0.007589835207909346, + 0.007004749495536089, + 0.005423125810921192, + 0.004804867785423994, + 0.004023519344627857, + 0.004913390148431063, + 0.005297187250107527, + 0.004765248857438564, + 0.006086781155318022, + 0.0056914277374744415, + 0.00681705679744482, + 0.006423471961170435, + 0.008240331895649433, + 0.0051772999577224255, + 0.005879580043256283, + 0.008967188186943531, + 0.005920921918004751, + 0.0047293915413320065, + 0.005535850301384926, + 0.007281644269824028, + 0.0031209043227136135 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 2 + ], + "selected_reward": -0.018178993836045265 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "slow, eye-level wide-angle shot gracefully pans from left to right, capturing a group of four joyful children, aged between six and nine, celebrating in a sun-drenched, historic cobblestone town square. The children are dressed in festive attire in the distinctive colors of the Polish flag. A young girl with blonde braids tied with a red ribbon, dressed in a plain white short-sleeved shirt and a lively red pleated skirt, sways with laughter. Beside her, a boy in a bright red polo shirt and crisp white shorts enthusiastically waves a small paper Polish flag on a thin wooden stick. Another girl wears a simple, elegant dress, split horizontally: a white top and a red bottom. The warm, golden light of the afternoon sun casts long, soft shadows on the stone, highlighting the cheerful, irreverent expressions on their faces. In the background, the charming, colorful facades of old European buildings appear softly out of focus, creating a beautiful bokeh effect that enhances the scene's depth. The overall atmosphere is one of pure, innocent patriotism and exuberant happiness, with the red and white of the vivid children’s clothing becoming the powerful, emotional centerpiece of the visual narrative.", + "seed": 28884, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000021/group_13/index_3040_seed_28884_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": -0.10540059357881545, + "vlm_logprob": -0.10540059357881545, + "vlm_yes_no_margin": 3.225, + "vlm_token_credit": -0.10540059357881545, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 eye-level wide-angle view that pans smoothly from left to right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are there exactly four children visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a sun-drenched historic cobblestone town square with old European buildings in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is one of the children a girl with blonde braids wearing a white shirt and a red pleated skirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is one of the children a boy in a red polo shirt and white shorts waving a small paper Polish flag on a stick?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.3263459801673889 + ], + "opposite_logprob_tokens": [ + -1.4513459205627441 + ], + "logprob_reward": -0.3263459801673889, + "yes_no_margin_reward": 1.125, + "token_credit_reward": -0.3263459801673889, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1637551635503769, + 0.08326032757759094, + 0.1414232850074768, + 0.11437494307756424, + 0.08069519698619843, + 0.05877978727221489, + 0.0962858572602272, + 0.13636305928230286, + 0.060761164873838425, + 0.06430118530988693 + ], + "credit_map": [ + 0.02110234461724758, + 0.02353651449084282, + 0.01781480200588703, + 0.015644973143935204, + 0.02209600992500782, + 0.02443072944879532, + 0.012904239818453789, + 0.026225551962852478, + 0.007831920869648457, + 0.015689847990870476, + 0.004692665301263332, + 0.009645181708037853, + 0.015164213255047798, + 0.010769383981823921, + 0.011352397501468658, + 0.008114716969430447, + 0.013837351463735104, + 0.018808715045452118, + 0.009785991162061691, + 0.015482048504054546, + 0.031090417876839638, + 0.019380124285817146, + 0.013847877271473408, + 0.019190747290849686, + 0.008538944646716118, + 0.009721898473799229, + 0.009680448099970818, + 0.007081447169184685, + 0.011387522332370281, + 0.015184078365564346, + 0.023891543969511986, + 0.028889060020446777, + 0.004385280888527632, + 0.009494883939623833, + 0.005790355149656534, + 0.007504432462155819, + 0.021570125594735146, + 0.01259776670485735, + 0.011851118877530098, + 0.00750123243778944, + 0.009802624583244324, + 0.005806621629744768, + 0.007259658072143793, + 0.005946420133113861, + 0.010236533358693123, + 0.007973152212798595, + 0.005743008106946945, + 0.006011769641190767, + 0.009623701684176922, + 0.013350429013371468, + 0.013617925345897675, + 0.009755013510584831, + 0.013865111395716667, + 0.014299590140581131, + 0.008867700584232807, + 0.012906377203762531, + 0.026286425068974495, + 0.015365172177553177, + 0.014898908324539661, + 0.013301176950335503, + 0.011335517279803753, + 0.021168440580368042, + 0.013578281737864017, + 0.020429134368896484, + 0.005649466533213854, + 0.008533449843525887, + 0.007870406843721867, + 0.00792127288877964, + 0.009573635645210743, + 0.008439517579972744, + 0.005943564232438803, + 0.006829853635281324, + 0.005840222351253033, + 0.008394531905651093, + 0.003922531381249428, + 0.005990623030811548, + 0.018728727474808693, + 0.006092467345297337, + 0.0068647004663944244, + 0.008467381820082664 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 7 + ], + "selected_reward": -0.3263459801673889 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 eye-level wide-angle view that pans smoothly from left to right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are there exactly four children visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a sun-drenched historic cobblestone town square with old European buildings in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is one of the children a girl with blonde braids wearing a white shirt and a red pleated skirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is one of the children a boy in a red polo shirt and white shorts waving a small paper Polish flag on a stick?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.16043917834758759 + ], + "opposite_logprob_tokens": [ + -1.910439133644104 + ], + "logprob_reward": -0.16043917834758759, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.16043917834758759, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08189491182565689, + 0.0439804382622242, + 0.11956275254487991, + 0.11160960048437119, + 0.07266081124544144, + 0.08301877230405807, + 0.29085248708724976, + 0.08715135604143143, + 0.045388396829366684, + 0.06388050317764282 + ], + "credit_map": [ + 0.010205715894699097, + 0.010094646364450455, + 0.0068359714932739735, + 0.011004804633557796, + 0.010803671553730965, + 0.01648976467549801, + 0.007282786536961794, + 0.009177552536129951, + 0.004198065027594566, + 0.008053680881857872, + 0.0031790847424417734, + 0.004878354724496603, + 0.006457589566707611, + 0.006318220403045416, + 0.005648933816701174, + 0.005246507935225964, + 0.014024389907717705, + 0.01754203997552395, + 0.012441342696547508, + 0.011980070732533932, + 0.025286288931965828, + 0.008237270638346672, + 0.012609613128006458, + 0.017441729083657265, + 0.006614934653043747, + 0.007083639968186617, + 0.0037058943416923285, + 0.006472120061516762, + 0.014073366299271584, + 0.010092975571751595, + 0.031005697324872017, + 0.03256097063422203, + 0.0050414265133440495, + 0.0079599404707551, + 0.00704991677775979, + 0.008168021216988564, + 0.01407718751579523, + 0.013128035701811314, + 0.009896954521536827, + 0.007339332718402147, + 0.0061498889699578285, + 0.00951836071908474, + 0.00861179269850254, + 0.011304591782391071, + 0.01941574551165104, + 0.008660032413899899, + 0.009500882588326931, + 0.009857471100986004, + 0.028228631243109703, + 0.04292520135641098, + 0.06116415560245514, + 0.025039048865437508, + 0.03554363176226616, + 0.034749336540699005, + 0.027283696457743645, + 0.035918787121772766, + 0.013568226248025894, + 0.010026403702795506, + 0.014581389725208282, + 0.008561653085052967, + 0.006514588836580515, + 0.010702433064579964, + 0.009206910617649555, + 0.013989747501909733, + 0.005531787872314453, + 0.004978971555829048, + 0.0061652702279388905, + 0.004315854050219059, + 0.008153853937983513, + 0.006197222974151373, + 0.004374290816485882, + 0.005671144463121891, + 0.005887294188141823, + 0.008462912403047085, + 0.006632157601416111, + 0.006622510962188244, + 0.0142003633081913, + 0.005600959528237581, + 0.007670009508728981, + 0.008804294280707836 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 2, + 3 + ], + "selected_reward": -0.16043917834758759 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 eye-level wide-angle view that pans smoothly from left to right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are there exactly four children visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a sun-drenched historic cobblestone town square with old European buildings in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is one of the children a girl with blonde braids wearing a white shirt and a red pleated skirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is one of the children a boy in a red polo shirt and white shorts waving a small paper Polish flag on a stick?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01608244515955448 + ], + "opposite_logprob_tokens": [ + -4.141082286834717 + ], + "logprob_reward": -0.01608244515955448, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.01608244515955448, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2070850431919098, + 0.09598664194345474, + 0.08005202561616898, + 0.05536355450749397, + 0.03091461956501007, + 0.04152626171708107, + 0.09087079763412476, + 0.04359001666307449, + 0.18036901950836182, + 0.17424198985099792 + ], + "credit_map": [ + 0.023060591891407967, + 0.026459772139787674, + 0.02023080550134182, + 0.01875961199402809, + 0.03209880366921425, + 0.038315851241350174, + 0.01880076341331005, + 0.02935885824263096, + 0.008920692838728428, + 0.014266112819314003, + 0.004181646276265383, + 0.009467446245253086, + 0.022296782582998276, + 0.018446218222379684, + 0.007643135264515877, + 0.010764614678919315, + 0.0058152563869953156, + 0.009997562505304813, + 0.006921150255948305, + 0.014811338856816292, + 0.015168377198278904, + 0.009340845048427582, + 0.006905707065016031, + 0.011091791093349457, + 0.003167920745909214, + 0.0032942674588412046, + 0.002593603916466236, + 0.0028894348070025444, + 0.006453611887991428, + 0.009215922094881535, + 0.013340131379663944, + 0.014408663846552372, + 0.0018053995445370674, + 0.004065289627760649, + 0.002112998627126217, + 0.0021743152756243944, + 0.008140414021909237, + 0.006806016433984041, + 0.003260578727349639, + 0.0025496080052107573, + 0.005126050673425198, + 0.0038029986899346113, + 0.003103018505498767, + 0.002183852018788457, + 0.006089483387768269, + 0.013209161348640919, + 0.003263560589402914, + 0.004748140927404165, + 0.008904369547963142, + 0.00857643224298954, + 0.014817011542618275, + 0.005281259771436453, + 0.011605752632021904, + 0.02233116142451763, + 0.007168234791606665, + 0.012186583131551743, + 0.007687731646001339, + 0.004300949163734913, + 0.0038299444131553173, + 0.0020010648295283318, + 0.008480188436806202, + 0.008203621953725815, + 0.0031858698930591345, + 0.005900649353861809, + 0.01637016423046589, + 0.021880362182855606, + 0.010379827581346035, + 0.017377957701683044, + 0.03878850117325783, + 0.03040635772049427, + 0.017384925857186317, + 0.027780940756201744, + 0.013230477459728718, + 0.015807325020432472, + 0.01340131089091301, + 0.015464537777006626, + 0.04791265353560448, + 0.02081240899860859, + 0.018520241603255272, + 0.029093051329255104 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 9 + ], + "selected_reward": -0.01608244515955448 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 eye-level wide-angle view that pans smoothly from left to right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are there exactly four children visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a sun-drenched historic cobblestone town square with old European buildings in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is one of the children a girl with blonde braids wearing a white shirt and a red pleated skirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is one of the children a boy in a red polo shirt and white shorts waving a small paper Polish flag on a stick?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0059540290385484695 + ], + "opposite_logprob_tokens": [ + -5.130954265594482 + ], + "logprob_reward": -0.0059540290385484695, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.0059540290385484695, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14108069241046906, + 0.1780775934457779, + 0.14980387687683105, + 0.0839681625366211, + 0.0633789673447609, + 0.08391038328409195, + 0.09597956389188766, + 0.02253933809697628, + 0.04314946010708809, + 0.13811194896697998 + ], + "credit_map": [ + 0.011928534135222435, + 0.019624562934041023, + 0.013221980072557926, + 0.011931321583688259, + 0.02294943295419216, + 0.028751729056239128, + 0.012150602415204048, + 0.02052254043519497, + 0.020291663706302643, + 0.031915631145238876, + 0.010728125460445881, + 0.01865694299340248, + 0.02619672380387783, + 0.016151517629623413, + 0.023363718762993813, + 0.030773276463150978, + 0.01606965996325016, + 0.023755794391036034, + 0.012704079039394855, + 0.00866150762885809, + 0.024226205423474312, + 0.029246436432003975, + 0.011621004901826382, + 0.02351919561624527, + 0.005744294263422489, + 0.004932105075567961, + 0.004202248528599739, + 0.005127064883708954, + 0.009625461883842945, + 0.006224050186574459, + 0.020989757031202316, + 0.027123183012008667, + 0.00479261064901948, + 0.009856104850769043, + 0.007357142865657806, + 0.008565383031964302, + 0.007307691965252161, + 0.006078159436583519, + 0.011418050155043602, + 0.008003832772374153, + 0.008020567707717419, + 0.01355995237827301, + 0.008967378176748753, + 0.0141067523509264, + 0.012637853622436523, + 0.008099001832306385, + 0.01044077705591917, + 0.00807809829711914, + 0.014437801204621792, + 0.014092124998569489, + 0.010032379999756813, + 0.00992283783853054, + 0.014668959192931652, + 0.013215959072113037, + 0.008814793080091476, + 0.010794715024530888, + 0.0031452460680156946, + 0.0024582287296652794, + 0.0027078231796622276, + 0.0027155994903296232, + 0.0032876073382794857, + 0.002740050433203578, + 0.003031676635146141, + 0.0024531094823032618, + 0.004908639006316662, + 0.0075046769343316555, + 0.005211245268583298, + 0.005928281228989363, + 0.006470572203397751, + 0.004791034385561943, + 0.0036013280041515827, + 0.004733686801046133, + 0.013684905134141445, + 0.01769818365573883, + 0.012606056407094002, + 0.01895618811249733, + 0.03253339231014252, + 0.012489119544625282, + 0.015097555704414845, + 0.015046549960970879 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 2, + 0 + ], + "selected_reward": -0.0059540290385484695 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 eye-level wide-angle view that pans smoothly from left to right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are there exactly four children visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a sun-drenched historic cobblestone town square with old European buildings in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is one of the children a girl with blonde braids wearing a white shirt and a red pleated skirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is one of the children a boy in a red polo shirt and white shorts waving a small paper Polish flag on a stick?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01818133518099785 + ], + "opposite_logprob_tokens": [ + -4.018181324005127 + ], + "logprob_reward": -0.01818133518099785, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.01818133518099785, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2951558530330658, + 0.07346490025520325, + 0.09839325398206711, + 0.06162448599934578, + 0.04834065958857536, + 0.10113460570573807, + 0.08918643742799759, + 0.05070872977375984, + 0.10477592051029205, + 0.07721514999866486 + ], + "credit_map": [ + 0.028866535052657127, + 0.03902824595570564, + 0.02453806810081005, + 0.02512899413704872, + 0.045970845967531204, + 0.056273460388183594, + 0.03094295598566532, + 0.044406745582818985, + 0.008359119296073914, + 0.014025002717971802, + 0.004145499784499407, + 0.00645896652713418, + 0.011930343694984913, + 0.007790621370077133, + 0.009222935885190964, + 0.01153241191059351, + 0.006724147591739893, + 0.012484170496463776, + 0.007764854468405247, + 0.011686107143759727, + 0.020564528182148933, + 0.01412989106029272, + 0.010387588292360306, + 0.014651970006525517, + 0.0061356076039373875, + 0.004564437083899975, + 0.0036067788023501635, + 0.0035327740479260683, + 0.006392472889274359, + 0.00791102834045887, + 0.012059264816343784, + 0.01742211915552616, + 0.003739646403118968, + 0.007031089160591364, + 0.0042536454275250435, + 0.0038209373597055674, + 0.006230226252228022, + 0.012713255360722542, + 0.0054248603992164135, + 0.00512699643149972, + 0.010440707206726074, + 0.014461862854659557, + 0.011245944537222385, + 0.012093542143702507, + 0.022111715748906136, + 0.009254985488951206, + 0.011047358624637127, + 0.010478493757545948, + 0.01753160171210766, + 0.01481894962489605, + 0.012015914544463158, + 0.008742648176848888, + 0.009876694530248642, + 0.010662122629582882, + 0.006579988170415163, + 0.008958514779806137, + 0.007836603559553623, + 0.005242216866463423, + 0.008341910317540169, + 0.004679616075009108, + 0.00489167682826519, + 0.007122411858290434, + 0.005765603389590979, + 0.006828689947724342, + 0.01148524321615696, + 0.020226309075951576, + 0.011777791194617748, + 0.01765904203057289, + 0.014303828589618206, + 0.008118820376694202, + 0.009900522418320179, + 0.011304360814392567, + 0.009348740801215172, + 0.009738708846271038, + 0.007348089478909969, + 0.010518232360482216, + 0.017598776146769524, + 0.006522100418806076, + 0.008256634697318077, + 0.007883863523602486 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 5 + ], + "selected_reward": -0.01818133518099785 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "slow, eye-level wide-angle shot gracefully pans from left to right, capturing a group of four joyful children, aged between six and nine, celebrating in a sun-drenched, historic cobblestone town square. The children are dressed in festive attire in the distinctive colors of the Polish flag. A young girl with blonde braids tied with a red ribbon, dressed in a plain white short-sleeved shirt and a lively red pleated skirt, sways with laughter. Beside her, a boy in a bright red polo shirt and crisp white shorts enthusiastically waves a small paper Polish flag on a thin wooden stick. Another girl wears a simple, elegant dress, split horizontally: a white top and a red bottom. The warm, golden light of the afternoon sun casts long, soft shadows on the stone, highlighting the cheerful, irreverent expressions on their faces. In the background, the charming, colorful facades of old European buildings appear softly out of focus, creating a beautiful bokeh effect that enhances the scene's depth. The overall atmosphere is one of pure, innocent patriotism and exuberant happiness, with the red and white of the vivid children’s clothing becoming the powerful, emotional centerpiece of the visual narrative.", + "seed": 28885, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000021/group_13/index_3040_seed_28885_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": -0.0960010931827128, + "vlm_logprob": -0.0960010931827128, + "vlm_yes_no_margin": 3.6, + "vlm_token_credit": -0.0960010931827128, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 eye-level wide-angle view that pans smoothly from left to right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are there exactly four children visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a sun-drenched historic cobblestone town square with old European buildings in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is one of the children a girl with blonde braids wearing a white shirt and a red pleated skirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is one of the children a boy in a red polo shirt and white shorts waving a small paper Polish flag on a stick?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.27714037895202637 + ], + "opposite_logprob_tokens": [ + -1.6521403789520264 + ], + "logprob_reward": -0.27714037895202637, + "yes_no_margin_reward": 1.375, + "token_credit_reward": -0.27714037895202637, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1250910758972168, + 0.03679119423031807, + 0.13785356283187866, + 0.09177637100219727, + 0.08601700514554977, + 0.08642448484897614, + 0.21022692322731018, + 0.1000208705663681, + 0.10162881761789322, + 0.024169694632291794 + ], + "credit_map": [ + 0.015609350986778736, + 0.02248305454850197, + 0.013472147285938263, + 0.010223347693681717, + 0.016593843698501587, + 0.015632975846529007, + 0.01589800976216793, + 0.015178348869085312, + 0.0032672055531293154, + 0.005841828882694244, + 0.004039755556732416, + 0.0028154682368040085, + 0.005082946736365557, + 0.00723033444955945, + 0.0034652012400329113, + 0.005048454273492098, + 0.013752145692706108, + 0.018691658973693848, + 0.009205959737300873, + 0.009423382580280304, + 0.030053861439228058, + 0.018318114802241325, + 0.020862624049186707, + 0.01754583977162838, + 0.01033618114888668, + 0.009031407535076141, + 0.01017799787223339, + 0.013165152631700039, + 0.013704081065952778, + 0.013112608343362808, + 0.010158421471714973, + 0.012090531177818775, + 0.015438126400113106, + 0.010379119776189327, + 0.012720983475446701, + 0.008440151810646057, + 0.01091085933148861, + 0.010711646638810635, + 0.008531083352863789, + 0.00888504832983017, + 0.012500684708356857, + 0.010398638434708118, + 0.013430214487016201, + 0.00904447678476572, + 0.0076255155727267265, + 0.017201049253344536, + 0.005798283964395523, + 0.010425626300275326, + 0.019584067165851593, + 0.01141718402504921, + 0.07041964679956436, + 0.009530112147331238, + 0.014775612391531467, + 0.02764078415930271, + 0.011805141344666481, + 0.045054372400045395, + 0.01009257324039936, + 0.009525353088974953, + 0.010762433521449566, + 0.007062519434839487, + 0.011276030912995338, + 0.020789580419659615, + 0.006710509769618511, + 0.023801889270544052, + 0.007528174668550491, + 0.007995047606527805, + 0.03444257378578186, + 0.005462747998535633, + 0.006388311740010977, + 0.02382262423634529, + 0.005923917051404715, + 0.01006542332470417, + 0.0029053010512143373, + 0.0027198889292776585, + 0.0036674323491752148, + 0.0020561562851071358, + 0.0029799442272633314, + 0.004526998847723007, + 0.002064184984192252, + 0.0032497879583388567 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 2, + 0 + ], + "selected_reward": -0.27714037895202637 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 eye-level wide-angle view that pans smoothly from left to right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are there exactly four children visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a sun-drenched historic cobblestone town square with old European buildings in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is one of the children a girl with blonde braids wearing a white shirt and a red pleated skirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is one of the children a boy in a red polo shirt and white shorts waving a small paper Polish flag on a stick?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.17995557188987732 + ], + "opposite_logprob_tokens": [ + -1.8049556016921997 + ], + "logprob_reward": -0.17995557188987732, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.17995557188987732, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0734834149479866, + 0.028335293754935265, + 0.28157278895378113, + 0.06839660555124283, + 0.027461837977170944, + 0.04149829223752022, + 0.0492098294198513, + 0.34382790327072144, + 0.07023868709802628, + 0.015975331887602806 + ], + "credit_map": [ + 0.01082103792577982, + 0.01187215931713581, + 0.009089178405702114, + 0.008155982941389084, + 0.007937437854707241, + 0.007804646622389555, + 0.007734953425824642, + 0.010068022646009922, + 0.0026341830380260944, + 0.004341245163232088, + 0.003879331052303314, + 0.0025058831088244915, + 0.0031714465003460646, + 0.00584800262004137, + 0.0022751283831894398, + 0.0036800773814320564, + 0.03234236314892769, + 0.03722097724676132, + 0.022989144548773766, + 0.015069731511175632, + 0.06973866373300552, + 0.03946489840745926, + 0.03180690109729767, + 0.03294013440608978, + 0.007008849177509546, + 0.008910978212952614, + 0.006671569310128689, + 0.008468206971883774, + 0.013079382479190826, + 0.009977126494050026, + 0.006296694278717041, + 0.007983801886439323, + 0.003594460431486368, + 0.003282640129327774, + 0.0031865686178207397, + 0.0030334715265780687, + 0.0031243402045220137, + 0.003324860241264105, + 0.004108885768800974, + 0.0038066126871854067, + 0.006621051114052534, + 0.005781041458249092, + 0.005894264671951532, + 0.004154704976826906, + 0.003955616615712643, + 0.007727758027613163, + 0.0021269682329148054, + 0.005236890632659197, + 0.004666808061301708, + 0.0027847415767610073, + 0.01215134747326374, + 0.003049678634852171, + 0.006658060476183891, + 0.0049968077801167965, + 0.004474078770726919, + 0.010428308509290218, + 0.027171451598405838, + 0.014779896475374699, + 0.021454328671097755, + 0.01270010694861412, + 0.016324728727340698, + 0.12860289216041565, + 0.00985066220164299, + 0.1129438728094101, + 0.0053343866020441055, + 0.00881200097501278, + 0.01929338276386261, + 0.005589825566858053, + 0.005828990135341883, + 0.012411062605679035, + 0.005795871373265982, + 0.007173176389187574, + 0.0013518028426915407, + 0.00263729365542531, + 0.001796374679543078, + 0.0025914134457707405, + 0.0016195171047002077, + 0.0016100809443742037, + 0.0021627857349812984, + 0.0022060652263462543 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 2, + 0 + ], + "selected_reward": -0.17995557188987732 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 eye-level wide-angle view that pans smoothly from left to right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are there exactly four children visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a sun-drenched historic cobblestone town square with old European buildings in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is one of the children a girl with blonde braids wearing a white shirt and a red pleated skirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is one of the children a boy in a red polo shirt and white shorts waving a small paper Polish flag on a stick?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00864359550178051 + ], + "opposite_logprob_tokens": [ + -4.758643627166748 + ], + "logprob_reward": -0.00864359550178051, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.00864359550178051, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1383918672800064, + 0.13418081402778625, + 0.09115120023488998, + 0.02606937102973461, + 0.07767278701066971, + 0.1861329823732376, + 0.07802356034517288, + 0.023088224232196808, + 0.1696503758430481, + 0.07563880831003189 + ], + "credit_map": [ + 0.02383265271782875, + 0.01989619992673397, + 0.014895908534526825, + 0.0099617475643754, + 0.013564558699727058, + 0.02635214477777481, + 0.014249846339225769, + 0.0156388022005558, + 0.018623199313879013, + 0.01996232010424137, + 0.008249112404882908, + 0.005288737826049328, + 0.017387688159942627, + 0.03586603328585625, + 0.012501576915383339, + 0.01630215346813202, + 0.01352388970553875, + 0.012384026311337948, + 0.010418813675642014, + 0.006892176810652018, + 0.01461282093077898, + 0.011515992693603039, + 0.010388793423771858, + 0.011414698325097561, + 0.0030067728366702795, + 0.0026791845448315144, + 0.003288887906819582, + 0.0035600876435637474, + 0.0029183924198150635, + 0.004297230858355761, + 0.003270290791988373, + 0.003048523562029004, + 0.015374741517007351, + 0.009016422554850578, + 0.01143177691847086, + 0.005221561528742313, + 0.010392471216619015, + 0.012276511639356613, + 0.006295329891145229, + 0.007663968950510025, + 0.02924828790128231, + 0.013385535217821598, + 0.04121413081884384, + 0.009524494409561157, + 0.017691342160105705, + 0.03689315915107727, + 0.010675542056560516, + 0.027500486001372337, + 0.007826895453035831, + 0.003977996297180653, + 0.02913033217191696, + 0.0028022623155266047, + 0.0028268410824239254, + 0.007968215271830559, + 0.002498712856322527, + 0.02099231258034706, + 0.0019607434514909983, + 0.002379361307248473, + 0.0035265448968857527, + 0.0022996612824499607, + 0.002961633959785104, + 0.003229257883504033, + 0.0015009287744760513, + 0.005230093840509653, + 0.02005811780691147, + 0.023259693756699562, + 0.034455180168151855, + 0.013796051032841206, + 0.013666368089616299, + 0.02221892960369587, + 0.015917208045721054, + 0.02627883478999138, + 0.009563092142343521, + 0.008732843212783337, + 0.009068026207387447, + 0.005127888638526201, + 0.011752444319427013, + 0.016797896474599838, + 0.005334347952157259, + 0.009262273088097572 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 8, + 0 + ], + "selected_reward": -0.00864359550178051 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 eye-level wide-angle view that pans smoothly from left to right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are there exactly four children visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a sun-drenched historic cobblestone town square with old European buildings in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is one of the children a girl with blonde braids wearing a white shirt and a red pleated skirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is one of the children a boy in a red polo shirt and white shorts waving a small paper Polish flag on a stick?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0031956578604876995 + ], + "opposite_logprob_tokens": [ + -5.753195762634277 + ], + "logprob_reward": -0.0031956578604876995, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.0031956578604876995, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16425444185733795, + 0.06837073713541031, + 0.09899913519620895, + 0.10903757065534592, + 0.0816478505730629, + 0.12205351144075394, + 0.10662595182657242, + 0.13103850185871124, + 0.06629068404436111, + 0.0516815148293972 + ], + "credit_map": [ + 0.02093484438955784, + 0.03509677201509476, + 0.01567601040005684, + 0.015909958630800247, + 0.016109397634863853, + 0.02095707319676876, + 0.02061779610812664, + 0.018952591344714165, + 0.006600324064493179, + 0.014395765028893948, + 0.005987206008285284, + 0.004488534294068813, + 0.008346340619027615, + 0.009798041544854641, + 0.00790195818990469, + 0.010852565057575703, + 0.010529338382184505, + 0.014819024130702019, + 0.008891967125236988, + 0.009951524436473846, + 0.02426675707101822, + 0.014179631136357784, + 0.006866188254207373, + 0.009494706057012081, + 0.009433070197701454, + 0.011334523558616638, + 0.010900076478719711, + 0.014633417129516602, + 0.016374796628952026, + 0.013710116036236286, + 0.020257966592907906, + 0.012393610551953316, + 0.00623337971046567, + 0.012466248124837875, + 0.009908621199429035, + 0.009410196915268898, + 0.012374182231724262, + 0.008795018307864666, + 0.01145199965685606, + 0.011008207686245441, + 0.012550784274935722, + 0.01228404138237238, + 0.018261602148413658, + 0.012548644095659256, + 0.017625650390982628, + 0.01645127311348915, + 0.013392282649874687, + 0.01893923617899418, + 0.007729301694780588, + 0.004636349622160196, + 0.04020736739039421, + 0.005201124120503664, + 0.005708800628781319, + 0.012867758050560951, + 0.004447342362254858, + 0.025827908888459206, + 0.013175644911825657, + 0.012781573459506035, + 0.023237688466906548, + 0.0076729836873710155, + 0.01256546750664711, + 0.039865847676992416, + 0.006896426901221275, + 0.014842874370515347, + 0.007198675535619259, + 0.0107218436896801, + 0.009398129768669605, + 0.006757090799510479, + 0.007657375652343035, + 0.006903541274368763, + 0.009354489855468273, + 0.008299537003040314, + 0.003921515308320522, + 0.008776118978857994, + 0.007626733742654324, + 0.004978202283382416, + 0.007212863303720951, + 0.0049568889662623405, + 0.006666937842965126, + 0.0075422534719109535 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 5 + ], + "selected_reward": -0.0031956578604876995 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 eye-level wide-angle view that pans smoothly from left to right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are there exactly four children visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a sun-drenched historic cobblestone town square with old European buildings in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is one of the children a girl with blonde braids wearing a white shirt and a red pleated skirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is one of the children a boy in a red polo shirt and white shorts waving a small paper Polish flag on a stick?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01107026170939207 + ], + "opposite_logprob_tokens": [ + -4.511070251464844 + ], + "logprob_reward": -0.01107026170939207, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.01107026170939207, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06886080652475357, + 0.0240379236638546, + 0.04057568311691284, + 0.052926234900951385, + 0.03643639385700226, + 0.0495489202439785, + 0.08339820802211761, + 0.19632813334465027, + 0.43485280871391296, + 0.013034897856414318 + ], + "credit_map": [ + 0.008732941932976246, + 0.0149285439401865, + 0.006121623329818249, + 0.007138629909604788, + 0.007575177121907473, + 0.0067602964118123055, + 0.009596296586096287, + 0.008007297292351723, + 0.002482035895809531, + 0.004733362700790167, + 0.004132529720664024, + 0.0019794621039181948, + 0.002507708268240094, + 0.003625171259045601, + 0.001659953617490828, + 0.0029176969546824694, + 0.004429824184626341, + 0.007063256576657295, + 0.0028730614576488733, + 0.002840634435415268, + 0.008836154825985432, + 0.0049217320047318935, + 0.0037878439761698246, + 0.005823174025863409, + 0.0047562457621097565, + 0.006826178170740604, + 0.004455490503460169, + 0.0077971648424863815, + 0.007160949986428022, + 0.008837000466883183, + 0.00643999082967639, + 0.006653210613876581, + 0.0030896204989403486, + 0.005116550251841545, + 0.0044347005896270275, + 0.0055773151107132435, + 0.0038862950168550014, + 0.003957304637879133, + 0.004213233012706041, + 0.006161373574286699, + 0.005253197159618139, + 0.00521137984469533, + 0.007363155949860811, + 0.004261758644133806, + 0.00755852647125721, + 0.006563409231603146, + 0.005036109127104282, + 0.008301382884383202, + 0.005841763690114021, + 0.004120304714888334, + 0.03115401417016983, + 0.004042606335133314, + 0.005040993448346853, + 0.0076424358412623405, + 0.003277846146374941, + 0.02227824367582798, + 0.016835222020745277, + 0.008119624108076096, + 0.01304863952100277, + 0.007257858291268349, + 0.008500948548316956, + 0.0765419453382492, + 0.004351701587438583, + 0.06167217344045639, + 0.03251400962471962, + 0.04815582558512688, + 0.13252367079257965, + 0.028777072206139565, + 0.03109072335064411, + 0.08836964517831802, + 0.030415579676628113, + 0.04300627484917641, + 0.0014269056264311075, + 0.0013356002746149898, + 0.0016768404748290777, + 0.002071850700303912, + 0.0013608008157461882, + 0.0014903150731697679, + 0.001895977882668376, + 0.0017766067758202553 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 6 + ], + "selected_reward": -0.01107026170939207 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "slow, eye-level wide-angle shot gracefully pans from left to right, capturing a group of four joyful children, aged between six and nine, celebrating in a sun-drenched, historic cobblestone town square. The children are dressed in festive attire in the distinctive colors of the Polish flag. A young girl with blonde braids tied with a red ribbon, dressed in a plain white short-sleeved shirt and a lively red pleated skirt, sways with laughter. Beside her, a boy in a bright red polo shirt and crisp white shorts enthusiastically waves a small paper Polish flag on a thin wooden stick. Another girl wears a simple, elegant dress, split horizontally: a white top and a red bottom. The warm, golden light of the afternoon sun casts long, soft shadows on the stone, highlighting the cheerful, irreverent expressions on their faces. In the background, the charming, colorful facades of old European buildings appear softly out of focus, creating a beautiful bokeh effect that enhances the scene's depth. The overall atmosphere is one of pure, innocent patriotism and exuberant happiness, with the red and white of the vivid children’s clothing becoming the powerful, emotional centerpiece of the visual narrative.", + "seed": 28886, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000021/group_13/index_3040_seed_28886_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": -0.11155235040932894, + "vlm_logprob": -0.11155235040932894, + "vlm_yes_no_margin": 3.225, + "vlm_token_credit": -0.11155235040932894, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 eye-level wide-angle view that pans smoothly from left to right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are there exactly four children visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a sun-drenched historic cobblestone town square with old European buildings in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is one of the children a girl with blonde braids wearing a white shirt and a red pleated skirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is one of the children a boy in a red polo shirt and white shorts waving a small paper Polish flag on a stick?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2369566708803177 + ], + "opposite_logprob_tokens": [ + -1.7369567155838013 + ], + "logprob_reward": -0.2369566708803177, + "yes_no_margin_reward": 1.5, + "token_credit_reward": -0.2369566708803177, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04817730933427811, + 0.03159015253186226, + 0.12900273501873016, + 0.02164355479180813, + 0.02262597344815731, + 0.022169485688209534, + 0.14532223343849182, + 0.41874757409095764, + 0.03828679770231247, + 0.12243419885635376 + ], + "credit_map": [ + 0.004772276617586613, + 0.006035288795828819, + 0.0033054077066481113, + 0.0055073765106499195, + 0.010924986563622952, + 0.0064125144854187965, + 0.004982875194400549, + 0.006236585788428783, + 0.0037516348529607058, + 0.0029473339673131704, + 0.0035388267133384943, + 0.004389097914099693, + 0.00423455098643899, + 0.004445589147508144, + 0.003921182826161385, + 0.004361932165920734, + 0.01183683704584837, + 0.007206926587969065, + 0.007640848867595196, + 0.005722931120544672, + 0.005309534724801779, + 0.03847060725092888, + 0.006543242838233709, + 0.04627181589603424, + 0.0019336147233843803, + 0.00222209095954895, + 0.0024226170498877764, + 0.002343607833608985, + 0.002578659914433956, + 0.003968763165175915, + 0.0025297002866864204, + 0.003644500393420458, + 0.00288440496660769, + 0.0028270273469388485, + 0.0022098636254668236, + 0.0028448542580008507, + 0.003976852633059025, + 0.002617373364046216, + 0.002469179453328252, + 0.002796416636556387, + 0.0034585341345518827, + 0.002419323893263936, + 0.002298501553013921, + 0.002016329439356923, + 0.003062111558392644, + 0.003166447626426816, + 0.0026478657964617014, + 0.003100372152402997, + 0.027954062446951866, + 0.017689626663923264, + 0.02587086521089077, + 0.01768917590379715, + 0.009785759262740612, + 0.011810537427663803, + 0.01622111350297928, + 0.018301082774996758, + 0.03380054235458374, + 0.03142879530787468, + 0.16096720099449158, + 0.02280230075120926, + 0.02083028480410576, + 0.020564991980791092, + 0.015506265684962273, + 0.1128472089767456, + 0.006079540587961674, + 0.004316654521971941, + 0.005629988387227058, + 0.0035582841373980045, + 0.00517651392146945, + 0.004003515001386404, + 0.004171363078057766, + 0.005350932944566011, + 0.014975924976170063, + 0.010029091499745846, + 0.034820638597011566, + 0.008832143619656563, + 0.012155951000750065, + 0.008043370209634304, + 0.014046240597963333, + 0.01953083835542202 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 6, + 2 + ], + "selected_reward": -0.2369566708803177 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 eye-level wide-angle view that pans smoothly from left to right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are there exactly four children visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a sun-drenched historic cobblestone town square with old European buildings in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is one of the children a girl with blonde braids wearing a white shirt and a red pleated skirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is one of the children a boy in a red polo shirt and white shorts waving a small paper Polish flag on a stick?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2813842296600342 + ], + "opposite_logprob_tokens": [ + -1.4063842296600342 + ], + "logprob_reward": -0.2813842296600342, + "yes_no_margin_reward": 1.125, + "token_credit_reward": -0.2813842296600342, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14643371105194092, + 0.06688828021287918, + 0.39349105954170227, + 0.1094997301697731, + 0.013658229261636734, + 0.023637104779481888, + 0.04316346347332001, + 0.14336764812469482, + 0.030797826126217842, + 0.02906293049454689 + ], + "credit_map": [ + 0.0138285206630826, + 0.01508871279656887, + 0.012272963300347328, + 0.020861536264419556, + 0.018576176837086678, + 0.033203449100255966, + 0.014262307435274124, + 0.018340054899454117, + 0.0071633257903158665, + 0.006744671147316694, + 0.00587398512288928, + 0.010964582674205303, + 0.009668567217886448, + 0.010588498786091805, + 0.007892036810517311, + 0.007992611266672611, + 0.0383419543504715, + 0.01681583561003208, + 0.023076534271240234, + 0.01203142013400793, + 0.012235467322170734, + 0.1301354318857193, + 0.014104695990681648, + 0.14674976468086243, + 0.00948891881853342, + 0.011520683765411377, + 0.012506186962127686, + 0.016836442053318024, + 0.013482806272804737, + 0.011531783267855644, + 0.012450158596038818, + 0.021682752296328545, + 0.001520106685347855, + 0.0013373381225392222, + 0.002107685199007392, + 0.0012238058261573315, + 0.0021170040126889944, + 0.0021748102735728025, + 0.001476164092309773, + 0.0017013156320899725, + 0.004235212691128254, + 0.0024503376334905624, + 0.0025180510710924864, + 0.0026350668631494045, + 0.004304169211536646, + 0.002466375008225441, + 0.0018230925779789686, + 0.0032048006542026997, + 0.008003032766282558, + 0.007781595457345247, + 0.006223559845238924, + 0.004909808747470379, + 0.004097695462405682, + 0.003130331402644515, + 0.004177187569439411, + 0.004840251058340073, + 0.011779491789638996, + 0.01138637401163578, + 0.04395773634314537, + 0.015129223465919495, + 0.011704517528414726, + 0.00813091080635786, + 0.0106456084176898, + 0.03063378296792507, + 0.003763581393286586, + 0.003242475213482976, + 0.003808701643720269, + 0.003180848201736808, + 0.005646576173603535, + 0.0036955284886062145, + 0.003441557055339217, + 0.0040185581892728806, + 0.0036942241713404655, + 0.0031717240344733, + 0.004641161300241947, + 0.002739404793828726, + 0.004223213531076908, + 0.004206482786685228, + 0.002850770950317383, + 0.0035359482280910015 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 7 + ], + "selected_reward": -0.2813842296600342 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 eye-level wide-angle view that pans smoothly from left to right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are there exactly four children visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a sun-drenched historic cobblestone town square with old European buildings in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is one of the children a girl with blonde braids wearing a white shirt and a red pleated skirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is one of the children a boy in a red polo shirt and white shorts waving a small paper Polish flag on a stick?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020595503970980644 + ], + "opposite_logprob_tokens": [ + -3.8955955505371094 + ], + "logprob_reward": -0.020595503970980644, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020595503970980644, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03174583613872528, + 0.019044427201151848, + 0.23807546496391296, + 0.018023431301116943, + 0.016934392973780632, + 0.015520740300416946, + 0.5926064848899841, + 0.029998818412423134, + 0.02561531960964203, + 0.012435177341103554 + ], + "credit_map": [ + 0.003373888786882162, + 0.003880762495100498, + 0.003806368913501501, + 0.0041707828640937805, + 0.0040267398580908775, + 0.004459143150597811, + 0.0033967364579439163, + 0.004631413146853447, + 0.0022883715573698282, + 0.0017934221541509032, + 0.0023602889850735664, + 0.002853009384125471, + 0.0023227669298648834, + 0.00241288635879755, + 0.0024906231556087732, + 0.002523059956729412, + 0.021574590355157852, + 0.009736641310155392, + 0.013144808821380138, + 0.005534633062779903, + 0.007657214999198914, + 0.07934670895338058, + 0.008367232978343964, + 0.09271363168954849, + 0.0017341016791760921, + 0.001587539678439498, + 0.001964051043614745, + 0.0023951204493641853, + 0.001963723683729768, + 0.002956876065582037, + 0.002026607049629092, + 0.0033954123500734568, + 0.002531732665374875, + 0.0022694398649036884, + 0.0022236532531678677, + 0.0019746010657399893, + 0.0022981707006692886, + 0.0018330507446080446, + 0.0017381825018674135, + 0.0020655628759413958, + 0.0020648210775107145, + 0.0014882651157677174, + 0.0021359529346227646, + 0.0017058984376490116, + 0.002480256836861372, + 0.0020747589878737926, + 0.0016833525151014328, + 0.0018874338129535317, + 0.10439237207174301, + 0.11054680496454239, + 0.07367297261953354, + 0.06774814426898956, + 0.04747060686349869, + 0.046274583786726, + 0.07776406407356262, + 0.06473689526319504, + 0.0034340298734605312, + 0.003129551187157631, + 0.006207216531038284, + 0.0019328008638694882, + 0.004663522355258465, + 0.0033135099802166224, + 0.001773857045918703, + 0.0055443295277655125, + 0.0034841657616198063, + 0.0023641616571694613, + 0.00669883843511343, + 0.0020913572516292334, + 0.002974099013954401, + 0.0025658169761300087, + 0.001733603305183351, + 0.00370327802374959, + 0.0018185137305408716, + 0.0016147603746503592, + 0.0012505392078310251, + 0.0012684193206951022, + 0.0019684750586748123, + 0.0013365150662139058, + 0.0016988582210615277, + 0.0014790965942665935 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 2, + 0 + ], + "selected_reward": -0.020595503970980644 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 eye-level wide-angle view that pans smoothly from left to right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are there exactly four children visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a sun-drenched historic cobblestone town square with old European buildings in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is one of the children a girl with blonde braids wearing a white shirt and a red pleated skirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is one of the children a boy in a red polo shirt and white shorts waving a small paper Polish flag on a stick?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00463874451816082 + ], + "opposite_logprob_tokens": [ + -5.379638671875 + ], + "logprob_reward": -0.00463874451816082, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.00463874451816082, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07199790328741074, + 0.03681660071015358, + 0.24694783985614777, + 0.05557325482368469, + 0.029325485229492188, + 0.01748622953891754, + 0.3484590947628021, + 0.12229429185390472, + 0.039487726986408234, + 0.03161155804991722 + ], + "credit_map": [ + 0.006425787229090929, + 0.010220145806670189, + 0.0089574521407485, + 0.011159874498844147, + 0.00833660364151001, + 0.00939854048192501, + 0.007688622456043959, + 0.009810876101255417, + 0.0036458417307585478, + 0.0032831199932843447, + 0.005201546009629965, + 0.005171849392354488, + 0.005501485429704189, + 0.0041044061072170734, + 0.00471932627260685, + 0.005189026240259409, + 0.025207193568348885, + 0.018450185656547546, + 0.014742942526936531, + 0.010328059084713459, + 0.015966074541211128, + 0.07007961720228195, + 0.014741579070687294, + 0.07743220031261444, + 0.00538609316572547, + 0.0068301050923764706, + 0.006461328826844692, + 0.008052675053477287, + 0.007073890417814255, + 0.0065499418415129185, + 0.006676645018160343, + 0.00854257307946682, + 0.004666019231081009, + 0.0036647929809987545, + 0.004111616872251034, + 0.0038239406421780586, + 0.0034058704040944576, + 0.00294364127330482, + 0.0029179954435676336, + 0.003791605355218053, + 0.0029163933359086514, + 0.0017921084072440863, + 0.002828353550285101, + 0.0019843261688947678, + 0.0025181605014950037, + 0.0014711193507537246, + 0.0016941146459430456, + 0.002281652996316552, + 0.05892364680767059, + 0.06786130368709564, + 0.042075105011463165, + 0.03670160099864006, + 0.03025401383638382, + 0.027391593903303146, + 0.04699202999472618, + 0.03825981914997101, + 0.008097732439637184, + 0.010906670242547989, + 0.04228764772415161, + 0.012334425933659077, + 0.007808611262589693, + 0.0057647861540317535, + 0.004875766579061747, + 0.03021864779293537, + 0.0042777033522725105, + 0.0031256298534572124, + 0.01164097711443901, + 0.005190317519009113, + 0.0027135664131492376, + 0.0026292321272194386, + 0.0036607631482183933, + 0.00624953955411911, + 0.0035375705920159817, + 0.004041315522044897, + 0.008145935833454132, + 0.0030027935281395912, + 0.003036657115444541, + 0.0022995450999587774, + 0.003526246640831232, + 0.004021491855382919 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 2, + 7 + ], + "selected_reward": -0.00463874451816082 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 eye-level wide-angle view that pans smoothly from left to right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are there exactly four children visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a sun-drenched historic cobblestone town square with old European buildings in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is one of the children a girl with blonde braids wearing a white shirt and a red pleated skirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is one of the children a boy in a red polo shirt and white shorts waving a small paper Polish flag on a stick?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014186603017151356 + ], + "opposite_logprob_tokens": [ + -4.264186382293701 + ], + "logprob_reward": -0.014186603017151356, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014186603017151356, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08890531957149506, + 0.054989006370306015, + 0.2315358817577362, + 0.055681340396404266, + 0.016857946291565895, + 0.035697340965270996, + 0.35424643754959106, + 0.09757022559642792, + 0.045034006237983704, + 0.019482480362057686 + ], + "credit_map": [ + 0.007455066312104464, + 0.01248896773904562, + 0.00894379336386919, + 0.013291322626173496, + 0.010187389329075813, + 0.017295412719249725, + 0.009283800609409809, + 0.009959563612937927, + 0.003926853183656931, + 0.003705926239490509, + 0.004602454137057066, + 0.0060734529979527, + 0.008888723328709602, + 0.01163255050778389, + 0.007976068183779716, + 0.008182978257536888, + 0.015087045729160309, + 0.02937091886997223, + 0.01772783137857914, + 0.015545574016869068, + 0.03212857246398926, + 0.04034624248743057, + 0.028045441955327988, + 0.05328426510095596, + 0.005690714810043573, + 0.00576718570664525, + 0.0061811101622879505, + 0.006821292918175459, + 0.0056789652444422245, + 0.007494548801332712, + 0.005683825351297855, + 0.012363692745566368, + 0.002019846346229315, + 0.002050437731668353, + 0.002597152953967452, + 0.002173135057091713, + 0.0025087047833949327, + 0.0016494381707161665, + 0.0017693061381578445, + 0.002089924644678831, + 0.003253958187997341, + 0.004207070451229811, + 0.005680363159626722, + 0.005922524258494377, + 0.004041682928800583, + 0.0029733742121607065, + 0.005066103767603636, + 0.004552262835204601, + 0.06559229642152786, + 0.06415878236293793, + 0.04457324370741844, + 0.039234258234500885, + 0.028124626725912094, + 0.028141168877482414, + 0.04567219316959381, + 0.03874989598989487, + 0.008343343622982502, + 0.007694449741393328, + 0.024521898478269577, + 0.016564015299081802, + 0.006948119029402733, + 0.00672114547342062, + 0.008002902381122112, + 0.018774356693029404, + 0.004473702050745487, + 0.0042240372858941555, + 0.014013348147273064, + 0.0050508142448961735, + 0.003238116856664419, + 0.0024966502096503973, + 0.0038320699241012335, + 0.007705268450081348, + 0.002271231496706605, + 0.0025407110806554556, + 0.003537328215315938, + 0.0021582436747848988, + 0.002087267814204097, + 0.0018002233700826764, + 0.002142340410500765, + 0.002945133950561285 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 2, + 7 + ], + "selected_reward": -0.014186603017151356 + } + ], + "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/0000021/group_14/index_tensor([549, 549])_group_14_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000021/group_14/index_tensor([549, 549])_group_14_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..d674538a07a210ca28a0c4b19e12f92e1ef7ed09 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000021/group_14/index_tensor([549, 549])_group_14_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "static, eye-level long shot captures the vast, silent expanse of an empty indoor basketball court. The polished hardwood floor gleams under the bright, even glow of the overhead gymnasium lights, its surface reflecting the stark white lines of a three-point arc. At the center of this solitude stands a solitary, athletic figure, his form sharply outlined against the shadowed, empty spectator stands in the background. He wears a vibrant purple basketball jersey with bold gold trim on the collar and stripes on the shorts, the moisture-wicking fabric clinging slightly to his thin, powerful frame. His brow furrowed in concentration, he follows a moment of quiet before suddenly whipping forward with explosive energy. He executes a swift and decisive low dribble, his body curling and unfurling with practiced grace as he sprints toward the basket, the basketball's movement blurring across the floor. The camera remains perfectly still, emphasizing the player's sudden bursts of speed and underscoring his singular dedication, a dynamic figure in purple and gold moving purposefully through the vast, silent space, creating an atmosphere of tranquility yet intense concentration.", + "seed": 679, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000021/group_14/index_549_seed_679_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": -0.28783386475406586, + "vlm_logprob": -0.28783386475406586, + "vlm_yes_no_margin": 4.025, + "vlm_token_credit": -0.28783386475406586, + "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 an empty indoor basketball court with a polished hardwood floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the player wearing a vibrant purple basketball jersey with gold trim on the collar and stripes on the shorts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the player execute a swift low dribble and sprint toward the basket?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the shadowed, empty spectator stands visible in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera remain perfectly still throughout the 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": [ + -0.013390842825174332 + ], + "opposite_logprob_tokens": [ + -5.638391017913818 + ], + "logprob_reward": -0.013390842825174332, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.013390842825174332, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3264921307563782, + 0.384464830160141, + 0.030852533876895905, + 0.025091925635933876, + 0.05222899466753006, + 0.021312400698661804, + 0.03355854004621506, + 0.05813627690076828, + 0.030011475086212158, + 0.037850841879844666 + ], + "credit_map": [ + 0.021726978942751884, + 0.011809642426669598, + 0.09796557575464249, + 0.11318796128034592, + 0.012975923717021942, + 0.012638555839657784, + 0.016041984781622887, + 0.040145520120859146, + 0.01706855185329914, + 0.013053403235971928, + 0.1593380719423294, + 0.1549593210220337, + 0.0075105344876646996, + 0.00850746687501669, + 0.012471380643546581, + 0.01155614573508501, + 0.002699405187740922, + 0.0029093774501234293, + 0.005964877083897591, + 0.004571949131786823, + 0.0034203289542347193, + 0.0023267832584679127, + 0.0033976861741393805, + 0.005562129430472851, + 0.002568062162026763, + 0.003493993543088436, + 0.003426604438573122, + 0.002698638942092657, + 0.004850124940276146, + 0.0025129064451903105, + 0.0024557134602218866, + 0.0030858837999403477, + 0.0056666163727641106, + 0.009116758592426777, + 0.007382097188383341, + 0.006806399207562208, + 0.005796157289296389, + 0.007754453923553228, + 0.004938541445881128, + 0.004767977632582188, + 0.0020540037658065557, + 0.003072549821808934, + 0.003686379874125123, + 0.002813630970194936, + 0.00343782315030694, + 0.0020915030036121607, + 0.002372449729591608, + 0.001784064224921167, + 0.0037798183038830757, + 0.0035465953405946493, + 0.007039680611342192, + 0.0060433028265833855, + 0.003740051994100213, + 0.003425363916903734, + 0.0031421261373907328, + 0.0028416067361831665, + 0.007860781624913216, + 0.004856653045862913, + 0.0076759010553359985, + 0.007498725317418575, + 0.008226164616644382, + 0.008283360861241817, + 0.006641760468482971, + 0.007092936895787716, + 0.0030430497135967016, + 0.0037668771110475063, + 0.004484970588237047, + 0.004117867909371853, + 0.0029723092447966337, + 0.003662291681393981, + 0.0037131039425730705, + 0.0042510079219937325, + 0.005178997293114662, + 0.004394496325403452, + 0.004019522573798895, + 0.005047770217061043, + 0.0045301420614123344, + 0.004569652024656534, + 0.004627597518265247, + 0.005482668522745371 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 7 + ], + "selected_reward": -0.013390842825174332 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene an empty indoor basketball court with a polished hardwood floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the player wearing a vibrant purple basketball jersey with gold trim on the collar and stripes on the shorts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the player execute a swift low dribble and sprint toward the basket?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the shadowed, empty spectator stands visible in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera remain perfectly still throughout the 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": [ + -0.008648204617202282 + ], + "opposite_logprob_tokens": [ + -4.75864839553833 + ], + "logprob_reward": -0.008648204617202282, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008648204617202282, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18381787836551666, + 0.3570990264415741, + 0.03782425448298454, + 0.018659111112356186, + 0.27952706813812256, + 0.02824697270989418, + 0.0197051540017128, + 0.03110485151410103, + 0.02113248035311699, + 0.022883225232362747 + ], + "credit_map": [ + 0.013515867292881012, + 0.008525371551513672, + 0.04242304712533951, + 0.061179839074611664, + 0.008775154128670692, + 0.006929158233106136, + 0.009591732174158096, + 0.03287772461771965, + 0.017754295840859413, + 0.007616675924509764, + 0.14409887790679932, + 0.08067020773887634, + 0.00619194470345974, + 0.014339999295771122, + 0.011181415989995003, + 0.07524560391902924, + 0.0038134639617055655, + 0.003787557827308774, + 0.005715250037610531, + 0.005447749048471451, + 0.005984975025057793, + 0.003960685804486275, + 0.00464570801705122, + 0.004468865226954222, + 0.0018449112540110946, + 0.001909450744278729, + 0.0020605886820703745, + 0.0018793321214616299, + 0.004473531618714333, + 0.002088494598865509, + 0.0020510689355432987, + 0.0023517324589192867, + 0.03265319764614105, + 0.0657753199338913, + 0.03047196753323078, + 0.031108718365430832, + 0.0214779544621706, + 0.05089501291513443, + 0.018774105235934258, + 0.028370777145028114, + 0.0024517900310456753, + 0.002845236100256443, + 0.005679962690919638, + 0.004289030563086271, + 0.004012011922895908, + 0.0025700244586914778, + 0.002521408023312688, + 0.0038775098510086536, + 0.001919256174005568, + 0.002192876534536481, + 0.002793263178318739, + 0.0025746235623955727, + 0.004008482675999403, + 0.001977544976398349, + 0.002232602098956704, + 0.0020065053831785917, + 0.003578697796911001, + 0.003336300142109394, + 0.004215457011014223, + 0.004579818341881037, + 0.0036479015834629536, + 0.003215402364730835, + 0.00387006183154881, + 0.0046612112782895565, + 0.0021905372850596905, + 0.00257404544390738, + 0.0032745320349931717, + 0.00343700242228806, + 0.0021152039989829063, + 0.002187514677643776, + 0.002111361362040043, + 0.003242282662540674, + 0.003644064534455538, + 0.0018378867534920573, + 0.0038705605547875166, + 0.0028027521912008524, + 0.00291424454189837, + 0.0022279962431639433, + 0.0026589178014546633, + 0.002926801797002554 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 4, + 0 + ], + "selected_reward": -0.008648204617202282 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene an empty indoor basketball court with a polished hardwood floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the player wearing a vibrant purple basketball jersey with gold trim on the collar and stripes on the shorts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the player execute a swift low dribble and sprint toward the basket?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the shadowed, empty spectator stands visible in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera remain perfectly still throughout the 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": [ + -0.008631658740341663 + ], + "opposite_logprob_tokens": [ + -4.758631706237793 + ], + "logprob_reward": -0.008631658740341663, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008631658740341663, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07313191145658493, + 0.4929794669151306, + 0.024814389646053314, + 0.024873558431863785, + 0.12484344840049744, + 0.07339084893465042, + 0.039765942841768265, + 0.0481308214366436, + 0.06314042955636978, + 0.03492918983101845 + ], + "credit_map": [ + 0.004946506582200527, + 0.004459943622350693, + 0.011865845881402493, + 0.016921596601605415, + 0.006691714283078909, + 0.004997229669243097, + 0.006585461087524891, + 0.016663622111082077, + 0.025054510682821274, + 0.010874221101403236, + 0.19756947457790375, + 0.10695107281208038, + 0.009118718095123768, + 0.01932607591152191, + 0.016009902581572533, + 0.10807552933692932, + 0.0022565992549061775, + 0.002322677755728364, + 0.004044818691909313, + 0.0033914989326149225, + 0.0041922833770513535, + 0.0022808671928942204, + 0.003090247744694352, + 0.0032353985588997602, + 0.002621475839987397, + 0.0026062498800456524, + 0.002960704965516925, + 0.003063706448301673, + 0.00377538800239563, + 0.0032464321702718735, + 0.003127078525722027, + 0.0034725256264209747, + 0.014533912762999535, + 0.029593488201498985, + 0.012360231950879097, + 0.013041689060628414, + 0.009868195280432701, + 0.023374024778604507, + 0.009221810847520828, + 0.012850108556449413, + 0.005333766806870699, + 0.006285985466092825, + 0.013450476340949535, + 0.01219966821372509, + 0.010333221405744553, + 0.004807477351278067, + 0.009852813556790352, + 0.011127443052828312, + 0.003593790577724576, + 0.003980390727519989, + 0.006937311962246895, + 0.007424726150929928, + 0.005407664459198713, + 0.004013713449239731, + 0.004294350743293762, + 0.0041139991953969, + 0.005301565397530794, + 0.005451862700283527, + 0.006690251175314188, + 0.0070453123189508915, + 0.007136365864425898, + 0.00472628278657794, + 0.005280077923089266, + 0.006499109324067831, + 0.004596640355885029, + 0.013035847805440426, + 0.007538666948676109, + 0.005939440336078405, + 0.00868456531316042, + 0.006413119379431009, + 0.007036891765892506, + 0.009895261377096176, + 0.004958585370332003, + 0.0030587813816964626, + 0.004197669215500355, + 0.0042058564722537994, + 0.004891726654022932, + 0.0033423721324652433, + 0.003975905478000641, + 0.0062982975505292416 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 4, + 5 + ], + "selected_reward": -0.008631658740341663 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene an empty indoor basketball court with a polished hardwood floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the player wearing a vibrant purple basketball jersey with gold trim on the collar and stripes on the shorts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the player execute a swift low dribble and sprint toward the basket?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the shadowed, empty spectator stands visible in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera remain perfectly still throughout the 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": [ + -0.00219533103518188 + ], + "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": [ + 0.0979110449552536, + 0.2779596745967865, + 0.019782651215791702, + 0.011334086768329144, + 0.4094177186489105, + 0.015897344797849655, + 0.007840141654014587, + 0.1081911101937294, + 0.04211537167429924, + 0.009550916031002998 + ], + "credit_map": [ + 0.00662097055464983, + 0.004468319937586784, + 0.028779851272702217, + 0.03474241867661476, + 0.002622668631374836, + 0.003458087332546711, + 0.004948761779814959, + 0.012269947677850723, + 0.024287035688757896, + 0.014922386035323143, + 0.08808554708957672, + 0.037923239171504974, + 0.011722947470843792, + 0.015558173879981041, + 0.012193743139505386, + 0.07326654344797134, + 0.00238966615870595, + 0.002243928611278534, + 0.0031376013066619635, + 0.002760990522801876, + 0.0016242379788309336, + 0.0021901573054492474, + 0.0019035886507481337, + 0.003532477654516697, + 0.0013241635169833899, + 0.0015179950278252363, + 0.0014637998538091779, + 0.001461516134440899, + 0.0018535383278504014, + 0.0010711430804803967, + 0.0011891957838088274, + 0.0014527327148243785, + 0.047053493559360504, + 0.09601932764053345, + 0.043944839388132095, + 0.04581199213862419, + 0.03185700997710228, + 0.07408980280160904, + 0.027938412502408028, + 0.042702797800302505, + 0.0014716078294441104, + 0.001838211202993989, + 0.0030418301466852427, + 0.002478664740920067, + 0.0016058532055467367, + 0.001548021100461483, + 0.0015012744115665555, + 0.002411880297586322, + 0.0007503699162043631, + 0.0010277738329023123, + 0.0010544380638748407, + 0.0010517612099647522, + 0.0013707263860851526, + 0.0007854747818782926, + 0.0010066799586638808, + 0.0007929160492494702, + 0.015370456501841545, + 0.013421046547591686, + 0.013270081952214241, + 0.016460677608847618, + 0.011529631912708282, + 0.014013608917593956, + 0.009264941327273846, + 0.014860638417303562, + 0.004576703067868948, + 0.0043184794485569, + 0.008053368888795376, + 0.004746835213154554, + 0.0035607831086963415, + 0.0049791219644248486, + 0.0039000497199594975, + 0.007980024442076683, + 0.0014760560588911176, + 0.0008915818179957569, + 0.0015150678809732199, + 0.001171046169474721, + 0.0010348435025662184, + 0.0009269568254239857, + 0.0010920295026153326, + 0.0014433328760787845 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 1, + 7 + ], + "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 scene an empty indoor basketball court with a polished hardwood floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the player wearing a vibrant purple basketball jersey with gold trim on the collar and stripes on the shorts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the player execute a swift low dribble and sprint toward the basket?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the shadowed, empty spectator stands visible in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera remain perfectly still throughout the 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": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.4063032865524292 + ], + "opposite_logprob_tokens": [ + -0.2813032567501068 + ], + "logprob_reward": -1.4063032865524292, + "yes_no_margin_reward": -1.125, + "token_credit_reward": -1.4063032865524292, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.1416107416152954, + 0.13605813682079315, + 0.05145900696516037, + 0.09004152566194534, + 0.3049725294113159, + 0.04482860118150711, + 0.05087665840983391, + 0.05155212804675102, + 0.04334050044417381, + 0.08526021987199783 + ], + "credit_map": [ + 0.0091306883841753, + 0.006260997615754604, + 0.04120725393295288, + 0.04602900519967079, + 0.012446409091353416, + 0.007353485561907291, + 0.008680975995957851, + 0.010501927696168423, + 0.010159043595194817, + 0.00662143062800169, + 0.04828701168298721, + 0.04391010105609894, + 0.006925409659743309, + 0.008363769389688969, + 0.005624832585453987, + 0.0061665368266403675, + 0.006492580287158489, + 0.005912110209465027, + 0.00855629425495863, + 0.007858934812247753, + 0.00492760818451643, + 0.005774743389338255, + 0.005335019435733557, + 0.006601715926080942, + 0.009165112860500813, + 0.009431522339582443, + 0.013864868320524693, + 0.015187930315732956, + 0.012537552043795586, + 0.009100942872464657, + 0.009145209565758705, + 0.01160837896168232, + 0.03478686884045601, + 0.07348596304655075, + 0.032510433346033096, + 0.03515724837779999, + 0.023760415613651276, + 0.054406892508268356, + 0.02026970684528351, + 0.03059498406946659, + 0.004215058870613575, + 0.0031875038985162973, + 0.007819284684956074, + 0.007776085287332535, + 0.007005359977483749, + 0.004450722131878138, + 0.0051271794363856316, + 0.0052474080584943295, + 0.004481461830437183, + 0.00585977965965867, + 0.009567459113895893, + 0.007329120300710201, + 0.006367853842675686, + 0.005891652312129736, + 0.005057252012193203, + 0.006322079803794622, + 0.007383470423519611, + 0.004418348427861929, + 0.005425680428743362, + 0.005592329427599907, + 0.006651881616562605, + 0.011353842914104462, + 0.005445792805403471, + 0.005280784796923399, + 0.0034603732638061047, + 0.007533717900514603, + 0.0050449795089662075, + 0.00795592088252306, + 0.004090876318514347, + 0.004251896403729916, + 0.006036562845110893, + 0.00496617192402482, + 0.009995758533477783, + 0.009483492001891136, + 0.009101583622395992, + 0.013290882110595703, + 0.010685164481401443, + 0.009720990434288979, + 0.008717282675206661, + 0.014265063218772411 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 1 + ], + "selected_reward": -1.4063032865524292 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "No" + ] + }, + { + "prompt": "static, eye-level long shot captures the vast, silent expanse of an empty indoor basketball court. The polished hardwood floor gleams under the bright, even glow of the overhead gymnasium lights, its surface reflecting the stark white lines of a three-point arc. At the center of this solitude stands a solitary, athletic figure, his form sharply outlined against the shadowed, empty spectator stands in the background. He wears a vibrant purple basketball jersey with bold gold trim on the collar and stripes on the shorts, the moisture-wicking fabric clinging slightly to his thin, powerful frame. His brow furrowed in concentration, he follows a moment of quiet before suddenly whipping forward with explosive energy. He executes a swift and decisive low dribble, his body curling and unfurling with practiced grace as he sprints toward the basket, the basketball's movement blurring across the floor. The camera remains perfectly still, emphasizing the player's sudden bursts of speed and underscoring his singular dedication, a dynamic figure in purple and gold moving purposefully through the vast, silent space, creating an atmosphere of tranquility yet intense concentration.", + "seed": 680, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000021/group_14/index_549_seed_680_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": -0.36672592754475775, + "vlm_logprob": -0.36672592754475775, + "vlm_yes_no_margin": 4.275, + "vlm_token_credit": -0.36672592754475775, + "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 an empty indoor basketball court with a polished hardwood floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the player wearing a vibrant purple basketball jersey with gold trim on the collar and stripes on the shorts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the player execute a swift low dribble and sprint toward the basket?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the shadowed, empty spectator stands visible in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera remain perfectly still throughout the 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": [ + -0.008339108899235725 + ], + "opposite_logprob_tokens": [ + -7.258338928222656 + ], + "logprob_reward": -0.008339108899235725, + "yes_no_margin_reward": 7.25, + "token_credit_reward": -0.008339108899235725, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16272951662540436, + 0.25453683733940125, + 0.07035063952207565, + 0.054486941546201706, + 0.04623522609472275, + 0.06932945549488068, + 0.11737073957920074, + 0.06517776101827621, + 0.08027718961238861, + 0.07950571924448013 + ], + "credit_map": [ + 0.013595794327557087, + 0.016049738973379135, + 0.022887403145432472, + 0.025073397904634476, + 0.021621787920594215, + 0.011036001145839691, + 0.025686224922537804, + 0.02677915059030056, + 0.012071027420461178, + 0.026780933141708374, + 0.03633652254939079, + 0.07013019919395447, + 0.022022295743227005, + 0.011405815370380878, + 0.031965505331754684, + 0.04382452368736267, + 0.006800801958888769, + 0.0057027763687074184, + 0.01199381984770298, + 0.01277236733585596, + 0.011959533207118511, + 0.004375750198960304, + 0.008992472663521767, + 0.007753114216029644, + 0.002684843959286809, + 0.005693655926734209, + 0.005175020545721054, + 0.01299623865634203, + 0.010650225915014744, + 0.004015525337308645, + 0.0072419410571455956, + 0.006029483862221241, + 0.0028328371699899435, + 0.005079824011772871, + 0.007415286730974913, + 0.006818622350692749, + 0.007158361840993166, + 0.0038881630171090364, + 0.005315935704857111, + 0.007726189214736223, + 0.006122662220150232, + 0.00829889252781868, + 0.008285674266517162, + 0.0108105493709445, + 0.009819862432777882, + 0.008817442692816257, + 0.009859935380518436, + 0.007314433343708515, + 0.02037988416850567, + 0.009476755745708942, + 0.01740751974284649, + 0.01553506962954998, + 0.012997177429497242, + 0.012627684511244297, + 0.013641024939715862, + 0.015305610373616219, + 0.00644714804366231, + 0.01230622734874487, + 0.00624106777831912, + 0.006462838500738144, + 0.009264453314244747, + 0.007071722764521837, + 0.009746377356350422, + 0.007637915667146444, + 0.004611151292920113, + 0.009846929460763931, + 0.01148973684757948, + 0.020444517955183983, + 0.005748603958636522, + 0.01049068570137024, + 0.009228738024830818, + 0.008416819386184216, + 0.004910518880933523, + 0.00957075972110033, + 0.008424089290201664, + 0.012848732993006706, + 0.008935455232858658, + 0.009808776900172234, + 0.014625953510403633, + 0.010381427593529224 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 6 + ], + "selected_reward": -0.008339108899235725 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene an empty indoor basketball court with a polished hardwood floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the player wearing a vibrant purple basketball jersey with gold trim on the collar and stripes on the shorts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the player execute a swift low dribble and sprint toward the basket?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the shadowed, empty spectator stands visible in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera remain perfectly still throughout the 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": [ + -0.00595793966203928 + ], + "opposite_logprob_tokens": [ + -5.130958080291748 + ], + "logprob_reward": -0.00595793966203928, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.00595793966203928, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15124671161174774, + 0.07080287486314774, + 0.0700453594326973, + 0.09350194782018661, + 0.11122030764818192, + 0.12482970952987671, + 0.1746314913034439, + 0.06858976185321808, + 0.09005296975374222, + 0.045078832656145096 + ], + "credit_map": [ + 0.011761724017560482, + 0.020532069727778435, + 0.01837931200861931, + 0.021200425922870636, + 0.02079864591360092, + 0.013804102316498756, + 0.02147628553211689, + 0.02329414337873459, + 0.007737290114164352, + 0.006398315541446209, + 0.012388925068080425, + 0.010957280173897743, + 0.007923480123281479, + 0.006054937373846769, + 0.009918765164911747, + 0.009423884563148022, + 0.004774072207510471, + 0.00987305399030447, + 0.00790922436863184, + 0.010214991867542267, + 0.010969801805913448, + 0.007146286312490702, + 0.010806412436068058, + 0.008351518772542477, + 0.009037170559167862, + 0.009050766937434673, + 0.020632978528738022, + 0.013893678784370422, + 0.008066540583968163, + 0.005649559665471315, + 0.010361370630562305, + 0.01680987887084484, + 0.012859709560871124, + 0.013657217845320702, + 0.02378251776099205, + 0.011379534378647804, + 0.01098722219467163, + 0.010104846209287643, + 0.010671173222362995, + 0.01777808368206024, + 0.00987187772989273, + 0.011626658961176872, + 0.025738324970006943, + 0.013924362137913704, + 0.01753970794379711, + 0.009557695128023624, + 0.01677694171667099, + 0.01979413814842701, + 0.018742375075817108, + 0.015470927581191063, + 0.027631407603621483, + 0.030949486419558525, + 0.014801772311329842, + 0.013389400206506252, + 0.025805648416280746, + 0.027840467169880867, + 0.009209323674440384, + 0.009183606132864952, + 0.0059983194805681705, + 0.008983743377029896, + 0.009235930629074574, + 0.007773095276206732, + 0.009053590707480907, + 0.009152153506875038, + 0.004980592057108879, + 0.00873805396258831, + 0.012433883734047413, + 0.014369167387485504, + 0.00927876029163599, + 0.013436087407171726, + 0.011244378052651882, + 0.015572050586342812, + 0.004954342730343342, + 0.006421870086342096, + 0.005427013151347637, + 0.0063575031235814095, + 0.0043974630534648895, + 0.0062272516079247, + 0.005243352148681879, + 0.006050036754459143 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 5 + ], + "selected_reward": -0.00595793966203928 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene an empty indoor basketball court with a polished hardwood floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the player wearing a vibrant purple basketball jersey with gold trim on the collar and stripes on the shorts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the player execute a swift low dribble and sprint toward the basket?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the shadowed, empty spectator stands visible in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera remain perfectly still throughout the 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": [ + -0.012530526146292686 + ], + "opposite_logprob_tokens": [ + -4.387530326843262 + ], + "logprob_reward": -0.012530526146292686, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012530526146292686, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11212172359228134, + 0.16462229192256927, + 0.08113320916891098, + 0.15546280145645142, + 0.10089954733848572, + 0.06907300651073456, + 0.11673756688833237, + 0.09404132515192032, + 0.04838996380567551, + 0.05751851201057434 + ], + "credit_map": [ + 0.006446933373808861, + 0.013802144676446915, + 0.01622447371482849, + 0.02209082804620266, + 0.010361175052821636, + 0.0063340310007333755, + 0.01662968285381794, + 0.02023245580494404, + 0.00725153274834156, + 0.01806178316473961, + 0.02169625833630562, + 0.04805770516395569, + 0.013429125770926476, + 0.006817126180976629, + 0.02113613858819008, + 0.028172628954052925, + 0.007130803540349007, + 0.008133449591696262, + 0.01334398239850998, + 0.009577623568475246, + 0.01194797083735466, + 0.0071481578052043915, + 0.01370238233357668, + 0.010148834437131882, + 0.00600597495213151, + 0.013139922171831131, + 0.027659742161631584, + 0.023355863988399506, + 0.03154592961072922, + 0.00890570692718029, + 0.020599620416760445, + 0.024250047281384468, + 0.005901804193854332, + 0.009666831232607365, + 0.018253035843372345, + 0.012847978621721268, + 0.01823517307639122, + 0.007928264327347279, + 0.012730009853839874, + 0.015336461365222931, + 0.006756688002496958, + 0.007774698548018932, + 0.008434871211647987, + 0.009651518426835537, + 0.009441071189939976, + 0.006927960552275181, + 0.012045138515532017, + 0.008041060529649258, + 0.014630257152020931, + 0.012723981402814388, + 0.02352047897875309, + 0.010440962389111519, + 0.011962749995291233, + 0.013281800784170628, + 0.012346234172582626, + 0.017831100150942802, + 0.019886713474988937, + 0.015929831191897392, + 0.010888433083891869, + 0.007865434512495995, + 0.008800025098025799, + 0.01008623093366623, + 0.009820614010095596, + 0.0107640465721488, + 0.0030733472667634487, + 0.004276657942682505, + 0.0052995034493505955, + 0.007107753772288561, + 0.006141732446849346, + 0.009347479790449142, + 0.006193958222866058, + 0.006949526723474264, + 0.005108522716909647, + 0.00502732815220952, + 0.006275018211454153, + 0.007057715207338333, + 0.004718780983239412, + 0.009821719489991665, + 0.008851741440594196, + 0.010657685808837414 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 3, + 6 + ], + "selected_reward": -0.012530526146292686 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene an empty indoor basketball court with a polished hardwood floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the player wearing a vibrant purple basketball jersey with gold trim on the collar and stripes on the shorts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the player execute a swift low dribble and sprint toward the basket?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the shadowed, empty spectator stands visible in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera remain perfectly still throughout the 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": [ + -0.0019377756398171186 + ], + "opposite_logprob_tokens": [ + -6.2519378662109375 + ], + "logprob_reward": -0.0019377756398171186, + "yes_no_margin_reward": 6.25, + "token_credit_reward": -0.0019377756398171186, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2396802455186844, + 0.1932738572359085, + 0.1301572024822235, + 0.06114191934466362, + 0.08568542450666428, + 0.05260203033685684, + 0.03188468515872955, + 0.057610806077718735, + 0.07005376368761063, + 0.07791008055210114 + ], + "credit_map": [ + 0.019254622980952263, + 0.029821230098605156, + 0.04098980128765106, + 0.034760285168886185, + 0.03044245019555092, + 0.01788467913866043, + 0.027895335108041763, + 0.03863183408975601, + 0.024966152384877205, + 0.023236986249685287, + 0.030064214020967484, + 0.034569792449474335, + 0.017985867336392403, + 0.023308971896767616, + 0.01868828386068344, + 0.02045358158648014, + 0.013947847299277782, + 0.014105003327131271, + 0.031421978026628494, + 0.016330145299434662, + 0.012375842779874802, + 0.014973619021475315, + 0.014865396544337273, + 0.012137368321418762, + 0.004267948679625988, + 0.0057731736451387405, + 0.008819807320833206, + 0.009848998859524727, + 0.007638320792466402, + 0.005968814250081778, + 0.0077113136649131775, + 0.011113540269434452, + 0.00672509940341115, + 0.014218151569366455, + 0.012210831977427006, + 0.007775282487273216, + 0.009718053974211216, + 0.011667195707559586, + 0.005692772567272186, + 0.017678041011095047, + 0.0029846581164747477, + 0.005692062899470329, + 0.00855852197855711, + 0.0068161701783537865, + 0.011598700657486916, + 0.004382709972560406, + 0.0055316174402832985, + 0.00703758979216218, + 0.002773202955722809, + 0.003693994600325823, + 0.005522143095731735, + 0.0043664914555847645, + 0.003292789915576577, + 0.0030062098521739244, + 0.004597624763846397, + 0.00463222898542881, + 0.010456295683979988, + 0.011953669600188732, + 0.0060598705895245075, + 0.005571869667619467, + 0.005464621819555759, + 0.005599204916507006, + 0.006851777900010347, + 0.00565349729731679, + 0.005098692141473293, + 0.00602748803794384, + 0.010746078565716743, + 0.01790591888129711, + 0.006104042287915945, + 0.008592721074819565, + 0.007436749059706926, + 0.008142078295350075, + 0.010103556327521801, + 0.011834590695798397, + 0.011013136245310307, + 0.009935581125319004, + 0.009994689375162125, + 0.006578677799552679, + 0.011204196140170097, + 0.007245655171573162 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.0019377756398171186 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene an empty indoor basketball court with a polished hardwood floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the player wearing a vibrant purple basketball jersey with gold trim on the collar and stripes on the shorts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the player execute a swift low dribble and sprint toward the basket?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the shadowed, empty spectator stands visible in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera remain perfectly still throughout the 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": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.8048642873764038 + ], + "opposite_logprob_tokens": [ + -0.17986425757408142 + ], + "logprob_reward": -1.8048642873764038, + "yes_no_margin_reward": -1.625, + "token_credit_reward": -1.8048642873764038, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.20741182565689087, + 0.11954154074192047, + 0.06309805810451508, + 0.04147198423743248, + 0.06537366658449173, + 0.08310522139072418, + 0.0897771567106247, + 0.1129198744893074, + 0.09788962453603745, + 0.11941107362508774 + ], + "credit_map": [ + 0.021992601454257965, + 0.025047026574611664, + 0.023946965113282204, + 0.029162051156163216, + 0.02413463406264782, + 0.01765918731689453, + 0.03182138130068779, + 0.03364798054099083, + 0.013470092788338661, + 0.011372297070920467, + 0.016990836709737778, + 0.02238341048359871, + 0.011985260993242264, + 0.011362490244209766, + 0.010953095741569996, + 0.021024052053689957, + 0.004945775028318167, + 0.007184631656855345, + 0.006805408280342817, + 0.009590798988938332, + 0.009938850998878479, + 0.005696701351553202, + 0.011412497609853745, + 0.007523396052420139, + 0.0022011937107890844, + 0.005751900840550661, + 0.006680771708488464, + 0.007623043842613697, + 0.0034796493127942085, + 0.0023978722747415304, + 0.005973598454147577, + 0.007363957818597555, + 0.004192267544567585, + 0.007645618170499802, + 0.011758903972804546, + 0.009778092615306377, + 0.008496078662574291, + 0.004645271226763725, + 0.006640446837991476, + 0.012216988950967789, + 0.007682060822844505, + 0.008768929168581963, + 0.011898806318640709, + 0.011572613380849361, + 0.01392383873462677, + 0.008517901413142681, + 0.010114481672644615, + 0.010626591742038727, + 0.008481411263346672, + 0.009001195430755615, + 0.013446711003780365, + 0.01280667632818222, + 0.013281659223139286, + 0.007581170182675123, + 0.014197568409144878, + 0.010980766266584396, + 0.016143430024385452, + 0.01908745989203453, + 0.013586430810391903, + 0.012679053470492363, + 0.012178927659988403, + 0.011455008760094643, + 0.011997961439192295, + 0.015791600570082664, + 0.006929518189281225, + 0.009590339846909046, + 0.016904933378100395, + 0.02308000810444355, + 0.007995068095624447, + 0.013037298806011677, + 0.010894604958593845, + 0.009457859210669994, + 0.011866101063787937, + 0.013251112774014473, + 0.01307591050863266, + 0.0179586261510849, + 0.011915554292500019, + 0.015503864735364914, + 0.016093343496322632, + 0.019746560603380203 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 9 + ], + "selected_reward": -1.8048642873764038 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "No" + ] + }, + { + "prompt": "static, eye-level long shot captures the vast, silent expanse of an empty indoor basketball court. The polished hardwood floor gleams under the bright, even glow of the overhead gymnasium lights, its surface reflecting the stark white lines of a three-point arc. At the center of this solitude stands a solitary, athletic figure, his form sharply outlined against the shadowed, empty spectator stands in the background. He wears a vibrant purple basketball jersey with bold gold trim on the collar and stripes on the shorts, the moisture-wicking fabric clinging slightly to his thin, powerful frame. His brow furrowed in concentration, he follows a moment of quiet before suddenly whipping forward with explosive energy. He executes a swift and decisive low dribble, his body curling and unfurling with practiced grace as he sprints toward the basket, the basketball's movement blurring across the floor. The camera remains perfectly still, emphasizing the player's sudden bursts of speed and underscoring his singular dedication, a dynamic figure in purple and gold moving purposefully through the vast, silent space, creating an atmosphere of tranquility yet intense concentration.", + "seed": 681, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000021/group_14/index_549_seed_681_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": -0.31003806749358775, + "vlm_logprob": -0.31003806749358775, + "vlm_yes_no_margin": 3.575, + "vlm_token_credit": -0.31003806749358775, + "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 an empty indoor basketball court with a polished hardwood floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the player wearing a vibrant purple basketball jersey with gold trim on the collar and stripes on the shorts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the player execute a swift low dribble and sprint toward the basket?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the shadowed, empty spectator stands visible in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera remain perfectly still throughout the 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": [ + -0.01925000175833702 + ], + "opposite_logprob_tokens": [ + -5.01924991607666 + ], + "logprob_reward": -0.01925000175833702, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.01925000175833702, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5647357106208801, + 0.03794364631175995, + 0.10920783877372742, + 0.021137041971087456, + 0.11364775896072388, + 0.051816221326589584, + 0.008360965177416801, + 0.024363860487937927, + 0.030915498733520508, + 0.03787146881222725 + ], + "credit_map": [ + 0.045747775584459305, + 0.01875733956694603, + 0.08779072761535645, + 0.042141590267419815, + 0.028761615976691246, + 0.08345203101634979, + 0.19509948790073395, + 0.0629851371049881, + 0.0035212840884923935, + 0.0018211640417575836, + 0.005225749686360359, + 0.0040118275210261345, + 0.002442903583869338, + 0.006286719348281622, + 0.010552111081779003, + 0.004081883002072573, + 0.00809931755065918, + 0.0055092633701860905, + 0.031992487609386444, + 0.012809207662940025, + 0.005883987061679363, + 0.007879417389631271, + 0.011615278199315071, + 0.025418885052204132, + 0.002525140531361103, + 0.0028126894030719995, + 0.002178880851715803, + 0.002928604604676366, + 0.002561819041147828, + 0.0020174947567284107, + 0.0035575127694755793, + 0.002554898615926504, + 0.01037913840264082, + 0.010255514644086361, + 0.018537253141403198, + 0.02965576946735382, + 0.006522380746901035, + 0.00924689695239067, + 0.01193776074796915, + 0.017113031819462776, + 0.005461071617901325, + 0.007817080244421959, + 0.006223645526915789, + 0.007555849384516478, + 0.004272261634469032, + 0.007001812104135752, + 0.005584991071373224, + 0.007899499498307705, + 0.0009691031300462782, + 0.0008989141206257045, + 0.0011797898914664984, + 0.0009947415674105287, + 0.0008579809800721705, + 0.0008589908247813582, + 0.001118731452152133, + 0.0014827130362391472, + 0.0017754402942955494, + 0.002197887748479843, + 0.003047299338504672, + 0.0033430797047913074, + 0.0031442209146916866, + 0.0038470576982945204, + 0.003280110890045762, + 0.003728762734681368, + 0.005926086567342281, + 0.005766337271779776, + 0.004406461492180824, + 0.003625786630436778, + 0.003096596337854862, + 0.0029669422656297684, + 0.002389222849160433, + 0.0027380622923374176, + 0.004428455606102943, + 0.005247593857347965, + 0.005741421598941088, + 0.006857241969555616, + 0.004503688309341669, + 0.002318846993148327, + 0.0037721211556345224, + 0.005002096761018038 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 2 + ], + "selected_reward": -0.01925000175833702 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene an empty indoor basketball court with a polished hardwood floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the player wearing a vibrant purple basketball jersey with gold trim on the collar and stripes on the shorts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the player execute a swift low dribble and sprint toward the basket?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the shadowed, empty spectator stands visible in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera remain perfectly still throughout the 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": [ + -0.011087356135249138 + ], + "opposite_logprob_tokens": [ + -4.511087417602539 + ], + "logprob_reward": -0.011087356135249138, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011087356135249138, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2908281981945038, + 0.0382850244641304, + 0.13713175058364868, + 0.06245808303356171, + 0.08208546042442322, + 0.06408049166202545, + 0.052956774830818176, + 0.027855159714818, + 0.095828577876091, + 0.14849048852920532 + ], + "credit_map": [ + 0.024631667882204056, + 0.01251997984945774, + 0.04446469619870186, + 0.02454988844692707, + 0.014531619846820831, + 0.040260639041662216, + 0.09522085636854172, + 0.03464885801076889, + 0.0028823453467339277, + 0.0025781127624213696, + 0.005976746324449778, + 0.004250438883900642, + 0.002762609627097845, + 0.004788583144545555, + 0.010358535684645176, + 0.004687647335231304, + 0.01213029120117426, + 0.008556269109249115, + 0.03244056552648544, + 0.015151512809097767, + 0.009156358428299427, + 0.017354609444737434, + 0.011792861856520176, + 0.03054927848279476, + 0.005512961186468601, + 0.004736308008432388, + 0.008521129377186298, + 0.009287948720157146, + 0.0057259369641542435, + 0.006787822116166353, + 0.010168763808906078, + 0.011717214249074459, + 0.005590499844402075, + 0.006044705864042044, + 0.012278866954147816, + 0.017374170944094658, + 0.006168432533740997, + 0.0051865908317267895, + 0.018728608265519142, + 0.01071358472108841, + 0.006116888951510191, + 0.006677169352769852, + 0.008256494998931885, + 0.009538012556731701, + 0.007442150730639696, + 0.008446522057056427, + 0.00894665066152811, + 0.008656603284180164, + 0.004279347136616707, + 0.0030492558144032955, + 0.007713911589235067, + 0.00774623965844512, + 0.0036175951827317476, + 0.005326014012098312, + 0.0072309281677007675, + 0.013993482105433941, + 0.0021277708001434803, + 0.0026819438207894564, + 0.0036446694284677505, + 0.0035373992286622524, + 0.0027819008100777864, + 0.003724822076037526, + 0.004578124266117811, + 0.004778527654707432, + 0.007325420621782541, + 0.00786732230335474, + 0.01020557526499033, + 0.025199489668011665, + 0.010592726990580559, + 0.008981019258499146, + 0.014601508155465126, + 0.01105551142245531, + 0.02398013323545456, + 0.022104252129793167, + 0.01570127159357071, + 0.02229021117091179, + 0.014804497361183167, + 0.007407164201140404, + 0.02096203900873661, + 0.02124091424047947 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 2 + ], + "selected_reward": -0.011087356135249138 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene an empty indoor basketball court with a polished hardwood floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the player wearing a vibrant purple basketball jersey with gold trim on the collar and stripes on the shorts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the player execute a swift low dribble and sprint toward the basket?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the shadowed, empty spectator stands visible in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera remain perfectly still throughout the 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": [ + -0.012531115673482418 + ], + "opposite_logprob_tokens": [ + -4.387531280517578 + ], + "logprob_reward": -0.012531115673482418, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012531115673482418, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09542587399482727, + 0.0699702724814415, + 0.47639065980911255, + 0.02853371389210224, + 0.08935066312551498, + 0.04236205294728279, + 0.06236006319522858, + 0.04386979714035988, + 0.03486403450369835, + 0.056872881948947906 + ], + "credit_map": [ + 0.010373981669545174, + 0.009770271368324757, + 0.013430409133434296, + 0.01758294366300106, + 0.008949872106313705, + 0.011282239109277725, + 0.011942832730710506, + 0.012093332596123219, + 0.006551531609147787, + 0.00298762577585876, + 0.008939754217863083, + 0.009599071927368641, + 0.004090003203600645, + 0.00815373845398426, + 0.02083955891430378, + 0.008808990009129047, + 0.04553468897938728, + 0.031962111592292786, + 0.11167678236961365, + 0.057975124567747116, + 0.02734473906457424, + 0.05053307116031647, + 0.04904584586620331, + 0.10231828689575195, + 0.0025433278642594814, + 0.002402480458840728, + 0.0028608895372599363, + 0.0042859879322350025, + 0.0032081797253340483, + 0.003551253816112876, + 0.004919664468616247, + 0.004761930555105209, + 0.005846414249390364, + 0.005286891479045153, + 0.012012429535388947, + 0.019294073805212975, + 0.00722620589658618, + 0.006329843774437904, + 0.02088455855846405, + 0.012470241636037827, + 0.003327038371935487, + 0.0036570969969034195, + 0.007438832428306341, + 0.006570018362253904, + 0.0031185108236968517, + 0.004624936264008284, + 0.0065021030604839325, + 0.007123513612896204, + 0.006208750884979963, + 0.005364138167351484, + 0.0070455120876431465, + 0.010707905516028404, + 0.004117674194276333, + 0.005668451078236103, + 0.010318280197679996, + 0.012929349206387997, + 0.004334553144872189, + 0.004479218740016222, + 0.006364891771227121, + 0.005906961392611265, + 0.004017074592411518, + 0.0066045536659657955, + 0.005639796610921621, + 0.006522744428366423, + 0.0042991600930690765, + 0.004162960685789585, + 0.0033402424305677414, + 0.007887864485383034, + 0.0034388171043246984, + 0.0032388928811997175, + 0.005528263747692108, + 0.0029678340069949627, + 0.007495479192584753, + 0.006968312431126833, + 0.008819647133350372, + 0.012278650887310505, + 0.006147478241473436, + 0.0032074619084596634, + 0.006507509388029575, + 0.005448343697935343 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 4 + ], + "selected_reward": -0.012531115673482418 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene an empty indoor basketball court with a polished hardwood floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the player wearing a vibrant purple basketball jersey with gold trim on the collar and stripes on the shorts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the player execute a swift low dribble and sprint toward the basket?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the shadowed, empty spectator stands visible in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera remain perfectly still throughout the 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": [ + -0.005248814355581999 + ], + "opposite_logprob_tokens": [ + -5.2552490234375 + ], + "logprob_reward": -0.005248814355581999, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005248814355581999, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.47648513317108154, + 0.12117927521467209, + 0.10060770809650421, + 0.0489581860601902, + 0.1704341024160385, + 0.022849664092063904, + 0.026541942730545998, + 0.010145680978894234, + 0.007737420964986086, + 0.015060867182910442 + ], + "credit_map": [ + 0.03783658519387245, + 0.01645510271191597, + 0.07430511713027954, + 0.03401019424200058, + 0.02491958811879158, + 0.06996007263660431, + 0.16718029975891113, + 0.05181821063160896, + 0.010886311531066895, + 0.0051805609837174416, + 0.018177200108766556, + 0.010388214141130447, + 0.006853012368083, + 0.017308512702584267, + 0.03893798589706421, + 0.013447490520775318, + 0.009381663054227829, + 0.006314700469374657, + 0.027783704921603203, + 0.009807760827243328, + 0.004616318270564079, + 0.010108713991940022, + 0.009873462840914726, + 0.022721396759152412, + 0.004099941346794367, + 0.0035632741637527943, + 0.008344044908881187, + 0.010923459194600582, + 0.0029166131280362606, + 0.0053254165686666965, + 0.0062771267257630825, + 0.007508315611630678, + 0.018554985523223877, + 0.020594503730535507, + 0.02292553335428238, + 0.04154176637530327, + 0.012869400903582573, + 0.02375723049044609, + 0.012563033029437065, + 0.0176276545971632, + 0.002535454463213682, + 0.0034270829055458307, + 0.0024547865614295006, + 0.0032285999041050673, + 0.001968802185729146, + 0.003357634646818042, + 0.002474664244800806, + 0.0034026417415589094, + 0.004144604783505201, + 0.00397529499605298, + 0.003944316878914833, + 0.004267086274921894, + 0.0023597439285367727, + 0.002588145202025771, + 0.002315401565283537, + 0.0029473542235791683, + 0.001079846522770822, + 0.0014200962614268064, + 0.0014246046775951982, + 0.0017038503428921103, + 0.0013882959028705955, + 0.0008158634300343692, + 0.001364521449431777, + 0.0009486037888564169, + 0.0005704231443814933, + 0.0006537216249853373, + 0.0010398688027635217, + 0.0021753176115453243, + 0.0008378672064282, + 0.0004908081609755754, + 0.0012625688686966896, + 0.0007068469421938062, + 0.003081599483266473, + 0.002454855479300022, + 0.002236702246591449, + 0.0025115504395216703, + 0.0012692813761532307, + 0.0007492317236028612, + 0.0016687290044501424, + 0.001088918885216117 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 1 + ], + "selected_reward": -0.005248814355581999 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene an empty indoor basketball court with a polished hardwood floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the player wearing a vibrant purple basketball jersey with gold trim on the collar and stripes on the shorts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the player execute a swift low dribble and sprint toward the basket?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the shadowed, empty spectator stands visible in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera remain perfectly still throughout the 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": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.502073049545288 + ], + "opposite_logprob_tokens": [ + -0.2520730197429657 + ], + "logprob_reward": -1.502073049545288, + "yes_no_margin_reward": -1.25, + "token_credit_reward": -1.502073049545288, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.3456193208694458, + 0.08894933015108109, + 0.04014037922024727, + 0.07616425305604935, + 0.10582572966814041, + 0.051366448402404785, + 0.050926681607961655, + 0.05077548697590828, + 0.032957930117845535, + 0.15727441012859344 + ], + "credit_map": [ + 0.02930605411529541, + 0.012518315576016903, + 0.05485248565673828, + 0.026193441823124886, + 0.018632927909493446, + 0.05396903306245804, + 0.11548284441232681, + 0.03466419875621796, + 0.00784375797957182, + 0.0032433723099529743, + 0.014361639507114887, + 0.0060228100046515465, + 0.004729531705379486, + 0.013893297873437405, + 0.029328428208827972, + 0.00952649861574173, + 0.003166642738506198, + 0.002839874243363738, + 0.007351408246904612, + 0.007725979201495647, + 0.0033989460207521915, + 0.0032212838996201754, + 0.005011710803955793, + 0.007424535695463419, + 0.007724604103714228, + 0.006274248939007521, + 0.011379617266356945, + 0.013768148608505726, + 0.006482481025159359, + 0.007869104854762554, + 0.010673175565898418, + 0.01199287362396717, + 0.01333546731621027, + 0.009929817169904709, + 0.01125385332852602, + 0.017583945766091347, + 0.00804330874234438, + 0.019311964511871338, + 0.015537232160568237, + 0.010830150917172432, + 0.005165227223187685, + 0.005567171145230532, + 0.007212589494884014, + 0.007607324048876762, + 0.0038006012327969074, + 0.004052378237247467, + 0.008975936099886894, + 0.008985219523310661, + 0.004439406096935272, + 0.005308210384100676, + 0.007470117881894112, + 0.006580093875527382, + 0.0057528261095285416, + 0.005436189938336611, + 0.0066716019064188, + 0.00926823541522026, + 0.004712192807346582, + 0.004367684479802847, + 0.008842498995363712, + 0.009125850163400173, + 0.005242705345153809, + 0.0060002729296684265, + 0.006760160438716412, + 0.005724117625504732, + 0.0045312135480344296, + 0.00404101237654686, + 0.004399957600980997, + 0.009129377081990242, + 0.0032138358801603317, + 0.002054365584626794, + 0.0033455430530011654, + 0.0022426231298595667, + 0.019419262185692787, + 0.022748829796910286, + 0.026328926905989647, + 0.03431050479412079, + 0.015699824318289757, + 0.009093532338738441, + 0.014579440467059612, + 0.015094099566340446 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 4 + ], + "selected_reward": -1.502073049545288 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "No" + ] + }, + { + "prompt": "static, eye-level long shot captures the vast, silent expanse of an empty indoor basketball court. The polished hardwood floor gleams under the bright, even glow of the overhead gymnasium lights, its surface reflecting the stark white lines of a three-point arc. At the center of this solitude stands a solitary, athletic figure, his form sharply outlined against the shadowed, empty spectator stands in the background. He wears a vibrant purple basketball jersey with bold gold trim on the collar and stripes on the shorts, the moisture-wicking fabric clinging slightly to his thin, powerful frame. His brow furrowed in concentration, he follows a moment of quiet before suddenly whipping forward with explosive energy. He executes a swift and decisive low dribble, his body curling and unfurling with practiced grace as he sprints toward the basket, the basketball's movement blurring across the floor. The camera remains perfectly still, emphasizing the player's sudden bursts of speed and underscoring his singular dedication, a dynamic figure in purple and gold moving purposefully through the vast, silent space, creating an atmosphere of tranquility yet intense concentration.", + "seed": 682, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000021/group_14/index_549_seed_682_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": -0.3218025671318173, + "vlm_logprob": -0.3218025671318173, + "vlm_yes_no_margin": 3.05, + "vlm_token_credit": -0.3218025671318173, + "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 an empty indoor basketball court with a polished hardwood floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the player wearing a vibrant purple basketball jersey with gold trim on the collar and stripes on the shorts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the player execute a swift low dribble and sprint toward the basket?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the shadowed, empty spectator stands visible in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera remain perfectly still throughout the 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": [ + -0.03418448939919472 + ], + "opposite_logprob_tokens": [ + -4.034184455871582 + ], + "logprob_reward": -0.03418448939919472, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.03418448939919472, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12996813654899597, + 0.0796051099896431, + 0.35768458247184753, + 0.03577616810798645, + 0.03831266611814499, + 0.08093015104532242, + 0.0354166217148304, + 0.02568424865603447, + 0.023924430832266808, + 0.19269783794879913 + ], + "credit_map": [ + 0.01353857945650816, + 0.010050253011286259, + 0.018732458353042603, + 0.01647605001926422, + 0.011181077919900417, + 0.017056580632925034, + 0.015287079848349094, + 0.027646074071526527, + 0.005764550529420376, + 0.007447395473718643, + 0.015406453050673008, + 0.009409591555595398, + 0.00791190192103386, + 0.008062477223575115, + 0.014285698533058167, + 0.011317051947116852, + 0.04122132062911987, + 0.03599847853183746, + 0.04556303843855858, + 0.03698589652776718, + 0.03455422446131706, + 0.06359539180994034, + 0.04299439117312431, + 0.05677183344960213, + 0.004527362063527107, + 0.0032591037452220917, + 0.005224228370934725, + 0.004130255896598101, + 0.002354210941120982, + 0.004502543713897467, + 0.004337680526077747, + 0.007440787274390459, + 0.005659544840455055, + 0.0034162465017288923, + 0.004392992239445448, + 0.0030319876968860626, + 0.009343323297798634, + 0.00489406194537878, + 0.004018608946353197, + 0.0035559022799134254, + 0.010295174084603786, + 0.00736875319853425, + 0.010404366999864578, + 0.00896158441901207, + 0.012361290864646435, + 0.008252589963376522, + 0.008589478209614754, + 0.01469691377133131, + 0.004240224603563547, + 0.004678709898144007, + 0.004306395538151264, + 0.005776916164904833, + 0.004373692441731691, + 0.004398879129439592, + 0.003717161249369383, + 0.003924645017832518, + 0.0027320447843521833, + 0.004284445196390152, + 0.0032473024912178516, + 0.002029712311923504, + 0.0027188463136553764, + 0.004197091329842806, + 0.003364834003150463, + 0.0031099740881472826, + 0.0033803980331867933, + 0.00369091983884573, + 0.0027922135777771473, + 0.003260332625359297, + 0.003394098486751318, + 0.0023199499119073153, + 0.0020098828244954348, + 0.0030766373965889215, + 0.025729598477482796, + 0.04339900240302086, + 0.02609204314649105, + 0.03607403486967087, + 0.016602037474513054, + 0.01776723377406597, + 0.013965181075036526, + 0.013068725354969501 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 9, + 0 + ], + "selected_reward": -0.03418448939919472 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene an empty indoor basketball court with a polished hardwood floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the player wearing a vibrant purple basketball jersey with gold trim on the collar and stripes on the shorts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the player execute a swift low dribble and sprint toward the basket?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the shadowed, empty spectator stands visible in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera remain perfectly still throughout the 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": [ + -0.012551834806799889 + ], + "opposite_logprob_tokens": [ + -4.387551784515381 + ], + "logprob_reward": -0.012551834806799889, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012551834806799889, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1591663658618927, + 0.05050713196396828, + 0.3090149462223053, + 0.11246537417173386, + 0.026154551655054092, + 0.06134543567895889, + 0.031221801415085793, + 0.0706104263663292, + 0.03414168581366539, + 0.14537228643894196 + ], + "credit_map": [ + 0.017716266214847565, + 0.02122625894844532, + 0.019025294110178947, + 0.019299356266856194, + 0.016132626682519913, + 0.018822839483618736, + 0.019393444061279297, + 0.027550285682082176, + 0.004713824484497309, + 0.003949709702283144, + 0.009734622202813625, + 0.005573473870754242, + 0.005819833371788263, + 0.00854400172829628, + 0.005538673140108585, + 0.006632991135120392, + 0.034759096801280975, + 0.030679650604724884, + 0.03904266282916069, + 0.030280711129307747, + 0.03094460628926754, + 0.05788222327828407, + 0.0382903628051281, + 0.047135621309280396, + 0.011724710464477539, + 0.012206241488456726, + 0.013374480418860912, + 0.013630508445203304, + 0.011666976846754551, + 0.013772601261734962, + 0.017685124650597572, + 0.018404725939035416, + 0.002808167366310954, + 0.0022666670847684145, + 0.003924703691154718, + 0.002866018796339631, + 0.004512098617851734, + 0.0034905693028122187, + 0.0034119985066354275, + 0.0028743266593664885, + 0.009729944169521332, + 0.005553225986659527, + 0.007216433994472027, + 0.006382863502949476, + 0.007212971802800894, + 0.010101149789988995, + 0.005873553454875946, + 0.009275286458432674, + 0.003291201777756214, + 0.003338470356538892, + 0.003115430474281311, + 0.004231730010360479, + 0.005404592026025057, + 0.00347396032884717, + 0.005659374408423901, + 0.0027070422656834126, + 0.0057752300053834915, + 0.005570481065660715, + 0.011278809048235416, + 0.010353217832744122, + 0.007826108485460281, + 0.007645612582564354, + 0.013374741189181805, + 0.008786221966147423, + 0.0035067317076027393, + 0.005198014434427023, + 0.0035888494458049536, + 0.005647904705256224, + 0.005586995277553797, + 0.00429494958370924, + 0.003198874182999134, + 0.003119365544989705, + 0.020158879458904266, + 0.030353957787156105, + 0.018961140885949135, + 0.026597198098897934, + 0.012081078253686428, + 0.014936202205717564, + 0.010470114648342133, + 0.011813707649707794 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 9 + ], + "selected_reward": -0.012551834806799889 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene an empty indoor basketball court with a polished hardwood floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the player wearing a vibrant purple basketball jersey with gold trim on the collar and stripes on the shorts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the player execute a swift low dribble and sprint toward the basket?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the shadowed, empty spectator stands visible in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera remain perfectly still throughout the 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": [ + -0.05494154617190361 + ], + "opposite_logprob_tokens": [ + -2.9299416542053223 + ], + "logprob_reward": -0.05494154617190361, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05494154617190361, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15091116726398468, + 0.13430161774158478, + 0.13570329546928406, + 0.04263580963015556, + 0.12733960151672363, + 0.049155279994010925, + 0.05200283229351044, + 0.0536343976855278, + 0.11141804605722427, + 0.14289794862270355 + ], + "credit_map": [ + 0.017310550436377525, + 0.01728222891688347, + 0.020465560257434845, + 0.01667867787182331, + 0.014907524921000004, + 0.021512329578399658, + 0.014868860132992268, + 0.027885431423783302, + 0.01622910611331463, + 0.011488157324492931, + 0.02098534069955349, + 0.011549064889550209, + 0.014614379033446312, + 0.029089346528053284, + 0.01616126485168934, + 0.014184956438839436, + 0.014407899230718613, + 0.011605736799538136, + 0.019014284014701843, + 0.012683387845754623, + 0.01337755098938942, + 0.02409667707979679, + 0.017160683870315552, + 0.023357069119811058, + 0.004234327003359795, + 0.00355209456756711, + 0.0058099632151424885, + 0.004388747271150351, + 0.0045095025561749935, + 0.0039504836313426495, + 0.00712955929338932, + 0.009061134420335293, + 0.014745447784662247, + 0.01111255120486021, + 0.017265040427446365, + 0.01759982854127884, + 0.018060600385069847, + 0.016270162537693977, + 0.018793217837810516, + 0.013492756523191929, + 0.004694676958024502, + 0.004858163185417652, + 0.007468640338629484, + 0.004976852796971798, + 0.0067941732704639435, + 0.006497220601886511, + 0.004859532229602337, + 0.009006020613014698, + 0.005099928006529808, + 0.005115365609526634, + 0.004777596797794104, + 0.00568803446367383, + 0.011318192817270756, + 0.006192806176841259, + 0.009557261131703854, + 0.004253647290170193, + 0.005029974039644003, + 0.00392950652167201, + 0.010875661857426167, + 0.00712719326838851, + 0.004571618977934122, + 0.006999612785875797, + 0.006469393614679575, + 0.008631434291601181, + 0.010601087473332882, + 0.012113903649151325, + 0.009567102417349815, + 0.012923363596200943, + 0.027042077854275703, + 0.014059433713555336, + 0.016086386516690254, + 0.009024694561958313, + 0.018251977860927582, + 0.02720760740339756, + 0.017076680436730385, + 0.023343771696090698, + 0.01598670706152916, + 0.015513624995946884, + 0.014182448387145996, + 0.011335134506225586 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 2 + ], + "selected_reward": -0.05494154617190361 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene an empty indoor basketball court with a polished hardwood floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the player wearing a vibrant purple basketball jersey with gold trim on the collar and stripes on the shorts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the player execute a swift low dribble and sprint toward the basket?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the shadowed, empty spectator stands visible in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera remain perfectly still throughout the 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": [ + -0.00525094848126173 + ], + "opposite_logprob_tokens": [ + -5.255250930786133 + ], + "logprob_reward": -0.00525094848126173, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.00525094848126173, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.129705548286438, + 0.19162291288375854, + 0.15144315361976624, + 0.022903159260749817, + 0.04402303323149681, + 0.04915477707982063, + 0.044171568006277084, + 0.03590592369437218, + 0.14588060975074768, + 0.18518929183483124 + ], + "credit_map": [ + 0.012555546127259731, + 0.012043529190123081, + 0.020101971924304962, + 0.012264631688594818, + 0.011337488889694214, + 0.015112640336155891, + 0.01793907769024372, + 0.02835066430270672, + 0.01504121907055378, + 0.01512086484581232, + 0.04923931881785393, + 0.012710114009678364, + 0.02165130339562893, + 0.026122141629457474, + 0.021631814539432526, + 0.030106130987405777, + 0.015922708436846733, + 0.014696438796818256, + 0.0202915221452713, + 0.013062406331300735, + 0.016558974981307983, + 0.029994899407029152, + 0.016108790412545204, + 0.024807414039969444, + 0.00322732119821012, + 0.001998741878196597, + 0.0027659013867378235, + 0.0031562699005007744, + 0.002100834157317877, + 0.003792926436290145, + 0.002556610619649291, + 0.0033045546151697636, + 0.0033218595199286938, + 0.0024479450657963753, + 0.013561395928263664, + 0.005271346308290958, + 0.00383415212854743, + 0.003545044921338558, + 0.0034161722287535667, + 0.008625117130577564, + 0.005032767076045275, + 0.004581877030432224, + 0.006241182796657085, + 0.0055738938972353935, + 0.008005959913134575, + 0.005864291917532682, + 0.006161208264529705, + 0.0076935915276408195, + 0.004439907148480415, + 0.006357559468597174, + 0.00552379572764039, + 0.007930709049105644, + 0.004596057813614607, + 0.0075996192172169685, + 0.003932134248316288, + 0.003791784169152379, + 0.003458780935034156, + 0.004065986257046461, + 0.007546457462012768, + 0.00514037162065506, + 0.0031356271356344223, + 0.004153217189013958, + 0.002991229062899947, + 0.005414254497736692, + 0.013670396991074085, + 0.02516191639006138, + 0.019460996612906456, + 0.030145378783345222, + 0.015239561907947063, + 0.015063428319990635, + 0.010374634526669979, + 0.01676429808139801, + 0.019360454753041267, + 0.02526000887155533, + 0.0246135126799345, + 0.03921211138367653, + 0.015294179320335388, + 0.02513071335852146, + 0.016967996954917908, + 0.019350318238139153 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 9, + 2 + ], + "selected_reward": -0.00525094848126173 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene an empty indoor basketball court with a polished hardwood floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the player wearing a vibrant purple basketball jersey with gold trim on the collar and stripes on the shorts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the player execute a swift low dribble and sprint toward the basket?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the shadowed, empty spectator stands visible in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera remain perfectly still throughout the 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": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.5020840167999268 + ], + "opposite_logprob_tokens": [ + -0.25208404660224915 + ], + "logprob_reward": -1.5020840167999268, + "yes_no_margin_reward": -1.25, + "token_credit_reward": -1.5020840167999268, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.11671019345521927, + 0.10382011532783508, + 0.19512146711349487, + 0.09006434679031372, + 0.160831019282341, + 0.1262091100215912, + 0.019993232563138008, + 0.058684442192316055, + 0.022828320041298866, + 0.10573773831129074 + ], + "credit_map": [ + 0.012580630369484425, + 0.013265848159790039, + 0.01298216637223959, + 0.01697123609483242, + 0.012899812310934067, + 0.016534460708498955, + 0.014802821911871433, + 0.016673218458890915, + 0.01080277655273676, + 0.008869082666933537, + 0.017173541709780693, + 0.01080650370568037, + 0.012135686352849007, + 0.01858619786798954, + 0.01304191816598177, + 0.012404419481754303, + 0.022182680666446686, + 0.018456069752573967, + 0.02513798512518406, + 0.018647443503141403, + 0.01999681070446968, + 0.038985319435596466, + 0.02185833267867565, + 0.029856830835342407, + 0.013132761232554913, + 0.007869365625083447, + 0.012996804900467396, + 0.008971882052719593, + 0.006171355489641428, + 0.015543579123914242, + 0.008746071718633175, + 0.016632528975605965, + 0.0158152487128973, + 0.01327962800860405, + 0.033659473061561584, + 0.01610022410750389, + 0.025260740891098976, + 0.018400097265839577, + 0.017430009320378304, + 0.020885605365037918, + 0.015564199537038803, + 0.014209272339940071, + 0.01735043153166771, + 0.01163418311625719, + 0.014181469567120075, + 0.019504263997077942, + 0.013490318320691586, + 0.020274989306926727, + 0.0017582933651283383, + 0.0020028476137667894, + 0.00269895326346159, + 0.002929670037701726, + 0.002244852716103196, + 0.0025218671653419733, + 0.003064890392124653, + 0.002771857427433133, + 0.005204721353948116, + 0.005833141505718231, + 0.007800595834851265, + 0.010860316455364227, + 0.005269853863865137, + 0.007696567103266716, + 0.008363116532564163, + 0.007656136993318796, + 0.002484483178704977, + 0.0027127377688884735, + 0.002310454612597823, + 0.0025372360832989216, + 0.003668476128950715, + 0.0024840040132403374, + 0.004015867132693529, + 0.002615061355754733, + 0.012160168029367924, + 0.01635311171412468, + 0.014139712788164616, + 0.020697370171546936, + 0.01171842310577631, + 0.010429805144667625, + 0.010817690752446651, + 0.009421462193131447 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 4, + 5 + ], + "selected_reward": -1.5020840167999268 + } + ], + "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/0000021/group_15/index_tensor([324, 324])_group_15_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000021/group_15/index_tensor([324, 324])_group_15_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..fbdc0eb9034a627da93887a13774f537c7aa6d62 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000021/group_15/index_tensor([324, 324])_group_15_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "An extreme close-up shot from a static, slightly elevated angle captures the dramatic moment when a raw egg meets a scorching surface. The camera is tightly focused on a dark, matte-textured cast iron skillet, its black surface coated with a shimmering layer of transparent, hot oil. A freshly cracked egg drops into the frame, and the pan hisses as its clear, sticky egg white makes contact. The transformation is instantaneous and captivating: the translucent white liquid instantly and violently coagulates into a solid, opaque white that spreads rapidly outward. The thinnest part of the perimeter cooks fastest, forming a delicate, lacy edge that crisps and curls to a perfect golden color. In the center of this rapidly solidifying base, the vibrant, orb of orange-yellow yolk drops with a gentle sway, its smooth surface catching bright, warm highlights, creating a culinary scene of focused concentration and intense satisfaction.", + "seed": 23321, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000021/group_15/index_324_seed_23321_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": -0.3652962000342086, + "vlm_logprob": -0.3652962000342086, + "vlm_yes_no_margin": 3.175, + "vlm_token_credit": -0.3652962000342086, + "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 of a dark, matte-textured cast iron skillet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the skillet surface coated with a shimmering layer of transparent, hot oil?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a raw egg drop into the frame and hiss upon contact with the hot surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the egg white instantly coagulate into a solid, opaque white that spreads outward with a golden, lacy, curled edge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a vibrant orange-yellow yolk visible in the center with a smooth surface catching bright, warm highlights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.3902031481266022 + ], + "opposite_logprob_tokens": [ + -1.1402031183242798 + ], + "logprob_reward": -0.3902031481266022, + "yes_no_margin_reward": 0.75, + "token_credit_reward": -0.3902031481266022, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3256663382053375, + 0.04301460459828377, + 0.12642180919647217, + 0.053419630974531174, + 0.021523654460906982, + 0.09934528172016144, + 0.20851248502731323, + 0.042100898921489716, + 0.061189208179712296, + 0.018806153908371925 + ], + "credit_map": [ + 0.05578204616904259, + 0.02717765048146248, + 0.03248396888375282, + 0.030849643051624298, + 0.06175410747528076, + 0.027785927057266235, + 0.05383291468024254, + 0.036000050604343414, + 0.0056916046887636185, + 0.00474544009193778, + 0.004774153232574463, + 0.005764076951891184, + 0.0050719426944851875, + 0.005701844114810228, + 0.004865943919867277, + 0.0063995919190347195, + 0.025047533214092255, + 0.008597936481237411, + 0.012882361188530922, + 0.0322025902569294, + 0.00831407681107521, + 0.011808660812675953, + 0.01572011597454548, + 0.01184852421283722, + 0.008075911551713943, + 0.0042546531185507774, + 0.005363864824175835, + 0.010524648241698742, + 0.0035996593069285154, + 0.007458044681698084, + 0.008054852485656738, + 0.006087993737310171, + 0.0027804423589259386, + 0.0026092110201716423, + 0.0032521113753318787, + 0.004080959130078554, + 0.0017250815872102976, + 0.002062227576971054, + 0.002883179346099496, + 0.0021304399706423283, + 0.014650912955403328, + 0.01374999713152647, + 0.016473716124892235, + 0.011696076020598412, + 0.008405057713389397, + 0.00634050415828824, + 0.01651621237397194, + 0.011512801051139832, + 0.03522290661931038, + 0.019011395052075386, + 0.02134971134364605, + 0.03204246610403061, + 0.015588154084980488, + 0.02932671085000038, + 0.02894236519932747, + 0.027028746902942657, + 0.005556098185479641, + 0.006965752225369215, + 0.005379076115787029, + 0.005367189645767212, + 0.00427299365401268, + 0.002956083742901683, + 0.0077953943982720375, + 0.003808306297287345, + 0.007100428454577923, + 0.004002653993666172, + 0.012221792712807655, + 0.015014426782727242, + 0.004760637879371643, + 0.0062016332522034645, + 0.0068352241069078445, + 0.005052404478192329, + 0.0026347448583692312, + 0.0014571546344086528, + 0.0021738281939178705, + 0.0028585612308233976, + 0.0018523361068218946, + 0.001630031387321651, + 0.0034576107282191515, + 0.002741884673014283 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 2 + ], + "selected_reward": -0.3902031481266022 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 dark, matte-textured cast iron skillet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the skillet surface coated with a shimmering layer of transparent, hot oil?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a raw egg drop into the frame and hiss upon contact with the hot surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the egg white instantly coagulate into a solid, opaque white that spreads outward with a golden, lacy, curled edge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a vibrant orange-yellow yolk visible in the center with a smooth surface catching bright, warm highlights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020625632256269455 + ], + "opposite_logprob_tokens": [ + -3.895625591278076 + ], + "logprob_reward": -0.020625632256269455, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020625632256269455, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.6628310084342957, + 0.08037002384662628, + 0.04811433330178261, + 0.0318947471678257, + 0.0226975716650486, + 0.024155905470252037, + 0.054767828434705734, + 0.022994352504611015, + 0.02770584262907505, + 0.024468358606100082 + ], + "credit_map": [ + 0.1078687384724617, + 0.057013094425201416, + 0.08450458198785782, + 0.06228889152407646, + 0.11702130734920502, + 0.04368215054273605, + 0.10446365177631378, + 0.08598864078521729, + 0.010514041408896446, + 0.005987229757010937, + 0.007445831783115864, + 0.015017460100352764, + 0.009051143191754818, + 0.00660906545817852, + 0.010996454395353794, + 0.014748805202543736, + 0.010372675023972988, + 0.005257606040686369, + 0.006039918400347233, + 0.006583376321941614, + 0.005194593220949173, + 0.004031188786029816, + 0.006493041757494211, + 0.0041419388726353645, + 0.007351515349000692, + 0.004395655356347561, + 0.004622333217412233, + 0.003328392980620265, + 0.002966222818940878, + 0.0026476928032934666, + 0.003680747700855136, + 0.0029021897353231907, + 0.0031132285948842764, + 0.0019644880667328835, + 0.004685579799115658, + 0.004245598800480366, + 0.0016673712525516748, + 0.0015204026130959392, + 0.003527193795889616, + 0.001973710721358657, + 0.005033009219914675, + 0.0024871344212442636, + 0.0030302582308650017, + 0.0029209181666374207, + 0.002335436176508665, + 0.0014946965966373682, + 0.00456745782867074, + 0.0022869957610964775, + 0.011977054178714752, + 0.004484748933464289, + 0.006117368582636118, + 0.008443119004368782, + 0.003465237095952034, + 0.007345541380345821, + 0.006580280605703592, + 0.006354481913149357, + 0.0027843713760375977, + 0.003560104640200734, + 0.0025301792193204165, + 0.0033050335478037596, + 0.002249352168291807, + 0.001567290397360921, + 0.0046641225926578045, + 0.0023338994942605495, + 0.00267613073810935, + 0.0020632343366742134, + 0.004893961828202009, + 0.005040645599365234, + 0.002658158540725708, + 0.002415710361674428, + 0.004681974183768034, + 0.0032760314643383026, + 0.0024586045183241367, + 0.0017975104274228215, + 0.00515397172421217, + 0.00498330919072032, + 0.0021006157621741295, + 0.001583529869094491, + 0.003269126871600747, + 0.003121693152934313 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 6 + ], + "selected_reward": -0.020625632256269455 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 dark, matte-textured cast iron skillet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the skillet surface coated with a shimmering layer of transparent, hot oil?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a raw egg drop into the frame and hiss upon contact with the hot surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the egg white instantly coagulate into a solid, opaque white that spreads outward with a golden, lacy, curled edge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a vibrant orange-yellow yolk visible in the center with a smooth surface catching bright, warm highlights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.4064891338348389 + ], + "opposite_logprob_tokens": [ + -0.28148913383483887 + ], + "logprob_reward": -1.4064891338348389, + "yes_no_margin_reward": -1.125, + "token_credit_reward": -1.4064891338348389, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.3557491600513458, + 0.059294458478689194, + 0.11920709162950516, + 0.039251770824193954, + 0.035131100565195084, + 0.0207112655043602, + 0.09327033162117004, + 0.036546576768159866, + 0.20791713893413544, + 0.032921142876148224 + ], + "credit_map": [ + 0.04957117885351181, + 0.03326389193534851, + 0.030552351847290993, + 0.04220994561910629, + 0.06344558298587799, + 0.01975243166089058, + 0.07554130256175995, + 0.04141245037317276, + 0.00603276863694191, + 0.003318768460303545, + 0.004970313981175423, + 0.011134011670947075, + 0.005285972263664007, + 0.013287524692714214, + 0.007295429240912199, + 0.007969667203724384, + 0.021383047103881836, + 0.013802092522382736, + 0.008372529409825802, + 0.02296161651611328, + 0.015492470003664494, + 0.009931519627571106, + 0.014817220158874989, + 0.012446587905287743, + 0.005521008279174566, + 0.003865137929096818, + 0.0027782716788351536, + 0.008710501715540886, + 0.003350174054503441, + 0.00413996446877718, + 0.006052698474377394, + 0.004834010265767574, + 0.004971303511410952, + 0.0036272311117500067, + 0.004545975476503372, + 0.004427928011864424, + 0.002992442110553384, + 0.004459424410015345, + 0.005480569787323475, + 0.00462622195482254, + 0.0019737198017537594, + 0.0019190944731235504, + 0.0030429770704358816, + 0.0031224756967276335, + 0.0017710073152557015, + 0.002207698067650199, + 0.004100641701370478, + 0.0025736491661518812, + 0.014042153023183346, + 0.007538059260696173, + 0.009076764807105064, + 0.014263754710555077, + 0.0066676330752670765, + 0.014989553950726986, + 0.01407476793974638, + 0.01261763833463192, + 0.004899932537227869, + 0.003961157985031605, + 0.0038800875190645456, + 0.006069735158234835, + 0.0029176489915698767, + 0.0044515118934214115, + 0.0055731236934661865, + 0.004793373867869377, + 0.01618621125817299, + 0.013084803707897663, + 0.050511717796325684, + 0.06075291708111763, + 0.015800541266798973, + 0.016854247078299522, + 0.023096103221178055, + 0.011630588211119175, + 0.0034899390302598476, + 0.004209218546748161, + 0.004118543118238449, + 0.0040736752562224865, + 0.0039309486746788025, + 0.0026767749805003405, + 0.005785538349300623, + 0.004636500496417284 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 2 + ], + "selected_reward": -1.4064891338348389 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 dark, matte-textured cast iron skillet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the skillet surface coated with a shimmering layer of transparent, hot oil?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a raw egg drop into the frame and hiss upon contact with the hot surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the egg white instantly coagulate into a solid, opaque white that spreads outward with a golden, lacy, curled edge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a vibrant orange-yellow yolk visible in the center with a smooth surface catching bright, warm highlights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008659076876938343 + ], + "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": [ + 0.2910134494304657, + 0.06581485271453857, + 0.09459370374679565, + 0.0399523489177227, + 0.04137907922267914, + 0.028253424912691116, + 0.12229771167039871, + 0.05795523151755333, + 0.2036089450120926, + 0.05513130873441696 + ], + "credit_map": [ + 0.04072713479399681, + 0.024813733994960785, + 0.026291433721780777, + 0.03365303575992584, + 0.05237598344683647, + 0.017620539292693138, + 0.06085239350795746, + 0.03467916324734688, + 0.006564129609614611, + 0.0038050133734941483, + 0.00453942222520709, + 0.011765100993216038, + 0.005835263058543205, + 0.01895678974688053, + 0.007651668041944504, + 0.006697465665638447, + 0.015434946864843369, + 0.00943455845117569, + 0.006334237288683653, + 0.021682435646653175, + 0.01153784990310669, + 0.008107554167509079, + 0.012508304789662361, + 0.009553801268339157, + 0.00516910245642066, + 0.004072893876582384, + 0.0038457734044641256, + 0.006330462172627449, + 0.0035163897555321455, + 0.004262836650013924, + 0.007382962852716446, + 0.005371922627091408, + 0.005993500351905823, + 0.0037624062970280647, + 0.006936935242265463, + 0.005944598466157913, + 0.0031144849490374327, + 0.003978644497692585, + 0.006683839950710535, + 0.00496467063203454, + 0.0045654126442968845, + 0.0019082118524238467, + 0.002893732627853751, + 0.003689592005684972, + 0.0024803774431347847, + 0.0025202762335538864, + 0.006549688056111336, + 0.003646133467555046, + 0.01979401521384716, + 0.010255396366119385, + 0.012399540282785892, + 0.018412725999951363, + 0.008472619578242302, + 0.018993040546774864, + 0.01703481748700142, + 0.016935555264353752, + 0.009505616500973701, + 0.004276431165635586, + 0.006456736940890551, + 0.009310624562203884, + 0.005668605677783489, + 0.004534171428531408, + 0.011384624056518078, + 0.006818415131419897, + 0.016135133802890778, + 0.012837455607950687, + 0.04965614899992943, + 0.0596267469227314, + 0.01531598623842001, + 0.01572599448263645, + 0.022695932537317276, + 0.011615531519055367, + 0.004407444968819618, + 0.0033058742992579937, + 0.01017800997942686, + 0.010866689495742321, + 0.005624261684715748, + 0.004452849272638559, + 0.008772876113653183, + 0.007523292675614357 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 6 + ], + "selected_reward": -0.008659076876938343 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 dark, matte-textured cast iron skillet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the skillet surface coated with a shimmering layer of transparent, hot oil?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a raw egg drop into the frame and hiss upon contact with the hot surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the egg white instantly coagulate into a solid, opaque white that spreads outward with a golden, lacy, curled edge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a vibrant orange-yellow yolk visible in the center with a smooth surface catching bright, warm highlights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0005040090763941407 + ], + "opposite_logprob_tokens": [ + -7.625504016876221 + ], + "logprob_reward": -0.0005040090763941407, + "yes_no_margin_reward": 7.625, + "token_credit_reward": -0.0005040090763941407, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.357058048248291, + 0.11065967381000519, + 0.05998455733060837, + 0.051114194095134735, + 0.04476174712181091, + 0.03489499166607857, + 0.09070926159620285, + 0.038693614304065704, + 0.14687494933605194, + 0.06524891406297684 + ], + "credit_map": [ + 0.0524718314409256, + 0.02976720966398716, + 0.029230695217847824, + 0.0546003133058548, + 0.06000954285264015, + 0.017583271488547325, + 0.0758872926235199, + 0.03750792518258095, + 0.0125268017873168, + 0.007012651301920414, + 0.009029190056025982, + 0.0234628077596426, + 0.011213268153369427, + 0.01636654883623123, + 0.012693790718913078, + 0.018354617059230804, + 0.00842223409563303, + 0.007378618232905865, + 0.005492428783327341, + 0.008368280716240406, + 0.00843755155801773, + 0.006667123641818762, + 0.007554367650300264, + 0.007663951255381107, + 0.0071827140636742115, + 0.0037956745363771915, + 0.003976296167820692, + 0.007728240452706814, + 0.004414759110659361, + 0.005293501075357199, + 0.011258313432335854, + 0.007464702241122723, + 0.005365834571421146, + 0.003151960438117385, + 0.006610362324863672, + 0.009679208509624004, + 0.003791903844103217, + 0.002525700954720378, + 0.009027747437357903, + 0.004609033465385437, + 0.0035974043421447277, + 0.0026994848158210516, + 0.00269308197312057, + 0.005222201347351074, + 0.0038265399634838104, + 0.003621702780947089, + 0.0097915418446064, + 0.0034430394880473614, + 0.013633684255182743, + 0.007239582948386669, + 0.0086824307218194, + 0.013133615255355835, + 0.006539295427501202, + 0.015014853328466415, + 0.013841868378221989, + 0.012623943388462067, + 0.0040340060368180275, + 0.003784989472478628, + 0.0049327039159834385, + 0.006897487211972475, + 0.0033344123512506485, + 0.0031554216984659433, + 0.0074803391471505165, + 0.005074257496744394, + 0.011031749658286572, + 0.007655079010874033, + 0.037770405411720276, + 0.04588064178824425, + 0.00978938676416874, + 0.009524324908852577, + 0.016808325424790382, + 0.008415059186518192, + 0.006266554817557335, + 0.005737231578677893, + 0.010127880610525608, + 0.012527893297374249, + 0.00727807218208909, + 0.005048802588135004, + 0.01092365849763155, + 0.0073388307355344296 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 1 + ], + "selected_reward": -0.0005040090763941407 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "No", + "Yes", + "Yes" + ] + }, + { + "prompt": "An extreme close-up shot from a static, slightly elevated angle captures the dramatic moment when a raw egg meets a scorching surface. The camera is tightly focused on a dark, matte-textured cast iron skillet, its black surface coated with a shimmering layer of transparent, hot oil. A freshly cracked egg drops into the frame, and the pan hisses as its clear, sticky egg white makes contact. The transformation is instantaneous and captivating: the translucent white liquid instantly and violently coagulates into a solid, opaque white that spreads rapidly outward. The thinnest part of the perimeter cooks fastest, forming a delicate, lacy edge that crisps and curls to a perfect golden color. In the center of this rapidly solidifying base, the vibrant, orb of orange-yellow yolk drops with a gentle sway, its smooth surface catching bright, warm highlights, creating a culinary scene of focused concentration and intense satisfaction.", + "seed": 23322, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000021/group_15/index_324_seed_23322_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": -0.4229741308488883, + "vlm_logprob": -0.4229741308488883, + "vlm_yes_no_margin": 3.1999999642372132, + "vlm_token_credit": -0.4229741308488883, + "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 of a dark, matte-textured cast iron skillet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the skillet surface coated with a shimmering layer of transparent, hot oil?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a raw egg drop into the frame and hiss upon contact with the hot surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the egg white instantly coagulate into a solid, opaque white that spreads outward with a golden, lacy, curled edge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a vibrant orange-yellow yolk visible in the center with a smooth surface catching bright, warm highlights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.38913530111312866 + ], + "opposite_logprob_tokens": [ + -1.1391353607177734 + ], + "logprob_reward": -0.38913530111312866, + "yes_no_margin_reward": 0.7500000596046448, + "token_credit_reward": -0.38913530111312866, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21148596704006195, + 0.07945200055837631, + 0.02089342102408409, + 0.06454131752252579, + 0.15291957557201385, + 0.10151524096727371, + 0.1309264451265335, + 0.11751861870288849, + 0.0758960098028183, + 0.04485132545232773 + ], + "credit_map": [ + 0.03561720997095108, + 0.021586274728178978, + 0.028145644813776016, + 0.022307783365249634, + 0.02133105881512165, + 0.03935106843709946, + 0.021679705008864403, + 0.021467234939336777, + 0.012544061057269573, + 0.006788855884224176, + 0.013649962842464447, + 0.006548506673425436, + 0.0072794584557414055, + 0.017818493768572807, + 0.005335106980055571, + 0.009487553499639034, + 0.0026002998929470778, + 0.0019135812763124704, + 0.004416292533278465, + 0.002163094002753496, + 0.0021081336308270693, + 0.0030785517301410437, + 0.0022587536368519068, + 0.0023547159507870674, + 0.007135548163205385, + 0.011378686875104904, + 0.012688823975622654, + 0.007609482388943434, + 0.008174112066626549, + 0.005146764684468508, + 0.005653804633766413, + 0.006754096131771803, + 0.01826784759759903, + 0.012224211357533932, + 0.017392203211784363, + 0.03035200759768486, + 0.009992905892431736, + 0.039806656539440155, + 0.012224704027175903, + 0.012659051455557346, + 0.01420549675822258, + 0.010149513371288776, + 0.022349396720528603, + 0.011803962290287018, + 0.008176375180482864, + 0.010896102525293827, + 0.007575641851872206, + 0.016358764842152596, + 0.010906706564128399, + 0.04469110071659088, + 0.006425066851079464, + 0.042090971022844315, + 0.005194217897951603, + 0.010261951945722103, + 0.006219183560460806, + 0.005137262400239706, + 0.008782174438238144, + 0.03581412136554718, + 0.010161299258470535, + 0.032911136746406555, + 0.0066893636249005795, + 0.004729081876575947, + 0.008115877397358418, + 0.010315579362213612, + 0.0075613814406096935, + 0.013030270114541054, + 0.00815792940557003, + 0.01674884557723999, + 0.006530617829412222, + 0.005732335615903139, + 0.006952241528779268, + 0.011182396672666073, + 0.00785946287214756, + 0.005703633651137352, + 0.0035846703685820103, + 0.006232272367924452, + 0.004586093593388796, + 0.004778169561177492, + 0.0038250640500336885, + 0.008281963877379894 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 6 + ], + "selected_reward": -0.38913530111312866 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 dark, matte-textured cast iron skillet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the skillet surface coated with a shimmering layer of transparent, hot oil?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a raw egg drop into the frame and hiss upon contact with the hot surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the egg white instantly coagulate into a solid, opaque white that spreads outward with a golden, lacy, curled edge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a vibrant orange-yellow yolk visible in the center with a smooth surface catching bright, warm highlights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01821727119386196 + ], + "opposite_logprob_tokens": [ + -4.018217086791992 + ], + "logprob_reward": -0.01821727119386196, + "yes_no_margin_reward": 3.999999761581421, + "token_credit_reward": -0.01821727119386196, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24980506300926208, + 0.03889806941151619, + 0.033970870077610016, + 0.030511820688843727, + 0.057521235197782516, + 0.03465866297483444, + 0.49053069949150085, + 0.028717534616589546, + 0.012620287016034126, + 0.022765807807445526 + ], + "credit_map": [ + 0.04828115925192833, + 0.03459225222468376, + 0.022677451372146606, + 0.018664365634322166, + 0.030782148241996765, + 0.06519675999879837, + 0.017482049763202667, + 0.01212884671986103, + 0.005499263294041157, + 0.004082062281668186, + 0.007129164878278971, + 0.0033729923889040947, + 0.004252877086400986, + 0.00713776983320713, + 0.0036177970468997955, + 0.003806138876825571, + 0.004694295581430197, + 0.002384118502959609, + 0.0057027884759008884, + 0.004867148119956255, + 0.0026392198633402586, + 0.009060491807758808, + 0.002331696217879653, + 0.002291104756295681, + 0.0037994838785380125, + 0.0025892897974699736, + 0.0032488126307725906, + 0.006970186252146959, + 0.003707379102706909, + 0.005141310393810272, + 0.0026416366454213858, + 0.002413721987977624, + 0.006438040174543858, + 0.010773700661957264, + 0.0040880655869841576, + 0.015437130816280842, + 0.003180566243827343, + 0.00763150816783309, + 0.00429571233689785, + 0.005676510278135538, + 0.006015555467456579, + 0.004002378322184086, + 0.003985395655035973, + 0.0033183391205966473, + 0.003631106112152338, + 0.004508215468376875, + 0.00304673844948411, + 0.006150930188596249, + 0.0434778593480587, + 0.16754350066184998, + 0.01817983016371727, + 0.1584281474351883, + 0.019539741799235344, + 0.04335986450314522, + 0.021643606945872307, + 0.018358120694756508, + 0.0025032192934304476, + 0.006085379980504513, + 0.003358074463903904, + 0.004059800878167152, + 0.00405236566439271, + 0.0018630886916071177, + 0.003405019175261259, + 0.003390583908185363, + 0.0012579293688759208, + 0.002246221760287881, + 0.0011317268945276737, + 0.0021542443428188562, + 0.0013476986205205321, + 0.0013308388879522681, + 0.0014722748892381787, + 0.001679350738413632, + 0.004241643939167261, + 0.0033729802817106247, + 0.002631287556141615, + 0.004439686890691519, + 0.0019013747805729508, + 0.0016936870524659753, + 0.0019844917114824057, + 0.0025006537325680256 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 4 + ], + "selected_reward": -0.01821727119386196 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 dark, matte-textured cast iron skillet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the skillet surface coated with a shimmering layer of transparent, hot oil?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a raw egg drop into the frame and hiss upon contact with the hot surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the egg white instantly coagulate into a solid, opaque white that spreads outward with a golden, lacy, curled edge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a vibrant orange-yellow yolk visible in the center with a smooth surface catching bright, warm highlights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.7016746997833252 + ], + "opposite_logprob_tokens": [ + -0.2016746997833252 + ], + "logprob_reward": -1.7016746997833252, + "yes_no_margin_reward": -1.5, + "token_credit_reward": -1.7016746997833252, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.39751943945884705, + 0.0968623235821724, + 0.059087108820676804, + 0.0526123121380806, + 0.06657364964485168, + 0.025749897584319115, + 0.10873986780643463, + 0.14595167338848114, + 0.026515483856201172, + 0.02038821205496788 + ], + "credit_map": [ + 0.07229837775230408, + 0.05813809856772423, + 0.03748473525047302, + 0.03872642293572426, + 0.042006392031908035, + 0.09768513590097427, + 0.02820463478565216, + 0.0229756198823452, + 0.014052142389118671, + 0.005862642545253038, + 0.015275525860488415, + 0.007162550929933786, + 0.008612190373241901, + 0.021994303911924362, + 0.010307393036782742, + 0.013595575466752052, + 0.00819856021553278, + 0.0036505882162600756, + 0.007357469759881496, + 0.010550649836659431, + 0.0053204926662147045, + 0.015088165178894997, + 0.004365673754364252, + 0.004555500112473965, + 0.006433494854718447, + 0.0055077350698411465, + 0.009220399893820286, + 0.0064941151067614555, + 0.009677675552666187, + 0.005238853394985199, + 0.004951431415975094, + 0.005088600795716047, + 0.008592713624238968, + 0.014496737159788609, + 0.005720069166272879, + 0.004952852614223957, + 0.004192771390080452, + 0.018437279388308525, + 0.0037043120246380568, + 0.006476911250501871, + 0.004089568741619587, + 0.0031738360412418842, + 0.003950728569179773, + 0.0028867539949715137, + 0.0024111983366310596, + 0.002452227519825101, + 0.002544781193137169, + 0.004240803886204958, + 0.015023335814476013, + 0.011803541332483292, + 0.006198184099048376, + 0.026770073920488358, + 0.005500239785760641, + 0.03187676891684532, + 0.005441367160528898, + 0.006126342341303825, + 0.011041094549000263, + 0.045018039643764496, + 0.007868994027376175, + 0.04076047241687775, + 0.007201125379651785, + 0.007695919368416071, + 0.009588495828211308, + 0.01677751913666725, + 0.002888833172619343, + 0.004295684397220612, + 0.003151235869154334, + 0.004639703314751387, + 0.0027956299018114805, + 0.0024941125884652138, + 0.003209848888218403, + 0.0030404336284846067, + 0.0029123954009264708, + 0.002989544766023755, + 0.002331131836399436, + 0.0029000663198530674, + 0.0022234090138226748, + 0.0020306191872805357, + 0.0025176662020385265, + 0.002483377465978265 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 6 + ], + "selected_reward": -1.7016746997833252 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 dark, matte-textured cast iron skillet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the skillet surface coated with a shimmering layer of transparent, hot oil?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a raw egg drop into the frame and hiss upon contact with the hot surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the egg white instantly coagulate into a solid, opaque white that spreads outward with a golden, lacy, curled edge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a vibrant orange-yellow yolk visible in the center with a smooth surface catching bright, warm highlights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005272175185382366 + ], + "opposite_logprob_tokens": [ + -5.255272388458252 + ], + "logprob_reward": -0.005272175185382366, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005272175185382366, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19470013678073883, + 0.06822917610406876, + 0.027704229578375816, + 0.039979513734579086, + 0.05695660784840584, + 0.032107677310705185, + 0.4932515323162079, + 0.046754319220781326, + 0.0280279703438282, + 0.012288820929825306 + ], + "credit_map": [ + 0.0309353768825531, + 0.029569802805781364, + 0.02002127654850483, + 0.020119236782193184, + 0.02191026695072651, + 0.04019402340054512, + 0.016947437077760696, + 0.015002721920609474, + 0.010239716619253159, + 0.004766138270497322, + 0.010675166733562946, + 0.0048189652152359486, + 0.005752005614340305, + 0.015731750056147575, + 0.006420470308512449, + 0.009824957698583603, + 0.00358205265365541, + 0.0024830896873027086, + 0.0046446300111711025, + 0.004254784435033798, + 0.0023186069447547197, + 0.005535798612982035, + 0.002075773663818836, + 0.0028094921726733446, + 0.004560903180390596, + 0.004220102913677692, + 0.00636654207482934, + 0.004167130682617426, + 0.007248934358358383, + 0.004360860213637352, + 0.0042451187036931515, + 0.004809919279068708, + 0.0064500137232244015, + 0.0077060069888830185, + 0.003925969358533621, + 0.013825200498104095, + 0.005023925099521875, + 0.01069577131420374, + 0.004749169573187828, + 0.004580553155392408, + 0.0037899171002209187, + 0.0034688690211623907, + 0.004525619093328714, + 0.003504469059407711, + 0.004259866662323475, + 0.0061797178350389, + 0.003712930018082261, + 0.002666291082277894, + 0.044072315096855164, + 0.16687829792499542, + 0.017675647512078285, + 0.16027848422527313, + 0.019597625359892845, + 0.0449124313890934, + 0.021225443109869957, + 0.018611261621117592, + 0.00407825643196702, + 0.012902367860078812, + 0.002429788699373603, + 0.011514564976096153, + 0.0025805530603975058, + 0.003114658407866955, + 0.003314435947686434, + 0.006819696165621281, + 0.0027728606946766376, + 0.0040467227809131145, + 0.0029645985923707485, + 0.004921726416796446, + 0.003171350806951523, + 0.0024743182584643364, + 0.003920454066246748, + 0.003755939891561866, + 0.0017315568402409554, + 0.0014931069454178214, + 0.0013689399929717183, + 0.0022099229972809553, + 0.0013390338281169534, + 0.0011800399515777826, + 0.0013368424260988832, + 0.0016293787630274892 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 1 + ], + "selected_reward": -0.005272175185382366 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 dark, matte-textured cast iron skillet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the skillet surface coated with a shimmering layer of transparent, hot oil?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a raw egg drop into the frame and hiss upon contact with the hot surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the egg white instantly coagulate into a solid, opaque white that spreads outward with a golden, lacy, curled edge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a vibrant orange-yellow yolk visible in the center with a smooth surface catching bright, warm highlights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0005712069687433541 + ], + "opposite_logprob_tokens": [ + -7.500571250915527 + ], + "logprob_reward": -0.0005712069687433541, + "yes_no_margin_reward": 7.5, + "token_credit_reward": -0.0005712069687433541, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3030616044998169, + 0.10811975598335266, + 0.0550033375620842, + 0.09657520055770874, + 0.09832926094532013, + 0.0528532937169075, + 0.06405555456876755, + 0.12171419709920883, + 0.05249032378196716, + 0.04779746010899544 + ], + "credit_map": [ + 0.033969275653362274, + 0.053308725357055664, + 0.03754976764321327, + 0.035642459988594055, + 0.038243941962718964, + 0.04071474075317383, + 0.03543802723288536, + 0.028194645419716835, + 0.0111174825578928, + 0.006889117881655693, + 0.026586584746837616, + 0.012014754116535187, + 0.008896809071302414, + 0.0178085844963789, + 0.013770745135843754, + 0.011035680770874023, + 0.007105827331542969, + 0.004267693962901831, + 0.006698210723698139, + 0.00964027363806963, + 0.004362782929092646, + 0.013619841076433659, + 0.003997622989118099, + 0.005311089102178812, + 0.011627476662397385, + 0.007146620657294989, + 0.009419558569788933, + 0.02286772057414055, + 0.012167786248028278, + 0.017598360776901245, + 0.008162792772054672, + 0.007584886625409126, + 0.012759965844452381, + 0.00891019869595766, + 0.009855001233518124, + 0.01700604520738125, + 0.005318519193679094, + 0.027470041066408157, + 0.0078117093071341515, + 0.009197769686579704, + 0.00791001133620739, + 0.00499696796759963, + 0.008784344419836998, + 0.006693457718938589, + 0.0054180980660021305, + 0.006569798570126295, + 0.004871938843280077, + 0.007608678657561541, + 0.008401867933571339, + 0.004667694680392742, + 0.006161368917673826, + 0.013444941490888596, + 0.004217210691422224, + 0.01793389394879341, + 0.004399747122079134, + 0.0048288325779139996, + 0.010666572488844395, + 0.03506176173686981, + 0.006370127201080322, + 0.03310130909085274, + 0.006504489108920097, + 0.007041521370410919, + 0.008132493123412132, + 0.014835917390882969, + 0.005803687032312155, + 0.008122887462377548, + 0.0056145708076655865, + 0.012539543211460114, + 0.006028991192579269, + 0.0032308134250342846, + 0.005880399607121944, + 0.005269428715109825, + 0.007249875925481319, + 0.006320708431303501, + 0.003309603314846754, + 0.008886167779564857, + 0.004436491057276726, + 0.004451761022210121, + 0.0040379841811954975, + 0.009104868397116661 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 1 + ], + "selected_reward": -0.0005712069687433541 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "No", + "Yes", + "Yes" + ] + }, + { + "prompt": "An extreme close-up shot from a static, slightly elevated angle captures the dramatic moment when a raw egg meets a scorching surface. The camera is tightly focused on a dark, matte-textured cast iron skillet, its black surface coated with a shimmering layer of transparent, hot oil. A freshly cracked egg drops into the frame, and the pan hisses as its clear, sticky egg white makes contact. The transformation is instantaneous and captivating: the translucent white liquid instantly and violently coagulates into a solid, opaque white that spreads rapidly outward. The thinnest part of the perimeter cooks fastest, forming a delicate, lacy edge that crisps and curls to a perfect golden color. In the center of this rapidly solidifying base, the vibrant, orb of orange-yellow yolk drops with a gentle sway, its smooth surface catching bright, warm highlights, creating a culinary scene of focused concentration and intense satisfaction.", + "seed": 23323, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000021/group_15/index_324_seed_23323_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": -0.02439574263407849, + "vlm_logprob": -0.02439574263407849, + "vlm_yes_no_margin": 5.125, + "vlm_token_credit": -0.02439574263407849, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 dark, matte-textured cast iron skillet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the skillet surface coated with a shimmering layer of transparent, hot oil?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a raw egg drop into the frame and hiss upon contact with the hot surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the egg white instantly coagulate into a solid, opaque white that spreads outward with a golden, lacy, curled edge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a vibrant orange-yellow yolk visible in the center with a smooth surface catching bright, warm highlights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.09486544877290726 + ], + "opposite_logprob_tokens": [ + -2.469865560531616 + ], + "logprob_reward": -0.09486544877290726, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.09486544877290726, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16544045507907867, + 0.2594224214553833, + 0.0698913037776947, + 0.10562305897474289, + 0.2071828991174698, + 0.053483746945858, + 0.031786009669303894, + 0.07711398601531982, + 0.010279236361384392, + 0.019776923581957817 + ], + "credit_map": [ + 0.0236586332321167, + 0.017701530829072, + 0.024450205266475677, + 0.019787302240729332, + 0.015756266191601753, + 0.01916886493563652, + 0.017355186864733696, + 0.027562476694583893, + 0.032820191234350204, + 0.015277431346476078, + 0.07897220551967621, + 0.02354171872138977, + 0.03530339151620865, + 0.04103558138012886, + 0.014811152592301369, + 0.017660735175013542, + 0.0074199894443154335, + 0.008034304715692997, + 0.020727723836898804, + 0.004141679964959621, + 0.008942285552620888, + 0.009415289387106895, + 0.006469809915870428, + 0.004740222357213497, + 0.012951018288731575, + 0.008449550718069077, + 0.016613055020570755, + 0.006453182082623243, + 0.02820657193660736, + 0.008061952888965607, + 0.019373372197151184, + 0.005514360964298248, + 0.02724730782210827, + 0.02004632167518139, + 0.05299430340528488, + 0.022615693509578705, + 0.020691895857453346, + 0.018046658486127853, + 0.018468735739588737, + 0.027071982622146606, + 0.007791688200086355, + 0.005993801634758711, + 0.01147510390728712, + 0.005201753694564104, + 0.004839725326746702, + 0.008062760345637798, + 0.004637680482119322, + 0.005481231026351452, + 0.004546497482806444, + 0.0039345212280750275, + 0.004500510636717081, + 0.0034642366226762533, + 0.003229263238608837, + 0.005320409778505564, + 0.0034431361127644777, + 0.0033474371302872896, + 0.011869399808347225, + 0.007277481723576784, + 0.007924480363726616, + 0.005564011633396149, + 0.010467682033777237, + 0.008301601745188236, + 0.00782267190515995, + 0.017886657267808914, + 0.0015002290019765496, + 0.0012294244952499866, + 0.0012630707351490855, + 0.0010759682627394795, + 0.0011465505231171846, + 0.001004158635623753, + 0.001333677675575018, + 0.0017261571483686566, + 0.003007213119417429, + 0.0019316618563607335, + 0.0018866563914343715, + 0.0019704762380570173, + 0.00305015011690557, + 0.00229281815700233, + 0.0028490375261753798, + 0.002788909012451768 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 4, + 0 + ], + "selected_reward": -0.09486544877290726 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 dark, matte-textured cast iron skillet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the skillet surface coated with a shimmering layer of transparent, hot oil?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a raw egg drop into the frame and hiss upon contact with the hot surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the egg white instantly coagulate into a solid, opaque white that spreads outward with a golden, lacy, curled edge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a vibrant orange-yellow yolk visible in the center with a smooth surface catching bright, warm highlights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0028473094571381807 + ], + "opposite_logprob_tokens": [ + -5.877847194671631 + ], + "logprob_reward": -0.0028473094571381807, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.0028473094571381807, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5444514155387878, + 0.09907212108373642, + 0.03034856542944908, + 0.06303805857896805, + 0.06356801837682724, + 0.08159028738737106, + 0.05208265781402588, + 0.0385461151599884, + 0.017505016177892685, + 0.00979776494204998 + ], + "credit_map": [ + 0.07293044030666351, + 0.05528741702437401, + 0.08282449096441269, + 0.06464037299156189, + 0.06159853935241699, + 0.06739188730716705, + 0.05408402532339096, + 0.08569422364234924, + 0.011835789307951927, + 0.00800785981118679, + 0.013682556338608265, + 0.009583810344338417, + 0.013025215826928616, + 0.019031120464205742, + 0.008054648526012897, + 0.01585111953318119, + 0.0033412366174161434, + 0.0020401086658239365, + 0.007662486284971237, + 0.0027514451649039984, + 0.0048193661496043205, + 0.0031754381489008665, + 0.0035408339463174343, + 0.0030176497530192137, + 0.005169925279915333, + 0.002466547768563032, + 0.0064925323240458965, + 0.0023383619263768196, + 0.024134768173098564, + 0.005828520283102989, + 0.013163736090064049, + 0.003443667432293296, + 0.007813898846507072, + 0.008954479359090328, + 0.013980451971292496, + 0.00794207863509655, + 0.00756520451977849, + 0.005235892720520496, + 0.005108687561005354, + 0.006967324763536453, + 0.011194062419235706, + 0.008837489411234856, + 0.016679935157299042, + 0.007906784303486347, + 0.007725871633738279, + 0.013819506391882896, + 0.006712216883897781, + 0.008714420720934868, + 0.00821203738451004, + 0.007020157761871815, + 0.007771460339426994, + 0.006528500001877546, + 0.005883990786969662, + 0.005628922488540411, + 0.004270477220416069, + 0.006767112296074629, + 0.005397025961428881, + 0.003040635958313942, + 0.01139494962990284, + 0.0030189501121640205, + 0.003491462441161275, + 0.003422671230509877, + 0.003032056149095297, + 0.005748362746089697, + 0.003196106292307377, + 0.0015054132090881467, + 0.0012555418070405722, + 0.0016411710530519485, + 0.0018680461216717958, + 0.0018488785717636347, + 0.001574347261339426, + 0.004615513142198324, + 0.0013500602217391133, + 0.0008185103652067482, + 0.0009403300937265158, + 0.0006236583576537669, + 0.0025094093289226294, + 0.0009647061815485358, + 0.001306003425270319, + 0.0012850878993049264 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 5 + ], + "selected_reward": -0.0028473094571381807 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 dark, matte-textured cast iron skillet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the skillet surface coated with a shimmering layer of transparent, hot oil?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a raw egg drop into the frame and hiss upon contact with the hot surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the egg white instantly coagulate into a solid, opaque white that spreads outward with a golden, lacy, curled edge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a vibrant orange-yellow yolk visible in the center with a smooth surface catching bright, warm highlights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020619675517082214 + ], + "opposite_logprob_tokens": [ + -3.8956196308135986 + ], + "logprob_reward": -0.020619675517082214, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020619675517082214, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11073729395866394, + 0.18308360874652863, + 0.04450700059533119, + 0.3374343812465668, + 0.04569971561431885, + 0.09129372239112854, + 0.07664088904857635, + 0.07087194919586182, + 0.009653831832110882, + 0.0300774946808815 + ], + "credit_map": [ + 0.014897428452968597, + 0.012423127889633179, + 0.014618715271353722, + 0.01635250449180603, + 0.01577889360487461, + 0.012043759226799011, + 0.013168038800358772, + 0.011454826220870018, + 0.015748100355267525, + 0.010173406451940536, + 0.04950281232595444, + 0.011648242361843586, + 0.014609047211706638, + 0.04878611862659454, + 0.009445594623684883, + 0.023170294240117073, + 0.005473203957080841, + 0.0044191456399858, + 0.00743522122502327, + 0.0054025244899094105, + 0.004911058582365513, + 0.006105129607021809, + 0.003773622214794159, + 0.006987093016505241, + 0.025182930752635002, + 0.016820745542645454, + 0.03830493241548538, + 0.0120314946398139, + 0.12663885951042175, + 0.02987287938594818, + 0.07152928411960602, + 0.017053255811333656, + 0.008854985237121582, + 0.0033096661791205406, + 0.009440689347684383, + 0.0031195629853755236, + 0.0096105532720685, + 0.004022621084004641, + 0.0036027461756020784, + 0.0037388866767287254, + 0.012112924829125404, + 0.008519775234162807, + 0.017818300053477287, + 0.008721396327018738, + 0.017238149419426918, + 0.00983648281544447, + 0.009120103903114796, + 0.007926588878035545, + 0.012392906472086906, + 0.01058508176356554, + 0.010541162453591824, + 0.009491514414548874, + 0.00704145897179842, + 0.009781251661479473, + 0.006226498167961836, + 0.01058101188391447, + 0.009885624051094055, + 0.005211776588112116, + 0.024371491745114326, + 0.0052589718252420425, + 0.006428558845072985, + 0.006303756032139063, + 0.004504410549998283, + 0.008907357230782509, + 0.001375385094434023, + 0.0009583515347912908, + 0.0009078410803340375, + 0.0011702333576977253, + 0.0015758599620312452, + 0.001065425924025476, + 0.0014685586793348193, + 0.0011321759084239602, + 0.004233676008880138, + 0.002288311952725053, + 0.002613992430269718, + 0.002311910968273878, + 0.006996815092861652, + 0.003448678180575371, + 0.004914980847388506, + 0.003269130364060402 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 1, + 0 + ], + "selected_reward": -0.020619675517082214 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 dark, matte-textured cast iron skillet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the skillet surface coated with a shimmering layer of transparent, hot oil?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a raw egg drop into the frame and hiss upon contact with the hot surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the egg white instantly coagulate into a solid, opaque white that spreads outward with a golden, lacy, curled edge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a vibrant orange-yellow yolk visible in the center with a smooth surface catching bright, warm highlights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003198272082954645 + ], + "opposite_logprob_tokens": [ + -5.753198146820068 + ], + "logprob_reward": -0.003198272082954645, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.003198272082954645, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14944559335708618, + 0.18681924045085907, + 0.07054240256547928, + 0.2567509114742279, + 0.056122779846191406, + 0.12597882747650146, + 0.08646095544099808, + 0.021167714148759842, + 0.03415851667523384, + 0.012553079053759575 + ], + "credit_map": [ + 0.020510204136371613, + 0.015952620655298233, + 0.020560268312692642, + 0.018655145540833473, + 0.018281321972608566, + 0.016971440985798836, + 0.01532051246613264, + 0.023194074630737305, + 0.01817743107676506, + 0.010386055335402489, + 0.0647183433175087, + 0.01383269764482975, + 0.018454954028129578, + 0.04228964447975159, + 0.008650708943605423, + 0.010309409350156784, + 0.0069472286850214005, + 0.003343575866892934, + 0.02307865396142006, + 0.00709136575460434, + 0.009938225150108337, + 0.008435042575001717, + 0.006848793476819992, + 0.004859509412199259, + 0.01837369240820408, + 0.01246866025030613, + 0.030049331486225128, + 0.0105333486571908, + 0.09339728206396103, + 0.025030076503753662, + 0.05082932114601135, + 0.016069207340478897, + 0.009418879635632038, + 0.008238546550273895, + 0.011717368848621845, + 0.007051621098071337, + 0.005066900048404932, + 0.0047326283529400826, + 0.004117554984986782, + 0.005779278464615345, + 0.016896171495318413, + 0.014497902244329453, + 0.026761557906866074, + 0.01279695238918066, + 0.01343486737459898, + 0.018741410225629807, + 0.010456953197717667, + 0.012393011711537838, + 0.011550885625183582, + 0.01235109381377697, + 0.013771656900644302, + 0.011169321835041046, + 0.01146824099123478, + 0.008510073646903038, + 0.006930458825081587, + 0.010709223337471485, + 0.002614287193864584, + 0.0018717799102887511, + 0.004713245667517185, + 0.002113667083904147, + 0.0029645650647580624, + 0.0024851250927895308, + 0.0016698490362614393, + 0.002735196379944682, + 0.004784506279975176, + 0.0031350539065897465, + 0.002851548371836543, + 0.003901609219610691, + 0.003206648165360093, + 0.0041192565113306046, + 0.003470785217359662, + 0.00868910737335682, + 0.001849401043727994, + 0.0013154034968465567, + 0.001677184016443789, + 0.0015540996100753546, + 0.0019095473689958453, + 0.0008822976378723979, + 0.0017209566431120038, + 0.001644188305363059 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 1, + 0 + ], + "selected_reward": -0.003198272082954645 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 dark, matte-textured cast iron skillet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the skillet surface coated with a shimmering layer of transparent, hot oil?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a raw egg drop into the frame and hiss upon contact with the hot surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the egg white instantly coagulate into a solid, opaque white that spreads outward with a golden, lacy, curled edge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a vibrant orange-yellow yolk visible in the center with a smooth surface catching bright, warm highlights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00044800734031014144 + ], + "opposite_logprob_tokens": [ + -7.750448226928711 + ], + "logprob_reward": -0.00044800734031014144, + "yes_no_margin_reward": 7.75, + "token_credit_reward": -0.00044800734031014144, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.25165966153144836, + 0.05575656145811081, + 0.06270799040794373, + 0.18718717992305756, + 0.16487570106983185, + 0.10524711012840271, + 0.05495616793632507, + 0.05044008046388626, + 0.04726223647594452, + 0.01990731805562973 + ], + "credit_map": [ + 0.03674238175153732, + 0.02793305553495884, + 0.03497314453125, + 0.03045664168894291, + 0.030430996790528297, + 0.029834793880581856, + 0.0248432494699955, + 0.036445390433073044, + 0.0052495370618999004, + 0.005246615968644619, + 0.013624854385852814, + 0.0066883377730846405, + 0.006354779936373234, + 0.007886610925197601, + 0.004660768900066614, + 0.006045056972652674, + 0.006113331764936447, + 0.006938238628208637, + 0.016683295369148254, + 0.0051017333753407, + 0.009435866959393024, + 0.007301297038793564, + 0.005823249462991953, + 0.005310976877808571, + 0.016459425911307335, + 0.0065546766854822636, + 0.015626518055796623, + 0.005893312860280275, + 0.0747155249118805, + 0.016141653060913086, + 0.04228575900197029, + 0.009510315023362637, + 0.02367723174393177, + 0.018700210377573967, + 0.03843119367957115, + 0.019798964262008667, + 0.015029204078018665, + 0.01424848847091198, + 0.013864569365978241, + 0.021125851199030876, + 0.013975868001580238, + 0.012047390453517437, + 0.02243022993206978, + 0.011456429958343506, + 0.009939489886164665, + 0.014693685807287693, + 0.010262317955493927, + 0.010441691614687443, + 0.008731410838663578, + 0.007363743614405394, + 0.007753674406558275, + 0.006712992675602436, + 0.006396239157766104, + 0.006004866678267717, + 0.004488400649279356, + 0.007504838053137064, + 0.008198415860533714, + 0.0038550691679120064, + 0.007181667257100344, + 0.003264445811510086, + 0.0074277641251683235, + 0.005214603152126074, + 0.005028221756219864, + 0.010269892401993275, + 0.0072114174254238605, + 0.004672240000218153, + 0.0031325325835496187, + 0.004259234294295311, + 0.004546904005110264, + 0.005181758664548397, + 0.005301556549966335, + 0.012956593185663223, + 0.004190455656498671, + 0.0022624963894486427, + 0.0030117458663880825, + 0.0019115254981443286, + 0.00247475178912282, + 0.0017539268592372537, + 0.0019555119797587395, + 0.0023469047155231237 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 4 + ], + "selected_reward": -0.00044800734031014144 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An extreme close-up shot from a static, slightly elevated angle captures the dramatic moment when a raw egg meets a scorching surface. The camera is tightly focused on a dark, matte-textured cast iron skillet, its black surface coated with a shimmering layer of transparent, hot oil. A freshly cracked egg drops into the frame, and the pan hisses as its clear, sticky egg white makes contact. The transformation is instantaneous and captivating: the translucent white liquid instantly and violently coagulates into a solid, opaque white that spreads rapidly outward. The thinnest part of the perimeter cooks fastest, forming a delicate, lacy edge that crisps and curls to a perfect golden color. In the center of this rapidly solidifying base, the vibrant, orb of orange-yellow yolk drops with a gentle sway, its smooth surface catching bright, warm highlights, creating a culinary scene of focused concentration and intense satisfaction.", + "seed": 23324, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000021/group_15/index_324_seed_23324_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": -0.15621510011842474, + "vlm_logprob": -0.15621510011842474, + "vlm_yes_no_margin": 3.875, + "vlm_token_credit": -0.15621510011842474, + "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 of a dark, matte-textured cast iron skillet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the skillet surface coated with a shimmering layer of transparent, hot oil?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a raw egg drop into the frame and hiss upon contact with the hot surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the egg white instantly coagulate into a solid, opaque white that spreads outward with a golden, lacy, curled edge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a vibrant orange-yellow yolk visible in the center with a smooth surface catching bright, warm highlights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0590030699968338 + ], + "opposite_logprob_tokens": [ + -2.9340031147003174 + ], + "logprob_reward": -0.0590030699968338, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.0590030699968338, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04506240785121918, + 0.01958438754081726, + 0.01256942842155695, + 0.037186380475759506, + 0.029712053015828133, + 0.009805806912481785, + 0.010862386785447598, + 0.058618221431970596, + 0.7424003481864929, + 0.03419860452413559 + ], + "credit_map": [ + 0.00269221980124712, + 0.002815169980749488, + 0.013266006484627724, + 0.0029651497025042772, + 0.002041357336565852, + 0.012117093428969383, + 0.0019160882802680135, + 0.00724932411685586, + 0.001776406541466713, + 0.0017791226273402572, + 0.002824930939823389, + 0.003259815275669098, + 0.0013508755946531892, + 0.0023066624999046326, + 0.0025044181384146214, + 0.0037821573205292225, + 0.0023921108804643154, + 0.001083230716176331, + 0.0020039204973727465, + 0.0011765911476686597, + 0.0009209606214426458, + 0.0023725780192762613, + 0.0011790856951847672, + 0.0014409513678401709, + 0.002725118538364768, + 0.003495857585221529, + 0.008349993266165257, + 0.003172989236190915, + 0.0023125018924474716, + 0.010995120741426945, + 0.0028850475791841745, + 0.0032497483771294355, + 0.003157893195748329, + 0.0037315282970666885, + 0.004871055018156767, + 0.0045410031452775, + 0.0022755588870495558, + 0.003444853238761425, + 0.0034757740795612335, + 0.004214385990053415, + 0.0012582369381561875, + 0.0013894753064960241, + 0.0011939011747017503, + 0.0016229519387707114, + 0.0008597965934313834, + 0.0010853968560695648, + 0.00124454649630934, + 0.001151501783169806, + 0.0008350920397788286, + 0.0008629996445961297, + 0.0007907677791081369, + 0.0006564926588907838, + 0.0008970404742285609, + 0.003770728362724185, + 0.0018981053726747632, + 0.0011511610355228186, + 0.0033829917665570974, + 0.002564050955697894, + 0.0017631194787099957, + 0.01048779021948576, + 0.006775521207600832, + 0.00711080152541399, + 0.019220983609557152, + 0.007312961854040623, + 0.017792804166674614, + 0.031756073236465454, + 0.014999504201114178, + 0.30926308035850525, + 0.037763264030218124, + 0.014927526004612446, + 0.2823270261287689, + 0.03357104957103729, + 0.001439777435734868, + 0.00211055064573884, + 0.0009684691904112697, + 0.015058498829603195, + 0.0014153675874695182, + 0.001154289348050952, + 0.009281386621296406, + 0.0027702627703547478 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 0 + ], + "selected_reward": -0.0590030699968338 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 dark, matte-textured cast iron skillet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the skillet surface coated with a shimmering layer of transparent, hot oil?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a raw egg drop into the frame and hiss upon contact with the hot surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the egg white instantly coagulate into a solid, opaque white that spreads outward with a golden, lacy, curled edge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a vibrant orange-yellow yolk visible in the center with a smooth surface catching bright, warm highlights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018221953883767128 + ], + "opposite_logprob_tokens": [ + -4.018221855163574 + ], + "logprob_reward": -0.018221953883767128, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018221953883767128, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14355796575546265, + 0.025196751579642296, + 0.009752487763762474, + 0.022717509418725967, + 0.009198152460157871, + 0.008011844009160995, + 0.009812609292566776, + 0.19800736010074615, + 0.5160582065582275, + 0.05768703669309616 + ], + "credit_map": [ + 0.011750821955502033, + 0.00886634923517704, + 0.03416510671377182, + 0.009115145541727543, + 0.0049691046588122845, + 0.03233112767338753, + 0.007850413210690022, + 0.03450991213321686, + 0.0030567380599677563, + 0.002641890197992325, + 0.0035473324824124575, + 0.0035169110633432865, + 0.0014216796262189746, + 0.00358598749153316, + 0.0029474960174411535, + 0.0044787186197936535, + 0.0015508168144151568, + 0.0008352259756065905, + 0.0016300504794344306, + 0.0011510440381243825, + 0.0007938762428238988, + 0.0014115014346316457, + 0.0011233610566705465, + 0.0012566125951707363, + 0.001931007020175457, + 0.002923340070992708, + 0.004048484843224287, + 0.002735340967774391, + 0.0018004528246819973, + 0.004693243652582169, + 0.002361547900363803, + 0.0022240937687456608, + 0.0011539605911821127, + 0.0010642888955771923, + 0.0014033839106559753, + 0.0011857027420774102, + 0.0009154422441497445, + 0.0012737360084429383, + 0.0011011642636731267, + 0.001100474619306624, + 0.0010492147412151098, + 0.0010349552612751722, + 0.0013343713944777846, + 0.0011003059335052967, + 0.0007426653755828738, + 0.0010077566839754581, + 0.0008247633231803775, + 0.0009178122272714972, + 0.0013051078421995044, + 0.000981993624009192, + 0.00134085095487535, + 0.0010954869212582707, + 0.0007717782282270491, + 0.0017568946350365877, + 0.0016880561597645283, + 0.0008724422659724951, + 0.010756443254649639, + 0.008910288102924824, + 0.004938969388604164, + 0.03593527153134346, + 0.02527712844312191, + 0.01557406596839428, + 0.07334265112876892, + 0.0232725627720356, + 0.01250701304525137, + 0.02196958102285862, + 0.010082215070724487, + 0.21536675095558167, + 0.026285922154784203, + 0.010114396922290325, + 0.19665247201919556, + 0.023079892620444298, + 0.003736928105354309, + 0.00237269583158195, + 0.0012734256451949477, + 0.015302231535315514, + 0.005043465178459883, + 0.0016958806663751602, + 0.022673403844237328, + 0.005589012522250414 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 0 + ], + "selected_reward": -0.018221953883767128 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 dark, matte-textured cast iron skillet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the skillet surface coated with a shimmering layer of transparent, hot oil?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a raw egg drop into the frame and hiss upon contact with the hot surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the egg white instantly coagulate into a solid, opaque white that spreads outward with a golden, lacy, curled edge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a vibrant orange-yellow yolk visible in the center with a smooth surface catching bright, warm highlights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.6936630010604858 + ], + "opposite_logprob_tokens": [ + -0.6936630010604858 + ], + "logprob_reward": -0.6936630010604858, + "yes_no_margin_reward": 0.0, + "token_credit_reward": -0.6936630010604858, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.2864575982093811, + 0.09652126580476761, + 0.02122943103313446, + 0.04098980501294136, + 0.019510533660650253, + 0.009318863041698933, + 0.01703105866909027, + 0.3261617124080658, + 0.1309191882610321, + 0.05186042934656143 + ], + "credit_map": [ + 0.018062930554151535, + 0.016216546297073364, + 0.07935741543769836, + 0.01694650389254093, + 0.007757055573165417, + 0.0731794536113739, + 0.013572173193097115, + 0.061365529894828796, + 0.013984059914946556, + 0.008845486678183079, + 0.02021566778421402, + 0.010990331880748272, + 0.005315762013196945, + 0.013447501696646214, + 0.007338482886552811, + 0.01638397015631199, + 0.003895498113706708, + 0.0023035777267068624, + 0.0030124064069241285, + 0.002320521278306842, + 0.00192097551189363, + 0.0033869571052491665, + 0.002212418708950281, + 0.002177076181396842, + 0.0024281248915940523, + 0.004049306269735098, + 0.008929718285799026, + 0.003673375118523836, + 0.002642424777150154, + 0.012151417322456837, + 0.003642852883785963, + 0.003472584765404463, + 0.001954005565494299, + 0.0025017368607223034, + 0.0026482832618057728, + 0.0028265255969017744, + 0.0016230419278144836, + 0.002553527010604739, + 0.002885127207264304, + 0.0025182876270264387, + 0.001319081406109035, + 0.0013610871974378824, + 0.0012650869321078062, + 0.0013234701473265886, + 0.0007434915751218796, + 0.0009953940752893686, + 0.0010707138571888208, + 0.001240538083948195, + 0.0015216792235150933, + 0.001961942296475172, + 0.0021561826579272747, + 0.0028796943370252848, + 0.001642608200199902, + 0.0026947297155857086, + 0.002328406320884824, + 0.0018458167323842645, + 0.016894184052944183, + 0.01421258132904768, + 0.00875720102339983, + 0.0640067607164383, + 0.04344569891691208, + 0.013698657974600792, + 0.13931982219219208, + 0.0258268341422081, + 0.01095562893897295, + 0.006230671890079975, + 0.00418601930141449, + 0.031845077872276306, + 0.012263156473636627, + 0.005469855852425098, + 0.05049256607890129, + 0.009476211853325367, + 0.0027085866313427687, + 0.0019181560492143035, + 0.0010661318665370345, + 0.01772330515086651, + 0.00380433420650661, + 0.001208542613312602, + 0.02055114507675171, + 0.0028802300803363323 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 8 + ], + "selected_reward": -0.6936630010604858 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 dark, matte-textured cast iron skillet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the skillet surface coated with a shimmering layer of transparent, hot oil?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a raw egg drop into the frame and hiss upon contact with the hot surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the egg white instantly coagulate into a solid, opaque white that spreads outward with a golden, lacy, curled edge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a vibrant orange-yellow yolk visible in the center with a smooth surface catching bright, warm highlights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009793566539883614 + ], + "opposite_logprob_tokens": [ + -4.634793758392334 + ], + "logprob_reward": -0.009793566539883614, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009793566539883614, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04004056006669998, + 0.019184717908501625, + 0.009410138241946697, + 0.01831907592713833, + 0.014942103065550327, + 0.006992718204855919, + 0.020571786910295486, + 0.41029220819473267, + 0.4267042279243469, + 0.033542416989803314 + ], + "credit_map": [ + 0.0025073541328310966, + 0.0029157039243727922, + 0.010605867952108383, + 0.0024070930667221546, + 0.001951165497303009, + 0.00960923358798027, + 0.0020045097917318344, + 0.008039630949497223, + 0.002091842470690608, + 0.001734486548230052, + 0.0037265082355588675, + 0.0020554917864501476, + 0.0013360022567212582, + 0.0036328400019556284, + 0.0017552003264427185, + 0.0028523458167910576, + 0.0015286029083654284, + 0.0010747866472229362, + 0.0012423665029928088, + 0.0013608088484033942, + 0.0007356530404649675, + 0.0013849753886461258, + 0.0011651756940409541, + 0.0009177694446407259, + 0.001359016983769834, + 0.0018707880517467856, + 0.003619591472670436, + 0.0020529835019260645, + 0.0014557864051312208, + 0.004686037078499794, + 0.0016551363514736295, + 0.0016197363147512078, + 0.0014835834736004472, + 0.0021473586093634367, + 0.0019854665733873844, + 0.0020770833361893892, + 0.0013912783470004797, + 0.0018619642360135913, + 0.002161492593586445, + 0.0018338754307478666, + 0.0008354466408491135, + 0.0009833895601332188, + 0.0010047043906524777, + 0.0013398202136158943, + 0.0006121926126070321, + 0.0007320938166230917, + 0.0008573327795602381, + 0.0006277382490225136, + 0.002295326441526413, + 0.0018941143061965704, + 0.0024877721443772316, + 0.0020434102043509483, + 0.0018441090360283852, + 0.005298464093357325, + 0.002598288469016552, + 0.0021103022154420614, + 0.014148985967040062, + 0.016926027834415436, + 0.017689533531665802, + 0.1011192575097084, + 0.03671516105532646, + 0.04365184158086777, + 0.15860600769519806, + 0.021435389295220375, + 0.011124380864202976, + 0.016990425065159798, + 0.008236405439674854, + 0.1705835610628128, + 0.0249093696475029, + 0.008453989401459694, + 0.16842113435268402, + 0.017984986305236816, + 0.002092486945912242, + 0.0014526252634823322, + 0.0009042345336638391, + 0.008585150353610516, + 0.0029105274006724358, + 0.0012148383539170027, + 0.013020890764892101, + 0.003361660987138748 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 0 + ], + "selected_reward": -0.009793566539883614 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 dark, matte-textured cast iron skillet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the skillet surface coated with a shimmering layer of transparent, hot oil?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a raw egg drop into the frame and hiss upon contact with the hot surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the egg white instantly coagulate into a solid, opaque white that spreads outward with a golden, lacy, curled edge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a vibrant orange-yellow yolk visible in the center with a smooth surface catching bright, warm highlights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0003939091111533344 + ], + "opposite_logprob_tokens": [ + -7.875393867492676 + ], + "logprob_reward": -0.0003939091111533344, + "yes_no_margin_reward": 7.875, + "token_credit_reward": -0.0003939091111533344, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08930622041225433, + 0.029845301061868668, + 0.012096795253455639, + 0.024848056957125664, + 0.009562401100993156, + 0.006902832072228193, + 0.007035756483674049, + 0.10975216329097748, + 0.6318925619125366, + 0.0787578895688057 + ], + "credit_map": [ + 0.006507021840661764, + 0.004710650537163019, + 0.025620635598897934, + 0.004900212865322828, + 0.0028346714098006487, + 0.02531777322292328, + 0.003680522320792079, + 0.015734735876321793, + 0.003266723360866308, + 0.0022035834845155478, + 0.004325252491980791, + 0.0038657851982861757, + 0.0014824487734586, + 0.003694382729008794, + 0.0031067735981196165, + 0.007900350727140903, + 0.0024007996544241905, + 0.0013745068572461605, + 0.0017902051331475377, + 0.0013073459267616272, + 0.0009970298269763589, + 0.001853501540608704, + 0.0009841352002695203, + 0.0013892716960981488, + 0.0018072525272145867, + 0.0036244273651391268, + 0.0033793337643146515, + 0.0034310754854232073, + 0.002248451579362154, + 0.004277967382222414, + 0.0031199471559375525, + 0.002959600882604718, + 0.0008962537976913154, + 0.0013443408533930779, + 0.0012396360980346799, + 0.001443347311578691, + 0.0009350581676699221, + 0.001204698346555233, + 0.0012695459881797433, + 0.0012295206543058157, + 0.0009022525628097355, + 0.000965817947871983, + 0.0008504394209012389, + 0.0011705607175827026, + 0.0008120698621496558, + 0.0006908069481141865, + 0.0008865179261192679, + 0.0006243666284717619, + 0.0008008695440366864, + 0.0005936801317147911, + 0.0009075472480617464, + 0.0007995234336704016, + 0.0008807373815216124, + 0.0009577312157489359, + 0.0014059257227927446, + 0.0006897414568811655, + 0.007769793272018433, + 0.005238301120698452, + 0.003977825865149498, + 0.013069930486381054, + 0.013474220409989357, + 0.019026365131139755, + 0.030150944367051125, + 0.017044775187969208, + 0.014718038029968739, + 0.026834845542907715, + 0.012329285964369774, + 0.26458001136779785, + 0.03212817385792732, + 0.01175021380186081, + 0.2413412630558014, + 0.028210720047354698, + 0.00364926690235734, + 0.005144035443663597, + 0.0023723451886326075, + 0.03441683202981949, + 0.0035988581366837025, + 0.002874191850423813, + 0.019928639754652977, + 0.006773718167096376 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 0 + ], + "selected_reward": -0.0003939091111533344 + } + ], + "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/0000021/group_2/index_tensor([6645, 6645])_group_2_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000021/group_2/index_tensor([6645, 6645])_group_2_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..721a4ca9801d9532a3f3b52bb10c66e2d223f305 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000021/group_2/index_tensor([6645, 6645])_group_2_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "low-angle, medium shot captures a silent coniferous forest blanketed in thick snow. In the diffuse light of a gloomy sky, a massive, full-grown white tiger explodes from behind a clump of snow-bent bushes. Its snow-white fur is streaked with deep black stripes. In its lunge, its muscular body stretches to its limit, its hind legs thrust forward, sending up a shower of powdery snow. The entire movement is imbued with a sense of primal, deadly power. Then, the scene shifts to an extreme close-up, focusing on its icy blue eyes, piercing with a single-minded killing instinct. The camera cuts again, focusing on its extended front paws. Four razor-sharp claws have fully emerged from their thick pads, gleaming like deadly ivory in the chill air. The entire sequence, executed in a breathtakingly short time, creates a heart-stopping tension and a savage beauty.", + "seed": 27340, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000021/group_2/index_6645_seed_27340_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": -0.02239142395555973, + "vlm_logprob": -0.02239142395555973, + "vlm_yes_no_margin": 4.725, + "vlm_token_credit": -0.02239142395555973, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 coniferous forest blanketed in thick snow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive, full-grown white tiger with snow-white fur and deep black stripes visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the tiger explode from behind a clump of snow-bent bushes in a leaping motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video include an extreme close-up shot of the tiger's icy blue eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video show an extended front paw with four razor-sharp claws emerging from the pad?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002916489727795124 + ], + "opposite_logprob_tokens": [ + -7.25291633605957 + ], + "logprob_reward": -0.002916489727795124, + "yes_no_margin_reward": 7.25, + "token_credit_reward": -0.002916489727795124, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10452090948820114, + 0.08156606554985046, + 0.10670112818479538, + 0.07202564179897308, + 0.0737854540348053, + 0.08991450816392899, + 0.10972882807254791, + 0.11043184995651245, + 0.14314547181129456, + 0.10818015038967133 + ], + "credit_map": [ + 0.012738034129142761, + 0.020322073251008987, + 0.011110503226518631, + 0.016855554655194283, + 0.01217006053775549, + 0.012633999809622765, + 0.008913783356547356, + 0.009776897728443146, + 0.005701758898794651, + 0.010295395739376545, + 0.01247891504317522, + 0.013025500811636448, + 0.01075334008783102, + 0.011640802957117558, + 0.011380430310964584, + 0.006289920769631863, + 0.007517565507441759, + 0.014240133576095104, + 0.010461440309882164, + 0.01587977446615696, + 0.025031527504324913, + 0.009990975260734558, + 0.01726037636399269, + 0.006319336127489805, + 0.0056281741708517075, + 0.011020045727491379, + 0.006636547856032848, + 0.012958432547748089, + 0.007638625334948301, + 0.010005667805671692, + 0.007570638787001371, + 0.010567502118647099, + 0.010617186315357685, + 0.006146441679447889, + 0.011714428663253784, + 0.009498880244791508, + 0.007288281340152025, + 0.01066200714558363, + 0.010657787322998047, + 0.007200444582849741, + 0.009920690208673477, + 0.009707421064376831, + 0.016239119693636894, + 0.011660933494567871, + 0.011784962378442287, + 0.011702653020620346, + 0.009940550662577152, + 0.00895818043500185, + 0.011041102930903435, + 0.0142471082508564, + 0.011157497763633728, + 0.016010597348213196, + 0.014758966863155365, + 0.01458127424120903, + 0.01585669443011284, + 0.012075583450496197, + 0.03677964583039284, + 0.009846328757703304, + 0.011603879742324352, + 0.00934549979865551, + 0.0066447011195123196, + 0.00819374993443489, + 0.010319158434867859, + 0.01769888401031494, + 0.0231010839343071, + 0.017976440489292145, + 0.01779993437230587, + 0.020846061408519745, + 0.01712724380195141, + 0.014631112106144428, + 0.01772168092429638, + 0.013941911049187183, + 0.01750035770237446, + 0.010873273946344852, + 0.02002486027777195, + 0.01601666957139969, + 0.01134426612406969, + 0.007424718700349331, + 0.013906072825193405, + 0.011089939624071121 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 6 + ], + "selected_reward": -0.002916489727795124 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 coniferous forest blanketed in thick snow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive, full-grown white tiger with snow-white fur and deep black stripes visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the tiger explode from behind a clump of snow-bent bushes in a leaping motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video include an extreme close-up shot of the tiger's icy blue eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video show an extended front paw with four razor-sharp claws emerging from the pad?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002498839981853962 + ], + "opposite_logprob_tokens": [ + -6.002498626708984 + ], + "logprob_reward": -0.002498839981853962, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.002498839981853962, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03857029229402542, + 0.0857180655002594, + 0.15226276218891144, + 0.033729810267686844, + 0.13560697436332703, + 0.15496575832366943, + 0.07515450567007065, + 0.06012984365224838, + 0.13510125875473022, + 0.1287606656551361 + ], + "credit_map": [ + 0.003998678643256426, + 0.004497795831412077, + 0.003720964305102825, + 0.004738011863082647, + 0.00835137814283371, + 0.005614896304905415, + 0.0038192563224583864, + 0.00382931437343359, + 0.004227271303534508, + 0.005918728653341532, + 0.006985224783420563, + 0.010587629862129688, + 0.011861261911690235, + 0.017684709280729294, + 0.02416209504008293, + 0.004291146527975798, + 0.010124233551323414, + 0.01856321468949318, + 0.01229961309581995, + 0.020542848855257034, + 0.03739989921450615, + 0.017406662926077843, + 0.026176657527685165, + 0.009749632328748703, + 0.0034379451535642147, + 0.00437243701890111, + 0.003528977045789361, + 0.004762224853038788, + 0.00540247093886137, + 0.003688226453959942, + 0.0047261593863368034, + 0.003811369650065899, + 0.010554106906056404, + 0.012985771521925926, + 0.017097877338528633, + 0.01333501748740673, + 0.023306740447878838, + 0.01841389201581478, + 0.02590242773294449, + 0.014011137187480927, + 0.024916695430874825, + 0.021131392568349838, + 0.027362896129488945, + 0.017234914004802704, + 0.016770729795098305, + 0.015696458518505096, + 0.019747771322727203, + 0.012104903347790241, + 0.006809095852077007, + 0.012456458993256092, + 0.007370265666395426, + 0.009547656401991844, + 0.01485413033515215, + 0.007002376485615969, + 0.010533195920288563, + 0.006581325549632311, + 0.012338473461568356, + 0.0053083631210029125, + 0.01239957008510828, + 0.005167163442820311, + 0.004497779067605734, + 0.010332311503589153, + 0.004074739292263985, + 0.0060114432126283646, + 0.021363303065299988, + 0.01756940968334675, + 0.022381866350769997, + 0.019253036007285118, + 0.01738565042614937, + 0.010949984192848206, + 0.016343221068382263, + 0.009854784235358238, + 0.03001309558749199, + 0.008539866656064987, + 0.03867439925670624, + 0.015242664143443108, + 0.010316789150238037, + 0.006722849328070879, + 0.011431815102696419, + 0.007819176651537418 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 2, + 4 + ], + "selected_reward": -0.002498839981853962 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 coniferous forest blanketed in thick snow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive, full-grown white tiger with snow-white fur and deep black stripes visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the tiger explode from behind a clump of snow-bent bushes in a leaping motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video include an extreme close-up shot of the tiger's icy blue eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video show an extended front paw with four razor-sharp claws emerging from the pad?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026343639940023422 + ], + "opposite_logprob_tokens": [ + -3.651343584060669 + ], + "logprob_reward": -0.026343639940023422, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026343639940023422, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10560250282287598, + 0.23512133955955505, + 0.056961819529533386, + 0.06415238976478577, + 0.08645981550216675, + 0.07407970726490021, + 0.08548162132501602, + 0.10668013989925385, + 0.15947581827640533, + 0.02598486840724945 + ], + "credit_map": [ + 0.008082178421318531, + 0.011402711272239685, + 0.00907114241272211, + 0.01714371144771576, + 0.02114388719201088, + 0.016511015594005585, + 0.012144896201789379, + 0.010102970525622368, + 0.008256038650870323, + 0.00969359464943409, + 0.017645111307501793, + 0.016942238435149193, + 0.017095282673835754, + 0.06625455617904663, + 0.0860242173075676, + 0.013210290111601353, + 0.005310020409524441, + 0.0071723731234669685, + 0.006791647523641586, + 0.007422052323818207, + 0.01133495382964611, + 0.006572054699063301, + 0.006648327223956585, + 0.005710387136787176, + 0.004506413824856281, + 0.007262031082063913, + 0.006679897662252188, + 0.00784402433782816, + 0.009658093564212322, + 0.010342605412006378, + 0.011236204765737057, + 0.006623119115829468, + 0.008028358221054077, + 0.007874907925724983, + 0.010632477700710297, + 0.007559644989669323, + 0.0151646938174963, + 0.01505567878484726, + 0.013681283220648766, + 0.00846278015524149, + 0.010374575853347778, + 0.011763105168938637, + 0.013996711932122707, + 0.008906759321689606, + 0.00692336168140173, + 0.005458544008433819, + 0.008523880504071712, + 0.008132766000926495, + 0.009670289233326912, + 0.009855681098997593, + 0.007540292572230101, + 0.014605279080569744, + 0.013694723136723042, + 0.007419980131089687, + 0.013737879693508148, + 0.00895750429481268, + 0.040754660964012146, + 0.0061165951192379, + 0.011344834230840206, + 0.005747442599385977, + 0.005345846991986036, + 0.009175736457109451, + 0.007944606244564056, + 0.02025042474269867, + 0.027018191292881966, + 0.018931249156594276, + 0.0235951766371727, + 0.02509903721511364, + 0.015773536637425423, + 0.015509678050875664, + 0.01754343882203102, + 0.016005510464310646, + 0.004189484287053347, + 0.0026939851231873035, + 0.004064817912876606, + 0.003548159496858716, + 0.002529699122533202, + 0.002341795712709427, + 0.0036946902982890606, + 0.002922235755249858 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 8, + 7 + ], + "selected_reward": -0.026343639940023422 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 coniferous forest blanketed in thick snow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive, full-grown white tiger with snow-white fur and deep black stripes visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the tiger explode from behind a clump of snow-bent bushes in a leaping motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video include an extreme close-up shot of the tiger's icy blue eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video show an extended front paw with four razor-sharp claws emerging from the pad?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.062021732330322266 + ], + "opposite_logprob_tokens": [ + -2.8120217323303223 + ], + "logprob_reward": -0.062021732330322266, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.062021732330322266, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.028982896357774734, + 0.026739515364170074, + 0.04904833436012268, + 0.022870320826768875, + 0.04521610587835312, + 0.09813371300697327, + 0.06766360998153687, + 0.30251073837280273, + 0.30358919501304626, + 0.055245641618967056 + ], + "credit_map": [ + 0.003059977199882269, + 0.002698424970731139, + 0.0030333222821354866, + 0.004860608838498592, + 0.0050257062539458275, + 0.003427843563258648, + 0.003728176699951291, + 0.0031488356180489063, + 0.0020531993359327316, + 0.002389573259279132, + 0.00547085190191865, + 0.003450533142313361, + 0.003246990730985999, + 0.005397655535489321, + 0.002173069631680846, + 0.002557641128078103, + 0.004267286043614149, + 0.006754720117896795, + 0.0035843143705278635, + 0.006270758807659149, + 0.011395370587706566, + 0.00519206328317523, + 0.007582994643598795, + 0.004000825807452202, + 0.0016598450019955635, + 0.003203827189281583, + 0.0038696262054145336, + 0.0030797331128269434, + 0.0031233238987624645, + 0.0027997756842523813, + 0.002990944543853402, + 0.0021432454232126474, + 0.0035526263527572155, + 0.0039551337249577045, + 0.005193122196942568, + 0.0044837817549705505, + 0.008700412698090076, + 0.007745896000415087, + 0.006541999988257885, + 0.005043135955929756, + 0.010594261810183525, + 0.011167186312377453, + 0.021040471270680428, + 0.009621421806514263, + 0.01152477040886879, + 0.01203162968158722, + 0.010407851077616215, + 0.011746121570467949, + 0.006902478169649839, + 0.011995784938335419, + 0.0057850610464811325, + 0.013945311307907104, + 0.008367453701794147, + 0.005166579969227314, + 0.009527948684990406, + 0.0059729889035224915, + 0.09209233522415161, + 0.021110067144036293, + 0.04712523892521858, + 0.022321859374642372, + 0.021004753187298775, + 0.03820240497589111, + 0.021368669345974922, + 0.03928539529442787, + 0.04255983233451843, + 0.035979434847831726, + 0.03350383788347244, + 0.05685923248529434, + 0.03048134222626686, + 0.04008195549249649, + 0.030350876972079277, + 0.03377267345786095, + 0.007166973780840635, + 0.0047047752887010574, + 0.008716248907148838, + 0.008241591043770313, + 0.008337294682860374, + 0.003805531421676278, + 0.00968928076326847, + 0.004583945032209158 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 5 + ], + "selected_reward": -0.062021732330322266 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 coniferous forest blanketed in thick snow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive, full-grown white tiger with snow-white fur and deep black stripes visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the tiger explode from behind a clump of snow-bent bushes in a leaping motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video include an extreme close-up shot of the tiger's icy blue eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video show an extended front paw with four razor-sharp claws emerging from the pad?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01817641779780388 + ], + "opposite_logprob_tokens": [ + -4.018176555633545 + ], + "logprob_reward": -0.01817641779780388, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.01817641779780388, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03440086171030998, + 0.05641477182507515, + 0.026591941714286804, + 0.04593442752957344, + 0.01929575763642788, + 0.562957763671875, + 0.0584736168384552, + 0.04063832759857178, + 0.13739337027072906, + 0.01789918541908264 + ], + "credit_map": [ + 0.002891932614147663, + 0.0037299171090126038, + 0.0028017335571348667, + 0.004319985397160053, + 0.007506502326577902, + 0.005326281767338514, + 0.0036876194644719362, + 0.004136888310313225, + 0.002594536170363426, + 0.003323902143165469, + 0.003765859641134739, + 0.00555372005328536, + 0.003731200937181711, + 0.016109341755509377, + 0.01905229315161705, + 0.0022839137818664312, + 0.0029369003605097532, + 0.003503759391605854, + 0.002870587632060051, + 0.0030669074039906263, + 0.004254197236150503, + 0.0027375619392842054, + 0.004059676546603441, + 0.00316235120408237, + 0.005110936239361763, + 0.005490918643772602, + 0.005283629521727562, + 0.0079196160659194, + 0.006573912687599659, + 0.005501571577042341, + 0.005223557353019714, + 0.004830292426049709, + 0.0018418619874864817, + 0.0018379593966528773, + 0.0028019172605127096, + 0.002607506001368165, + 0.002784338081255555, + 0.0035507269203662872, + 0.0019912461284548044, + 0.0018802026752382517, + 0.07397309690713882, + 0.06999251991510391, + 0.12273941189050674, + 0.053368113934993744, + 0.061874132603406906, + 0.06190873682498932, + 0.05809490755200386, + 0.061006832867860794, + 0.0043526398949325085, + 0.007955647073686123, + 0.004435817711055279, + 0.010392404161393642, + 0.011847862042486668, + 0.005114665720611811, + 0.009143847972154617, + 0.005230729468166828, + 0.014042040333151817, + 0.003386522876098752, + 0.005709055811166763, + 0.0027498260606080294, + 0.002294866368174553, + 0.004138544667512178, + 0.0027947023045271635, + 0.0055227698758244514, + 0.018575066700577736, + 0.0142530407756567, + 0.01466695312410593, + 0.0235856082290411, + 0.014483417384326458, + 0.017597129568457603, + 0.01836155168712139, + 0.015870610252022743, + 0.0021573398262262344, + 0.0019116521580144763, + 0.002272832440212369, + 0.002350148279219866, + 0.0019372900715097785, + 0.002613725606352091, + 0.0021637582685798407, + 0.0024924392346292734 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 8, + 6 + ], + "selected_reward": -0.01817641779780388 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle, medium shot captures a silent coniferous forest blanketed in thick snow. In the diffuse light of a gloomy sky, a massive, full-grown white tiger explodes from behind a clump of snow-bent bushes. Its snow-white fur is streaked with deep black stripes. In its lunge, its muscular body stretches to its limit, its hind legs thrust forward, sending up a shower of powdery snow. The entire movement is imbued with a sense of primal, deadly power. Then, the scene shifts to an extreme close-up, focusing on its icy blue eyes, piercing with a single-minded killing instinct. The camera cuts again, focusing on its extended front paws. Four razor-sharp claws have fully emerged from their thick pads, gleaming like deadly ivory in the chill air. The entire sequence, executed in a breathtakingly short time, creates a heart-stopping tension and a savage beauty.", + "seed": 27341, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000021/group_2/index_6645_seed_27341_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": -0.02828351780772209, + "vlm_logprob": -0.02828351780772209, + "vlm_yes_no_margin": 4.325, + "vlm_token_credit": -0.02828351780772209, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 coniferous forest blanketed in thick snow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive, full-grown white tiger with snow-white fur and deep black stripes visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the tiger explode from behind a clump of snow-bent bushes in a leaping motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video include an extreme close-up shot of the tiger's icy blue eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video show an extended front paw with four razor-sharp claws emerging from the pad?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007378470152616501 + ], + "opposite_logprob_tokens": [ + -7.382378578186035 + ], + "logprob_reward": -0.007378470152616501, + "yes_no_margin_reward": 7.375, + "token_credit_reward": -0.007378470152616501, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06067660078406334, + 0.06623043119907379, + 0.06286372244358063, + 0.03799807280302048, + 0.2218695729970932, + 0.06857789307832718, + 0.2126842439174652, + 0.12595190107822418, + 0.044234730303287506, + 0.09891290217638016 + ], + "credit_map": [ + 0.007721773348748684, + 0.008421103470027447, + 0.010405249893665314, + 0.007029549218714237, + 0.0064433240331709385, + 0.006526283919811249, + 0.006366707384586334, + 0.007762601133435965, + 0.006506424397230148, + 0.010389814153313637, + 0.007588746026158333, + 0.012267072685062885, + 0.00650993175804615, + 0.00786925945430994, + 0.005885271355509758, + 0.009213906712830067, + 0.005119020119309425, + 0.00597289577126503, + 0.0072031812742352486, + 0.00609583081677556, + 0.011821347288787365, + 0.007753998041152954, + 0.008649410679936409, + 0.010248024947941303, + 0.004236122593283653, + 0.006209869869053364, + 0.004505896009504795, + 0.004416515585035086, + 0.005277753341943026, + 0.004082816187292337, + 0.005733213387429714, + 0.0035358816385269165, + 0.010186814703047276, + 0.019562097266316414, + 0.02374320849776268, + 0.03674265742301941, + 0.026768596842885017, + 0.04235035181045532, + 0.023751061409711838, + 0.03876475989818573, + 0.00422399677336216, + 0.010557285510003567, + 0.0053199464455246925, + 0.009478163905441761, + 0.010083051398396492, + 0.008939419873058796, + 0.012377834878861904, + 0.00759818684309721, + 0.017508583143353462, + 0.0262397862970829, + 0.02591513842344284, + 0.03090081736445427, + 0.0218135304749012, + 0.031114157289266586, + 0.02697630226612091, + 0.032215896993875504, + 0.011969620361924171, + 0.01331484504044056, + 0.015469370409846306, + 0.013650967739522457, + 0.01416986808180809, + 0.021901993080973625, + 0.014967299997806549, + 0.020507926121354103, + 0.004045774694532156, + 0.00874530803412199, + 0.0035734933335334063, + 0.004900761879980564, + 0.005841219797730446, + 0.004722115583717823, + 0.004684455692768097, + 0.007721596397459507, + 0.010337632149457932, + 0.017893176525831223, + 0.010005131363868713, + 0.013621936552226543, + 0.012311023660004139, + 0.008157224394381046, + 0.014198402874171734, + 0.012388363480567932 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 6, + 7 + ], + "selected_reward": -0.007378470152616501 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 coniferous forest blanketed in thick snow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive, full-grown white tiger with snow-white fur and deep black stripes visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the tiger explode from behind a clump of snow-bent bushes in a leaping motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video include an extreme close-up shot of the tiger's icy blue eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video show an extended front paw with four razor-sharp claws emerging from the pad?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008648913353681564 + ], + "opposite_logprob_tokens": [ + -4.758648872375488 + ], + "logprob_reward": -0.008648913353681564, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008648913353681564, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.057015594094991684, + 0.022554345428943634, + 0.051783084869384766, + 0.03809688240289688, + 0.1935427039861679, + 0.09270945191383362, + 0.04318179935216904, + 0.2834946811199188, + 0.049660131335258484, + 0.16796132922172546 + ], + "credit_map": [ + 0.0062641482800245285, + 0.010244281962513924, + 0.009451799094676971, + 0.005435007158666849, + 0.006247004494071007, + 0.006457328796386719, + 0.0067209843546152115, + 0.0061950390227139, + 0.0029921012464910746, + 0.0034632496535778046, + 0.002956944750621915, + 0.0032128619495779276, + 0.0019480197224766016, + 0.002583612222224474, + 0.0027437377721071243, + 0.002653819741681218, + 0.004831905942410231, + 0.007971296086907387, + 0.007123402785509825, + 0.008571051061153412, + 0.005339528899639845, + 0.005558749660849571, + 0.004339483566582203, + 0.00804766546934843, + 0.002869732677936554, + 0.005080437287688255, + 0.004231257364153862, + 0.005689373705536127, + 0.005134870298206806, + 0.003830668283626437, + 0.0052210427820682526, + 0.006039496511220932, + 0.005824335850775242, + 0.007249967195093632, + 0.01214277371764183, + 0.012062527239322662, + 0.009807842783629894, + 0.06908155232667923, + 0.009564477950334549, + 0.06780923157930374, + 0.007483969908207655, + 0.008807403966784477, + 0.008146574720740318, + 0.007677137386053801, + 0.011145039461553097, + 0.01885223016142845, + 0.010538308881223202, + 0.02005879022181034, + 0.003357896814122796, + 0.005177091341465712, + 0.004891334567219019, + 0.005003994796425104, + 0.004227489698678255, + 0.007566747255623341, + 0.005100979935377836, + 0.007856265641748905, + 0.02918355166912079, + 0.02656068280339241, + 0.032721661031246185, + 0.02040218375623226, + 0.025822972878813744, + 0.05441843718290329, + 0.03199293464422226, + 0.06239224970340729, + 0.004034455865621567, + 0.007694812957197428, + 0.00573233375325799, + 0.007064647041261196, + 0.005742114968597889, + 0.005540926940739155, + 0.007961886003613472, + 0.0058889552019536495, + 0.015519359149038792, + 0.036929741501808167, + 0.016988134011626244, + 0.02102678455412388, + 0.022342417389154434, + 0.013266857713460922, + 0.024708010256290436, + 0.01718001253902912 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 4, + 9 + ], + "selected_reward": -0.008648913353681564 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 coniferous forest blanketed in thick snow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive, full-grown white tiger with snow-white fur and deep black stripes visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the tiger explode from behind a clump of snow-bent bushes in a leaping motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video include an extreme close-up shot of the tiger's icy blue eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video show an extended front paw with four razor-sharp claws emerging from the pad?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04306419938802719 + ], + "opposite_logprob_tokens": [ + -3.1680641174316406 + ], + "logprob_reward": -0.04306419938802719, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04306419938802719, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05911637470126152, + 0.0925309807062149, + 0.10046914964914322, + 0.15863275527954102, + 0.1483922004699707, + 0.054368045181035995, + 0.09268967062234879, + 0.11472085118293762, + 0.027678705751895905, + 0.15140119194984436 + ], + "credit_map": [ + 0.00682632252573967, + 0.00892492663115263, + 0.007919219322502613, + 0.0069089545868337154, + 0.006639413069933653, + 0.006515697576105595, + 0.008409042842686176, + 0.006972804199904203, + 0.009524337016046047, + 0.01785019040107727, + 0.009549438953399658, + 0.015262426808476448, + 0.008385936729609966, + 0.009028731845319271, + 0.009736698120832443, + 0.013193229213356972, + 0.006679938640445471, + 0.013799705542623997, + 0.018946953117847443, + 0.022733943536877632, + 0.006584860384464264, + 0.008643326349556446, + 0.008595716208219528, + 0.014484724961221218, + 0.01138300634920597, + 0.0189296193420887, + 0.015997199341654778, + 0.025669945403933525, + 0.026121150702238083, + 0.016351481899619102, + 0.027331186458468437, + 0.016849184408783913, + 0.013206643052399158, + 0.0161896925419569, + 0.023755976930260658, + 0.017561936751008034, + 0.018283652141690254, + 0.016246125102043152, + 0.022583235055208206, + 0.020564956590533257, + 0.0028279265388846397, + 0.0042473855428397655, + 0.004237127490341663, + 0.005469462368637323, + 0.0064615351147949696, + 0.011953805573284626, + 0.008294092491269112, + 0.010876714251935482, + 0.006884097587317228, + 0.01123106013983488, + 0.010022848844528198, + 0.01279749721288681, + 0.009336636401712894, + 0.013754262588918209, + 0.013340231962502003, + 0.015323050320148468, + 0.01519020926207304, + 0.0120050935074687, + 0.018118202686309814, + 0.012963966466486454, + 0.013825678266584873, + 0.01325099915266037, + 0.017572330310940742, + 0.011794373393058777, + 0.0021429797634482384, + 0.004991143476217985, + 0.0031582643277943134, + 0.004717755597084761, + 0.0029956442303955555, + 0.0028992879670113325, + 0.0039958106353878975, + 0.002777822082862258, + 0.014479871839284897, + 0.03093015030026436, + 0.0167775247246027, + 0.02115298993885517, + 0.0200035460293293, + 0.012835529632866383, + 0.02002803608775139, + 0.015193559229373932 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 9, + 4 + ], + "selected_reward": -0.04306419938802719 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 coniferous forest blanketed in thick snow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive, full-grown white tiger with snow-white fur and deep black stripes visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the tiger explode from behind a clump of snow-bent bushes in a leaping motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video include an extreme close-up shot of the tiger's icy blue eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video show an extended front paw with four razor-sharp claws emerging from the pad?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04864047095179558 + ], + "opposite_logprob_tokens": [ + -3.048640489578247 + ], + "logprob_reward": -0.04864047095179558, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04864047095179558, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.023588163778185844, + 0.015432638116180897, + 0.0566096268594265, + 0.025086525827646255, + 0.10165679454803467, + 0.07553344964981079, + 0.12043207883834839, + 0.1913592368364334, + 0.07260984927415848, + 0.3176916539669037 + ], + "credit_map": [ + 0.002593264915049076, + 0.003533543786033988, + 0.003066485049203038, + 0.0024583751801401377, + 0.0027957248967140913, + 0.003235022071748972, + 0.003170969896018505, + 0.002734777983278036, + 0.001161486143246293, + 0.0016902267234399915, + 0.0017421126831322908, + 0.0019061287166550756, + 0.001424547634087503, + 0.0023946149740368128, + 0.0019518746994435787, + 0.0031616457272320986, + 0.003779037157073617, + 0.007609898224473, + 0.006159791257232428, + 0.008307071402668953, + 0.010886150412261486, + 0.006564699113368988, + 0.009016687981784344, + 0.004286289215087891, + 0.0025581999216228724, + 0.003527268534526229, + 0.002979201264679432, + 0.0040422528982162476, + 0.0034318859688937664, + 0.002816001418977976, + 0.003468654118478298, + 0.002263061935082078, + 0.005552302580326796, + 0.00977292563766241, + 0.013910846784710884, + 0.020905470475554466, + 0.012448551133275032, + 0.01394582074135542, + 0.013463703915476799, + 0.011657175607979298, + 0.004944593645632267, + 0.01115391030907631, + 0.005813623312860727, + 0.010253018699586391, + 0.011004584841430187, + 0.009340934455394745, + 0.014736843295395374, + 0.008285940624773502, + 0.012093128636479378, + 0.01421456877142191, + 0.014977658167481422, + 0.016744066029787064, + 0.011770152486860752, + 0.01770663633942604, + 0.01386898010969162, + 0.019056878983974457, + 0.013800205662846565, + 0.015713270753622055, + 0.01297206524759531, + 0.014246029779314995, + 0.01348472572863102, + 0.050277844071388245, + 0.015877285972237587, + 0.05498780682682991, + 0.008206428028643131, + 0.010938792489469051, + 0.007792813237756491, + 0.008306930772960186, + 0.00787670724093914, + 0.008355558849871159, + 0.009276713244616985, + 0.011855901218950748, + 0.027903975918889046, + 0.06227385997772217, + 0.03659398481249809, + 0.036414846777915955, + 0.040745120495557785, + 0.030060822144150734, + 0.03885594382882118, + 0.044843096286058426 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 7, + 6 + ], + "selected_reward": -0.04864047095179558 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 coniferous forest blanketed in thick snow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive, full-grown white tiger with snow-white fur and deep black stripes visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the tiger explode from behind a clump of snow-bent bushes in a leaping motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video include an extreme close-up shot of the tiger's icy blue eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video show an extended front paw with four razor-sharp claws emerging from the pad?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.033685535192489624 + ], + "opposite_logprob_tokens": [ + -3.4086854457855225 + ], + "logprob_reward": -0.033685535192489624, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.033685535192489624, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04094498232007027, + 0.03546309843659401, + 0.06115678325295448, + 0.07853929698467255, + 0.16317804157733917, + 0.03526376932859421, + 0.22345463931560516, + 0.13704882562160492, + 0.08829998224973679, + 0.13665056228637695 + ], + "credit_map": [ + 0.005872076377272606, + 0.006202887278050184, + 0.00558839738368988, + 0.00405975291505456, + 0.005325319245457649, + 0.004743926227092743, + 0.0031321861315518618, + 0.006020434200763702, + 0.003959381487220526, + 0.0077054849825799465, + 0.004030861891806126, + 0.004888987168669701, + 0.003467544913291931, + 0.0031435457058250904, + 0.005218757316470146, + 0.003048535203561187, + 0.00510959280654788, + 0.009240265935659409, + 0.009791671298444271, + 0.013894584961235523, + 0.00566163519397378, + 0.004872191697359085, + 0.005766560323536396, + 0.006820280570536852, + 0.005260992329567671, + 0.007879515178501606, + 0.009305533953011036, + 0.014000310562551022, + 0.011896410025656223, + 0.006866600830107927, + 0.013881931081414223, + 0.009447996504604816, + 0.007783898618072271, + 0.012029014527797699, + 0.01791699044406414, + 0.02034631371498108, + 0.016019122675061226, + 0.030432632192969322, + 0.017196159809827805, + 0.04145392030477524, + 0.003107899334281683, + 0.003955829422920942, + 0.003955983556807041, + 0.00489353621378541, + 0.0038765391800552607, + 0.0050874375738203526, + 0.005735859740525484, + 0.0046506840735673904, + 0.01801457814872265, + 0.03434119373559952, + 0.027386967092752457, + 0.03755566477775574, + 0.02806953713297844, + 0.02394486777484417, + 0.03293726593255997, + 0.02120455913245678, + 0.00914658885449171, + 0.012111269868910313, + 0.009565959684550762, + 0.01120305247604847, + 0.009013929404318333, + 0.03642849996685982, + 0.01084231398999691, + 0.03873720392584801, + 0.008395635522902012, + 0.01280655525624752, + 0.011809295043349266, + 0.012235701084136963, + 0.009184427559375763, + 0.010253826156258583, + 0.012865005992352962, + 0.010749544017016888, + 0.013947493396699429, + 0.024998510256409645, + 0.015643365681171417, + 0.02109566330909729, + 0.01848798803985119, + 0.012579371221363544, + 0.01554159540683031, + 0.014356582425534725 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 4, + 7 + ], + "selected_reward": -0.033685535192489624 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle, medium shot captures a silent coniferous forest blanketed in thick snow. In the diffuse light of a gloomy sky, a massive, full-grown white tiger explodes from behind a clump of snow-bent bushes. Its snow-white fur is streaked with deep black stripes. In its lunge, its muscular body stretches to its limit, its hind legs thrust forward, sending up a shower of powdery snow. The entire movement is imbued with a sense of primal, deadly power. Then, the scene shifts to an extreme close-up, focusing on its icy blue eyes, piercing with a single-minded killing instinct. The camera cuts again, focusing on its extended front paws. Four razor-sharp claws have fully emerged from their thick pads, gleaming like deadly ivory in the chill air. The entire sequence, executed in a breathtakingly short time, creates a heart-stopping tension and a savage beauty.", + "seed": 27342, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000021/group_2/index_6645_seed_27342_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": -0.015612086001783609, + "vlm_logprob": -0.015612086001783609, + "vlm_yes_no_margin": 5.05, + "vlm_token_credit": -0.015612086001783609, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 coniferous forest blanketed in thick snow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive, full-grown white tiger with snow-white fur and deep black stripes visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the tiger explode from behind a clump of snow-bent bushes in a leaping motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video include an extreme close-up shot of the tiger's icy blue eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video show an extended front paw with four razor-sharp claws emerging from the pad?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008441363461315632 + ], + "opposite_logprob_tokens": [ + -7.13344144821167 + ], + "logprob_reward": -0.008441363461315632, + "yes_no_margin_reward": 7.125, + "token_credit_reward": -0.008441363461315632, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1679781973361969, + 0.06676069647073746, + 0.08550938218832016, + 0.0627090334892273, + 0.048750121146440506, + 0.09146934747695923, + 0.12992580235004425, + 0.07549253106117249, + 0.12686917185783386, + 0.14453570544719696 + ], + "credit_map": [ + 0.018722176551818848, + 0.02211209014058113, + 0.014122662134468555, + 0.021262967959046364, + 0.021894510835409164, + 0.024705884978175163, + 0.017656350508332253, + 0.027501575648784637, + 0.008549734950065613, + 0.008998497389256954, + 0.006557854823768139, + 0.01020929217338562, + 0.009476188570261002, + 0.009990151040256023, + 0.005311961285769939, + 0.007667017634958029, + 0.008416996337473392, + 0.01354820467531681, + 0.007656297646462917, + 0.018015390262007713, + 0.010728188790380955, + 0.011164488270878792, + 0.008588891476392746, + 0.007390920072793961, + 0.006673324853181839, + 0.005741537548601627, + 0.006109162699431181, + 0.011544986627995968, + 0.00862663984298706, + 0.008885247632861137, + 0.006540277041494846, + 0.008587859570980072, + 0.0028633077163249254, + 0.00434477161616087, + 0.0060470225289464, + 0.005107575561851263, + 0.005655059590935707, + 0.012869367375969887, + 0.0056839678436517715, + 0.006179045420140028, + 0.006076135206967592, + 0.009789155796170235, + 0.010055912658572197, + 0.009466785937547684, + 0.01949269138276577, + 0.013858471065759659, + 0.014498457312583923, + 0.008231734856963158, + 0.009934978559613228, + 0.01127364207059145, + 0.028438515961170197, + 0.01234263926744461, + 0.01353782881051302, + 0.03078516758978367, + 0.009631730616092682, + 0.013981296680867672, + 0.009393032640218735, + 0.011226627975702286, + 0.011859551072120667, + 0.007280435413122177, + 0.006341993808746338, + 0.012177513912320137, + 0.00840667076408863, + 0.008806711994111538, + 0.02036386914551258, + 0.011475942097604275, + 0.03299558162689209, + 0.01114565972238779, + 0.014381442219018936, + 0.006265686359256506, + 0.02362164668738842, + 0.006619335152208805, + 0.01855369657278061, + 0.01106992270797491, + 0.01956125907599926, + 0.014730589464306831, + 0.021594755351543427, + 0.013261799700558186, + 0.032371263951063156, + 0.01339242234826088 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 6 + ], + "selected_reward": -0.008441363461315632 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 coniferous forest blanketed in thick snow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive, full-grown white tiger with snow-white fur and deep black stripes visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the tiger explode from behind a clump of snow-bent bushes in a leaping motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video include an extreme close-up shot of the tiger's icy blue eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video show an extended front paw with four razor-sharp claws emerging from the pad?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004650491289794445 + ], + "opposite_logprob_tokens": [ + -5.379650592803955 + ], + "logprob_reward": -0.004650491289794445, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004650491289794445, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16702428460121155, + 0.06087450683116913, + 0.03212183713912964, + 0.057473283261060715, + 0.026692159473896027, + 0.037522487342357635, + 0.08180364966392517, + 0.29233428835868835, + 0.1965118795633316, + 0.04764161631464958 + ], + "credit_map": [ + 0.022438563406467438, + 0.025845853611826897, + 0.01662580855190754, + 0.017744068056344986, + 0.022643301635980606, + 0.024098189547657967, + 0.009646277874708176, + 0.027982210740447044, + 0.007807211484760046, + 0.009409934282302856, + 0.005379906855523586, + 0.008272607810795307, + 0.009214187040925026, + 0.008505119010806084, + 0.004513595253229141, + 0.007771950215101242, + 0.0026726480573415756, + 0.004788181744515896, + 0.002609058516100049, + 0.007184224668890238, + 0.004388265311717987, + 0.004427918698638678, + 0.003242407226935029, + 0.002809132682159543, + 0.0038276230916380882, + 0.0071161710657179356, + 0.004877492319792509, + 0.011664850637316704, + 0.008309452794492245, + 0.008303960785269737, + 0.007397580426186323, + 0.00597615260630846, + 0.0023051409516483545, + 0.0026458895299583673, + 0.0027904000598937273, + 0.004037640057504177, + 0.0031601539812982082, + 0.004054834134876728, + 0.003250670153647661, + 0.00444742850959301, + 0.003541631158441305, + 0.00327743380330503, + 0.004898799583315849, + 0.0035050224978476763, + 0.005492085125297308, + 0.004557377193123102, + 0.007892887108027935, + 0.004357248544692993, + 0.007267236243933439, + 0.010861175134778023, + 0.007700328715145588, + 0.008391417562961578, + 0.012807086110115051, + 0.010108865797519684, + 0.01033005677163601, + 0.01433748472481966, + 0.032522693276405334, + 0.07291413843631744, + 0.03376823291182518, + 0.023580608889460564, + 0.02015974558889866, + 0.04639707878232002, + 0.024986637756228447, + 0.03800515830516815, + 0.03218315169215202, + 0.015778537839651108, + 0.05685291811823845, + 0.017261246219277382, + 0.01402227021753788, + 0.00832319539040327, + 0.040744848549366, + 0.011345715261995792, + 0.0073831407353281975, + 0.005516303237527609, + 0.007345116231590509, + 0.0045444234274327755, + 0.007806952111423016, + 0.00473976181820035, + 0.006138589698821306, + 0.004167329519987106 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 8, + 0 + ], + "selected_reward": -0.004650491289794445 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 coniferous forest blanketed in thick snow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive, full-grown white tiger with snow-white fur and deep black stripes visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the tiger explode from behind a clump of snow-bent bushes in a leaping motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video include an extreme close-up shot of the tiger's icy blue eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video show an extended front paw with four razor-sharp claws emerging from the pad?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011080989614129066 + ], + "opposite_logprob_tokens": [ + -4.511081218719482 + ], + "logprob_reward": -0.011080989614129066, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011080989614129066, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15655721724033356, + 0.162602961063385, + 0.030451267957687378, + 0.02588510513305664, + 0.036902666091918945, + 0.07752212136983871, + 0.14007863402366638, + 0.10528683662414551, + 0.21964290738105774, + 0.04507025331258774 + ], + "credit_map": [ + 0.022511927410960197, + 0.020277611911296844, + 0.014986537396907806, + 0.01317742932587862, + 0.02077348716557026, + 0.024824421852827072, + 0.00993236806243658, + 0.03007344901561737, + 0.022436540573835373, + 0.024169014766812325, + 0.01590205915272236, + 0.02277505025267601, + 0.02138126641511917, + 0.022044533863663673, + 0.013633884489536285, + 0.02026062086224556, + 0.0030101167503744364, + 0.004557113628834486, + 0.002251643454656005, + 0.006582696922123432, + 0.004544459283351898, + 0.003868363332003355, + 0.003157256869599223, + 0.002479621209204197, + 0.002813551342114806, + 0.0027805930003523827, + 0.003056894987821579, + 0.005706608295440674, + 0.002874070079997182, + 0.003017480717971921, + 0.0030122988391667604, + 0.002623609732836485, + 0.002798902103677392, + 0.004260313231498003, + 0.00408775731921196, + 0.006313940975815058, + 0.0055938707664608955, + 0.004454163368791342, + 0.0046854279935359955, + 0.004708292428404093, + 0.005739132408052683, + 0.008091774769127369, + 0.014032424427568913, + 0.007235023193061352, + 0.011422696523368359, + 0.014498201198875904, + 0.009477788582444191, + 0.007025084458291531, + 0.013042197562754154, + 0.01868385821580887, + 0.021788612008094788, + 0.014584346674382687, + 0.025194933637976646, + 0.016471518203616142, + 0.016175754368305206, + 0.014137429185211658, + 0.012693753466010094, + 0.011594881303608418, + 0.02102542668581009, + 0.01103463675826788, + 0.007008720189332962, + 0.01755375601351261, + 0.009116376750171185, + 0.015259294770658016, + 0.018793711438775063, + 0.024105122312903404, + 0.03223509341478348, + 0.016776757314801216, + 0.033226534724235535, + 0.02021162211894989, + 0.06048008054494858, + 0.013813999481499195, + 0.006242858711630106, + 0.006077337544411421, + 0.006260269787162542, + 0.003857747884467244, + 0.006244042422622442, + 0.005157647654414177, + 0.005314846057444811, + 0.005915503948926926 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 1, + 0 + ], + "selected_reward": -0.011080989614129066 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 coniferous forest blanketed in thick snow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive, full-grown white tiger with snow-white fur and deep black stripes visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the tiger explode from behind a clump of snow-bent bushes in a leaping motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video include an extreme close-up shot of the tiger's icy blue eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video show an extended front paw with four razor-sharp claws emerging from the pad?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04863865301012993 + ], + "opposite_logprob_tokens": [ + -3.0486385822296143 + ], + "logprob_reward": -0.04863865301012993, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04863865301012993, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.043906714767217636, + 0.025176575407385826, + 0.008267207071185112, + 0.015706868842244148, + 0.02179720252752304, + 0.0328345000743866, + 0.14329703152179718, + 0.04542098566889763, + 0.6336750984191895, + 0.029917754232883453 + ], + "credit_map": [ + 0.0052169798873364925, + 0.006049629766494036, + 0.004594506695866585, + 0.004190254956483841, + 0.0064483326859772205, + 0.006955398712307215, + 0.002450208645313978, + 0.008001407608389854, + 0.0032719236332923174, + 0.004072804003953934, + 0.0023755726870149374, + 0.0035185504239052534, + 0.003498709760606289, + 0.0034991675056517124, + 0.0021237567998468876, + 0.0028160950168967247, + 0.0009193369769491255, + 0.001267573214136064, + 0.0010194290662184358, + 0.001032617292366922, + 0.0009754315833561122, + 0.0014796897303313017, + 0.0007060556672513485, + 0.0008670740062370896, + 0.0015895682154223323, + 0.001764697371982038, + 0.0018173420103266835, + 0.0021628697868436575, + 0.0018464192980900407, + 0.00250031566247344, + 0.001924601849168539, + 0.0021010565105825663, + 0.001758898375555873, + 0.002763967728242278, + 0.0025318681728094816, + 0.0038630813360214233, + 0.0030172637198120356, + 0.0033147064968943596, + 0.0027019719127565622, + 0.0018454458331689239, + 0.0019142362289130688, + 0.0021753706969320774, + 0.007388222962617874, + 0.00180986023042351, + 0.0030183405615389347, + 0.008884336799383163, + 0.004981451202183962, + 0.0026626852340996265, + 0.012132001109421253, + 0.016401346772909164, + 0.028195377439260483, + 0.013391280546784401, + 0.01932448521256447, + 0.02731304056942463, + 0.012637061066925526, + 0.013902445323765278, + 0.005307137966156006, + 0.004313236102461815, + 0.008517819456756115, + 0.0035214845556765795, + 0.003321971744298935, + 0.009814045391976833, + 0.004232823848724365, + 0.006392468698322773, + 0.09741249680519104, + 0.04389874264597893, + 0.1885012984275818, + 0.0461408793926239, + 0.04502040147781372, + 0.021438341587781906, + 0.15730412304401398, + 0.03395886346697807, + 0.0041049448773264885, + 0.002770491410046816, + 0.004078486934304237, + 0.002769357757642865, + 0.0038921344093978405, + 0.00423637218773365, + 0.005216415040194988, + 0.002849554643034935 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 6, + 7 + ], + "selected_reward": -0.04863865301012993 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 coniferous forest blanketed in thick snow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive, full-grown white tiger with snow-white fur and deep black stripes visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the tiger explode from behind a clump of snow-bent bushes in a leaping motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video include an extreme close-up shot of the tiger's icy blue eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video show an extended front paw with four razor-sharp claws emerging from the pad?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"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": [ + 0.05540401488542557, + 0.07276738435029984, + 0.029598765075206757, + 0.052195560187101364, + 0.037073686718940735, + 0.07089316844940186, + 0.08748966455459595, + 0.17464715242385864, + 0.2328670471906662, + 0.18706361949443817 + ], + "credit_map": [ + 0.00750864390283823, + 0.007001264952123165, + 0.004945788998156786, + 0.01038335170596838, + 0.006280845962464809, + 0.007712777704000473, + 0.004749728832393885, + 0.006821618415415287, + 0.009386559017002583, + 0.011793646961450577, + 0.006424634717404842, + 0.009201005101203918, + 0.011347677558660507, + 0.010672517120838165, + 0.005729837343096733, + 0.008211510255932808, + 0.002280954271554947, + 0.003369709476828575, + 0.002148978877812624, + 0.004567083902657032, + 0.006175342481583357, + 0.005366926081478596, + 0.0026341110933572054, + 0.003055661218240857, + 0.006566240452229977, + 0.004610027652233839, + 0.004996233619749546, + 0.005826478358358145, + 0.0073740421794354916, + 0.00737422052770853, + 0.004855117294937372, + 0.010593201965093613, + 0.0033965916372835636, + 0.003373885527253151, + 0.004223441705107689, + 0.005141817033290863, + 0.004567941650748253, + 0.007554964162409306, + 0.0044487700797617435, + 0.004366276320070028, + 0.007008674554526806, + 0.0075663309544324875, + 0.009771730750799179, + 0.0077025750651955605, + 0.010879778303205967, + 0.010345756076276302, + 0.011752997525036335, + 0.005865324288606644, + 0.006784058641642332, + 0.007730050943791866, + 0.02609199471771717, + 0.005719333421438932, + 0.006133484188467264, + 0.018708141520619392, + 0.006515141110867262, + 0.009807459078729153, + 0.014006742276251316, + 0.04645143449306488, + 0.02450256422162056, + 0.016379915177822113, + 0.013194072991609573, + 0.019197067245841026, + 0.01901116780936718, + 0.02190418541431427, + 0.04502122849225998, + 0.02129439078271389, + 0.06518537551164627, + 0.016001205891370773, + 0.018992740660905838, + 0.01962299272418022, + 0.025912903249263763, + 0.020836198702454567, + 0.023477444425225258, + 0.023498307913541794, + 0.0197268296033144, + 0.016394436359405518, + 0.03538437932729721, + 0.023010464385151863, + 0.02497105300426483, + 0.0206006970256567 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 7 + ], + "selected_reward": -0.005248932633548975 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle, medium shot captures a silent coniferous forest blanketed in thick snow. In the diffuse light of a gloomy sky, a massive, full-grown white tiger explodes from behind a clump of snow-bent bushes. Its snow-white fur is streaked with deep black stripes. In its lunge, its muscular body stretches to its limit, its hind legs thrust forward, sending up a shower of powdery snow. The entire movement is imbued with a sense of primal, deadly power. Then, the scene shifts to an extreme close-up, focusing on its icy blue eyes, piercing with a single-minded killing instinct. The camera cuts again, focusing on its extended front paws. Four razor-sharp claws have fully emerged from their thick pads, gleaming like deadly ivory in the chill air. The entire sequence, executed in a breathtakingly short time, creates a heart-stopping tension and a savage beauty.", + "seed": 27343, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000021/group_2/index_6645_seed_27343_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": -0.016863813158124685, + "vlm_logprob": -0.016863813158124685, + "vlm_yes_no_margin": 4.825, + "vlm_token_credit": -0.016863813158124685, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 coniferous forest blanketed in thick snow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive, full-grown white tiger with snow-white fur and deep black stripes visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the tiger explode from behind a clump of snow-bent bushes in a leaping motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video include an extreme close-up shot of the tiger's icy blue eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video show an extended front paw with four razor-sharp claws emerging from the pad?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007298710290342569 + ], + "opposite_logprob_tokens": [ + -7.507298946380615 + ], + "logprob_reward": -0.007298710290342569, + "yes_no_margin_reward": 7.5, + "token_credit_reward": -0.007298710290342569, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20180165767669678, + 0.0858166366815567, + 0.10390479117631912, + 0.10892503708600998, + 0.058125682175159454, + 0.0497581884264946, + 0.1155323013663292, + 0.07450835406780243, + 0.10806473344564438, + 0.09356261789798737 + ], + "credit_map": [ + 0.018340421840548515, + 0.02538357302546501, + 0.0266700629144907, + 0.02987905591726303, + 0.03512534126639366, + 0.02566061168909073, + 0.021219950169324875, + 0.019522638991475105, + 0.006156153045594692, + 0.011025059968233109, + 0.004694699309766293, + 0.010946108028292656, + 0.013611975125968456, + 0.01382047776132822, + 0.01238767709583044, + 0.013174479827284813, + 0.005281314719468355, + 0.007105846423655748, + 0.004339332226663828, + 0.0052513377740979195, + 0.008529890328645706, + 0.01223183423280716, + 0.02621980383992195, + 0.03494542837142944, + 0.006617015227675438, + 0.011486221104860306, + 0.006041189190000296, + 0.015999173745512962, + 0.013973149470984936, + 0.015208600088953972, + 0.02167326770722866, + 0.017926417291164398, + 0.009278688579797745, + 0.007984638214111328, + 0.004216915462166071, + 0.006174459122121334, + 0.008397453464567661, + 0.006580631714314222, + 0.008475033566355705, + 0.007017860654741526, + 0.00510746706277132, + 0.003985775634646416, + 0.004911767318844795, + 0.008397860452532768, + 0.005982700269669294, + 0.005765692330896854, + 0.0058442093431949615, + 0.009762711822986603, + 0.008813843131065369, + 0.01042962446808815, + 0.013858732767403126, + 0.008960902690887451, + 0.01117306761443615, + 0.02849559858441353, + 0.008406585082411766, + 0.02539394609630108, + 0.007849159650504589, + 0.011488002724945545, + 0.00579107366502285, + 0.006472093518823385, + 0.013536197133362293, + 0.009820583276450634, + 0.010583971627056599, + 0.008967269212007523, + 0.007551724091172218, + 0.0042579383589327335, + 0.009262182749807835, + 0.005434755701571703, + 0.007353451102972031, + 0.03289976343512535, + 0.008188930340111256, + 0.03311597555875778, + 0.01766921579837799, + 0.008569400757551193, + 0.014080092310905457, + 0.010330098681151867, + 0.00843790266662836, + 0.00867516826838255, + 0.011444809846580029, + 0.014355912804603577 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 3 + ], + "selected_reward": -0.007298710290342569 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 coniferous forest blanketed in thick snow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive, full-grown white tiger with snow-white fur and deep black stripes visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the tiger explode from behind a clump of snow-bent bushes in a leaping motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video include an extreme close-up shot of the tiger's icy blue eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video show an extended front paw with four razor-sharp claws emerging from the pad?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004111765883862972 + ], + "opposite_logprob_tokens": [ + -5.5041117668151855 + ], + "logprob_reward": -0.004111765883862972, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004111765883862972, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08339055627584457, + 0.0546523742377758, + 0.037061095237731934, + 0.029287362471222878, + 0.02209232561290264, + 0.07637311518192291, + 0.16400787234306335, + 0.0902489647269249, + 0.3895629942417145, + 0.053323354572057724 + ], + "credit_map": [ + 0.008385999128222466, + 0.012196856550872326, + 0.0076610613614320755, + 0.019364504143595695, + 0.009397774003446102, + 0.011300948448479176, + 0.00847178976982832, + 0.006611621007323265, + 0.006657542195171118, + 0.005804043263196945, + 0.003682284615933895, + 0.007725891191512346, + 0.011316805146634579, + 0.007070664782077074, + 0.0071113877929747105, + 0.005283755715936422, + 0.0026613855734467506, + 0.004571783356368542, + 0.0021342646796256304, + 0.006038513034582138, + 0.005540269427001476, + 0.005223504733294249, + 0.005934313405305147, + 0.004957062657922506, + 0.0028309240005910397, + 0.0028243432752788067, + 0.0026644300669431686, + 0.003881503129377961, + 0.004090611357241869, + 0.0042261830531060696, + 0.004570626653730869, + 0.004198742099106312, + 0.002339596627280116, + 0.0021636146120727062, + 0.0021464598830789328, + 0.0030233992729336023, + 0.003067762590944767, + 0.0029523309785872698, + 0.0032178182154893875, + 0.0031813443638384342, + 0.006485534366220236, + 0.011455945670604706, + 0.005411388352513313, + 0.00832831859588623, + 0.018026873469352722, + 0.00737261027097702, + 0.010535487905144691, + 0.00875695701688528, + 0.010882658883929253, + 0.012677078135311604, + 0.015001707710325718, + 0.012328794226050377, + 0.014594973996281624, + 0.04295601695775986, + 0.01249003503471613, + 0.04307660460472107, + 0.00937782134860754, + 0.009601621888577938, + 0.013565407134592533, + 0.00977335311472416, + 0.01336322445422411, + 0.011666183359920979, + 0.01208377256989479, + 0.010817582719027996, + 0.021163810044527054, + 0.017219344154000282, + 0.02169605903327465, + 0.016732582822442055, + 0.020919527858495712, + 0.13339677453041077, + 0.020033210515975952, + 0.13840167224407196, + 0.007233820389956236, + 0.006499901879578829, + 0.009596690535545349, + 0.0052036396227777, + 0.006128168199211359, + 0.004727659281343222, + 0.007116522639989853, + 0.0068169510923326015 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 6, + 7 + ], + "selected_reward": -0.004111765883862972 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 coniferous forest blanketed in thick snow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive, full-grown white tiger with snow-white fur and deep black stripes visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the tiger explode from behind a clump of snow-bent bushes in a leaping motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video include an extreme close-up shot of the tiger's icy blue eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video show an extended front paw with four razor-sharp claws emerging from the pad?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026349911466240883 + ], + "opposite_logprob_tokens": [ + -3.6513500213623047 + ], + "logprob_reward": -0.026349911466240883, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026349911466240883, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13816089928150177, + 0.1003410667181015, + 0.09331528097391129, + 0.12714634835720062, + 0.06241511180996895, + 0.059067580848932266, + 0.11216749995946884, + 0.06570494174957275, + 0.18830999732017517, + 0.0533713772892952 + ], + "credit_map": [ + 0.02016742154955864, + 0.016187705099582672, + 0.021301880478858948, + 0.019308075308799744, + 0.016698481515049934, + 0.016758384183049202, + 0.014234114438295364, + 0.013504808768630028, + 0.007012173067778349, + 0.012584421783685684, + 0.010920917615294456, + 0.010470355860888958, + 0.013699966482818127, + 0.01874459907412529, + 0.01162657793611288, + 0.015282033011317253, + 0.005581855773925781, + 0.009090239182114601, + 0.006163459271192551, + 0.017947658896446228, + 0.012602058239281178, + 0.013333726674318314, + 0.013143894262611866, + 0.015452374704182148, + 0.007177903316915035, + 0.013007525354623795, + 0.00704965740442276, + 0.013511582277715206, + 0.02510163187980652, + 0.015358909033238888, + 0.031553350389003754, + 0.014385771006345749, + 0.0049458397552371025, + 0.007515300065279007, + 0.00435804994776845, + 0.010985036380589008, + 0.009309755638241768, + 0.007232304196804762, + 0.009137424640357494, + 0.008931389078497887, + 0.006184662692248821, + 0.00910333264619112, + 0.0052528176456689835, + 0.00961331743746996, + 0.01104812603443861, + 0.006211332511156797, + 0.006731509696692228, + 0.004922472406178713, + 0.007999417372047901, + 0.014898824505507946, + 0.010031656362116337, + 0.015391144901514053, + 0.013355880044400692, + 0.015557908453047276, + 0.014548404142260551, + 0.02038424089550972, + 0.008743084035813808, + 0.006034568417817354, + 0.008503234945237637, + 0.0052334824576973915, + 0.005887266714125872, + 0.01361948810517788, + 0.007513497024774551, + 0.010170304216444492, + 0.008536268025636673, + 0.006514469161629677, + 0.007119797170162201, + 0.008891907520592213, + 0.007875892333686352, + 0.06881885975599289, + 0.009201562963426113, + 0.07135121524333954, + 0.009181831032037735, + 0.004121528007090092, + 0.012200761586427689, + 0.005341001320630312, + 0.004673581570386887, + 0.004839715547859669, + 0.006721435114741325, + 0.006291515659540892 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 3 + ], + "selected_reward": -0.026349911466240883 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 coniferous forest blanketed in thick snow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive, full-grown white tiger with snow-white fur and deep black stripes visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the tiger explode from behind a clump of snow-bent bushes in a leaping motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video include an extreme close-up shot of the tiger's icy blue eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video show an extended front paw with four razor-sharp claws emerging from the pad?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023279689252376556 + ], + "opposite_logprob_tokens": [ + -3.7732796669006348 + ], + "logprob_reward": -0.023279689252376556, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023279689252376556, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.02637588419020176, + 0.01883157715201378, + 0.030865609645843506, + 0.03579180687665939, + 0.03859233483672142, + 0.03935394063591957, + 0.10191933065652847, + 0.27379530668258667, + 0.19125467538833618, + 0.2432195097208023 + ], + "credit_map": [ + 0.002689134795218706, + 0.0031241453252732754, + 0.0032574983779340982, + 0.003644530661404133, + 0.00426447531208396, + 0.003080694703385234, + 0.0033035073429346085, + 0.003011900233104825, + 0.0023239434231072664, + 0.002875284058973193, + 0.0017327748937532306, + 0.002508078468963504, + 0.0018812029156833887, + 0.002619331469759345, + 0.0027280778158456087, + 0.0021628858521580696, + 0.0023044736590236425, + 0.003484569024294615, + 0.0020437845960259438, + 0.003985823132097721, + 0.0034206388518214226, + 0.0060739051550626755, + 0.004035762511193752, + 0.005516653414815664, + 0.0021252010483294725, + 0.004299899563193321, + 0.002270145108923316, + 0.005768412724137306, + 0.004136971198022366, + 0.004553105682134628, + 0.00557273905724287, + 0.007065335288643837, + 0.005286086350679398, + 0.002803131239488721, + 0.003926552832126617, + 0.004729534965008497, + 0.004992998205125332, + 0.004786944482475519, + 0.00606349715963006, + 0.006003595422953367, + 0.006202552001923323, + 0.0037258253432810307, + 0.005474770907312632, + 0.006774047389626503, + 0.0034589378628879786, + 0.00535106286406517, + 0.003557776566594839, + 0.004808968398720026, + 0.007340048439800739, + 0.010855152271687984, + 0.009105187840759754, + 0.00941361952573061, + 0.011498762294650078, + 0.022892603650689125, + 0.009089077822864056, + 0.02172488532960415, + 0.02791641280055046, + 0.029629815369844437, + 0.03095129504799843, + 0.026147030293941498, + 0.047854579985141754, + 0.03923623636364937, + 0.03623221069574356, + 0.035827748477458954, + 0.015428006649017334, + 0.01842041313648224, + 0.02076774649322033, + 0.01892288401722908, + 0.01707090623676777, + 0.03899678960442543, + 0.017460111528635025, + 0.04418783262372017, + 0.04216675087809563, + 0.015027021989226341, + 0.048896048218011856, + 0.023420846089720726, + 0.017716290429234505, + 0.024453170597553253, + 0.03110675700008869, + 0.0404326431453228 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 9, + 8 + ], + "selected_reward": -0.023279689252376556 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a coniferous forest blanketed in thick snow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive, full-grown white tiger with snow-white fur and deep black stripes visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the tiger explode from behind a clump of snow-bent bushes in a leaping motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video include an extreme close-up shot of the tiger's icy blue eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video show an extended front paw with four razor-sharp claws emerging from the pad?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023278988897800446 + ], + "opposite_logprob_tokens": [ + -3.7732789516448975 + ], + "logprob_reward": -0.023278988897800446, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023278988897800446, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07817396521568298, + 0.05236024782061577, + 0.04242977127432823, + 0.040108710527420044, + 0.04493523761630058, + 0.0540456585586071, + 0.16902963817119598, + 0.15628042817115784, + 0.1413416564464569, + 0.22129477560520172 + ], + "credit_map": [ + 0.010139308869838715, + 0.010955369099974632, + 0.011544491164386272, + 0.00880499742925167, + 0.011752142570912838, + 0.009277808479964733, + 0.007390616461634636, + 0.008309222757816315, + 0.005111902952194214, + 0.006470308639109135, + 0.004200357478111982, + 0.008737707510590553, + 0.008378498256206512, + 0.007820258848369122, + 0.007646102458238602, + 0.003995103295892477, + 0.002905916655436158, + 0.0041162543930113316, + 0.0028546478133648634, + 0.007368588820099831, + 0.0038617923855781555, + 0.007921448908746243, + 0.008069386705756187, + 0.005331730004400015, + 0.0029644565656781197, + 0.0036958663258701563, + 0.003401042427867651, + 0.0041230591014027596, + 0.007326621562242508, + 0.005087913013994694, + 0.00865351315587759, + 0.004856230691075325, + 0.006086151115596294, + 0.00507204607129097, + 0.003820871701464057, + 0.006239597219973803, + 0.004991724621504545, + 0.005600358825176954, + 0.00593221839517355, + 0.007192264776676893, + 0.0048219067975878716, + 0.007139123976230621, + 0.004470539279282093, + 0.005714131519198418, + 0.012348719872534275, + 0.006020242348313332, + 0.007192832417786121, + 0.006338156759738922, + 0.010753664188086987, + 0.01625756174325943, + 0.014283747412264347, + 0.016139565035700798, + 0.01809372380375862, + 0.03881393000483513, + 0.012745261192321777, + 0.041942156851291656, + 0.022170789539813995, + 0.011134638451039791, + 0.015081595629453659, + 0.01121545396745205, + 0.015990717336535454, + 0.03953886032104492, + 0.014177828095853329, + 0.026970520615577698, + 0.01290989201515913, + 0.019618291407823563, + 0.018764417618513107, + 0.01673566736280918, + 0.0186616238206625, + 0.019438063725829124, + 0.018427863717079163, + 0.016785820946097374, + 0.031015802174806595, + 0.01301337219774723, + 0.06611832976341248, + 0.020528031513094902, + 0.016993746161460876, + 0.019466936588287354, + 0.028897136449813843, + 0.02526138722896576 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 6, + 7 + ], + "selected_reward": -0.023278988897800446 + } + ], + "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/0000021/group_3/index_tensor([7672, 7672])_group_3_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000021/group_3/index_tensor([7672, 7672])_group_3_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..c8f12f9b874057816fb4cfbffa1f3f511e759c7d --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000021/group_3/index_tensor([7672, 7672])_group_3_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "low-angle, wide-angle shot, the camera panning erratically to follow the violent movements, captures two enormous crocodiles locked in a fierce battle in the heart of a swampy hellscape. One creature, its skin the color of damp obsidian, is a behemoth, its ancient, scarred scales gleaming in the dim, oppressive light. Its jaws clamp down on the other, a slightly smaller but equally ferocious rival, its deep green-black body entwining and lashing out with primal force. They churn the murky, algae-choked water into a violent foam, their powerful tails lashing the surface in a savage struggle for dominance. The realistic physics simulation renders every powerful splash and spray with astonishing clarity, as geysers of brown water erupt with each thrashing motion. As one crocodile sprints savagely in a death roll, a slight motion blur follows the incredibly fast turns, emphasizing the raw, chaotic power of the confrontation. The scene is illuminated by a sickly light filtering through the dense, gnarled cypress canopy, creating a terrifying atmosphere of primal brutality and inner struggle.", + "seed": 33098, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000021/group_3/index_7672_seed_33098_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": -0.4240318536758423, + "vlm_logprob": -0.4240318536758423, + "vlm_yes_no_margin": 1.0, + "vlm_token_credit": -0.4240318536758423, + "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 enormous crocodiles visible locked in a battle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is one crocodile's skin the color of damp obsidian with ancient, scarred scales?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the other crocodile slightly smaller with a deep green-black body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a swampy hellscape with murky, algae-choked water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan erratically to follow the violent movements?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.525869607925415 + ], + "opposite_logprob_tokens": [ + -0.900869607925415 + ], + "logprob_reward": -0.525869607925415, + "yes_no_margin_reward": 0.375, + "token_credit_reward": -0.525869607925415, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16480521857738495, + 0.049133893102407455, + 0.057877302169799805, + 0.13999438285827637, + 0.08534295856952667, + 0.15282145142555237, + 0.16124005615711212, + 0.049027640372514725, + 0.09199140965938568, + 0.04776567220687866 + ], + "credit_map": [ + 0.020266827195882797, + 0.024884356185793877, + 0.023594040423631668, + 0.016457024961709976, + 0.027973758056759834, + 0.023533927276730537, + 0.013474419713020325, + 0.014620868489146233, + 0.00270021497271955, + 0.003337146481499076, + 0.0030417575035244226, + 0.0033091718796640635, + 0.01341673918068409, + 0.01281615998595953, + 0.005199882201850414, + 0.005312821362167597, + 0.005417944863438606, + 0.0056910207495093346, + 0.010481923818588257, + 0.008161180652678013, + 0.006098561454564333, + 0.006075282581150532, + 0.008333415724337101, + 0.007617971859872341, + 0.017829380929470062, + 0.02002418413758278, + 0.014325817115604877, + 0.018517019227147102, + 0.018641285598278046, + 0.02492096833884716, + 0.015548248775303364, + 0.010187480598688126, + 0.01061017531901598, + 0.010239602997899055, + 0.012295085936784744, + 0.012471312656998634, + 0.008136301301419735, + 0.00963340513408184, + 0.009512358345091343, + 0.012444715015590191, + 0.022451359778642654, + 0.025240838527679443, + 0.014377974905073643, + 0.01845964975655079, + 0.010214250534772873, + 0.028465112671256065, + 0.011769403703510761, + 0.021842868998646736, + 0.026995128020644188, + 0.02067188359797001, + 0.01854810304939747, + 0.014408405870199203, + 0.015087855979800224, + 0.029058771207928658, + 0.016356125473976135, + 0.020113782957196236, + 0.007507257163524628, + 0.006609307136386633, + 0.006142798811197281, + 0.005376958288252354, + 0.004631138872355223, + 0.007740096189081669, + 0.0034687791485339403, + 0.007551305461674929, + 0.009478824213147163, + 0.015185455791652203, + 0.01068780105561018, + 0.012906578369438648, + 0.013366766273975372, + 0.011273659765720367, + 0.009403427131474018, + 0.009688897989690304, + 0.004912293050438166, + 0.00874997302889824, + 0.003345871577039361, + 0.006706223823130131, + 0.006633011158555746, + 0.007722797803580761, + 0.005531612318009138, + 0.004163888283073902 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 5 + ], + "selected_reward": -0.525869607925415 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" 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 enormous crocodiles visible locked in a battle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is one crocodile's skin the color of damp obsidian with ancient, scarred scales?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the other crocodile slightly smaller with a deep green-black body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a swampy hellscape with murky, algae-choked water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan erratically to follow the violent movements?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07903791218996048 + ], + "opposite_logprob_tokens": [ + -2.57903790473938 + ], + "logprob_reward": -0.07903791218996048, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07903791218996048, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20965352654457092, + 0.06971103698015213, + 0.04134351387619972, + 0.13367880880832672, + 0.06952980905771255, + 0.09512500464916229, + 0.1703232079744339, + 0.09546097368001938, + 0.04581240937113762, + 0.06936172395944595 + ], + "credit_map": [ + 0.024136310443282127, + 0.029421977698802948, + 0.030506080016493797, + 0.024973148480057716, + 0.04145333915948868, + 0.029565805569291115, + 0.014608001336455345, + 0.014988861046731472, + 0.004623078275471926, + 0.007390997838228941, + 0.005065197125077248, + 0.006636220496147871, + 0.01882132887840271, + 0.02039836347103119, + 0.004000999499112368, + 0.0027748539578169584, + 0.004349490161985159, + 0.0049444264732301235, + 0.006495287176221609, + 0.0064868382178246975, + 0.006028760224580765, + 0.004752123728394508, + 0.0050006769597530365, + 0.003285908605903387, + 0.020041655749082565, + 0.012592483311891556, + 0.012671067379415035, + 0.01639070361852646, + 0.01633014716207981, + 0.03251877799630165, + 0.014671767130494118, + 0.00846220925450325, + 0.010166789405047894, + 0.006954733747988939, + 0.010388270951807499, + 0.010029153898358345, + 0.006986453663557768, + 0.007767781149595976, + 0.008070699870586395, + 0.009165927767753601, + 0.01530105248093605, + 0.014150434173643589, + 0.010424193926155567, + 0.011518392711877823, + 0.006905944086611271, + 0.01576799526810646, + 0.007321542594581842, + 0.013735451735556126, + 0.026768147945404053, + 0.028498418629169464, + 0.019434204325079918, + 0.020802652463316917, + 0.015564142726361752, + 0.021589461714029312, + 0.014967973344027996, + 0.022698191925883293, + 0.017932135611772537, + 0.016438059508800507, + 0.009867350570857525, + 0.007570234127342701, + 0.0077697234228253365, + 0.012392999604344368, + 0.006874969229102135, + 0.016615498811006546, + 0.007575439754873514, + 0.008527419529855251, + 0.004481207579374313, + 0.005637194495648146, + 0.006687941029667854, + 0.004996830131858587, + 0.004222478251904249, + 0.0036838988307863474, + 0.008174275048077106, + 0.010168299078941345, + 0.005381901282817125, + 0.007251212373375893, + 0.0120240468531847, + 0.014029254205524921, + 0.007612632587552071, + 0.004720102995634079 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 3 + ], + "selected_reward": -0.07903791218996048 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" 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 enormous crocodiles visible locked in a battle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is one crocodile's skin the color of damp obsidian with ancient, scarred scales?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the other crocodile slightly smaller with a deep green-black body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a swampy hellscape with murky, algae-choked water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan erratically to follow the violent movements?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11293809860944748 + ], + "opposite_logprob_tokens": [ + -2.237938165664673 + ], + "logprob_reward": -0.11293809860944748, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11293809860944748, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20969222486019135, + 0.13673162460327148, + 0.04400051385164261, + 0.0602632500231266, + 0.08158739656209946, + 0.1126636415719986, + 0.16171501576900482, + 0.06378483772277832, + 0.05633543059229851, + 0.07322608679533005 + ], + "credit_map": [ + 0.025778118520975113, + 0.027753500267863274, + 0.03484464809298515, + 0.02372242510318756, + 0.03607147932052612, + 0.02843433991074562, + 0.01660875417292118, + 0.016478965058922768, + 0.007169701624661684, + 0.010017018765211105, + 0.010877467691898346, + 0.015812858939170837, + 0.030416501685976982, + 0.032094087451696396, + 0.013832003809511662, + 0.016511986032128334, + 0.004576785955578089, + 0.006556676235049963, + 0.004626763518899679, + 0.005404322408139706, + 0.007059372030198574, + 0.005288188345730305, + 0.005136171355843544, + 0.00535223726183176, + 0.007027285639196634, + 0.007155437488108873, + 0.009437911212444305, + 0.009739991277456284, + 0.006994121707975864, + 0.007281784899532795, + 0.00835080724209547, + 0.0042759133502841, + 0.010241893120110035, + 0.0087282108142972, + 0.012750412337481976, + 0.015070551075041294, + 0.00973106175661087, + 0.0065553761087358, + 0.01058986410498619, + 0.007920024916529655, + 0.017056677490472794, + 0.02056550420820713, + 0.00794362835586071, + 0.013967047445476055, + 0.010958925820887089, + 0.022204730659723282, + 0.006526430137455463, + 0.013440693728625774, + 0.021795041859149933, + 0.02505517564713955, + 0.02093237079679966, + 0.02145131677389145, + 0.017153004184365273, + 0.02075866051018238, + 0.016017669811844826, + 0.018551776185631752, + 0.01087742019444704, + 0.010855288244783878, + 0.007613523863255978, + 0.004599394742399454, + 0.005439921282231808, + 0.009333164431154728, + 0.005153591278940439, + 0.009912535548210144, + 0.0074923718348145485, + 0.011110897175967693, + 0.006025956477969885, + 0.005463053938001394, + 0.00890449620783329, + 0.006917449180036783, + 0.004875697661191225, + 0.005545508582144976, + 0.008604534901678562, + 0.010791107080876827, + 0.006061248481273651, + 0.01005321554839611, + 0.009978711605072021, + 0.010809935629367828, + 0.01167984213680029, + 0.005247493740171194 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 1 + ], + "selected_reward": -0.11293809860944748 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" 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 enormous crocodiles visible locked in a battle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is one crocodile's skin the color of damp obsidian with ancient, scarred scales?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the other crocodile slightly smaller with a deep green-black body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a swampy hellscape with murky, algae-choked water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan erratically to follow the violent movements?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.5761635899543762 + ], + "opposite_logprob_tokens": [ + -0.8261635899543762 + ], + "logprob_reward": -0.5761635899543762, + "yes_no_margin_reward": 0.25, + "token_credit_reward": -0.5761635899543762, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1814490258693695, + 0.03985648229718208, + 0.0336361788213253, + 0.20244456827640533, + 0.03356333449482918, + 0.1892164796590805, + 0.14475607872009277, + 0.04325828328728676, + 0.06040830910205841, + 0.07141128182411194 + ], + "credit_map": [ + 0.018272925168275833, + 0.033386848866939545, + 0.016763988882303238, + 0.017841598019003868, + 0.035386547446250916, + 0.02949018031358719, + 0.013631504029035568, + 0.0166754312813282, + 0.00386004033498466, + 0.006656538229435682, + 0.004537631291896105, + 0.004553427919745445, + 0.006148169282823801, + 0.005091781262308359, + 0.004180267918854952, + 0.004828627221286297, + 0.0032116661313921213, + 0.005539611913263798, + 0.004724942613393068, + 0.005285287741571665, + 0.0044741639867424965, + 0.004155048169195652, + 0.0033438256941735744, + 0.002901635831221938, + 0.029002459719777107, + 0.021512500941753387, + 0.020647546276450157, + 0.02380429208278656, + 0.023282429203391075, + 0.046843331307172775, + 0.023605698719620705, + 0.013746329583227634, + 0.003787621855735779, + 0.003906591795384884, + 0.003972322680056095, + 0.005181773100048304, + 0.0037526236847043037, + 0.004977979231625795, + 0.004623458255082369, + 0.0033609666861593723, + 0.028737399727106094, + 0.024502184242010117, + 0.023416729643940926, + 0.023870116099715233, + 0.013683011755347252, + 0.031048618257045746, + 0.015874823555350304, + 0.02808360755443573, + 0.020029466599225998, + 0.027549181133508682, + 0.022366417571902275, + 0.019768666476011276, + 0.01427029725164175, + 0.01571773923933506, + 0.010726507753133774, + 0.014327810145914555, + 0.006598494481295347, + 0.006204349920153618, + 0.0047444188967347145, + 0.004182284232228994, + 0.006443451624363661, + 0.007074548397213221, + 0.0032245179172605276, + 0.004786224104464054, + 0.009256701916456223, + 0.013920038938522339, + 0.005934885703027248, + 0.006699681282043457, + 0.010159468278288841, + 0.0051818303763866425, + 0.005086848977953196, + 0.004168858751654625, + 0.010648641735315323, + 0.014576787129044533, + 0.004262594971805811, + 0.006398329045623541, + 0.012879841029644012, + 0.013191604055464268, + 0.00465972488746047, + 0.004793768748641014 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 5, + 0 + ], + "selected_reward": -0.5761635899543762 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" 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 enormous crocodiles visible locked in a battle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is one crocodile's skin the color of damp obsidian with ancient, scarred scales?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the other crocodile slightly smaller with a deep green-black body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a swampy hellscape with murky, algae-choked water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan erratically to follow the violent movements?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"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": [ + -0.8261500597000122 + ], + "opposite_logprob_tokens": [ + -0.5761500597000122 + ], + "logprob_reward": -0.8261500597000122, + "yes_no_margin_reward": -0.25, + "token_credit_reward": -0.8261500597000122, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.08582442253828049, + 0.17799368500709534, + 0.05729956179857254, + 0.11424015462398529, + 0.05027265101671219, + 0.12414047867059708, + 0.1055399626493454, + 0.11371805518865585, + 0.10872258991003036, + 0.06224845349788666 + ], + "credit_map": [ + 0.008868834003806114, + 0.011388474144041538, + 0.013662015087902546, + 0.011473719961941242, + 0.012763438746333122, + 0.012269762344658375, + 0.008103483356535435, + 0.0072947037406265736, + 0.01062886044383049, + 0.017773358151316643, + 0.013574355281889439, + 0.011490162461996078, + 0.044458549469709396, + 0.04448648542165756, + 0.01896577514708042, + 0.016616152599453926, + 0.0052710468880832195, + 0.009424643591046333, + 0.009242489002645016, + 0.007607315666973591, + 0.008301028981804848, + 0.0071453494019806385, + 0.006073861848562956, + 0.004233825020492077, + 0.014786133542656898, + 0.01950051449239254, + 0.009262423031032085, + 0.01695220172405243, + 0.019191842526197433, + 0.015139228664338589, + 0.011439993046224117, + 0.007967832498252392, + 0.006002459675073624, + 0.004860535729676485, + 0.005837146192789078, + 0.0048349881544709206, + 0.0066581955179572105, + 0.005631884094327688, + 0.00831021461635828, + 0.00813723262399435, + 0.018105868250131607, + 0.0212512519210577, + 0.010271408595144749, + 0.01610257290303707, + 0.009338066913187504, + 0.022193899378180504, + 0.00913158431649208, + 0.017745843157172203, + 0.012815923430025578, + 0.013365345075726509, + 0.010028079152107239, + 0.012192511931061745, + 0.01216980442404747, + 0.017540154978632927, + 0.014497890137135983, + 0.012930258177220821, + 0.01654975488781929, + 0.016975972801446915, + 0.015418533235788345, + 0.00973832979798317, + 0.013316451571881771, + 0.011200332082808018, + 0.009635998867452145, + 0.02088269591331482, + 0.017341474071145058, + 0.03275933861732483, + 0.008000108413398266, + 0.0076232487335801125, + 0.01799306832253933, + 0.009818353690207005, + 0.007174622733145952, + 0.008012382313609123, + 0.009192246943712234, + 0.018222693353891373, + 0.00355165870860219, + 0.005173386074602604, + 0.010709947906434536, + 0.008027829229831696, + 0.0037633555475622416, + 0.0036073389928787947 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 5, + 3 + ], + "selected_reward": -0.8261500597000122 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "No" + ] + }, + { + "prompt": "low-angle, wide-angle shot, the camera panning erratically to follow the violent movements, captures two enormous crocodiles locked in a fierce battle in the heart of a swampy hellscape. One creature, its skin the color of damp obsidian, is a behemoth, its ancient, scarred scales gleaming in the dim, oppressive light. Its jaws clamp down on the other, a slightly smaller but equally ferocious rival, its deep green-black body entwining and lashing out with primal force. They churn the murky, algae-choked water into a violent foam, their powerful tails lashing the surface in a savage struggle for dominance. The realistic physics simulation renders every powerful splash and spray with astonishing clarity, as geysers of brown water erupt with each thrashing motion. As one crocodile sprints savagely in a death roll, a slight motion blur follows the incredibly fast turns, emphasizing the raw, chaotic power of the confrontation. The scene is illuminated by a sickly light filtering through the dense, gnarled cypress canopy, creating a terrifying atmosphere of primal brutality and inner struggle.", + "seed": 33099, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000021/group_3/index_7672_seed_33099_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": -0.22931153811514376, + "vlm_logprob": -0.22931153811514376, + "vlm_yes_no_margin": 2.175, + "vlm_token_credit": -0.22931153811514376, + "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 enormous crocodiles visible locked in a battle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is one crocodile's skin the color of damp obsidian with ancient, scarred scales?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the other crocodile slightly smaller with a deep green-black body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a swampy hellscape with murky, algae-choked water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan erratically to follow the violent movements?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.16307035088539124 + ], + "opposite_logprob_tokens": [ + -1.9130703210830688 + ], + "logprob_reward": -0.16307035088539124, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.16307035088539124, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3355214595794678, + 0.09694148600101471, + 0.06446867436170578, + 0.06127254292368889, + 0.04819346219301224, + 0.07418932020664215, + 0.037997204810380936, + 0.14943476021289825, + 0.07142013311386108, + 0.06056094169616699 + ], + "credit_map": [ + 0.054798904806375504, + 0.07054018974304199, + 0.037168558686971664, + 0.037710271775722504, + 0.030153930187225342, + 0.0360141396522522, + 0.047927677631378174, + 0.021207790821790695, + 0.015170897357165813, + 0.009729133918881416, + 0.00914122723042965, + 0.009263777174055576, + 0.012855062261223793, + 0.011844240128993988, + 0.012542243115603924, + 0.016394905745983124, + 0.008385870605707169, + 0.010423089377582073, + 0.012581707909703255, + 0.009494212456047535, + 0.0053862882778048515, + 0.005768164061009884, + 0.006453309208154678, + 0.0059760333970189095, + 0.00961377751082182, + 0.010650165379047394, + 0.006334687117487192, + 0.0064401342533528805, + 0.006385729648172855, + 0.01201539859175682, + 0.004704260732978582, + 0.0051283943466842175, + 0.0064442353323102, + 0.0057169594801962376, + 0.004144143313169479, + 0.003605206497013569, + 0.006211393047124147, + 0.00635245256125927, + 0.009495815262198448, + 0.006223258562386036, + 0.008377309888601303, + 0.008691615425050259, + 0.00888828095048666, + 0.014350835233926773, + 0.007793434429913759, + 0.008052092976868153, + 0.01062061358243227, + 0.007415135391056538, + 0.007588099222630262, + 0.0050827558152377605, + 0.004075216129422188, + 0.0032419776543974876, + 0.005252548959106207, + 0.004429289605468512, + 0.0040828874334692955, + 0.004244430456310511, + 0.014663935638964176, + 0.028674762696027756, + 0.008960849605500698, + 0.015408678911626339, + 0.026614056900143623, + 0.020007997751235962, + 0.019457951188087463, + 0.015646517276763916, + 0.008306914940476418, + 0.011860984377563, + 0.0054645356722176075, + 0.0060126264579594135, + 0.010161705315113068, + 0.009588406421244144, + 0.01179996132850647, + 0.008225005120038986, + 0.004495990928262472, + 0.010137115605175495, + 0.0036021554842591286, + 0.005534634459763765, + 0.00933076348155737, + 0.009730610996484756, + 0.011525924317538738, + 0.006203750614076853 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 1 + ], + "selected_reward": -0.16307035088539124 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" 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 enormous crocodiles visible locked in a battle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is one crocodile's skin the color of damp obsidian with ancient, scarred scales?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the other crocodile slightly smaller with a deep green-black body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a swampy hellscape with murky, algae-choked water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan erratically to follow the violent movements?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02332068420946598 + ], + "opposite_logprob_tokens": [ + -3.7733206748962402 + ], + "logprob_reward": -0.02332068420946598, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.02332068420946598, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22115568816661835, + 0.06715695559978485, + 0.0708787813782692, + 0.1738489866256714, + 0.10021844506263733, + 0.0787988156080246, + 0.0744922012090683, + 0.10514315217733383, + 0.05405150726437569, + 0.05425550788640976 + ], + "credit_map": [ + 0.032134003937244415, + 0.038109567016363144, + 0.021709276363253593, + 0.021060382947325706, + 0.026750104501843452, + 0.026051385328173637, + 0.031015723943710327, + 0.024325255304574966, + 0.01099464762955904, + 0.010162047110497952, + 0.008238319307565689, + 0.0063752043060958385, + 0.008329004049301147, + 0.008794478140771389, + 0.00613571098074317, + 0.008127542212605476, + 0.008663051761686802, + 0.00897535402327776, + 0.013838105835020542, + 0.010385327972471714, + 0.006708231288939714, + 0.006453253794461489, + 0.00780948530882597, + 0.008045968599617481, + 0.028086723759770393, + 0.05216167867183685, + 0.016081342473626137, + 0.0154347512871027, + 0.01771444082260132, + 0.019096480682492256, + 0.0105753717944026, + 0.014698192477226257, + 0.014413547702133656, + 0.0190839022397995, + 0.008586592972278595, + 0.007593623828142881, + 0.01259029470384121, + 0.015070866793394089, + 0.013666965998709202, + 0.009212648496031761, + 0.009470155462622643, + 0.01422182284295559, + 0.008507411926984787, + 0.011699811555445194, + 0.006504514254629612, + 0.012578693218529224, + 0.008583255112171173, + 0.007233147043734789, + 0.01681901142001152, + 0.010544726625084877, + 0.007785798516124487, + 0.007182559464126825, + 0.013027293607592583, + 0.007426092866808176, + 0.0054117352701723576, + 0.006294979713857174, + 0.012779798358678818, + 0.02991347946226597, + 0.005222220905125141, + 0.007068069651722908, + 0.0157941784709692, + 0.016101516783237457, + 0.007417723070830107, + 0.01084616594016552, + 0.007721344940364361, + 0.009298005141317844, + 0.0041981772519648075, + 0.0044892337173223495, + 0.009726395830512047, + 0.008561818860471249, + 0.005666006822139025, + 0.004390527959913015, + 0.004392967093735933, + 0.007661004550755024, + 0.0027384618297219276, + 0.004212181083858013, + 0.013051126152276993, + 0.01194685511291027, + 0.005854797083884478, + 0.004398116376250982 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 7 + ], + "selected_reward": -0.02332068420946598 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" 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 enormous crocodiles visible locked in a battle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is one crocodile's skin the color of damp obsidian with ancient, scarred scales?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the other crocodile slightly smaller with a deep green-black body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a swampy hellscape with murky, algae-choked water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan erratically to follow the violent movements?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03371123969554901 + ], + "opposite_logprob_tokens": [ + -3.4087111949920654 + ], + "logprob_reward": -0.03371123969554901, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03371123969554901, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3010343611240387, + 0.09354045987129211, + 0.06254574656486511, + 0.19150207936763763, + 0.021901633590459824, + 0.07908198237419128, + 0.04962703585624695, + 0.07481052726507187, + 0.08740530908107758, + 0.03855086490511894 + ], + "credit_map": [ + 0.04157494008541107, + 0.04025277867913246, + 0.03410905599594116, + 0.03618953749537468, + 0.03224419429898262, + 0.028292490169405937, + 0.06464670598506927, + 0.023724650964140892, + 0.017125234007835388, + 0.007096727378666401, + 0.006865996867418289, + 0.005727124400436878, + 0.01213669404387474, + 0.0147244269028306, + 0.018064633011817932, + 0.011799628846347332, + 0.008396346122026443, + 0.007662939839065075, + 0.00798526406288147, + 0.007830594666302204, + 0.007535567507147789, + 0.007996588945388794, + 0.00716770114377141, + 0.007970744743943214, + 0.03215699642896652, + 0.06402852386236191, + 0.01430509239435196, + 0.012620732188224792, + 0.02045883983373642, + 0.02061106264591217, + 0.011746549978852272, + 0.015574268996715546, + 0.002839272376149893, + 0.0033699702471494675, + 0.0018880740972235799, + 0.0020319027826189995, + 0.003182078944519162, + 0.0027707968838512897, + 0.0028939347248524427, + 0.0029256027191877365, + 0.00822821632027626, + 0.01720271073281765, + 0.005215763114392757, + 0.006349580362439156, + 0.0040770163759589195, + 0.024754680693149567, + 0.004958029370754957, + 0.00829598680138588, + 0.0111882034689188, + 0.00720932288095355, + 0.004622157663106918, + 0.003982249181717634, + 0.007221026811748743, + 0.005876253824681044, + 0.003245058236643672, + 0.006282764952629805, + 0.008378694765269756, + 0.01947268657386303, + 0.004767134785652161, + 0.005578350275754929, + 0.013899245299398899, + 0.009742428548634052, + 0.005475237034261227, + 0.007496752310544252, + 0.009203104302287102, + 0.013000454753637314, + 0.00897027738392353, + 0.009336471557617188, + 0.011249096132814884, + 0.01105264201760292, + 0.016486650332808495, + 0.008106619119644165, + 0.004370192997157574, + 0.006084851920604706, + 0.0021560140885412693, + 0.003013184294104576, + 0.008523128926753998, + 0.007494224701076746, + 0.004009975586086512, + 0.0028992914594709873 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 1 + ], + "selected_reward": -0.03371123969554901 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" 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 enormous crocodiles visible locked in a battle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is one crocodile's skin the color of damp obsidian with ancient, scarred scales?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the other crocodile slightly smaller with a deep green-black body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a swampy hellscape with murky, algae-choked water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan erratically to follow the violent movements?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10031107813119888 + ], + "opposite_logprob_tokens": [ + -2.350311040878296 + ], + "logprob_reward": -0.10031107813119888, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10031107813119888, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1607358157634735, + 0.05157366394996643, + 0.06178794428706169, + 0.2762538194656372, + 0.0468955896794796, + 0.03727464750409126, + 0.09716207534074783, + 0.18098396062850952, + 0.04024014249444008, + 0.04709231108427048 + ], + "credit_map": [ + 0.041173771023750305, + 0.029810752719640732, + 0.014691528864204884, + 0.011853961274027824, + 0.01633763685822487, + 0.023611893877387047, + 0.013838211074471474, + 0.009418068453669548, + 0.0073691909201443195, + 0.004987621679902077, + 0.004307982511818409, + 0.004035443067550659, + 0.0077627249993383884, + 0.007201792672276497, + 0.00674137007445097, + 0.009167538024485111, + 0.008726646192371845, + 0.009610154666006565, + 0.011633938178420067, + 0.00832535419613123, + 0.004891523625701666, + 0.005386296194046736, + 0.007186222355812788, + 0.006027819123119116, + 0.046514324843883514, + 0.08970580250024796, + 0.020155925303697586, + 0.015523114241659641, + 0.028841473162174225, + 0.03438971936702728, + 0.018153168261051178, + 0.02297031134366989, + 0.007379485759884119, + 0.008961498737335205, + 0.0036221493501216173, + 0.0037773081567138433, + 0.006914661265909672, + 0.007126773241907358, + 0.004873804748058319, + 0.004239910282194614, + 0.004093626979738474, + 0.0058355568908154964, + 0.004169064108282328, + 0.0066091138869524, + 0.003365876618772745, + 0.004814686253666878, + 0.004133721813559532, + 0.004253001883625984, + 0.021109187975525856, + 0.013494996353983879, + 0.010452724993228912, + 0.008940070867538452, + 0.016383886337280273, + 0.010091766715049744, + 0.007163969799876213, + 0.009525489993393421, + 0.025824591517448425, + 0.045397307723760605, + 0.006539243273437023, + 0.00829876959323883, + 0.04020875319838524, + 0.029321864247322083, + 0.012207044288516045, + 0.013186398893594742, + 0.005306190811097622, + 0.007023751735687256, + 0.0031691899057477713, + 0.0024266839027404785, + 0.006829605437815189, + 0.005817660596221685, + 0.0064531355164945126, + 0.003213929245248437, + 0.003629893995821476, + 0.008624876849353313, + 0.002898728009313345, + 0.0024175087455660105, + 0.011265598237514496, + 0.011268998496234417, + 0.004753678105771542, + 0.0022330323699861765 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 7, + 0 + ], + "selected_reward": -0.10031107813119888 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" 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 enormous crocodiles visible locked in a battle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is one crocodile's skin the color of damp obsidian with ancient, scarred scales?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the other crocodile slightly smaller with a deep green-black body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a swampy hellscape with murky, algae-choked water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan erratically to follow the violent movements?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"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": [ + -0.8261443376541138 + ], + "opposite_logprob_tokens": [ + -0.5761443376541138 + ], + "logprob_reward": -0.8261443376541138, + "yes_no_margin_reward": -0.25, + "token_credit_reward": -0.8261443376541138, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.21738256514072418, + 0.07291045784950256, + 0.051304902881383896, + 0.16100764274597168, + 0.15475714206695557, + 0.10188987851142883, + 0.05625234916806221, + 0.09362825006246567, + 0.03447999805212021, + 0.05638681352138519 + ], + "credit_map": [ + 0.045047562569379807, + 0.029875094071030617, + 0.019885459914803505, + 0.03209414333105087, + 0.02737721987068653, + 0.027057088911533356, + 0.015305901877582073, + 0.020740078762173653, + 0.01325055118650198, + 0.00599625613540411, + 0.009423328563570976, + 0.007902400568127632, + 0.00834670104086399, + 0.008219870738685131, + 0.010510769672691822, + 0.009260582737624645, + 0.006607165094465017, + 0.006149597931653261, + 0.005721478722989559, + 0.006336102727800608, + 0.005943411961197853, + 0.006610846146941185, + 0.007767507806420326, + 0.006168794352561235, + 0.02656938135623932, + 0.049667902290821075, + 0.01005575805902481, + 0.011447775177657604, + 0.017437033355236053, + 0.022136496379971504, + 0.011743443086743355, + 0.01194984745234251, + 0.023382525891065598, + 0.028451474383473396, + 0.01506773941218853, + 0.014118174090981483, + 0.01824258826673031, + 0.022029710933566093, + 0.02020687237381935, + 0.013258065097033978, + 0.016853632405400276, + 0.017795348539948463, + 0.012132511474192142, + 0.007121154572814703, + 0.007259081583470106, + 0.018817342817783356, + 0.012962350621819496, + 0.008948465809226036, + 0.005773523356765509, + 0.007580810692161322, + 0.006581311579793692, + 0.0038771899417042732, + 0.01048948708921671, + 0.008903531357645988, + 0.0041673444211483, + 0.008879149332642555, + 0.010293442755937576, + 0.021534372121095657, + 0.005698826629668474, + 0.011395839042961597, + 0.01592033915221691, + 0.012126650661230087, + 0.010284370742738247, + 0.006374404300004244, + 0.004030711017549038, + 0.0062806252390146255, + 0.0026920735836029053, + 0.0031300929840654135, + 0.006317895837128162, + 0.005225207656621933, + 0.0036358556244522333, + 0.0031675403006374836, + 0.006770807318389416, + 0.009152449667453766, + 0.004250970669090748, + 0.006142919417470694, + 0.007158915512263775, + 0.006985538639128208, + 0.009350691922008991, + 0.006574524100869894 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 4 + ], + "selected_reward": -0.8261443376541138 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "No" + ] + }, + { + "prompt": "low-angle, wide-angle shot, the camera panning erratically to follow the violent movements, captures two enormous crocodiles locked in a fierce battle in the heart of a swampy hellscape. One creature, its skin the color of damp obsidian, is a behemoth, its ancient, scarred scales gleaming in the dim, oppressive light. Its jaws clamp down on the other, a slightly smaller but equally ferocious rival, its deep green-black body entwining and lashing out with primal force. They churn the murky, algae-choked water into a violent foam, their powerful tails lashing the surface in a savage struggle for dominance. The realistic physics simulation renders every powerful splash and spray with astonishing clarity, as geysers of brown water erupt with each thrashing motion. As one crocodile sprints savagely in a death roll, a slight motion blur follows the incredibly fast turns, emphasizing the raw, chaotic power of the confrontation. The scene is illuminated by a sickly light filtering through the dense, gnarled cypress canopy, creating a terrifying atmosphere of primal brutality and inner struggle.", + "seed": 33100, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000021/group_3/index_7672_seed_33100_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": -0.16711126565933226, + "vlm_logprob": -0.16711126565933226, + "vlm_yes_no_margin": 2.15, + "vlm_token_credit": -0.16711126565933226, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" 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 enormous crocodiles visible locked in a battle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is one crocodile's skin the color of damp obsidian with ancient, scarred scales?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the other crocodile slightly smaller with a deep green-black body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a swampy hellscape with murky, algae-choked water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan erratically to follow the violent movements?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2568666934967041 + ], + "opposite_logprob_tokens": [ + -1.506866693496704 + ], + "logprob_reward": -0.2568666934967041, + "yes_no_margin_reward": 1.25, + "token_credit_reward": -0.2568666934967041, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14210182428359985, + 0.10979461669921875, + 0.05889525264501572, + 0.040611036121845245, + 0.2181285321712494, + 0.07305599749088287, + 0.06536638736724854, + 0.08251729607582092, + 0.053728118538856506, + 0.155800923705101 + ], + "credit_map": [ + 0.014840266667306423, + 0.015288974158465862, + 0.011112631298601627, + 0.01617695949971676, + 0.01913287304341793, + 0.023381488397717476, + 0.0225361455231905, + 0.019632486626505852, + 0.009931689128279686, + 0.012296090833842754, + 0.01382008008658886, + 0.011543484404683113, + 0.015628736466169357, + 0.015935014933347702, + 0.019393835216760635, + 0.01124569308012724, + 0.0066039077937603, + 0.007111314684152603, + 0.007233500014990568, + 0.008439216762781143, + 0.006872115191072226, + 0.00897457730025053, + 0.007974380627274513, + 0.005686243996024132, + 0.003921543247997761, + 0.0036150391679257154, + 0.00782453641295433, + 0.004755998495966196, + 0.005378513131290674, + 0.005819966085255146, + 0.006127653177827597, + 0.003167787566781044, + 0.02367841638624668, + 0.023484468460083008, + 0.0350966714322567, + 0.05254630744457245, + 0.018950533121824265, + 0.02245485968887806, + 0.02254224568605423, + 0.019375042989850044, + 0.007143223658204079, + 0.008188656531274319, + 0.005151721648871899, + 0.010997419245541096, + 0.01219792477786541, + 0.011242127977311611, + 0.01187125314027071, + 0.006263671442866325, + 0.0052668810822069645, + 0.007075032684952021, + 0.00700085936114192, + 0.005647290963679552, + 0.010617676191031933, + 0.015879355370998383, + 0.005711542442440987, + 0.008167754858732224, + 0.009443371556699276, + 0.01499338448047638, + 0.015300544910132885, + 0.010208917781710625, + 0.01186694297939539, + 0.009523062035441399, + 0.005801326595246792, + 0.005379756446927786, + 0.008245761506259441, + 0.009588992223143578, + 0.007065959740430117, + 0.004748847801238298, + 0.006250156555324793, + 0.0056328545324504375, + 0.005577224772423506, + 0.0066183265298604965, + 0.012327010743319988, + 0.03285963460803032, + 0.013805144466459751, + 0.014469752088189125, + 0.023558303713798523, + 0.022053351625800133, + 0.010647895745933056, + 0.02607983537018299 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 9, + 0 + ], + "selected_reward": -0.2568666934967041 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" 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 enormous crocodiles visible locked in a battle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is one crocodile's skin the color of damp obsidian with ancient, scarred scales?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the other crocodile slightly smaller with a deep green-black body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a swampy hellscape with murky, algae-choked water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan erratically to follow the violent movements?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.043100737035274506 + ], + "opposite_logprob_tokens": [ + -3.1681008338928223 + ], + "logprob_reward": -0.043100737035274506, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.043100737035274506, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2590426504611969, + 0.1695951223373413, + 0.06278558820486069, + 0.07084376364946365, + 0.07508005946874619, + 0.04709663614630699, + 0.06629445403814316, + 0.05270816758275032, + 0.10413966327905655, + 0.09241385757923126 + ], + "credit_map": [ + 0.04937794804573059, + 0.051565177738666534, + 0.014015459455549717, + 0.022211436182260513, + 0.026889832690358162, + 0.037384141236543655, + 0.028656259179115295, + 0.028942406177520752, + 0.014048855751752853, + 0.017658328637480736, + 0.02032473124563694, + 0.020370619371533394, + 0.02632296085357666, + 0.024495474994182587, + 0.02659534476697445, + 0.019778801128268242, + 0.005316037684679031, + 0.008519673720002174, + 0.008287152275443077, + 0.007564925588667393, + 0.008490001782774925, + 0.009327275678515434, + 0.007934375666081905, + 0.00734614348039031, + 0.007756427396088839, + 0.006921500898897648, + 0.010290248319506645, + 0.00782209262251854, + 0.009411466307938099, + 0.010795319452881813, + 0.010402309708297253, + 0.00744440034031868, + 0.01226868573576212, + 0.010303118266165257, + 0.006151006557047367, + 0.008749857544898987, + 0.00932407844811678, + 0.013786004856228828, + 0.009203968569636345, + 0.005293345544487238, + 0.006814141292124987, + 0.0060789878480136395, + 0.0043233176693320274, + 0.00870007835328579, + 0.00671162037178874, + 0.004588050302118063, + 0.007184657268226147, + 0.0026957797817885876, + 0.0052824742160737514, + 0.006292570382356644, + 0.005227755755186081, + 0.003568108193576336, + 0.009723814204335213, + 0.014491318725049496, + 0.0139081459492445, + 0.007800271734595299, + 0.004788236692547798, + 0.006460133008658886, + 0.006329104769974947, + 0.007082051131874323, + 0.007816432975232601, + 0.008118825033307076, + 0.005840845871716738, + 0.006272540427744389, + 0.017414594069123268, + 0.022104911506175995, + 0.014507549814879894, + 0.009760749526321888, + 0.014756666496396065, + 0.013541381806135178, + 0.0059011853300035, + 0.006152629852294922, + 0.009943525306880474, + 0.0175461508333683, + 0.011160015128552914, + 0.009541724808514118, + 0.018953535705804825, + 0.00939584244042635, + 0.008174048736691475, + 0.007699012756347656 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 8 + ], + "selected_reward": -0.043100737035274506 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" 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 enormous crocodiles visible locked in a battle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is one crocodile's skin the color of damp obsidian with ancient, scarred scales?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the other crocodile slightly smaller with a deep green-black body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a swampy hellscape with murky, algae-choked water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan erratically to follow the violent movements?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020586393773555756 + ], + "opposite_logprob_tokens": [ + -3.8955864906311035 + ], + "logprob_reward": -0.020586393773555756, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020586393773555756, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2869037389755249, + 0.13877204060554504, + 0.06450363248586655, + 0.0356927327811718, + 0.08467792719602585, + 0.07549242675304413, + 0.10561363399028778, + 0.06246587261557579, + 0.06327686458826065, + 0.08260112255811691 + ], + "credit_map": [ + 0.0446171760559082, + 0.04496801272034645, + 0.02722928114235401, + 0.02134876325726509, + 0.030246613547205925, + 0.037876930087804794, + 0.04530239850282669, + 0.03531460091471672, + 0.012801632285118103, + 0.01851707324385643, + 0.015111060813069344, + 0.01676941104233265, + 0.02228846773505211, + 0.018886692821979523, + 0.013987747021019459, + 0.02040996588766575, + 0.005912372376769781, + 0.010629796423017979, + 0.007969279773533344, + 0.006001899018883705, + 0.0096108578145504, + 0.009070845320820808, + 0.006939722690731287, + 0.008368865586817265, + 0.003935321234166622, + 0.0042030829936265945, + 0.003819988574832678, + 0.00415764981880784, + 0.005533665418624878, + 0.0051856255158782005, + 0.004830033052712679, + 0.004027368035167456, + 0.010464743711054325, + 0.00915401428937912, + 0.011993291787803173, + 0.018883178010582924, + 0.0064124478958547115, + 0.01233216654509306, + 0.007962048053741455, + 0.0074760462157428265, + 0.010772235691547394, + 0.010864677838981152, + 0.004575854167342186, + 0.007061596494168043, + 0.022591326385736465, + 0.009600086137652397, + 0.0057274093851447105, + 0.00429924763739109, + 0.00907624326646328, + 0.011015820316970348, + 0.008675078861415386, + 0.00749534647911787, + 0.014385195448994637, + 0.021487262099981308, + 0.021228132769465446, + 0.012250557541847229, + 0.005808006972074509, + 0.006617490667849779, + 0.00983321201056242, + 0.0074707199819386005, + 0.009242554195225239, + 0.010079458355903625, + 0.007433224003762007, + 0.005981207825243473, + 0.007263106293976307, + 0.010913979262113571, + 0.010054117068648338, + 0.006481740158051252, + 0.010188058950006962, + 0.008163671009242535, + 0.004729392472654581, + 0.005482807755470276, + 0.010391097515821457, + 0.01886649802327156, + 0.00969234760850668, + 0.007166972383856773, + 0.011208266951143742, + 0.010432976298034191, + 0.006795819383114576, + 0.008047153241932392 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 6 + ], + "selected_reward": -0.020586393773555756 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" 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 enormous crocodiles visible locked in a battle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is one crocodile's skin the color of damp obsidian with ancient, scarred scales?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the other crocodile slightly smaller with a deep green-black body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a swampy hellscape with murky, algae-choked water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan erratically to follow the violent movements?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.20155708491802216 + ], + "opposite_logprob_tokens": [ + -1.7015570402145386 + ], + "logprob_reward": -0.20155708491802216, + "yes_no_margin_reward": 1.5, + "token_credit_reward": -0.20155708491802216, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11333959549665451, + 0.12710675597190857, + 0.06239788234233856, + 0.09682077169418335, + 0.07559531927108765, + 0.07860762625932693, + 0.17967362701892853, + 0.14575590193271637, + 0.04539787396788597, + 0.07530461251735687 + ], + "credit_map": [ + 0.0165183674544096, + 0.014397419057786465, + 0.010536315850913525, + 0.01072965282946825, + 0.013441549614071846, + 0.01640423573553562, + 0.015889473259449005, + 0.01542259007692337, + 0.01154475286602974, + 0.012635619379580021, + 0.011919129639863968, + 0.015674453228712082, + 0.019539259374141693, + 0.016686510294675827, + 0.016054054722189903, + 0.02305297553539276, + 0.007870402187108994, + 0.007686402648687363, + 0.006626949179917574, + 0.006403819657862186, + 0.010531696490943432, + 0.009517211467027664, + 0.006206371821463108, + 0.007555036805570126, + 0.006625651381909847, + 0.008048808202147484, + 0.013715403154492378, + 0.014804108068346977, + 0.0117914704605937, + 0.01388590969145298, + 0.018374469131231308, + 0.0095749506726861, + 0.011826303787529469, + 0.008322157897055149, + 0.005484062246978283, + 0.008571721613407135, + 0.008931695483624935, + 0.01664453186094761, + 0.010201367549598217, + 0.005613488610833883, + 0.014117885380983353, + 0.011469841934740543, + 0.0053178295493125916, + 0.013386656530201435, + 0.009708687663078308, + 0.007491820026189089, + 0.01156676933169365, + 0.00554813863709569, + 0.01649666205048561, + 0.015090785920619965, + 0.008945086039602757, + 0.009702075272798538, + 0.028526000678539276, + 0.04868680611252785, + 0.03649391606450081, + 0.015732301399111748, + 0.011383426375687122, + 0.012211140245199203, + 0.018830491229891777, + 0.017768913879990578, + 0.01885042153298855, + 0.022711753845214844, + 0.024889707565307617, + 0.01911005936563015, + 0.006646332331001759, + 0.00708359619602561, + 0.0075111715123057365, + 0.00457099499180913, + 0.005664541851729155, + 0.004944392014294863, + 0.004806706216186285, + 0.004170142579823732, + 0.007856874726712704, + 0.012641200795769691, + 0.009811658412218094, + 0.010457189753651619, + 0.015672536566853523, + 0.007722512818872929, + 0.006906186696141958, + 0.0042364648543298244 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 7, + 1 + ], + "selected_reward": -0.20155708491802216 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" 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 enormous crocodiles visible locked in a battle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is one crocodile's skin the color of damp obsidian with ancient, scarred scales?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the other crocodile slightly smaller with a deep green-black body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a swampy hellscape with murky, algae-choked water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan erratically to follow the violent movements?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.31344541907310486 + ], + "opposite_logprob_tokens": [ + -1.3134454488754272 + ], + "logprob_reward": -0.31344541907310486, + "yes_no_margin_reward": 1.0, + "token_credit_reward": -0.31344541907310486, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05470889434218407, + 0.11357728391885757, + 0.1159733235836029, + 0.10860342532396317, + 0.1567584127187729, + 0.09091878682374954, + 0.04154834896326065, + 0.2317580282688141, + 0.026877386495471, + 0.0592760406434536 + ], + "credit_map": [ + 0.0075717587023973465, + 0.009340958669781685, + 0.007774706929922104, + 0.005986467003822327, + 0.0061662327498197556, + 0.006912211887538433, + 0.004880733322352171, + 0.006075823679566383, + 0.011016960255801678, + 0.01043735072016716, + 0.012840128503739834, + 0.013347197324037552, + 0.019085440784692764, + 0.016560468822717667, + 0.012094059959053993, + 0.018195681273937225, + 0.013101335614919662, + 0.01765313185751438, + 0.012482975609600544, + 0.012157456949353218, + 0.016048312187194824, + 0.016326872631907463, + 0.01295471005141735, + 0.015248538926243782, + 0.009040752425789833, + 0.010494301095604897, + 0.015294035896658897, + 0.014758704230189323, + 0.012826196849346161, + 0.014116336591541767, + 0.019169190898537636, + 0.012903906404972076, + 0.01951405592262745, + 0.018891872838139534, + 0.014789791777729988, + 0.019468342885375023, + 0.02813151851296425, + 0.02378484420478344, + 0.023470789194107056, + 0.008707207627594471, + 0.009365960024297237, + 0.012286057695746422, + 0.006949604954570532, + 0.013466713018715382, + 0.019649973139166832, + 0.012107483111321926, + 0.010542292147874832, + 0.00655070086941123, + 0.0052620465867221355, + 0.004864788148552179, + 0.0028799520805478096, + 0.0041955821216106415, + 0.005885554477572441, + 0.005903066601604223, + 0.008503854274749756, + 0.0040535046719014645, + 0.02231748029589653, + 0.03160801902413368, + 0.03224685415625572, + 0.02530434913933277, + 0.03923514857888222, + 0.03171372041106224, + 0.026855159550905228, + 0.022477300837635994, + 0.002591664670035243, + 0.004100800957530737, + 0.004225634504109621, + 0.0028966099489480257, + 0.004267269279807806, + 0.0034504993818700314, + 0.002749974373728037, + 0.002594933845102787, + 0.005454091355204582, + 0.006544674281030893, + 0.00885070487856865, + 0.008946249261498451, + 0.007781228516250849, + 0.007768307812511921, + 0.006931179668754339, + 0.0069996053352952 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 4, + 2 + ], + "selected_reward": -0.31344541907310486 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle, wide-angle shot, the camera panning erratically to follow the violent movements, captures two enormous crocodiles locked in a fierce battle in the heart of a swampy hellscape. One creature, its skin the color of damp obsidian, is a behemoth, its ancient, scarred scales gleaming in the dim, oppressive light. Its jaws clamp down on the other, a slightly smaller but equally ferocious rival, its deep green-black body entwining and lashing out with primal force. They churn the murky, algae-choked water into a violent foam, their powerful tails lashing the surface in a savage struggle for dominance. The realistic physics simulation renders every powerful splash and spray with astonishing clarity, as geysers of brown water erupt with each thrashing motion. As one crocodile sprints savagely in a death roll, a slight motion blur follows the incredibly fast turns, emphasizing the raw, chaotic power of the confrontation. The scene is illuminated by a sickly light filtering through the dense, gnarled cypress canopy, creating a terrifying atmosphere of primal brutality and inner struggle.", + "seed": 33101, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000021/group_3/index_7672_seed_33101_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": -0.1862863540649414, + "vlm_logprob": -0.1862863540649414, + "vlm_yes_no_margin": 1.95, + "vlm_token_credit": -0.1862863540649414, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" 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 enormous crocodiles visible locked in a battle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is one crocodile's skin the color of damp obsidian with ancient, scarred scales?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the other crocodile slightly smaller with a deep green-black body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a swampy hellscape with murky, algae-choked water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan erratically to follow the violent movements?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.39074090123176575 + ], + "opposite_logprob_tokens": [ + -1.1407408714294434 + ], + "logprob_reward": -0.39074090123176575, + "yes_no_margin_reward": 0.75, + "token_credit_reward": -0.39074090123176575, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19638410210609436, + 0.06229589506983757, + 0.04986754432320595, + 0.1207529604434967, + 0.02710239589214325, + 0.18116219341754913, + 0.09321537613868713, + 0.029009222984313965, + 0.20474818348884583, + 0.035462118685245514 + ], + "credit_map": [ + 0.01893359050154686, + 0.028485605493187904, + 0.021793274208903313, + 0.019869515672326088, + 0.033199477940797806, + 0.02491452358663082, + 0.027722854167222977, + 0.021465273573994637, + 0.005534762050956488, + 0.008284704759716988, + 0.007131623104214668, + 0.007811954244971275, + 0.012897285632789135, + 0.006819215137511492, + 0.006980780512094498, + 0.006835563573986292, + 0.004900910425931215, + 0.004995953757315874, + 0.006416891235858202, + 0.006025801878422499, + 0.008803041651844978, + 0.007230664603412151, + 0.005406429525464773, + 0.006087849847972393, + 0.008788984268903732, + 0.02762829326093197, + 0.004262540023773909, + 0.018731236457824707, + 0.026715688407421112, + 0.023300137370824814, + 0.0068068490363657475, + 0.004519233945757151, + 0.0036351692397147417, + 0.003130431519821286, + 0.003983491100370884, + 0.0030675046145915985, + 0.005178940482437611, + 0.0030484951566904783, + 0.0030476769898086786, + 0.002010688418522477, + 0.0338427759706974, + 0.020338034257292747, + 0.01125267893075943, + 0.01388780027627945, + 0.02469126507639885, + 0.028288863599300385, + 0.03392106667160988, + 0.014939709566533566, + 0.009185203351080418, + 0.006788521073758602, + 0.003074126085266471, + 0.020605331286787987, + 0.01246854942291975, + 0.025535723194479942, + 0.009623745456337929, + 0.005934170912951231, + 0.0024234766606241465, + 0.004718168638646603, + 0.002374660922214389, + 0.004356388933956623, + 0.004031873308122158, + 0.00575258256867528, + 0.0028144835960119963, + 0.0025375899858772755, + 0.013670402579009533, + 0.02912665717303753, + 0.008927165530622005, + 0.010258217342197895, + 0.0773141086101532, + 0.037495505064725876, + 0.012403933331370354, + 0.015552198514342308, + 0.0026422301307320595, + 0.004164494574069977, + 0.003524743253365159, + 0.0026017248164862394, + 0.004264544695615768, + 0.008114614523947239, + 0.0035528112202882767, + 0.006596959661692381 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 5 + ], + "selected_reward": -0.39074090123176575 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" 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 enormous crocodiles visible locked in a battle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is one crocodile's skin the color of damp obsidian with ancient, scarred scales?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the other crocodile slightly smaller with a deep green-black body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a swampy hellscape with murky, algae-choked water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan erratically to follow the violent movements?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.062094442546367645 + ], + "opposite_logprob_tokens": [ + -2.8120944499969482 + ], + "logprob_reward": -0.062094442546367645, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.062094442546367645, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20501117408275604, + 0.038012221455574036, + 0.043007250875234604, + 0.215994730591774, + 0.028963206335902214, + 0.10484500974416733, + 0.08992870151996613, + 0.04146619886159897, + 0.10665405541658401, + 0.12611743807792664 + ], + "credit_map": [ + 0.019429417327046394, + 0.03313649818301201, + 0.021896708756685257, + 0.02159838005900383, + 0.02990773506462574, + 0.02628118358552456, + 0.02999134361743927, + 0.022769909352064133, + 0.0038404238875955343, + 0.008909009397029877, + 0.0022869508247822523, + 0.002267693867906928, + 0.010761874727904797, + 0.0048046354204416275, + 0.0026195223908871412, + 0.002522111404687166, + 0.003919610287994146, + 0.00423573749139905, + 0.004715403076261282, + 0.00594518193975091, + 0.008486389182507992, + 0.005933129228651524, + 0.004799012094736099, + 0.004972788505256176, + 0.016277983784675598, + 0.037907663732767105, + 0.006527260411530733, + 0.042942486703395844, + 0.04321148619055748, + 0.05155164375901222, + 0.009944302029907703, + 0.007631909567862749, + 0.003966054413467646, + 0.0033209072425961494, + 0.004477380774915218, + 0.0033594074193388224, + 0.0038217089604586363, + 0.0038013553712517023, + 0.0034336172975599766, + 0.0027827746234834194, + 0.01576627418398857, + 0.014116915874183178, + 0.011113238520920277, + 0.009277978911995888, + 0.017532488331198692, + 0.012030608020722866, + 0.017431674525141716, + 0.00757582625374198, + 0.011228011921048164, + 0.006500230636447668, + 0.002662974176928401, + 0.01655033603310585, + 0.012848315760493279, + 0.027613770216703415, + 0.007659698370844126, + 0.004865364637225866, + 0.004024185705929995, + 0.005626368802040815, + 0.0029851498547941446, + 0.004236555192619562, + 0.007718909066170454, + 0.007534155622124672, + 0.005532529205083847, + 0.003808344481512904, + 0.008716668002307415, + 0.01631924696266651, + 0.003863012185320258, + 0.003961754962801933, + 0.039739374071359634, + 0.016850324347615242, + 0.005753467790782452, + 0.011450216174125671, + 0.010401256382465363, + 0.007671920582652092, + 0.009253046475350857, + 0.03152360022068024, + 0.011570541188120842, + 0.03483746200799942, + 0.013567466288805008, + 0.007292142603546381 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 9 + ], + "selected_reward": -0.062094442546367645 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" 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 enormous crocodiles visible locked in a battle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is one crocodile's skin the color of damp obsidian with ancient, scarred scales?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the other crocodile slightly smaller with a deep green-black body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a swampy hellscape with murky, algae-choked water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan erratically to follow the violent movements?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.038113825023174286 + ], + "opposite_logprob_tokens": [ + -3.288113832473755 + ], + "logprob_reward": -0.038113825023174286, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.038113825023174286, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13478028774261475, + 0.015346815809607506, + 0.02948674187064171, + 0.23376356065273285, + 0.05370871722698212, + 0.3529307544231415, + 0.06293974071741104, + 0.0342717245221138, + 0.03354974091053009, + 0.04922192171216011 + ], + "credit_map": [ + 0.013664564117789268, + 0.02024216577410698, + 0.014947295188903809, + 0.013438675552606583, + 0.021994341164827347, + 0.017780307680368423, + 0.018159599974751472, + 0.014553336426615715, + 0.0019222330302000046, + 0.0019093144219368696, + 0.002065758453682065, + 0.001460638246499002, + 0.0022682116832584143, + 0.002762230345979333, + 0.0014841665979474783, + 0.0014742636121809483, + 0.0025686149019747972, + 0.0029445441905409098, + 0.002230651443824172, + 0.005471526179462671, + 0.004310167394578457, + 0.003750985022634268, + 0.005154841113835573, + 0.0030554113909602165, + 0.01671164482831955, + 0.03564414009451866, + 0.007097250781953335, + 0.054306842386722565, + 0.03971950337290764, + 0.061746351420879364, + 0.010595197789371014, + 0.007942630909383297, + 0.005938882939517498, + 0.005777408368885517, + 0.006556410808116198, + 0.00562316132709384, + 0.010319259017705917, + 0.008286966010928154, + 0.00680305901914835, + 0.004403566941618919, + 0.07116908580064774, + 0.04417593404650688, + 0.03649327903985977, + 0.025167137384414673, + 0.060624342411756516, + 0.05193100497126579, + 0.043444618582725525, + 0.01992536522448063, + 0.015962403267621994, + 0.005975024309009314, + 0.0022248809691518545, + 0.0038384601939469576, + 0.007717215456068516, + 0.022771187126636505, + 0.0023329646792262793, + 0.002117602387443185, + 0.0029909564182162285, + 0.0042360564693808556, + 0.0026353111024945974, + 0.004356703720986843, + 0.006557830609381199, + 0.0054334430024027824, + 0.004864148795604706, + 0.0031972783617675304, + 0.0031473515555262566, + 0.004603502806276083, + 0.003630709135904908, + 0.0026774185243993998, + 0.009586763568222523, + 0.005194235127419233, + 0.0026737854350358248, + 0.002035972662270069, + 0.004333086311817169, + 0.00402959156781435, + 0.003390486817806959, + 0.01415245234966278, + 0.003896726295351982, + 0.009988178499042988, + 0.004655756521970034, + 0.00477564474567771 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 3, + 0 + ], + "selected_reward": -0.038113825023174286 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" 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 enormous crocodiles visible locked in a battle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is one crocodile's skin the color of damp obsidian with ancient, scarred scales?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the other crocodile slightly smaller with a deep green-black body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a swampy hellscape with murky, algae-choked water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan erratically to follow the violent movements?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.12703587114810944 + ], + "opposite_logprob_tokens": [ + -2.1270358562469482 + ], + "logprob_reward": -0.12703587114810944, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.12703587114810944, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.188600093126297, + 0.012685944326221943, + 0.02898026444017887, + 0.2045218050479889, + 0.03305372595787048, + 0.1103152334690094, + 0.16844654083251953, + 0.031405847519636154, + 0.018537066876888275, + 0.20345348119735718 + ], + "credit_map": [ + 0.019782880321145058, + 0.027738535776734352, + 0.020004186779260635, + 0.017855031415820122, + 0.03157168999314308, + 0.02873990871012211, + 0.02489209733903408, + 0.018015751615166664, + 0.0014095318038016558, + 0.0022440562024712563, + 0.0009935108246281743, + 0.0008680108585394919, + 0.002641925122588873, + 0.002150212647393346, + 0.0010907649993896484, + 0.0012879313435405493, + 0.002881888998672366, + 0.0026914216578006744, + 0.0037372938822954893, + 0.003628327278420329, + 0.004740542266517878, + 0.0046461764723062515, + 0.003118265885859728, + 0.0035363484639674425, + 0.016699878498911858, + 0.03339838236570358, + 0.005967572331428528, + 0.038021355867385864, + 0.04700368642807007, + 0.04521869868040085, + 0.010679194703698158, + 0.007533038966357708, + 0.003912605810910463, + 0.003806989174336195, + 0.00472196564078331, + 0.003541212063282728, + 0.00687636760994792, + 0.005022610537707806, + 0.0028879635501652956, + 0.002284014830365777, + 0.015403896570205688, + 0.012880941852927208, + 0.010420037433505058, + 0.008504767902195454, + 0.027846360579133034, + 0.015496926382184029, + 0.013586987741291523, + 0.006175316870212555, + 0.01913272589445114, + 0.011491449549794197, + 0.005140985362231731, + 0.03409775346517563, + 0.022678015753626823, + 0.049783170223236084, + 0.01600911095738411, + 0.010113324970006943, + 0.0027473806403577328, + 0.004932640586048365, + 0.002411594381555915, + 0.0034743770956993103, + 0.004622689448297024, + 0.006183810997754335, + 0.004023951478302479, + 0.0030094049870967865, + 0.0018361101392656565, + 0.0030208565294742584, + 0.001242628088220954, + 0.0015343929408118129, + 0.0043396600522100925, + 0.002242481801658869, + 0.0016141803935170174, + 0.002706756116822362, + 0.016947511583566666, + 0.011885176412761211, + 0.015522642061114311, + 0.05700412392616272, + 0.01660492643713951, + 0.052776485681533813, + 0.02367655374109745, + 0.0090360501781106 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 9, + 0 + ], + "selected_reward": -0.12703587114810944 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" 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 enormous crocodiles visible locked in a battle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is one crocodile's skin the color of damp obsidian with ancient, scarred scales?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the other crocodile slightly smaller with a deep green-black body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a swampy hellscape with murky, algae-choked water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan erratically to follow the violent movements?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.3134467303752899 + ], + "opposite_logprob_tokens": [ + -1.3134467601776123 + ], + "logprob_reward": -0.3134467303752899, + "yes_no_margin_reward": 1.0, + "token_credit_reward": -0.3134467303752899, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.053124841302633286, + 0.04315371811389923, + 0.017955033108592033, + 0.07140754163265228, + 0.12021049857139587, + 0.1694178432226181, + 0.08973336964845657, + 0.010282290168106556, + 0.20530344545841217, + 0.2194114774465561 + ], + "credit_map": [ + 0.006074708886444569, + 0.006542535033077002, + 0.005906257312744856, + 0.004641575273126364, + 0.01073672715574503, + 0.006748783402144909, + 0.006960312370210886, + 0.005513940006494522, + 0.004348342306911945, + 0.007020395714789629, + 0.004601878579705954, + 0.004467013757675886, + 0.01034278329461813, + 0.0046727778390049934, + 0.004566783085465431, + 0.0031337409745901823, + 0.0024367566220462322, + 0.0020642036106437445, + 0.0021009715273976326, + 0.0026068035513162613, + 0.002757916459813714, + 0.002092973329126835, + 0.0021440491545945406, + 0.0017513589700683951, + 0.00488289212808013, + 0.013166818767786026, + 0.0030141703318804502, + 0.012467271648347378, + 0.015555648133158684, + 0.01244803611189127, + 0.005811687093228102, + 0.004061013925820589, + 0.02266036719083786, + 0.011792897246778011, + 0.009881635196506977, + 0.012500148266553879, + 0.02564256452023983, + 0.016541998833417892, + 0.012153464369475842, + 0.009037411771714687, + 0.037229958921670914, + 0.02422458492219448, + 0.01021071057766676, + 0.013113098219037056, + 0.021398324519395828, + 0.02658885344862938, + 0.025870012119412422, + 0.01078229770064354, + 0.01155804842710495, + 0.0062562162056565285, + 0.0027816984802484512, + 0.015797164291143417, + 0.012490266002714634, + 0.028284838423132896, + 0.007567958906292915, + 0.004997167736291885, + 0.0008137584663927555, + 0.001850214321166277, + 0.0009532264084555209, + 0.0015268755378201604, + 0.0012623242801055312, + 0.0011607934720814228, + 0.001884352182969451, + 0.0008307445095852017, + 0.014783688820898533, + 0.029672035947442055, + 0.0065994723699986935, + 0.007112087216228247, + 0.0775887593626976, + 0.03480861335992813, + 0.01326768472790718, + 0.02147109992802143, + 0.019737577065825462, + 0.013889101333916187, + 0.015851614996790886, + 0.06277377158403397, + 0.018216866999864578, + 0.060039032250642776, + 0.019658951088786125, + 0.009244555607438087 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 5 + ], + "selected_reward": -0.3134467303752899 + } + ], + "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/0000021/group_4/index_tensor([3494, 3494])_group_4_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000021/group_4/index_tensor([3494, 3494])_group_4_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..07045e0e6e18dfdf6eff87bb14557bc64687f6b4 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000021/group_4/index_tensor([3494, 3494])_group_4_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "close-up shot from a level angle, the camera held still to capture the drastic change about to take place. At the center of the frame is a dense digital cover, interwoven with countless luminescent emerald binary codes and ancient magical runes shimmering with a mysterious golden glow. These streams of data flow vertically like cascading cascades, forming a complex barrier suspended against a pure black void. Suddenly, a blinding beam of pure white energy erupts from the very center of the frame, instantly cleaving the digital cover with the force of thunder. Following this devastating blast of light, the entire cover disintegrates into countless luminous fragments of code and runes. These fragments radiate outward at high speed, leaving bright trails in the air and forming a spectacular digital storm. The entire scene, rendered in real-time by the engine, presents a cinematic CG feel with ultra-high-definition resolution, making every scattered number and rune clearly visible with sharp edges. The burst of strong light creates extremely rich volumetric lighting effects in the darkness, and holy beams penetrate the diffuse digital dust, injecting unparalleled visual impact and epic atmosphere into this moment of digital destruction full of power and beauty.", + "seed": 24950, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000021/group_4/index_3494_seed_24950_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": -0.08468791646882892, + "vlm_logprob": -0.08468791646882892, + "vlm_yes_no_margin": 3.625, + "vlm_token_credit": -0.08468791646882892, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a close-up from a level angle held still?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a dense digital cover interwoven with emerald binary codes and golden magical runes visible at the center?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a blinding beam of pure white energy erupt from the center and cleave the digital cover?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the disintegrated fragments radiate outward at high speed leaving bright trails and forming a digital storm?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a pure black void with rich volumetric lighting effects and sharp edges on the numbers and runes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.34759005904197693 + ], + "opposite_logprob_tokens": [ + -1.3475900888442993 + ], + "logprob_reward": -0.34759005904197693, + "yes_no_margin_reward": 1.0, + "token_credit_reward": -0.34759005904197693, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09562687575817108, + 0.3111828863620758, + 0.06306859105825424, + 0.14208953082561493, + 0.07097987830638885, + 0.041267599910497665, + 0.029067952185869217, + 0.06190449371933937, + 0.060011811554431915, + 0.12480039894580841 + ], + "credit_map": [ + 0.013562467880547047, + 0.01579073816537857, + 0.009137868881225586, + 0.01093591470271349, + 0.009627756662666798, + 0.015075257048010826, + 0.008977774530649185, + 0.012519098818302155, + 0.03298140689730644, + 0.03549943491816521, + 0.025844590738415718, + 0.025115087628364563, + 0.06188115477561951, + 0.03687041997909546, + 0.06272979080677032, + 0.03026099130511284, + 0.004960119258612394, + 0.004710418172180653, + 0.017887629568576813, + 0.006232313811779022, + 0.004205244593322277, + 0.005377952475100756, + 0.010755863972008228, + 0.008939053863286972, + 0.017941506579518318, + 0.01148755382746458, + 0.030805418267846107, + 0.01602928526699543, + 0.012924568727612495, + 0.01967749185860157, + 0.016579775139689445, + 0.016643935814499855, + 0.006110356654971838, + 0.016744181513786316, + 0.007024173159152269, + 0.006820295471698046, + 0.006899320520460606, + 0.00867965817451477, + 0.012258356437087059, + 0.006443536840379238, + 0.003529392881318927, + 0.007207914255559444, + 0.0048876260407269, + 0.005405420437455177, + 0.005653703585267067, + 0.003607225138694048, + 0.006862929556518793, + 0.004113386385142803, + 0.001928212819620967, + 0.002778120106086135, + 0.002325537148863077, + 0.006811854895204306, + 0.002503512194380164, + 0.0019215873908251524, + 0.007925129495561123, + 0.002873998833820224, + 0.005493907257914543, + 0.0029662929009646177, + 0.003891551401466131, + 0.016506116837263107, + 0.00512100150808692, + 0.0028596685733646154, + 0.02028757892549038, + 0.004778374917805195, + 0.0043257419019937515, + 0.0073249563574790955, + 0.00781018054112792, + 0.010965224355459213, + 0.00826217606663704, + 0.0052820113487541676, + 0.009494061581790447, + 0.006547456607222557, + 0.014042232185602188, + 0.015764696523547173, + 0.02029176987707615, + 0.013081260956823826, + 0.02098163776099682, + 0.01740044169127941, + 0.00841147918254137, + 0.014826890081167221 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 3, + 9 + ], + "selected_reward": -0.34759005904197693 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a close-up from a level angle held still?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a dense digital cover interwoven with emerald binary codes and golden magical runes visible at the center?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a blinding beam of pure white energy erupt from the center and cleave the digital cover?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the disintegrated fragments radiate outward at high speed leaving bright trails and forming a digital storm?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a pure black void with rich volumetric lighting effects and sharp edges on the numbers and runes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01613089069724083 + ], + "opposite_logprob_tokens": [ + -4.1411309242248535 + ], + "logprob_reward": -0.01613089069724083, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.01613089069724083, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09823376685380936, + 0.14599457383155823, + 0.04622529074549675, + 0.14346972107887268, + 0.07328380644321442, + 0.09447523951530457, + 0.028247185051441193, + 0.2249508798122406, + 0.07257269322872162, + 0.07254683971405029 + ], + "credit_map": [ + 0.010275037959218025, + 0.010537518188357353, + 0.010319066233932972, + 0.01477001141756773, + 0.009574363008141518, + 0.014264147728681564, + 0.017261499539017677, + 0.011232119053602219, + 0.013855405151844025, + 0.015633122995495796, + 0.012350966222584248, + 0.012588161043822765, + 0.028444679453969002, + 0.01978510618209839, + 0.03056216426193714, + 0.012774968519806862, + 0.004172929096966982, + 0.003825145773589611, + 0.005322697106748819, + 0.006205372512340546, + 0.00353857665322721, + 0.01013239286839962, + 0.004370348993688822, + 0.008657827042043209, + 0.015298888087272644, + 0.01346854493021965, + 0.01935594528913498, + 0.01641165278851986, + 0.014885351993143559, + 0.018569132313132286, + 0.028154682368040085, + 0.017325516790151596, + 0.0072593106888234615, + 0.01234524603933096, + 0.009619376622140408, + 0.011574977077543736, + 0.006783565506339073, + 0.0062517584301531315, + 0.010729281231760979, + 0.008720289915800095, + 0.007266939617693424, + 0.0102086765691638, + 0.008838267996907234, + 0.0190101470798254, + 0.008586316369473934, + 0.006166711915284395, + 0.025548698380589485, + 0.00884947832673788, + 0.002502255607396364, + 0.002989384112879634, + 0.0034870950039476156, + 0.004236252047121525, + 0.003508396679535508, + 0.002602748805657029, + 0.004491116851568222, + 0.0044299340806901455, + 0.009888336062431335, + 0.008696679025888443, + 0.008720657788217068, + 0.08098065108060837, + 0.012662476859986782, + 0.008859662339091301, + 0.0831226035952568, + 0.012019787915050983, + 0.004778394941240549, + 0.011428920552134514, + 0.006812199484556913, + 0.013617128133773804, + 0.009429304860532284, + 0.004158238880336285, + 0.01561334915459156, + 0.0067351520992815495, + 0.010910231620073318, + 0.01079468335956335, + 0.008045668713748455, + 0.005647388752549887, + 0.01652960106730461, + 0.010318486019968987, + 0.0047454992309212685, + 0.005555277224630117 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 1, + 3 + ], + "selected_reward": -0.01613089069724083 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a close-up from a level angle held still?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a dense digital cover interwoven with emerald binary codes and golden magical runes visible at the center?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a blinding beam of pure white energy erupt from the center and cleave the digital cover?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the disintegrated fragments radiate outward at high speed leaving bright trails and forming a digital storm?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a pure black void with rich volumetric lighting effects and sharp edges on the numbers and runes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004101199563592672 + ], + "opposite_logprob_tokens": [ + -5.504101276397705 + ], + "logprob_reward": -0.004101199563592672, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004101199563592672, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08663944154977798, + 0.08336188644170761, + 0.08952502906322479, + 0.13441890478134155, + 0.08991546928882599, + 0.041851501911878586, + 0.046924054622650146, + 0.23287208378314972, + 0.12342577427625656, + 0.07106590270996094 + ], + "credit_map": [ + 0.007925591431558132, + 0.017419278621673584, + 0.008002713322639465, + 0.011419734917581081, + 0.011989686638116837, + 0.013586272485554218, + 0.008805216290056705, + 0.007490938063710928, + 0.009242778643965721, + 0.010026381351053715, + 0.006232992745935917, + 0.007111190352588892, + 0.015295288525521755, + 0.012857351452112198, + 0.014812963083386421, + 0.007782935164868832, + 0.01042263861745596, + 0.008215639740228653, + 0.02029823698103428, + 0.009987973608076572, + 0.0077501991763710976, + 0.012802443467080593, + 0.01115982886403799, + 0.008888060227036476, + 0.015121258795261383, + 0.011097928509116173, + 0.025378571823239326, + 0.01938401535153389, + 0.014913488179445267, + 0.012587853707373142, + 0.021384458988904953, + 0.014551318250596523, + 0.007527773734182119, + 0.020250655710697174, + 0.008602894842624664, + 0.01157732680439949, + 0.009156359359622002, + 0.0074530672281980515, + 0.018157869577407837, + 0.007189520169049501, + 0.003476750571280718, + 0.004986908286809921, + 0.003966128453612328, + 0.007743736729025841, + 0.004711788147687912, + 0.0028524284716695547, + 0.00927977729588747, + 0.0048339832574129105, + 0.0033421488478779793, + 0.004280981607735157, + 0.0039056050591170788, + 0.010561122559010983, + 0.004720438737422228, + 0.003226069500669837, + 0.013063154183328152, + 0.0038245320320129395, + 0.021811287850141525, + 0.012071224860846996, + 0.011867750436067581, + 0.0645049586892128, + 0.015612010844051838, + 0.0077204108238220215, + 0.0830782800912857, + 0.016206132248044014, + 0.0076252250000834465, + 0.030863570049405098, + 0.0076172929257154465, + 0.031058460474014282, + 0.013272172771394253, + 0.0070468648336827755, + 0.019696801900863647, + 0.006245380267500877, + 0.00682677561417222, + 0.008601648733019829, + 0.00952716451138258, + 0.007432703394442797, + 0.013244880363345146, + 0.008227959275245667, + 0.007700236514210701, + 0.009504526853561401 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 3, + 8 + ], + "selected_reward": -0.004101199563592672 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a close-up from a level angle held still?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a dense digital cover interwoven with emerald binary codes and golden magical runes visible at the center?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a blinding beam of pure white energy erupt from the center and cleave the digital cover?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the disintegrated fragments radiate outward at high speed leaving bright trails and forming a digital storm?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a pure black void with rich volumetric lighting effects and sharp edges on the numbers and runes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012541357427835464 + ], + "opposite_logprob_tokens": [ + -4.3875412940979 + ], + "logprob_reward": -0.012541357427835464, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012541357427835464, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05951699614524841, + 0.04698704555630684, + 0.08614658564329147, + 0.1503322720527649, + 0.09125557541847229, + 0.04616163298487663, + 0.04194049537181854, + 0.18453726172447205, + 0.20986029505729675, + 0.08326184004545212 + ], + "credit_map": [ + 0.008078833110630512, + 0.009990807622671127, + 0.005409909412264824, + 0.008613898418843746, + 0.005693682469427586, + 0.008043196052312851, + 0.007518209051340818, + 0.006168459542095661, + 0.004518382716923952, + 0.006864861585199833, + 0.005220402963459492, + 0.00598076730966568, + 0.00589684396982193, + 0.006915213540196419, + 0.0056410133838653564, + 0.005949558224529028, + 0.00810066144913435, + 0.005385177209973335, + 0.02227005548775196, + 0.010140517726540565, + 0.007222408894449472, + 0.009685388766229153, + 0.016054587438702583, + 0.007287793792784214, + 0.01640351489186287, + 0.014658262021839619, + 0.026656292378902435, + 0.02019627019762993, + 0.013593176379799843, + 0.015588776208460331, + 0.027589039877057076, + 0.01564694009721279, + 0.007644845172762871, + 0.010974155738949776, + 0.007071096915751696, + 0.016768228262662888, + 0.01109587587416172, + 0.007016672752797604, + 0.023669568821787834, + 0.0070151337422430515, + 0.004097188822925091, + 0.007126500364392996, + 0.004931106697767973, + 0.007493893150240183, + 0.005266895983368158, + 0.0031554200686514378, + 0.009923437610268593, + 0.004167189821600914, + 0.0035146158188581467, + 0.004296597559005022, + 0.0048001310788095, + 0.007716737687587738, + 0.005064026452600956, + 0.0033988626673817635, + 0.008751478977501392, + 0.004398043267428875, + 0.006736869923770428, + 0.0069238874129951, + 0.005592294968664646, + 0.07077192515134811, + 0.009368091821670532, + 0.0038606019224971533, + 0.07261121273040771, + 0.008672380819916725, + 0.012360824272036552, + 0.04928220435976982, + 0.012040125206112862, + 0.05837226286530495, + 0.01946067251265049, + 0.01086009293794632, + 0.037007469683885574, + 0.010476639494299889, + 0.009957726113498211, + 0.007971042767167091, + 0.01434199046343565, + 0.010919101536273956, + 0.009875601157546043, + 0.010477404110133648, + 0.007545893546193838, + 0.01217307522892952 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 3 + ], + "selected_reward": -0.012541357427835464 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a close-up from a level angle held still?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a dense digital cover interwoven with emerald binary codes and golden magical runes visible at the center?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a blinding beam of pure white energy erupt from the center and cleave the digital cover?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the disintegrated fragments radiate outward at high speed leaving bright trails and forming a digital storm?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a pure black void with rich volumetric lighting effects and sharp edges on the numbers and runes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"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": [ + 0.151273712515831, + 0.23032599687576294, + 0.11784610897302628, + 0.10988786816596985, + 0.02929745987057686, + 0.02208847925066948, + 0.01261626835912466, + 0.09435904026031494, + 0.13322743773460388, + 0.09907757490873337 + ], + "credit_map": [ + 0.014180661179125309, + 0.023074744269251823, + 0.012807941064238548, + 0.02303042821586132, + 0.018398886546492577, + 0.02057037502527237, + 0.02051972597837448, + 0.01869095116853714, + 0.020436551421880722, + 0.02304108254611492, + 0.01926662027835846, + 0.018178174272179604, + 0.04936540126800537, + 0.02605934627354145, + 0.05347316339612007, + 0.020505672320723534, + 0.011851557530462742, + 0.010162729769945145, + 0.010942325927317142, + 0.013174321502447128, + 0.01663944683969021, + 0.01490581501275301, + 0.02859107404947281, + 0.01157884206622839, + 0.011224428191781044, + 0.010135489515960217, + 0.013336765579879284, + 0.014382721856236458, + 0.00972899328917265, + 0.015622756443917751, + 0.02191760390996933, + 0.013539115898311138, + 0.0026956016663461924, + 0.003925203811377287, + 0.0046978904865682125, + 0.0036596497520804405, + 0.0030189543031156063, + 0.0036487916950136423, + 0.004362473264336586, + 0.003288897220045328, + 0.0023651071824133396, + 0.0025420666206628084, + 0.003612856613472104, + 0.0032390926498919725, + 0.0025038865860551596, + 0.0025599754881113768, + 0.0030420690309256315, + 0.0022234246134757996, + 0.0011802812805399299, + 0.0019229237223044038, + 0.0013646116713061929, + 0.0014781045028939843, + 0.001950390636920929, + 0.0011467811418697238, + 0.001617478090338409, + 0.001955697312951088, + 0.005334489978849888, + 0.006615583319216967, + 0.00358177674934268, + 0.03699926659464836, + 0.003906582482159138, + 0.0034845485351979733, + 0.030788855627179146, + 0.003647930221632123, + 0.00813957117497921, + 0.03684237226843834, + 0.005681691691279411, + 0.03705541417002678, + 0.013093800283968449, + 0.006085303612053394, + 0.022247670218348503, + 0.004081603139638901, + 0.01073718722909689, + 0.01471709180623293, + 0.013223685324192047, + 0.00957291480153799, + 0.02121584862470627, + 0.014360317029058933, + 0.005952736362814903, + 0.00929779652506113 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 8 + ], + "selected_reward": -0.04307607561349869 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "close-up shot from a level angle, the camera held still to capture the drastic change about to take place. At the center of the frame is a dense digital cover, interwoven with countless luminescent emerald binary codes and ancient magical runes shimmering with a mysterious golden glow. These streams of data flow vertically like cascading cascades, forming a complex barrier suspended against a pure black void. Suddenly, a blinding beam of pure white energy erupts from the very center of the frame, instantly cleaving the digital cover with the force of thunder. Following this devastating blast of light, the entire cover disintegrates into countless luminous fragments of code and runes. These fragments radiate outward at high speed, leaving bright trails in the air and forming a spectacular digital storm. The entire scene, rendered in real-time by the engine, presents a cinematic CG feel with ultra-high-definition resolution, making every scattered number and rune clearly visible with sharp edges. The burst of strong light creates extremely rich volumetric lighting effects in the darkness, and holy beams penetrate the diffuse digital dust, injecting unparalleled visual impact and epic atmosphere into this moment of digital destruction full of power and beauty.", + "seed": 24951, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000021/group_4/index_3494_seed_24951_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": -0.1298425728455186, + "vlm_logprob": -0.1298425728455186, + "vlm_yes_no_margin": 3.075, + "vlm_token_credit": -0.1298425728455186, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a close-up from a level angle held still?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a dense digital cover interwoven with emerald binary codes and golden magical runes visible at the center?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a blinding beam of pure white energy erupt from the center and cleave the digital cover?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the disintegrated fragments radiate outward at high speed leaving bright trails and forming a digital storm?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a pure black void with rich volumetric lighting effects and sharp edges on the numbers and runes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.371966689825058 + ], + "opposite_logprob_tokens": [ + -1.2469667196273804 + ], + "logprob_reward": -0.371966689825058, + "yes_no_margin_reward": 0.875, + "token_credit_reward": -0.371966689825058, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1277305781841278, + 0.02984619140625, + 0.15252195298671722, + 0.06723637878894806, + 0.04154335707426071, + 0.0952228531241417, + 0.10828248411417007, + 0.2645880877971649, + 0.044703409075737, + 0.06832467764616013 + ], + "credit_map": [ + 0.013625597581267357, + 0.026077471673488617, + 0.017058314755558968, + 0.015386906452476978, + 0.012650525197386742, + 0.018537571653723717, + 0.011384748853743076, + 0.013009445741772652, + 0.0020482547115534544, + 0.0035507604479789734, + 0.004521247930824757, + 0.0037514674477279186, + 0.003728107549250126, + 0.004234776832163334, + 0.0033838904928416014, + 0.00462768878787756, + 0.011479091830551624, + 0.03293176740407944, + 0.007345395162701607, + 0.018242523074150085, + 0.03509298339486122, + 0.01060436014086008, + 0.026585951447486877, + 0.010239883325994015, + 0.007707573007792234, + 0.008025885559618473, + 0.0049449666403234005, + 0.010575691238045692, + 0.006292082369327545, + 0.008813410997390747, + 0.011118464171886444, + 0.009758302010595798, + 0.003920807968825102, + 0.003808210603892803, + 0.003284214064478874, + 0.005898403003811836, + 0.0034987886901944876, + 0.005767907481640577, + 0.005650716833770275, + 0.009714307263493538, + 0.007827757857739925, + 0.004712643101811409, + 0.006531996186822653, + 0.014916865155100822, + 0.006736054550856352, + 0.02156580053269863, + 0.017738964408636093, + 0.015192772261798382, + 0.006571720354259014, + 0.006620082072913647, + 0.007003491278737783, + 0.026562226936221123, + 0.00440575834363699, + 0.01784702204167843, + 0.030935268849134445, + 0.008336914703249931, + 0.00772668095305562, + 0.046263545751571655, + 0.015142278745770454, + 0.100344717502594, + 0.014472578652203083, + 0.007431310135871172, + 0.06483671069145203, + 0.008370276540517807, + 0.003564880695194006, + 0.004983994178473949, + 0.005143113899976015, + 0.007445237599313259, + 0.006145944353193045, + 0.003992054611444473, + 0.006783781107515097, + 0.006644404027611017, + 0.008921568281948566, + 0.006121956277638674, + 0.009045880287885666, + 0.00648636044934392, + 0.006587376818060875, + 0.006806360092014074, + 0.011122608557343483, + 0.01323256641626358 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 2, + 0 + ], + "selected_reward": -0.371966689825058 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a close-up from a level angle held still?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a dense digital cover interwoven with emerald binary codes and golden magical runes visible at the center?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a blinding beam of pure white energy erupt from the center and cleave the digital cover?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the disintegrated fragments radiate outward at high speed leaving bright trails and forming a digital storm?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a pure black void with rich volumetric lighting effects and sharp edges on the numbers and runes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029844719916582108 + ], + "opposite_logprob_tokens": [ + -3.5298447608947754 + ], + "logprob_reward": -0.029844719916582108, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029844719916582108, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0994967669248581, + 0.034047555178403854, + 0.4505966901779175, + 0.02047273889183998, + 0.020427575334906578, + 0.1365601271390915, + 0.0521613284945488, + 0.14252370595932007, + 0.02124292589724064, + 0.022470571100711823 + ], + "credit_map": [ + 0.0075494167394936085, + 0.017143448814749718, + 0.015572651289403439, + 0.011772331781685352, + 0.010998454876244068, + 0.012512417510151863, + 0.009634501300752163, + 0.014313547872006893, + 0.0025652830954641104, + 0.004109532106667757, + 0.004856013227254152, + 0.004671692848205566, + 0.004816786851733923, + 0.004165683872997761, + 0.00487869419157505, + 0.003983871545642614, + 0.033117782324552536, + 0.0991872027516365, + 0.020880457013845444, + 0.050051409751176834, + 0.10673090070486069, + 0.032032158225774765, + 0.07515310496091843, + 0.03344365581870079, + 0.0017367268446832895, + 0.0027629840187728405, + 0.0014707184163853526, + 0.0035374602302908897, + 0.0023642275482416153, + 0.0016390258679166436, + 0.0028896001167595387, + 0.004071996081620455, + 0.002035169629380107, + 0.002346538472920656, + 0.0013871105620637536, + 0.004510569851845503, + 0.002404699334874749, + 0.0018253279849886894, + 0.002860559383407235, + 0.0030575997661799192, + 0.009632909670472145, + 0.004098530393093824, + 0.005953328683972359, + 0.028707345947623253, + 0.007682120427489281, + 0.027884142473340034, + 0.03574710339307785, + 0.016854654997587204, + 0.004313532263040543, + 0.009060777723789215, + 0.003731722477823496, + 0.008204762823879719, + 0.004084776621311903, + 0.006441768258810043, + 0.011824561282992363, + 0.004499428439885378, + 0.004079738166183233, + 0.013544448651373386, + 0.008443388156592846, + 0.05680812895298004, + 0.00576477125287056, + 0.0035270547959953547, + 0.04394131153821945, + 0.006414872128516436, + 0.0023494085762649775, + 0.0012564562493935227, + 0.005430323071777821, + 0.00439769821241498, + 0.0017422129167243838, + 0.0017766111996024847, + 0.0020497962832450867, + 0.0022404207848012447, + 0.0026264896150678396, + 0.0017626389162614942, + 0.002840259578078985, + 0.0037336349487304688, + 0.0014304067008197308, + 0.0022652591578662395, + 0.0032843779772520065, + 0.004527504555881023 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 7, + 5 + ], + "selected_reward": -0.029844719916582108 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a close-up from a level angle held still?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a dense digital cover interwoven with emerald binary codes and golden magical runes visible at the center?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a blinding beam of pure white energy erupt from the center and cleave the digital cover?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the disintegrated fragments radiate outward at high speed leaving bright trails and forming a digital storm?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a pure black void with rich volumetric lighting effects and sharp edges on the numbers and runes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0036182431504130363 + ], + "opposite_logprob_tokens": [ + -5.628618240356445 + ], + "logprob_reward": -0.0036182431504130363, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.0036182431504130363, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08626727759838104, + 0.04756435006856918, + 0.06998881697654724, + 0.033240750432014465, + 0.01822590082883835, + 0.26298415660858154, + 0.14388883113861084, + 0.2292141169309616, + 0.03791813924908638, + 0.07070761173963547 + ], + "credit_map": [ + 0.009778240695595741, + 0.01852462626993656, + 0.011389307677745819, + 0.01018724124878645, + 0.008329812437295914, + 0.01167517900466919, + 0.007301486562937498, + 0.009081387892365456, + 0.002971276640892029, + 0.006738258060067892, + 0.00634489580988884, + 0.005055525805801153, + 0.005563780199736357, + 0.003778836689889431, + 0.004630565643310547, + 0.012481210753321648, + 0.004976163152605295, + 0.0143749313428998, + 0.004098943900316954, + 0.008385543711483479, + 0.015737904235720634, + 0.006157775409519672, + 0.00997087825089693, + 0.00628667464479804, + 0.0017299343599006534, + 0.002317517064511776, + 0.0019062445499002934, + 0.010057639330625534, + 0.0027562046889215708, + 0.0022440459579229355, + 0.009325616993010044, + 0.0029035485349595547, + 0.002067216206341982, + 0.0015216503525152802, + 0.0019380822777748108, + 0.0029999359976500273, + 0.0017055778298527002, + 0.0021760750096291304, + 0.0033050139900296926, + 0.002512349747121334, + 0.018393075093626976, + 0.008231965824961662, + 0.012445610947906971, + 0.0543600358068943, + 0.013763791881501675, + 0.05724477395415306, + 0.06748169660568237, + 0.03106321208178997, + 0.010385191068053246, + 0.005741458851844072, + 0.0067631579004228115, + 0.032317545264959335, + 0.0064622447825968266, + 0.03372917324304581, + 0.03658866137266159, + 0.011901401914656162, + 0.007450031582266092, + 0.03527732938528061, + 0.011584263294935226, + 0.09437581151723862, + 0.009978551417589188, + 0.006009403150528669, + 0.05846567451953888, + 0.006073040422052145, + 0.0032260254956781864, + 0.0022788899950683117, + 0.010241947136819363, + 0.008548933081328869, + 0.0025194110348820686, + 0.002575563034042716, + 0.004448153544217348, + 0.004079214762896299, + 0.005799305159598589, + 0.006459986325353384, + 0.00955379568040371, + 0.00963929109275341, + 0.005995503161102533, + 0.005818346049636602, + 0.01290799118578434, + 0.0145333930850029 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 7, + 6 + ], + "selected_reward": -0.0036182431504130363 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a close-up from a level angle held still?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a dense digital cover interwoven with emerald binary codes and golden magical runes visible at the center?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a blinding beam of pure white energy erupt from the center and cleave the digital cover?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the disintegrated fragments radiate outward at high speed leaving bright trails and forming a digital storm?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a pure black void with rich volumetric lighting effects and sharp edges on the numbers and runes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018186485394835472 + ], + "opposite_logprob_tokens": [ + -4.018186569213867 + ], + "logprob_reward": -0.018186485394835472, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018186485394835472, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0419468879699707, + 0.018147936090826988, + 0.03235708549618721, + 0.01673770695924759, + 0.02512780949473381, + 0.18185044825077057, + 0.2641054689884186, + 0.306305468082428, + 0.012190052308142185, + 0.10123118758201599 + ], + "credit_map": [ + 0.005097802262753248, + 0.006880585104227066, + 0.005746141541749239, + 0.0036470317281782627, + 0.006229840684682131, + 0.005934730637818575, + 0.0040733348578214645, + 0.004337420221418142, + 0.0011874267365783453, + 0.0017190770013257861, + 0.002585234120488167, + 0.0032516068313270807, + 0.002410389482975006, + 0.0017913583433255553, + 0.002267292235046625, + 0.0029355501756072044, + 0.0028211460448801517, + 0.0063835633918643, + 0.001302696531638503, + 0.004071641713380814, + 0.006967741064727306, + 0.0019914007280021906, + 0.004971023183315992, + 0.003847873769700527, + 0.0013039938639849424, + 0.002367550041526556, + 0.002243858063593507, + 0.0018615936860442162, + 0.0019970980938524008, + 0.0014454303309321404, + 0.0028074763249605894, + 0.0027107063215225935, + 0.0025687350425869226, + 0.002491878578439355, + 0.0029952507466077805, + 0.004112234804779291, + 0.0025506888050585985, + 0.00247950735501945, + 0.00507229333743453, + 0.0028572194278240204, + 0.012747065164148808, + 0.005687932949513197, + 0.00856922846287489, + 0.037688225507736206, + 0.009448274038732052, + 0.03928930312395096, + 0.04685386270284653, + 0.021566547453403473, + 0.01917496882379055, + 0.010369461961090565, + 0.012500331737101078, + 0.06040985882282257, + 0.012576546519994736, + 0.058383259922266006, + 0.06937036663293839, + 0.021320665255188942, + 0.008935725316405296, + 0.05173743516206741, + 0.022833211347460747, + 0.11221880465745926, + 0.015620001591742039, + 0.007261042948812246, + 0.07601172477006912, + 0.011687511578202248, + 0.0009683849639259279, + 0.0011698585003614426, + 0.003081971313804388, + 0.0019033438293263316, + 0.0014823313103988767, + 0.0008929785690270364, + 0.0012590013211593032, + 0.0014321820344775915, + 0.008270973339676857, + 0.0101008927449584, + 0.01337576936930418, + 0.013085180893540382, + 0.009751713834702969, + 0.008077356033027172, + 0.018886709585785866, + 0.019682597368955612 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 6, + 5 + ], + "selected_reward": -0.018186485394835472 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a close-up from a level angle held still?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a dense digital cover interwoven with emerald binary codes and golden magical runes visible at the center?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a blinding beam of pure white energy erupt from the center and cleave the digital cover?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the disintegrated fragments radiate outward at high speed leaving bright trails and forming a digital storm?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a pure black void with rich volumetric lighting effects and sharp edges on the numbers and runes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.22559672594070435 + ], + "opposite_logprob_tokens": [ + -1.6005966663360596 + ], + "logprob_reward": -0.22559672594070435, + "yes_no_margin_reward": 1.375, + "token_credit_reward": -0.22559672594070435, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2881641089916229, + 0.02714216709136963, + 0.31617286801338196, + 0.06455922871828079, + 0.06149900704622269, + 0.012495565228164196, + 0.09422466903924942, + 0.09016861021518707, + 0.029408875852823257, + 0.016164885833859444 + ], + "credit_map": [ + 0.02221330814063549, + 0.05574869364500046, + 0.03600262850522995, + 0.03240052983164787, + 0.045617636293172836, + 0.032865166664123535, + 0.035955023020505905, + 0.02736111544072628, + 0.0015659395139664412, + 0.004893969744443893, + 0.0027338999789208174, + 0.0027333865873515606, + 0.006619824096560478, + 0.0026367134414613247, + 0.0031093303114175797, + 0.002849105978384614, + 0.025166992098093033, + 0.06669838726520538, + 0.014827672392129898, + 0.04008479416370392, + 0.07126180082559586, + 0.0205227080732584, + 0.05610765144228935, + 0.02150288224220276, + 0.007879397831857204, + 0.005966220051050186, + 0.010035396553575993, + 0.005810444243252277, + 0.006305634509772062, + 0.00861257966607809, + 0.008743718266487122, + 0.011205843649804592, + 0.007943380624055862, + 0.00586865795776248, + 0.007665444165468216, + 0.009044189006090164, + 0.0054207188077270985, + 0.007439802400767803, + 0.01113245077431202, + 0.006984370294958353, + 0.0010127436835318804, + 0.001007760176435113, + 0.0012679987121373415, + 0.0022896386217325926, + 0.0012984978966414928, + 0.0011772216530516744, + 0.002398158423602581, + 0.002043546410277486, + 0.006229344289749861, + 0.03347037732601166, + 0.003942376002669334, + 0.007129913195967674, + 0.008749382570385933, + 0.009704307653009892, + 0.019791599363088608, + 0.005207371432334185, + 0.007717201020568609, + 0.004241800867021084, + 0.021190334111452103, + 0.02770548313856125, + 0.004005203023552895, + 0.004876005928963423, + 0.014571186155080795, + 0.005861392244696617, + 0.005062030162662268, + 0.00300040515139699, + 0.004081410355865955, + 0.00287553365342319, + 0.0032114414498209953, + 0.003552280366420746, + 0.0028449343517422676, + 0.004780839197337627, + 0.0015748083824291825, + 0.0019017683807760477, + 0.0019654405768960714, + 0.001842446275986731, + 0.0016640881076455116, + 0.0014182382728904486, + 0.0029140522237867117, + 0.00288404431194067 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 6 + ], + "selected_reward": -0.22559672594070435 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "close-up shot from a level angle, the camera held still to capture the drastic change about to take place. At the center of the frame is a dense digital cover, interwoven with countless luminescent emerald binary codes and ancient magical runes shimmering with a mysterious golden glow. These streams of data flow vertically like cascading cascades, forming a complex barrier suspended against a pure black void. Suddenly, a blinding beam of pure white energy erupts from the very center of the frame, instantly cleaving the digital cover with the force of thunder. Following this devastating blast of light, the entire cover disintegrates into countless luminous fragments of code and runes. These fragments radiate outward at high speed, leaving bright trails in the air and forming a spectacular digital storm. The entire scene, rendered in real-time by the engine, presents a cinematic CG feel with ultra-high-definition resolution, making every scattered number and rune clearly visible with sharp edges. The burst of strong light creates extremely rich volumetric lighting effects in the darkness, and holy beams penetrate the diffuse digital dust, injecting unparalleled visual impact and epic atmosphere into this moment of digital destruction full of power and beauty.", + "seed": 24952, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000021/group_4/index_3494_seed_24952_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": -0.15323232505470513, + "vlm_logprob": -0.15323232505470513, + "vlm_yes_no_margin": 3.0, + "vlm_token_credit": -0.15323232505470513, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a close-up from a level angle held still?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a dense digital cover interwoven with emerald binary codes and golden magical runes visible at the center?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a blinding beam of pure white energy erupt from the center and cleave the digital cover?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the disintegrated fragments radiate outward at high speed leaving bright trails and forming a digital storm?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a pure black void with rich volumetric lighting effects and sharp edges on the numbers and runes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.5596418380737305 + ], + "opposite_logprob_tokens": [ + -0.9346418380737305 + ], + "logprob_reward": -0.5596418380737305, + "yes_no_margin_reward": 0.375, + "token_credit_reward": -0.5596418380737305, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1517287641763687, + 0.18286938965320587, + 0.12784914672374725, + 0.037178706377744675, + 0.15423348546028137, + 0.055459100753068924, + 0.06715483218431473, + 0.1067415103316307, + 0.08787906914949417, + 0.02890593931078911 + ], + "credit_map": [ + 0.010958828032016754, + 0.01817816123366356, + 0.019403085112571716, + 0.022585855796933174, + 0.013697542250156403, + 0.018342765048146248, + 0.020804349333047867, + 0.027758179232478142, + 0.0076967282220721245, + 0.03552398085594177, + 0.01089111715555191, + 0.03921324014663696, + 0.04716929420828819, + 0.011860636062920094, + 0.01940370351076126, + 0.011110699735581875, + 0.007666696794331074, + 0.020492874085903168, + 0.006900571286678314, + 0.013060658238828182, + 0.014453992247581482, + 0.02756262570619583, + 0.007591277826577425, + 0.030120445415377617, + 0.002633048687130213, + 0.001966226613149047, + 0.005774658173322678, + 0.002385577419772744, + 0.0023670876398682594, + 0.007947116158902645, + 0.003262939630076289, + 0.010842051357030869, + 0.01998688280582428, + 0.02637924626469612, + 0.01629794016480446, + 0.015112632885575294, + 0.02903655357658863, + 0.018930790945887566, + 0.014187917113304138, + 0.014301519840955734, + 0.004393484443426132, + 0.009108031168580055, + 0.003861390519887209, + 0.005404151044785976, + 0.015602486208081245, + 0.00884436909109354, + 0.00483415462076664, + 0.003411033656448126, + 0.004789919592440128, + 0.005281038582324982, + 0.0108298659324646, + 0.004786623176187277, + 0.01309004332870245, + 0.013920311816036701, + 0.008761161006987095, + 0.005695867817848921, + 0.005410981364548206, + 0.00773730268701911, + 0.005467154551297426, + 0.006322773639112711, + 0.01018134132027626, + 0.024499084800481796, + 0.011424997821450233, + 0.03569788113236427, + 0.002994486363604665, + 0.0055869026109576225, + 0.014140009880065918, + 0.03331417962908745, + 0.009209041483700275, + 0.0042979102581739426, + 0.006347951013594866, + 0.011988583020865917, + 0.002310769399628043, + 0.003091834019869566, + 0.0029014782048761845, + 0.0036456517409533262, + 0.0035415294114500284, + 0.00613048393279314, + 0.0031657726503908634, + 0.004118419252336025 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 4, + 0 + ], + "selected_reward": -0.5596418380737305 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a close-up from a level angle held still?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a dense digital cover interwoven with emerald binary codes and golden magical runes visible at the center?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a blinding beam of pure white energy erupt from the center and cleave the digital cover?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the disintegrated fragments radiate outward at high speed leaving bright trails and forming a digital storm?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a pure black void with rich volumetric lighting effects and sharp edges on the numbers and runes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026405641809105873 + ], + "opposite_logprob_tokens": [ + -3.6514055728912354 + ], + "logprob_reward": -0.026405641809105873, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026405641809105873, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07731901854276657, + 0.33875739574432373, + 0.16984973847866058, + 0.026761356741189957, + 0.05689804255962372, + 0.06500395387411118, + 0.043962907046079636, + 0.1382884830236435, + 0.06367578357458115, + 0.019483258947730064 + ], + "credit_map": [ + 0.0050579942762851715, + 0.013971849344670773, + 0.008734621107578278, + 0.01619924046099186, + 0.013159321621060371, + 0.005792991258203983, + 0.00892975926399231, + 0.005473253317177296, + 0.01395382545888424, + 0.057269807904958725, + 0.01856888085603714, + 0.0682915449142456, + 0.09348969161510468, + 0.02346017025411129, + 0.03680073842406273, + 0.026922767981886864, + 0.01095670461654663, + 0.04085680842399597, + 0.008540743961930275, + 0.02186301350593567, + 0.021035414189100266, + 0.023941954597830772, + 0.010419457219541073, + 0.032235659658908844, + 0.002463596174493432, + 0.0030640794429928064, + 0.003812222508713603, + 0.002008137060329318, + 0.00321183237247169, + 0.004534658510237932, + 0.002388506196439266, + 0.005278329364955425, + 0.004764179233461618, + 0.00514973932877183, + 0.007392126601189375, + 0.004906854126602411, + 0.011670765466988087, + 0.00825261790305376, + 0.006399767007678747, + 0.00836199801415205, + 0.005832333583384752, + 0.012340258806943893, + 0.004370901267975569, + 0.0065333060920238495, + 0.015456789173185825, + 0.009735792875289917, + 0.006110570393502712, + 0.004624010529369116, + 0.0026861787773668766, + 0.006765096448361874, + 0.004039717372506857, + 0.003656682325527072, + 0.010511940345168114, + 0.008192524313926697, + 0.004332365468144417, + 0.003778408747166395, + 0.01043021585792303, + 0.006283878348767757, + 0.008365195244550705, + 0.011681763455271721, + 0.00932164117693901, + 0.041635725647211075, + 0.018111921846866608, + 0.03245816379785538, + 0.0021553265396505594, + 0.003853344591334462, + 0.014761783182621002, + 0.02461886778473854, + 0.006048594135791063, + 0.002722743898630142, + 0.004623860586434603, + 0.004891270771622658, + 0.0019498757319524884, + 0.002055675722658634, + 0.002758883871138096, + 0.0026415607426315546, + 0.001466527464799583, + 0.0023082303814589977, + 0.003498942358419299, + 0.002803565002977848 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 2, + 7 + ], + "selected_reward": -0.026405641809105873 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a close-up from a level angle held still?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a dense digital cover interwoven with emerald binary codes and golden magical runes visible at the center?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a blinding beam of pure white energy erupt from the center and cleave the digital cover?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the disintegrated fragments radiate outward at high speed leaving bright trails and forming a digital storm?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a pure black void with rich volumetric lighting effects and sharp edges on the numbers and runes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011078868061304092 + ], + "opposite_logprob_tokens": [ + -4.511078834533691 + ], + "logprob_reward": -0.011078868061304092, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011078868061304092, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09463842213153839, + 0.5109418630599976, + 0.24252548813819885, + 0.016856012865900993, + 0.020040346309542656, + 0.02317131496965885, + 0.007234152872115374, + 0.03514988347887993, + 0.04243428632616997, + 0.00700824661180377 + ], + "credit_map": [ + 0.007140881847590208, + 0.011837356723845005, + 0.019343560561537743, + 0.014806954190135002, + 0.006652122363448143, + 0.009370324201881886, + 0.009270826354622841, + 0.016216395422816277, + 0.021046191453933716, + 0.08807669579982758, + 0.027150247246026993, + 0.10307307541370392, + 0.1393343061208725, + 0.03570498526096344, + 0.05551960691809654, + 0.04103679209947586, + 0.016676221042871475, + 0.07152395695447922, + 0.02048419415950775, + 0.037280015647411346, + 0.03457104042172432, + 0.023118095472455025, + 0.0168283861130476, + 0.022043582051992416, + 0.0011684723431244493, + 0.0019951912108808756, + 0.0018831552006304264, + 0.0015612725401297212, + 0.002316133351996541, + 0.0024952259846031666, + 0.002099792705848813, + 0.003336769063025713, + 0.0023302931804209948, + 0.002199474023655057, + 0.003094688057899475, + 0.0012361792614683509, + 0.002897001802921295, + 0.003268044674769044, + 0.0016532876761630177, + 0.0033613773994147778, + 0.0019760113209486008, + 0.004584331531077623, + 0.0016956962645053864, + 0.002632021438330412, + 0.004677389282733202, + 0.003329657716676593, + 0.002272274112328887, + 0.0020039337687194347, + 0.0005885960417799652, + 0.000950877380091697, + 0.0007963521056808531, + 0.0006271362653933465, + 0.0016254446236416698, + 0.0013229704927653074, + 0.0005296517047099769, + 0.0007931244326755404, + 0.0025534392334520817, + 0.003670564154163003, + 0.0020814912859350443, + 0.003327044192701578, + 0.005953273735940456, + 0.007011365611106157, + 0.007836399599909782, + 0.0027163065969944, + 0.0016544829122722149, + 0.0016344543546438217, + 0.013960780575871468, + 0.016787376254796982, + 0.0029873393941670656, + 0.001421069260686636, + 0.002517229877412319, + 0.0014715504366904497, + 0.0006675269105471671, + 0.0006565728690475225, + 0.0010911996942013502, + 0.0011164001189172268, + 0.0006914322730153799, + 0.0009933969704434276, + 0.0007231523632071912, + 0.0010685655288398266 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 2, + 0 + ], + "selected_reward": -0.011078868061304092 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a close-up from a level angle held still?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a dense digital cover interwoven with emerald binary codes and golden magical runes visible at the center?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a blinding beam of pure white energy erupt from the center and cleave the digital cover?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the disintegrated fragments radiate outward at high speed leaving bright trails and forming a digital storm?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a pure black void with rich volumetric lighting effects and sharp edges on the numbers and runes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008640876971185207 + ], + "opposite_logprob_tokens": [ + -4.758640766143799 + ], + "logprob_reward": -0.008640876971185207, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008640876971185207, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08259853720664978, + 0.1587805598974228, + 0.2738659977912903, + 0.028235727921128273, + 0.054792482405900955, + 0.07570123672485352, + 0.038034599274396896, + 0.03887810930609703, + 0.22576531767845154, + 0.02334739826619625 + ], + "credit_map": [ + 0.006556411273777485, + 0.007897152565419674, + 0.0175005029886961, + 0.009161300025880337, + 0.010419770143926144, + 0.008765214122831821, + 0.009522772394120693, + 0.012775424867868423, + 0.007159629371017218, + 0.028464147821068764, + 0.009187495335936546, + 0.03204143047332764, + 0.04162493720650673, + 0.011294432915747166, + 0.016996560618281364, + 0.012011944316327572, + 0.01693868264555931, + 0.07293695956468582, + 0.01278749294579029, + 0.0377955362200737, + 0.03258318081498146, + 0.034031402319669724, + 0.015193058177828789, + 0.051599714905023575, + 0.0021716735791414976, + 0.002152126282453537, + 0.003241446800529957, + 0.0025832108221948147, + 0.0033209079410880804, + 0.0044382899068295956, + 0.004104978404939175, + 0.006223097909241915, + 0.004828138742595911, + 0.004971156362444162, + 0.010442760773003101, + 0.004621278028935194, + 0.00856268871575594, + 0.008028019219636917, + 0.005113690625876188, + 0.008224750868976116, + 0.007661567069590092, + 0.01115497574210167, + 0.006333641242235899, + 0.006172737572342157, + 0.017601927742362022, + 0.01164780929684639, + 0.007514242082834244, + 0.007614344358444214, + 0.0023702369071543217, + 0.005329041741788387, + 0.002168666571378708, + 0.0038289211224764585, + 0.00980823952704668, + 0.0069620124995708466, + 0.003954288084059954, + 0.003613195149227977, + 0.0029543517157435417, + 0.004202612675726414, + 0.0027325006667524576, + 0.0032174407970160246, + 0.004056659061461687, + 0.010434710420668125, + 0.0053284685127437115, + 0.005951372440904379, + 0.006926433648914099, + 0.009885251522064209, + 0.071443110704422, + 0.09851635247468948, + 0.011934719048440456, + 0.00673084519803524, + 0.013034919276833534, + 0.0072937035001814365, + 0.002072117989882827, + 0.0036054272204637527, + 0.0023217557463794947, + 0.0027171247638761997, + 0.0028614632319658995, + 0.004163593519479036, + 0.0022572800517082214, + 0.0033486371394246817 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 8, + 1 + ], + "selected_reward": -0.008640876971185207 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a close-up from a level angle held still?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a dense digital cover interwoven with emerald binary codes and golden magical runes visible at the center?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a blinding beam of pure white energy erupt from the center and cleave the digital cover?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the disintegrated fragments radiate outward at high speed leaving bright trails and forming a digital storm?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a pure black void with rich volumetric lighting effects and sharp edges on the numbers and runes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.16039440035820007 + ], + "opposite_logprob_tokens": [ + -1.9103944301605225 + ], + "logprob_reward": -0.16039440035820007, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.16039440035820007, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.32128581404685974, + 0.22005586326122284, + 0.1277264952659607, + 0.055823784321546555, + 0.06422323733568192, + 0.04639458283782005, + 0.03476761281490326, + 0.07134493440389633, + 0.050786085426807404, + 0.007591634523123503 + ], + "credit_map": [ + 0.025934329256415367, + 0.05591503903269768, + 0.03988579288125038, + 0.06344524770975113, + 0.048196472227573395, + 0.02031782455742359, + 0.03176061809062958, + 0.03583047538995743, + 0.010240177623927593, + 0.030359555035829544, + 0.012380357831716537, + 0.03747163712978363, + 0.06395775824785233, + 0.016146250069141388, + 0.02725783921778202, + 0.022242296487092972, + 0.008918814361095428, + 0.03770650550723076, + 0.007041038945317268, + 0.019981637597084045, + 0.01845209300518036, + 0.014420759864151478, + 0.009505931288003922, + 0.011699700728058815, + 0.007135619875043631, + 0.0025240767281502485, + 0.00749345263466239, + 0.0023571173660457134, + 0.0030941807199269533, + 0.0130529273301363, + 0.0032177185639739037, + 0.01694868877530098, + 0.007806919515132904, + 0.010369362309575081, + 0.00882619060575962, + 0.005626381374895573, + 0.011114156804978848, + 0.008742008358240128, + 0.004938703496009111, + 0.006799515802413225, + 0.00377690140157938, + 0.009682576172053814, + 0.0029428498819470406, + 0.004557877313345671, + 0.0110435476526618, + 0.006880174390971661, + 0.004443677142262459, + 0.003066978184506297, + 0.0034526584204286337, + 0.004117810633033514, + 0.004509171936661005, + 0.005537031684070826, + 0.004572965204715729, + 0.005422456189990044, + 0.0035056511405855417, + 0.0036498699337244034, + 0.0050347172655165195, + 0.003235047683119774, + 0.004333626013249159, + 0.005195349454879761, + 0.004503319971263409, + 0.02171729877591133, + 0.004948240239173174, + 0.02237733080983162, + 0.001701541361398995, + 0.0030432541389018297, + 0.010757018812000751, + 0.01895977184176445, + 0.006146694533526897, + 0.002776138484477997, + 0.0036592173855751753, + 0.003742448054254055, + 0.0009040031000040472, + 0.0007558641955256462, + 0.0013130834558978677, + 0.0013280643615871668, + 0.000652037444524467, + 0.0008057292434386909, + 0.0007907265680842102, + 0.0010421258630231023 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.16039440035820007 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "close-up shot from a level angle, the camera held still to capture the drastic change about to take place. At the center of the frame is a dense digital cover, interwoven with countless luminescent emerald binary codes and ancient magical runes shimmering with a mysterious golden glow. These streams of data flow vertically like cascading cascades, forming a complex barrier suspended against a pure black void. Suddenly, a blinding beam of pure white energy erupts from the very center of the frame, instantly cleaving the digital cover with the force of thunder. Following this devastating blast of light, the entire cover disintegrates into countless luminous fragments of code and runes. These fragments radiate outward at high speed, leaving bright trails in the air and forming a spectacular digital storm. The entire scene, rendered in real-time by the engine, presents a cinematic CG feel with ultra-high-definition resolution, making every scattered number and rune clearly visible with sharp edges. The burst of strong light creates extremely rich volumetric lighting effects in the darkness, and holy beams penetrate the diffuse digital dust, injecting unparalleled visual impact and epic atmosphere into this moment of digital destruction full of power and beauty.", + "seed": 24953, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000021/group_4/index_3494_seed_24953_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": -0.14084766265004872, + "vlm_logprob": -0.14084766265004872, + "vlm_yes_no_margin": 3.35, + "vlm_token_credit": -0.14084766265004872, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a close-up from a level angle held still?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a dense digital cover interwoven with emerald binary codes and golden magical runes visible at the center?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a blinding beam of pure white energy erupt from the center and cleave the digital cover?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the disintegrated fragments radiate outward at high speed leaving bright trails and forming a digital storm?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a pure black void with rich volumetric lighting effects and sharp edges on the numbers and runes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.5447090864181519 + ], + "opposite_logprob_tokens": [ + -0.9197090864181519 + ], + "logprob_reward": -0.5447090864181519, + "yes_no_margin_reward": 0.375, + "token_credit_reward": -0.5447090864181519, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16488610208034515, + 0.21017883718013763, + 0.07029353827238083, + 0.034007202833890915, + 0.08129023015499115, + 0.08082050085067749, + 0.06745737046003342, + 0.10240613669157028, + 0.08762526512145996, + 0.10103483498096466 + ], + "credit_map": [ + 0.01825992576777935, + 0.03113345429301262, + 0.022016236558556557, + 0.017585307359695435, + 0.009815743193030357, + 0.04238046333193779, + 0.010190717875957489, + 0.013504250906407833, + 0.01913437992334366, + 0.03880578652024269, + 0.011515193618834019, + 0.00735104363411665, + 0.011015032418072224, + 0.07843036204576492, + 0.01210546400398016, + 0.03182157874107361, + 0.00881606712937355, + 0.00871081929653883, + 0.011516271159052849, + 0.006579020991921425, + 0.007069895975291729, + 0.011198624968528748, + 0.006808863021433353, + 0.009593973867595196, + 0.003135864157229662, + 0.003934870008379221, + 0.003781846258789301, + 0.00383925111964345, + 0.0044242627918720245, + 0.004621434025466442, + 0.0036695189774036407, + 0.0066001578234136105, + 0.00675294641405344, + 0.016360798850655556, + 0.0062118517234921455, + 0.00952279195189476, + 0.006323434878140688, + 0.008486182428896427, + 0.010643595829606056, + 0.016988633200526237, + 0.007593377958983183, + 0.009554916992783546, + 0.008180303499102592, + 0.011498875916004181, + 0.01045005489140749, + 0.010219068266451359, + 0.011804216541349888, + 0.011519682593643665, + 0.0062997727654874325, + 0.008715522475540638, + 0.009028445929288864, + 0.008543500676751137, + 0.008279602974653244, + 0.007340404205024242, + 0.009825721383094788, + 0.009424402378499508, + 0.019983859732747078, + 0.0122159318998456, + 0.009656165726482868, + 0.01576344482600689, + 0.011842992156744003, + 0.00985946785658598, + 0.01514163613319397, + 0.007942640222609043, + 0.015118472278118134, + 0.008438212797045708, + 0.014604715630412102, + 0.007151544559746981, + 0.011928458698093891, + 0.01209455169737339, + 0.008512375876307487, + 0.00977693498134613, + 0.015428348444402218, + 0.022255729883909225, + 0.010506800375878811, + 0.007834513671696186, + 0.013391849584877491, + 0.010876631364226341, + 0.007777029648423195, + 0.012963936664164066 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 7 + ], + "selected_reward": -0.5447090864181519 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a close-up from a level angle held still?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a dense digital cover interwoven with emerald binary codes and golden magical runes visible at the center?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a blinding beam of pure white energy erupt from the center and cleave the digital cover?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the disintegrated fragments radiate outward at high speed leaving bright trails and forming a digital storm?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a pure black void with rich volumetric lighting effects and sharp edges on the numbers and runes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020624814555048943 + ], + "opposite_logprob_tokens": [ + -3.895624876022339 + ], + "logprob_reward": -0.020624814555048943, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020624814555048943, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4426526427268982, + 0.23045268654823303, + 0.0393470861017704, + 0.040222082287073135, + 0.033900026232004166, + 0.06241925060749054, + 0.034154947847127914, + 0.04737907275557518, + 0.031247489154338837, + 0.038224734365940094 + ], + "credit_map": [ + 0.04562251642346382, + 0.06280194222927094, + 0.05426297336816788, + 0.03496125340461731, + 0.02741326205432415, + 0.1315750628709793, + 0.03153126686811447, + 0.05448438972234726, + 0.018326422199606895, + 0.06090161204338074, + 0.012145966291427612, + 0.008559622801840305, + 0.013415820896625519, + 0.07908380776643753, + 0.013392322696745396, + 0.02462710626423359, + 0.0060428958386182785, + 0.005150388460606337, + 0.007312721572816372, + 0.00347872544080019, + 0.00341822043992579, + 0.004833885934203863, + 0.0035110781900584698, + 0.005599171854555607, + 0.004277301486581564, + 0.006432411726564169, + 0.003998218569904566, + 0.004649212118238211, + 0.006927244830876589, + 0.004771862179040909, + 0.005360241513699293, + 0.0038055928889662027, + 0.005170314572751522, + 0.0030085514299571514, + 0.0034306433517485857, + 0.00385640119202435, + 0.0034091374836862087, + 0.004531941842287779, + 0.004070046357810497, + 0.006422989536076784, + 0.007146931253373623, + 0.006973680574446917, + 0.0066097029484808445, + 0.0080148009583354, + 0.008555020205676556, + 0.007613860070705414, + 0.007909654639661312, + 0.009595595300197601, + 0.004244362935423851, + 0.0037150257267057896, + 0.0031796128023415804, + 0.004357889760285616, + 0.005474725738167763, + 0.004100873600691557, + 0.004129708744585514, + 0.004952749237418175, + 0.007519966457039118, + 0.004947588313370943, + 0.0037793784867972136, + 0.007921908982098103, + 0.0044064209796488285, + 0.004127426538616419, + 0.010330145247280598, + 0.004346233326941729, + 0.004077612422406673, + 0.0034937586169689894, + 0.004927358124405146, + 0.002728719264268875, + 0.0035401019267737865, + 0.0037356321699917316, + 0.0036843554116785526, + 0.005059953313320875, + 0.006090829614549875, + 0.00683882599696517, + 0.0036569833755493164, + 0.003832436865195632, + 0.005936999339610338, + 0.004014418460428715, + 0.003912350162863731, + 0.003941892646253109 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 5 + ], + "selected_reward": -0.020624814555048943 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a close-up from a level angle held still?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a dense digital cover interwoven with emerald binary codes and golden magical runes visible at the center?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a blinding beam of pure white energy erupt from the center and cleave the digital cover?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the disintegrated fragments radiate outward at high speed leaving bright trails and forming a digital storm?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a pure black void with rich volumetric lighting effects and sharp edges on the numbers and runes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0031949449330568314 + ], + "opposite_logprob_tokens": [ + -5.753194808959961 + ], + "logprob_reward": -0.0031949449330568314, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.0031949449330568314, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.264116495847702, + 0.24304261803627014, + 0.0751914530992508, + 0.06199154257774353, + 0.03809736669063568, + 0.06450919806957245, + 0.039422761648893356, + 0.1303117871284485, + 0.0250246562063694, + 0.058292124420404434 + ], + "credit_map": [ + 0.02473929338157177, + 0.03457598015666008, + 0.031152866780757904, + 0.021090714260935783, + 0.03047463670372963, + 0.07149574905633926, + 0.02419568970799446, + 0.02639155462384224, + 0.027026690542697906, + 0.01592891849577427, + 0.030177459120750427, + 0.013857584446668625, + 0.02289901301264763, + 0.07707349956035614, + 0.013807362876832485, + 0.04227207973599434, + 0.009381623938679695, + 0.009794201701879501, + 0.011165855452418327, + 0.007476001512259245, + 0.00837817881256342, + 0.00907143484801054, + 0.008657167665660381, + 0.011266984045505524, + 0.010230254381895065, + 0.007469832431524992, + 0.0065094646997749805, + 0.005949085578322411, + 0.007523251231759787, + 0.00958091951906681, + 0.006327493581920862, + 0.008401245810091496, + 0.0045981938019394875, + 0.0051793502643704414, + 0.004179477225989103, + 0.003417262341827154, + 0.003734308760613203, + 0.003931284882128239, + 0.006324269808828831, + 0.006733221001923084, + 0.004879206884652376, + 0.006960980128496885, + 0.0064407652243971825, + 0.00943785347044468, + 0.009073082357645035, + 0.006116238422691822, + 0.011643986217677593, + 0.009957083500921726, + 0.002678758231922984, + 0.005584167316555977, + 0.0032317962031811476, + 0.006103089079260826, + 0.006255748216062784, + 0.0063049509190022945, + 0.0046835108660161495, + 0.004580739885568619, + 0.015533968806266785, + 0.01396973803639412, + 0.00884714350104332, + 0.02311912178993225, + 0.015374825336039066, + 0.00931550096720457, + 0.031186005100607872, + 0.012965485453605652, + 0.0033469076734036207, + 0.002476287307217717, + 0.004015558864921331, + 0.002424748847261071, + 0.002719621639698744, + 0.0034088692627847195, + 0.0023708392400294542, + 0.004261824302375317, + 0.012467730790376663, + 0.006732998415827751, + 0.00691091688349843, + 0.0049217091873288155, + 0.006114882417023182, + 0.007620805874466896, + 0.005586089566349983, + 0.007936990819871426 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 7 + ], + "selected_reward": -0.0031949449330568314 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a close-up from a level angle held still?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a dense digital cover interwoven with emerald binary codes and golden magical runes visible at the center?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a blinding beam of pure white energy erupt from the center and cleave the digital cover?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the disintegrated fragments radiate outward at high speed leaving bright trails and forming a digital storm?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a pure black void with rich volumetric lighting effects and sharp edges on the numbers and runes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"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": [ + 0.2441670447587967, + 0.32594558596611023, + 0.1032755970954895, + 0.028909631073474884, + 0.049968961626291275, + 0.035382695496082306, + 0.037366144359111786, + 0.08605840057134628, + 0.03733676299452782, + 0.0515892468392849 + ], + "credit_map": [ + 0.0234085526317358, + 0.031742241233587265, + 0.02787097729742527, + 0.02767917700111866, + 0.021704306825995445, + 0.05047658830881119, + 0.02362327091395855, + 0.03766193613409996, + 0.028227893635630608, + 0.05434759333729744, + 0.01188010722398758, + 0.009275199845433235, + 0.035514868795871735, + 0.1253957599401474, + 0.02118895761668682, + 0.040115199983119965, + 0.015075214207172394, + 0.014904658310115337, + 0.01932346820831299, + 0.00887236651033163, + 0.008620471693575382, + 0.013574971817433834, + 0.008225064724683762, + 0.014679374173283577, + 0.0024469243362545967, + 0.005467026028782129, + 0.0025925070513039827, + 0.0035230224020779133, + 0.004035081248730421, + 0.0031896631699055433, + 0.0038944350089877844, + 0.0037609729915857315, + 0.0065388912335038185, + 0.005817095749080181, + 0.004873459227383137, + 0.005472373217344284, + 0.00417967326939106, + 0.006775872781872749, + 0.006425783969461918, + 0.009885811246931553, + 0.0026966442819684744, + 0.0039353868924081326, + 0.003877153154462576, + 0.005109652876853943, + 0.0053275884129107, + 0.004223442170768976, + 0.005418004933744669, + 0.004794823471456766, + 0.0032236736733466387, + 0.00501554599031806, + 0.0032754528801888227, + 0.005676967557519674, + 0.004655375611037016, + 0.004248839803040028, + 0.006260117515921593, + 0.005010167136788368, + 0.01031959056854248, + 0.007921533659100533, + 0.006592041812837124, + 0.013747641816735268, + 0.010310317389667034, + 0.0072148581966757774, + 0.019824760034680367, + 0.010127660818397999, + 0.005666014738380909, + 0.0038908447604626417, + 0.005495978519320488, + 0.004256399814039469, + 0.006003811489790678, + 0.004587583243846893, + 0.004210000392049551, + 0.0032261318992823362, + 0.011018529534339905, + 0.0063935816287994385, + 0.005974090192466974, + 0.004462840501219034, + 0.005466639995574951, + 0.00644852826371789, + 0.005192285869270563, + 0.0066327499225735664 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 2 + ], + "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 a close-up from a level angle held still?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a dense digital cover interwoven with emerald binary codes and golden magical runes visible at the center?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a blinding beam of pure white energy erupt from the center and cleave the digital cover?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the disintegrated fragments radiate outward at high speed leaving bright trails and forming a digital storm?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a pure black void with rich volumetric lighting effects and sharp edges on the numbers and runes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.127069890499115 + ], + "opposite_logprob_tokens": [ + -2.1270699501037598 + ], + "logprob_reward": -0.127069890499115, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.127069890499115, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13385628163814545, + 0.3250449001789093, + 0.05178271234035492, + 0.12170135229825974, + 0.057955965399742126, + 0.06645773351192474, + 0.08437567949295044, + 0.057428181171417236, + 0.028534136712551117, + 0.07286304235458374 + ], + "credit_map": [ + 0.016430208459496498, + 0.012177675031125546, + 0.02011839486658573, + 0.01612991653382778, + 0.01801004819571972, + 0.025742465630173683, + 0.014510244131088257, + 0.010737337172031403, + 0.0277020875364542, + 0.07757268846035004, + 0.01654922589659691, + 0.010070343501865864, + 0.016793223097920418, + 0.11531487107276917, + 0.018866201862692833, + 0.0421762652695179, + 0.0061581372283399105, + 0.008615031838417053, + 0.007031009066849947, + 0.004563508089631796, + 0.007945768535137177, + 0.0059149861335754395, + 0.006299567874521017, + 0.0052547091618180275, + 0.018068449571728706, + 0.017286278307437897, + 0.012380369007587433, + 0.013865753076970577, + 0.017191356047987938, + 0.014245284721255302, + 0.01466776430606842, + 0.013996104709804058, + 0.005387795623391867, + 0.012511256150901318, + 0.004460553172975779, + 0.008657557889819145, + 0.005486236419528723, + 0.004207917954772711, + 0.007543290499597788, + 0.009701363742351532, + 0.006825101096183062, + 0.00714691961184144, + 0.008785979822278023, + 0.00830701645463705, + 0.006173739675432444, + 0.00798268523067236, + 0.009851876646280289, + 0.011384418234229088, + 0.011530183255672455, + 0.009458471089601517, + 0.008775975555181503, + 0.010750479064881802, + 0.009357672184705734, + 0.01077223289757967, + 0.012210563756525517, + 0.011520111002027988, + 0.008898843079805374, + 0.006081968545913696, + 0.0048745498061180115, + 0.010699957609176636, + 0.004602296277880669, + 0.0060377358458936214, + 0.010493683628737926, + 0.005739150568842888, + 0.004248759709298611, + 0.0024748716969043016, + 0.00429080706089735, + 0.0029107287991791964, + 0.0026863855309784412, + 0.004033976700156927, + 0.002767546335235238, + 0.005121062509715557, + 0.011881534941494465, + 0.014970069751143456, + 0.007532019633799791, + 0.00589927239343524, + 0.011433886364102364, + 0.007788616698235273, + 0.005982594098895788, + 0.007375056389719248 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 3 + ], + "selected_reward": -0.127069890499115 + } + ], + "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/0000021/group_5/index_tensor([3252, 3252])_group_5_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000021/group_5/index_tensor([3252, 3252])_group_5_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..6b5e55a722495b1f572b04439ebc181b58f3793e --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000021/group_5/index_tensor([3252, 3252])_group_5_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "cinematic long shot, shot from a low angle, the camera smoothly pans right, following the flight path, to capture a flock of geese in a meticulously ordered V-formation, passing low over the mirror-smooth surface of a lake. The last rays of sunset paint the sky a gradient of hues from fiery orange to soft pink-purple. This warm light precisely outlines the golden edges of each goose's wings and casts a warm glow on the white feathers on their bellies. In exquisitely realistic detail, their powerful wings beat gracefully in sync, each gray-brown and white feather trembling slightly under the influence of aerodynamic forces. The tips of the wings almost touch the water, and each beat creates tiny, rapidly spreading ripples on the surface. On the distant horizon, the outlines of the rolling mountains are dyed into deep silhouettes, forming a sharp contrast with the colorful sky. The whole picture is full of a tranquil, magnificent and slightly poetic atmosphere, perfectly presenting the grandeur of nature and the dynamic beauty of life.", + "seed": 31888, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000021/group_5/index_3252_seed_31888_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": -0.019904146157205104, + "vlm_logprob": -0.019904146157205104, + "vlm_yes_no_margin": 4.5, + "vlm_token_credit": -0.019904146157205104, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a cinematic long shot featuring a flock of geese in a V-formation?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera shot from a low angle and smoothly panning right to follow the geese?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the geese fly low over a mirror-smooth lake surface with wing tips almost touching the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the sky painted with a sunset gradient from fiery orange to soft pink-purple, casting golden light on the geese?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are deep silhouette outlines of rolling mountains visible on the distant horizon contrasting with the sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.043322108685970306 + ], + "opposite_logprob_tokens": [ + -3.2933220863342285 + ], + "logprob_reward": -0.043322108685970306, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.043322108685970306, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.46217817068099976, + 0.09585961699485779, + 0.028482139110565186, + 0.09924487769603729, + 0.05901642516255379, + 0.036814894527196884, + 0.10492373257875443, + 0.03159884735941887, + 0.01428135670721531, + 0.06760001182556152 + ], + "credit_map": [ + 0.05498383194208145, + 0.0757175013422966, + 0.03831712529063225, + 0.0466604083776474, + 0.06730181723833084, + 0.11837245523929596, + 0.035428401082754135, + 0.025396600365638733, + 0.00938248261809349, + 0.014239266514778137, + 0.014049791730940342, + 0.008942211978137493, + 0.011970366351306438, + 0.011602728627622128, + 0.0110683124512434, + 0.01460445299744606, + 0.004569895565509796, + 0.0038482726085931063, + 0.0019278796389698982, + 0.00310412491671741, + 0.0033803354017436504, + 0.004544613417237997, + 0.0034964391961693764, + 0.0036105792969465256, + 0.013392950408160686, + 0.01147533766925335, + 0.010477669537067413, + 0.008514832705259323, + 0.02090560831129551, + 0.011053629219532013, + 0.011778190732002258, + 0.011646662838757038, + 0.009037836454808712, + 0.01974860206246376, + 0.0032541744876652956, + 0.0027040052227675915, + 0.005398311652243137, + 0.010621271096169949, + 0.004072170704603195, + 0.004180058371275663, + 0.004141807556152344, + 0.005704074166715145, + 0.004407593980431557, + 0.00423055374994874, + 0.00464975368231535, + 0.005543559789657593, + 0.00394674576818943, + 0.004190804902464151, + 0.007882491685450077, + 0.02982897311449051, + 0.006558434572070837, + 0.006127785425633192, + 0.030724134296178818, + 0.009456537663936615, + 0.007136230822652578, + 0.007209138013422489, + 0.004341796971857548, + 0.006462125573307276, + 0.002453746274113655, + 0.0025410056114196777, + 0.004608372692018747, + 0.005511304829269648, + 0.002768688602373004, + 0.0029118070378899574, + 0.002767808735370636, + 0.0017368084518238902, + 0.001431889715604484, + 0.0011689969105646014, + 0.0013404288329184055, + 0.002224087016656995, + 0.0015097069554030895, + 0.0021016302052885294, + 0.006609092932194471, + 0.0075913299806416035, + 0.014365216717123985, + 0.008230587467551231, + 0.005166708957403898, + 0.008480490185320377, + 0.008460692130029202, + 0.008695894852280617 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 3 + ], + "selected_reward": -0.043322108685970306 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a cinematic long shot featuring a flock of geese in a V-formation?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera shot from a low angle and smoothly panning right to follow the geese?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the geese fly low over a mirror-smooth lake surface with wing tips almost touching the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the sky painted with a sunset gradient from fiery orange to soft pink-purple, casting golden light on the geese?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are deep silhouette outlines of rolling mountains visible on the distant horizon contrasting with the sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029808735474944115 + ], + "opposite_logprob_tokens": [ + -3.529808759689331 + ], + "logprob_reward": -0.029808735474944115, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029808735474944115, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16930145025253296, + 0.11529906839132309, + 0.15360461175441742, + 0.18437595665454865, + 0.10692641139030457, + 0.07227511703968048, + 0.04250252991914749, + 0.055457860231399536, + 0.03083146922290325, + 0.0694255605340004 + ], + "credit_map": [ + 0.01670043356716633, + 0.02403305470943451, + 0.01679764688014984, + 0.02268267050385475, + 0.0222060214728117, + 0.0344967320561409, + 0.017797593027353287, + 0.014587294310331345, + 0.014420055784285069, + 0.01564808376133442, + 0.013028552755713463, + 0.010386601090431213, + 0.011298836208879948, + 0.015196863561868668, + 0.015393639914691448, + 0.019926441833376884, + 0.02151782624423504, + 0.0363968126475811, + 0.015109162777662277, + 0.013224045746028423, + 0.018625259399414062, + 0.01855665072798729, + 0.011405391618609428, + 0.018769459798932076, + 0.020011352375149727, + 0.053216494619846344, + 0.011306334286928177, + 0.009107164107263088, + 0.028130285441875458, + 0.04040834680199623, + 0.013217436149716377, + 0.008978537283837795, + 0.012272588908672333, + 0.02582644671201706, + 0.007811481598764658, + 0.007418384309858084, + 0.010928265750408173, + 0.010776122100651264, + 0.008262873627245426, + 0.023630257695913315, + 0.00626257574185729, + 0.012595283798873425, + 0.008724803104996681, + 0.007327025756239891, + 0.011030755005776882, + 0.011003046296536922, + 0.007138290908187628, + 0.008193337358534336, + 0.005615627858787775, + 0.0039777494966983795, + 0.0034831841476261616, + 0.0035442549269646406, + 0.008371369913220406, + 0.006618198473006487, + 0.005010029301047325, + 0.0058821155689656734, + 0.00616325531154871, + 0.010669607669115067, + 0.006660488434135914, + 0.004941554274410009, + 0.008186611346900463, + 0.008023574948310852, + 0.005038488656282425, + 0.005774280056357384, + 0.004561916459351778, + 0.004147232510149479, + 0.0024832917843014, + 0.0018575135618448257, + 0.004473389126360416, + 0.004744211211800575, + 0.0030767701100558043, + 0.005487144459038973, + 0.010303732939064503, + 0.008591474033892155, + 0.009740404784679413, + 0.00627580052241683, + 0.006881204433739185, + 0.014764634892344475, + 0.0049926359206438065, + 0.007875669747591019 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 2 + ], + "selected_reward": -0.029808735474944115 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a cinematic long shot featuring a flock of geese in a V-formation?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera shot from a low angle and smoothly panning right to follow the geese?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the geese fly low over a mirror-smooth lake surface with wing tips almost touching the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the sky painted with a sunset gradient from fiery orange to soft pink-purple, casting golden light on the geese?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are deep silhouette outlines of rolling mountains visible on the distant horizon contrasting with the sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020587796345353127 + ], + "opposite_logprob_tokens": [ + -3.895587682723999 + ], + "logprob_reward": -0.020587796345353127, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020587796345353127, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08109001815319061, + 0.06894593685865402, + 0.11854758113622665, + 0.0956554263830185, + 0.03942422196269035, + 0.21644403040409088, + 0.25270357728004456, + 0.09237595647573471, + 0.016992395743727684, + 0.017820898443460464 + ], + "credit_map": [ + 0.009231023490428925, + 0.011517117731273174, + 0.008472125045955181, + 0.01257284078747034, + 0.00863988883793354, + 0.015132876113057137, + 0.010233687236905098, + 0.005290455650538206, + 0.008342951536178589, + 0.010932627134025097, + 0.008946306072175503, + 0.006882041692733765, + 0.007899267598986626, + 0.008096752688288689, + 0.006817768793553114, + 0.011028217151761055, + 0.014191258698701859, + 0.028546225279569626, + 0.010617806576192379, + 0.009804651141166687, + 0.01727638766169548, + 0.011117386631667614, + 0.014332801103591919, + 0.012661068700253963, + 0.011180685833096504, + 0.02157542109489441, + 0.007787028327584267, + 0.006584954913705587, + 0.01850028522312641, + 0.01354350708425045, + 0.010078346356749535, + 0.00640520453453064, + 0.0048635415732860565, + 0.010282007977366447, + 0.0031704227440059185, + 0.002960709622129798, + 0.004148280248045921, + 0.0047840033657848835, + 0.002989316824823618, + 0.006225941237062216, + 0.016347253695130348, + 0.05222124606370926, + 0.019221220165491104, + 0.01954532414674759, + 0.04970254749059677, + 0.021657776087522507, + 0.019452324137091637, + 0.018296340480446815, + 0.015515066683292389, + 0.08950978517532349, + 0.010229078121483326, + 0.01192115992307663, + 0.08457410335540771, + 0.014880776405334473, + 0.014238105155527592, + 0.011835502460598946, + 0.01872354932129383, + 0.02551829069852829, + 0.004811558872461319, + 0.005221362225711346, + 0.016517313197255135, + 0.010399790480732918, + 0.0041871508583426476, + 0.0069969417527318, + 0.003250905778259039, + 0.0023740993347018957, + 0.0016629222081974149, + 0.0015797052765265107, + 0.001819743774831295, + 0.002959198784083128, + 0.0016345700714737177, + 0.0017112493515014648, + 0.002149650128558278, + 0.0021301943343132734, + 0.0027434760704636574, + 0.002244764007627964, + 0.0016934885643422604, + 0.0026242684107273817, + 0.0020692101679742336, + 0.002165848156437278 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 5, + 2 + ], + "selected_reward": -0.020587796345353127 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a cinematic long shot featuring a flock of geese in a V-formation?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera shot from a low angle and smoothly panning right to follow the geese?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the geese fly low over a mirror-smooth lake surface with wing tips almost touching the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the sky painted with a sunset gradient from fiery orange to soft pink-purple, casting golden light on the geese?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are deep silhouette outlines of rolling mountains visible on the distant horizon contrasting with the sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0017131189815700054 + ], + "opposite_logprob_tokens": [ + -6.376713275909424 + ], + "logprob_reward": -0.0017131189815700054, + "yes_no_margin_reward": 6.375, + "token_credit_reward": -0.0017131189815700054, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2332872599363327, + 0.03375503420829773, + 0.0472632460296154, + 0.11329014599323273, + 0.10580261796712875, + 0.12975338101387024, + 0.22713454067707062, + 0.04881773889064789, + 0.026612302288413048, + 0.034283772110939026 + ], + "credit_map": [ + 0.020406020805239677, + 0.02382148988544941, + 0.0209841076284647, + 0.023564698174595833, + 0.05025240406394005, + 0.054913364350795746, + 0.02263803966343403, + 0.016707133501768112, + 0.006731155328452587, + 0.003802374005317688, + 0.0033016586676239967, + 0.0031525103840976954, + 0.0037173281889408827, + 0.0039938767440617085, + 0.003990515600889921, + 0.005065614357590675, + 0.007493104785680771, + 0.008348562754690647, + 0.0037397167179733515, + 0.004030577372759581, + 0.004535789601504803, + 0.009088711813092232, + 0.005517018027603626, + 0.004509765654802322, + 0.012177329510450363, + 0.018270766362547874, + 0.010638491250574589, + 0.009473874233663082, + 0.02578710950911045, + 0.013469909317791462, + 0.012326576747000217, + 0.01114608533680439, + 0.01717723160982132, + 0.029549015685915947, + 0.007407476659864187, + 0.007351478561758995, + 0.012675582431256771, + 0.014522725716233253, + 0.007061895914375782, + 0.010057203471660614, + 0.01937389001250267, + 0.019879532977938652, + 0.013303818181157112, + 0.015604082494974136, + 0.015348349697887897, + 0.01940724439918995, + 0.011274198070168495, + 0.015562251210212708, + 0.013444246724247932, + 0.08729071170091629, + 0.008578489534556866, + 0.009945062920451164, + 0.08394315093755722, + 0.010004735551774502, + 0.006601701956242323, + 0.007326435297727585, + 0.005640038754791021, + 0.008077513426542282, + 0.0051337373442947865, + 0.006140009965747595, + 0.005522347055375576, + 0.009280318394303322, + 0.0041404711082577705, + 0.00488330377265811, + 0.003937702160328627, + 0.0034825042821466923, + 0.0027311902958899736, + 0.0021582855843007565, + 0.0045921690762043, + 0.004603010136634111, + 0.0021598448511213064, + 0.002947596600279212, + 0.005755552556365728, + 0.0035002119839191437, + 0.004643189255148172, + 0.003176057245582342, + 0.004366717766970396, + 0.0055245221592485905, + 0.003454255871474743, + 0.0038632669020444155 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 5 + ], + "selected_reward": -0.0017131189815700054 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a cinematic long shot featuring a flock of geese in a V-formation?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera shot from a low angle and smoothly panning right to follow the geese?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the geese fly low over a mirror-smooth lake surface with wing tips almost touching the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the sky painted with a sunset gradient from fiery orange to soft pink-purple, casting golden light on the geese?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are deep silhouette outlines of rolling mountains visible on the distant horizon contrasting with the sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004088971298187971 + ], + "opposite_logprob_tokens": [ + -5.504088878631592 + ], + "logprob_reward": -0.004088971298187971, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004088971298187971, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07912995666265488, + 0.03189009800553322, + 0.01888258010149002, + 0.14672082662582397, + 0.04392899572849274, + 0.18694821000099182, + 0.14932262897491455, + 0.03337821736931801, + 0.04028589650988579, + 0.2695125639438629 + ], + "credit_map": [ + 0.011533372104167938, + 0.011008966714143753, + 0.006988548208028078, + 0.011710242368280888, + 0.012303879484534264, + 0.0122448755428195, + 0.007164770271629095, + 0.006175297778099775, + 0.004442368634045124, + 0.004290917422622442, + 0.004577533807605505, + 0.0031628008000552654, + 0.0035101298708468676, + 0.003452880308032036, + 0.004428517539054155, + 0.004024949390441179, + 0.0024948029313236475, + 0.003406951669603586, + 0.0015586952213197947, + 0.0013173578772693872, + 0.0025372228119522333, + 0.0032309351954609156, + 0.0021716859191656113, + 0.0021649280097335577, + 0.01828421838581562, + 0.020137907937169075, + 0.013787210918962955, + 0.01085957046598196, + 0.03012879006564617, + 0.025714265182614326, + 0.014422707259654999, + 0.013386148028075695, + 0.006684105843305588, + 0.010397111065685749, + 0.0035871746949851513, + 0.0034746851306408644, + 0.006137542426586151, + 0.006435320246964693, + 0.0033369504380971193, + 0.0038761054165661335, + 0.01428070105612278, + 0.05273747444152832, + 0.011644676327705383, + 0.012602880597114563, + 0.04823549464344978, + 0.016058381646871567, + 0.01805071532726288, + 0.013337895274162292, + 0.007868342101573944, + 0.05662035197019577, + 0.005129450000822544, + 0.006239461246877909, + 0.05523616448044777, + 0.00856868363916874, + 0.004674611147493124, + 0.004985569976270199, + 0.003285080660134554, + 0.0049005537293851376, + 0.004152837209403515, + 0.0038986308500170708, + 0.0037377418484538794, + 0.005350450985133648, + 0.004423819482326508, + 0.0036291046999394894, + 0.006795617286115885, + 0.007297765463590622, + 0.003962989896535873, + 0.0028562473598867655, + 0.0058291759341955185, + 0.00652390718460083, + 0.0031735398806631565, + 0.0038466525729745626, + 0.034678999334573746, + 0.03359072282910347, + 0.0277868565171957, + 0.02314225770533085, + 0.03492436558008194, + 0.06517049670219421, + 0.022752154618501663, + 0.027466721832752228 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 5, + 6 + ], + "selected_reward": -0.004088971298187971 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "cinematic long shot, shot from a low angle, the camera smoothly pans right, following the flight path, to capture a flock of geese in a meticulously ordered V-formation, passing low over the mirror-smooth surface of a lake. The last rays of sunset paint the sky a gradient of hues from fiery orange to soft pink-purple. This warm light precisely outlines the golden edges of each goose's wings and casts a warm glow on the white feathers on their bellies. In exquisitely realistic detail, their powerful wings beat gracefully in sync, each gray-brown and white feather trembling slightly under the influence of aerodynamic forces. The tips of the wings almost touch the water, and each beat creates tiny, rapidly spreading ripples on the surface. On the distant horizon, the outlines of the rolling mountains are dyed into deep silhouettes, forming a sharp contrast with the colorful sky. The whole picture is full of a tranquil, magnificent and slightly poetic atmosphere, perfectly presenting the grandeur of nature and the dynamic beauty of life.", + "seed": 31889, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000021/group_5/index_3252_seed_31889_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": -0.024850771366618573, + "vlm_logprob": -0.024850771366618573, + "vlm_yes_no_margin": 4.25, + "vlm_token_credit": -0.024850771366618573, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a cinematic long shot featuring a flock of geese in a V-formation?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera shot from a low angle and smoothly panning right to follow the geese?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the geese fly low over a mirror-smooth lake surface with wing tips almost touching the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the sky painted with a sunset gradient from fiery orange to soft pink-purple, casting golden light on the geese?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are deep silhouette outlines of rolling mountains visible on the distant horizon contrasting with the sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02691148966550827 + ], + "opposite_logprob_tokens": [ + -3.776911497116089 + ], + "logprob_reward": -0.02691148966550827, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.02691148966550827, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.42298662662506104, + 0.08842793107032776, + 0.09021356701850891, + 0.028619226068258286, + 0.11893575638532639, + 0.04623197764158249, + 0.11363058537244797, + 0.0266902893781662, + 0.032123830169439316, + 0.032140281051397324 + ], + "credit_map": [ + 0.030227841809391975, + 0.03791278973221779, + 0.013068962842226028, + 0.01755070872604847, + 0.16147971153259277, + 0.06536565721035004, + 0.026551242917776108, + 0.07082970440387726, + 0.0072523485869169235, + 0.009129893966019154, + 0.009952054359018803, + 0.007409554906189442, + 0.023206418380141258, + 0.014260947704315186, + 0.010407879017293453, + 0.006808830890804529, + 0.014300950802862644, + 0.01691594533622265, + 0.007388371508568525, + 0.008169807493686676, + 0.015463868156075478, + 0.010155826807022095, + 0.007843594998121262, + 0.00997519213706255, + 0.004591833800077438, + 0.0049025071784853935, + 0.003554616356268525, + 0.0024493872188031673, + 0.0029589785262942314, + 0.004128906410187483, + 0.003016113769263029, + 0.003016882110387087, + 0.014044034294784069, + 0.016914663836359978, + 0.013360823504626751, + 0.012929776683449745, + 0.015165839344263077, + 0.018021361902356148, + 0.015323897823691368, + 0.013175356201827526, + 0.006897337269037962, + 0.01069849357008934, + 0.006548404693603516, + 0.004867543000727892, + 0.00638918811455369, + 0.0033334060572087765, + 0.0034569043200463057, + 0.004040702246129513, + 0.007629802916198969, + 0.01195321511477232, + 0.005027029663324356, + 0.005680351983755827, + 0.04371850937604904, + 0.018494240939617157, + 0.016700422391295433, + 0.004427019506692886, + 0.002624662360176444, + 0.003709056181833148, + 0.0034035402350127697, + 0.0036409462336450815, + 0.003471646923571825, + 0.002675004070624709, + 0.0034731768537312746, + 0.003692258382216096, + 0.005534048657864332, + 0.0032125920988619328, + 0.0038701724261045456, + 0.0019270580960437655, + 0.005732654593884945, + 0.004235481843352318, + 0.0025319417472928762, + 0.005079880356788635, + 0.004026224836707115, + 0.0052108438685536385, + 0.0034158711787313223, + 0.0030792702455073595, + 0.004577786661684513, + 0.00404610438272357, + 0.003967344295233488, + 0.003816835582256317 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 6 + ], + "selected_reward": -0.02691148966550827 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a cinematic long shot featuring a flock of geese in a V-formation?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera shot from a low angle and smoothly panning right to follow the geese?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the geese fly low over a mirror-smooth lake surface with wing tips almost touching the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the sky painted with a sunset gradient from fiery orange to soft pink-purple, casting golden light on the geese?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are deep silhouette outlines of rolling mountains visible on the distant horizon contrasting with the sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03370651230216026 + ], + "opposite_logprob_tokens": [ + -3.4087064266204834 + ], + "logprob_reward": -0.03370651230216026, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03370651230216026, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11349181085824966, + 0.032772816717624664, + 0.36588189005851746, + 0.08112553507089615, + 0.08372095972299576, + 0.08050383627414703, + 0.1344662755727768, + 0.04703094810247421, + 0.030119353905320168, + 0.030886506661772728 + ], + "credit_map": [ + 0.01362142525613308, + 0.017113665118813515, + 0.012174663133919239, + 0.014540852047502995, + 0.01394478976726532, + 0.01566241681575775, + 0.014630374498665333, + 0.01180363167077303, + 0.004340064246207476, + 0.004312447737902403, + 0.003915686625987291, + 0.003790548536926508, + 0.004417869262397289, + 0.004778739530593157, + 0.004229404032230377, + 0.0029880600050091743, + 0.05846462398767471, + 0.06729603558778763, + 0.035554688423871994, + 0.032193105667829514, + 0.03475909307599068, + 0.05324844643473625, + 0.03700670227408409, + 0.047359220683574677, + 0.01597006246447563, + 0.010402068495750427, + 0.00785750336945057, + 0.008068785071372986, + 0.011601862497627735, + 0.011310575529932976, + 0.006673962343484163, + 0.009240728802978992, + 0.01117528136819601, + 0.015313522890210152, + 0.005820524878799915, + 0.0057720281183719635, + 0.013150122947990894, + 0.01933198794722557, + 0.00808860082179308, + 0.005068901460617781, + 0.01132464874535799, + 0.027340400964021683, + 0.004564168397337198, + 0.005796333309262991, + 0.009494127705693245, + 0.009106273762881756, + 0.00475709792226553, + 0.008120784536004066, + 0.008671613410115242, + 0.012344827875494957, + 0.006812089588493109, + 0.008190103806555271, + 0.05071668699383736, + 0.019616257399320602, + 0.022758599370718002, + 0.005356106441468, + 0.0048714494332671165, + 0.015092683024704456, + 0.00845666229724884, + 0.002621883526444435, + 0.006237823516130447, + 0.00427146116271615, + 0.0032703550532460213, + 0.002208636375144124, + 0.004656977951526642, + 0.005304302554577589, + 0.0025538115296512842, + 0.0018451635260134935, + 0.006396627984941006, + 0.004194836597889662, + 0.0025363152381032705, + 0.0026313213165849447, + 0.004367296118289232, + 0.003974519204348326, + 0.00275502260774374, + 0.0033524869941174984, + 0.005734639707952738, + 0.0050112344324588776, + 0.0035873379092663527, + 0.0021039689891040325 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 6, + 0 + ], + "selected_reward": -0.03370651230216026 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a cinematic long shot featuring a flock of geese in a V-formation?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera shot from a low angle and smoothly panning right to follow the geese?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the geese fly low over a mirror-smooth lake surface with wing tips almost touching the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the sky painted with a sunset gradient from fiery orange to soft pink-purple, casting golden light on the geese?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are deep silhouette outlines of rolling mountains visible on the distant horizon contrasting with the sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05496400222182274 + ], + "opposite_logprob_tokens": [ + -2.929964065551758 + ], + "logprob_reward": -0.05496400222182274, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05496400222182274, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1951473355293274, + 0.05193702504038811, + 0.22274239361286163, + 0.02437528222799301, + 0.06580613553524017, + 0.08931928128004074, + 0.19233576953411102, + 0.09254514425992966, + 0.04821862652897835, + 0.0175730362534523 + ], + "credit_map": [ + 0.015672562643885612, + 0.02405078336596489, + 0.014652573503553867, + 0.012420817278325558, + 0.060948487371206284, + 0.029911303892731667, + 0.020667649805545807, + 0.016823144629597664, + 0.0047601922415196896, + 0.0065480307675898075, + 0.007040513213723898, + 0.004932515788823366, + 0.009623979218304157, + 0.00748145068064332, + 0.006486047059297562, + 0.005064293276518583, + 0.04307004436850548, + 0.03260190784931183, + 0.020421208813786507, + 0.016799448058009148, + 0.03380872681736946, + 0.023647885769605637, + 0.021332770586013794, + 0.03106040693819523, + 0.004029710777103901, + 0.004173111170530319, + 0.002656286582350731, + 0.001683168695308268, + 0.0039109112694859505, + 0.0037380526773631573, + 0.0019206057768315077, + 0.0022634363267570734, + 0.00913273449987173, + 0.013721893541514874, + 0.003696597646921873, + 0.005245325155556202, + 0.007315975148230791, + 0.01567559316754341, + 0.004618693143129349, + 0.006399320904165506, + 0.009847071953117847, + 0.02552221156656742, + 0.006795327179133892, + 0.0066639878787100315, + 0.009120648726820946, + 0.011685921810567379, + 0.007646999321877956, + 0.012037113308906555, + 0.012343664653599262, + 0.01801243983209133, + 0.008295048959553242, + 0.009130082093179226, + 0.07556074112653732, + 0.029449928551912308, + 0.032453227788209915, + 0.007090629078447819, + 0.0076385182328522205, + 0.020185386762022972, + 0.01373649388551712, + 0.009682604111731052, + 0.01081738993525505, + 0.006800444796681404, + 0.013122987933456898, + 0.010561315342783928, + 0.006658349186182022, + 0.007936463691294193, + 0.004004249814897776, + 0.00359945185482502, + 0.009426756761968136, + 0.008567530661821365, + 0.0038411023560911417, + 0.00418472196906805, + 0.002377868629992008, + 0.0031551627907902002, + 0.0012686297995969653, + 0.0014367573894560337, + 0.0028443108312785625, + 0.003033821005374193, + 0.0016606758581474423, + 0.0017958107637241483 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 6 + ], + "selected_reward": -0.05496400222182274 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a cinematic long shot featuring a flock of geese in a V-formation?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera shot from a low angle and smoothly panning right to follow the geese?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the geese fly low over a mirror-smooth lake surface with wing tips almost touching the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the sky painted with a sunset gradient from fiery orange to soft pink-purple, casting golden light on the geese?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are deep silhouette outlines of rolling mountains visible on the distant horizon contrasting with the sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0019418209558352828 + ], + "opposite_logprob_tokens": [ + -6.251941680908203 + ], + "logprob_reward": -0.0019418209558352828, + "yes_no_margin_reward": 6.25, + "token_credit_reward": -0.0019418209558352828, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3441358208656311, + 0.07684224098920822, + 0.07716592401266098, + 0.04486231878399849, + 0.0898197740316391, + 0.09704415500164032, + 0.10526184737682343, + 0.048679448664188385, + 0.03955494612455368, + 0.07663353532552719 + ], + "credit_map": [ + 0.022069351747632027, + 0.029617857187986374, + 0.009125001728534698, + 0.014911097474396229, + 0.13059277832508087, + 0.059174858033657074, + 0.02412763610482216, + 0.054517243057489395, + 0.005886940285563469, + 0.011839030310511589, + 0.00987185351550579, + 0.00855338480323553, + 0.010346776805818081, + 0.010454813949763775, + 0.010078663006424904, + 0.009810778312385082, + 0.00795084610581398, + 0.018787862733006477, + 0.0075837415643036366, + 0.0070075467228889465, + 0.010284475050866604, + 0.008436008356511593, + 0.008380778133869171, + 0.008734670467674732, + 0.006894279271364212, + 0.0064159478060901165, + 0.005066338926553726, + 0.005109093617647886, + 0.004759110044687986, + 0.006566336378455162, + 0.004472971428185701, + 0.005578243173658848, + 0.016293300315737724, + 0.025135252624750137, + 0.006614665500819683, + 0.009108292870223522, + 0.009170075878500938, + 0.010886592790484428, + 0.0068338667042553425, + 0.005777719430625439, + 0.010775127448141575, + 0.020666517317295074, + 0.009527301415801048, + 0.008277646265923977, + 0.011988304555416107, + 0.015034268610179424, + 0.010311832651495934, + 0.010463154874742031, + 0.007671299856156111, + 0.011247552931308746, + 0.004759876057505608, + 0.004433727823197842, + 0.035894975066185, + 0.014379419386386871, + 0.02017749473452568, + 0.006697503384202719, + 0.007154371123760939, + 0.011477458290755749, + 0.004420265089720488, + 0.004367999732494354, + 0.006626565009355545, + 0.0058091674000024796, + 0.004103456623852253, + 0.0047201719135046005, + 0.006099728401750326, + 0.007564743049442768, + 0.0035578622482717037, + 0.0029305533971637487, + 0.00667233020067215, + 0.004719304386526346, + 0.0028057650197297335, + 0.005204657558351755, + 0.010722123086452484, + 0.010798281989991665, + 0.004813492763787508, + 0.004114788491278887, + 0.02280452288687229, + 0.012414010241627693, + 0.006571517325937748, + 0.004394800402224064 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 5 + ], + "selected_reward": -0.0019418209558352828 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a cinematic long shot featuring a flock of geese in a V-formation?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera shot from a low angle and smoothly panning right to follow the geese?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the geese fly low over a mirror-smooth lake surface with wing tips almost touching the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the sky painted with a sunset gradient from fiery orange to soft pink-purple, casting golden light on the geese?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are deep silhouette outlines of rolling mountains visible on the distant horizon contrasting with the sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0067300316877663136 + ], + "opposite_logprob_tokens": [ + -5.006730079650879 + ], + "logprob_reward": -0.0067300316877663136, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.0067300316877663136, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.31196466088294983, + 0.05816452577710152, + 0.04854834824800491, + 0.02359537035226822, + 0.02761191315948963, + 0.10033509135246277, + 0.045778509229421616, + 0.04089534282684326, + 0.02490375004708767, + 0.3182024657726288 + ], + "credit_map": [ + 0.01978331245481968, + 0.028886325657367706, + 0.007983813993632793, + 0.009871171787381172, + 0.12474684417247772, + 0.0456756129860878, + 0.017549565061926842, + 0.05746804177761078, + 0.0044351862743496895, + 0.006106474902480841, + 0.007820003665983677, + 0.0045798299834132195, + 0.013034917414188385, + 0.008233596570789814, + 0.008694745600223541, + 0.005259774159640074, + 0.0050258333794772625, + 0.013618318364024162, + 0.0037095581647008657, + 0.0029189002234488726, + 0.010509848594665527, + 0.005998195149004459, + 0.0030452918726950884, + 0.0037224057596176863, + 0.0053568193688988686, + 0.0033007031306624413, + 0.0018911842489615083, + 0.001884590950794518, + 0.003614903660491109, + 0.0032626704778522253, + 0.0019573948811739683, + 0.002327102469280362, + 0.0036289584822952747, + 0.004593075253069401, + 0.001939978450536728, + 0.002807458397001028, + 0.004734233487397432, + 0.005297957453876734, + 0.002361470367759466, + 0.0022487810347229242, + 0.008670817129313946, + 0.018195029348134995, + 0.011093361303210258, + 0.008481512777507305, + 0.01183186937123537, + 0.01609090529382229, + 0.012777010910212994, + 0.013194586150348186, + 0.003893829183652997, + 0.006061104126274586, + 0.003322602016851306, + 0.00266942847520113, + 0.01190995704382658, + 0.006508087273687124, + 0.008664724417030811, + 0.002748778322711587, + 0.004381303209811449, + 0.009613854810595512, + 0.00539437448605895, + 0.0038284151814877987, + 0.005125754978507757, + 0.003903632052242756, + 0.004083259962499142, + 0.004564747214317322, + 0.004163062199950218, + 0.004155660048127174, + 0.0023749268148094416, + 0.0017575916135683656, + 0.0035278666764497757, + 0.003810798516497016, + 0.0016797870630398393, + 0.0034340566489845514, + 0.0420546755194664, + 0.04986882209777832, + 0.023249581456184387, + 0.022406568750739098, + 0.07517586648464203, + 0.05434442684054375, + 0.02722976543009281, + 0.023872748017311096 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 5 + ], + "selected_reward": -0.0067300316877663136 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "cinematic long shot, shot from a low angle, the camera smoothly pans right, following the flight path, to capture a flock of geese in a meticulously ordered V-formation, passing low over the mirror-smooth surface of a lake. The last rays of sunset paint the sky a gradient of hues from fiery orange to soft pink-purple. This warm light precisely outlines the golden edges of each goose's wings and casts a warm glow on the white feathers on their bellies. In exquisitely realistic detail, their powerful wings beat gracefully in sync, each gray-brown and white feather trembling slightly under the influence of aerodynamic forces. The tips of the wings almost touch the water, and each beat creates tiny, rapidly spreading ripples on the surface. On the distant horizon, the outlines of the rolling mountains are dyed into deep silhouettes, forming a sharp contrast with the colorful sky. The whole picture is full of a tranquil, magnificent and slightly poetic atmosphere, perfectly presenting the grandeur of nature and the dynamic beauty of life.", + "seed": 31890, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000021/group_5/index_3252_seed_31890_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": -0.014163771807216109, + "vlm_logprob": -0.014163771807216109, + "vlm_yes_no_margin": 4.799999952316284, + "vlm_token_credit": -0.014163771807216109, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a cinematic long shot featuring a flock of geese in a V-formation?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera shot from a low angle and smoothly panning right to follow the geese?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the geese fly low over a mirror-smooth lake surface with wing tips almost touching the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the sky painted with a sunset gradient from fiery orange to soft pink-purple, casting golden light on the geese?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are deep silhouette outlines of rolling mountains visible on the distant horizon contrasting with the sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02344144694507122 + ], + "opposite_logprob_tokens": [ + -4.023441314697266 + ], + "logprob_reward": -0.02344144694507122, + "yes_no_margin_reward": 3.999999761581421, + "token_credit_reward": -0.02344144694507122, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.7198326587677002, + 0.03860624134540558, + 0.032348111271858215, + 0.03417930006980896, + 0.04856369271874428, + 0.023007746785879135, + 0.03866992145776749, + 0.0180506594479084, + 0.017321791499853134, + 0.029419850558042526 + ], + "credit_map": [ + 0.03846787288784981, + 0.026431715115904808, + 0.033202093094587326, + 0.03613191470503807, + 0.2056942880153656, + 0.2044650763273239, + 0.062375422567129135, + 0.11306431144475937, + 0.00343067804351449, + 0.005374066531658173, + 0.00408617127686739, + 0.0036029531620442867, + 0.00727939885109663, + 0.0032874930184334517, + 0.0057259672321379185, + 0.005819516722112894, + 0.003254561685025692, + 0.005902115721255541, + 0.0035541076213121414, + 0.0036049147602170706, + 0.005779536906629801, + 0.003590669482946396, + 0.003530994290485978, + 0.0031312170904129744, + 0.003849621629342437, + 0.00717141292989254, + 0.0034656694624572992, + 0.003577667288482189, + 0.004286026116460562, + 0.004762808792293072, + 0.004035156220197678, + 0.003030944848433137, + 0.008127234876155853, + 0.007204160559922457, + 0.0035705009941011667, + 0.00457770237699151, + 0.005673173349350691, + 0.008612614125013351, + 0.006197164300829172, + 0.004601147025823593, + 0.0029584208969026804, + 0.002733321627601981, + 0.0023227580823004246, + 0.0025728153996169567, + 0.0028214254416525364, + 0.0023506612051278353, + 0.0031773557420819998, + 0.004070988390594721, + 0.0029093516059219837, + 0.010739562101662159, + 0.003416702151298523, + 0.003260512137785554, + 0.00551911536604166, + 0.0038325837813317776, + 0.004888225346803665, + 0.004103872925043106, + 0.0017557133687660098, + 0.0027565325144678354, + 0.003128268988803029, + 0.0024629549589008093, + 0.0021038262639194727, + 0.0017118649557232857, + 0.0024801611434668303, + 0.0016513388836756349, + 0.0017869394505396485, + 0.0015191317070275545, + 0.0025397350545972586, + 0.00264365435577929, + 0.002225040690973401, + 0.0016594122862443328, + 0.002433472080156207, + 0.0025144079700112343, + 0.003745666239410639, + 0.008555063046514988, + 0.00389951653778553, + 0.0030049162451177835, + 0.0022368880454450846, + 0.002096227603033185, + 0.002570349955931306, + 0.003311226377263665 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 6 + ], + "selected_reward": -0.02344144694507122 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a cinematic long shot featuring a flock of geese in a V-formation?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera shot from a low angle and smoothly panning right to follow the geese?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the geese fly low over a mirror-smooth lake surface with wing tips almost touching the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the sky painted with a sunset gradient from fiery orange to soft pink-purple, casting golden light on the geese?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are deep silhouette outlines of rolling mountains visible on the distant horizon contrasting with the sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02635188400745392 + ], + "opposite_logprob_tokens": [ + -3.6513519287109375 + ], + "logprob_reward": -0.02635188400745392, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.02635188400745392, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.32075265049934387, + 0.06261266022920609, + 0.07620596885681152, + 0.04479891061782837, + 0.038284845650196075, + 0.11413775384426117, + 0.155803844332695, + 0.05504903569817543, + 0.1084783598780632, + 0.023875990882515907 + ], + "credit_map": [ + 0.02337309904396534, + 0.024689042940735817, + 0.017643596976995468, + 0.018387937918305397, + 0.08017748594284058, + 0.08142530173063278, + 0.024616651237010956, + 0.050439536571502686, + 0.005989869125187397, + 0.00621721101924777, + 0.006876625120639801, + 0.005956993903964758, + 0.013141624629497528, + 0.007567542605102062, + 0.009135459549725056, + 0.007727331947535276, + 0.0059530120342969894, + 0.00958604272454977, + 0.010420707054436207, + 0.012310744263231754, + 0.009913777001202106, + 0.006466194521635771, + 0.01311384979635477, + 0.00844164751470089, + 0.004870409145951271, + 0.005571742542088032, + 0.004905261564999819, + 0.005343776196241379, + 0.005920279771089554, + 0.007504939567297697, + 0.0060490830801427364, + 0.004633419681340456, + 0.007569044828414917, + 0.004792904946953058, + 0.0026514194905757904, + 0.0031725715380162, + 0.004944234620779753, + 0.007475919555872679, + 0.0035686036571860313, + 0.004110144916921854, + 0.01393486000597477, + 0.014266454614698887, + 0.01086148526519537, + 0.011778388172388077, + 0.01805109903216362, + 0.018885144963860512, + 0.010678944177925587, + 0.01568138599395752, + 0.015209712088108063, + 0.040312111377716064, + 0.012519926764070988, + 0.00899178721010685, + 0.02441369742155075, + 0.022686250507831573, + 0.014118175953626633, + 0.01755218580365181, + 0.005980680231004953, + 0.010936425067484379, + 0.007330293767154217, + 0.0059958514757454395, + 0.0071811978705227375, + 0.005560095421969891, + 0.006756216287612915, + 0.005308275111019611, + 0.01073412038385868, + 0.012022211216390133, + 0.018195312470197678, + 0.017040478065609932, + 0.011765892617404461, + 0.012749945744872093, + 0.013759569264948368, + 0.012210829183459282, + 0.0024743019603192806, + 0.00290752318687737, + 0.0027231103740632534, + 0.002142620272934437, + 0.003195198718458414, + 0.004766385070979595, + 0.002446273574605584, + 0.0032205767929553986 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 5 + ], + "selected_reward": -0.02635188400745392 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a cinematic long shot featuring a flock of geese in a V-formation?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera shot from a low angle and smoothly panning right to follow the geese?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the geese fly low over a mirror-smooth lake surface with wing tips almost touching the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the sky painted with a sunset gradient from fiery orange to soft pink-purple, casting golden light on the geese?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are deep silhouette outlines of rolling mountains visible on the distant horizon contrasting with the sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016075992956757545 + ], + "opposite_logprob_tokens": [ + -4.14107608795166 + ], + "logprob_reward": -0.016075992956757545, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016075992956757545, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3432900607585907, + 0.06978143751621246, + 0.11513669788837433, + 0.09661771357059479, + 0.1622926890850067, + 0.08957482129335403, + 0.06094297394156456, + 0.020983746275305748, + 0.02340385876595974, + 0.017975986003875732 + ], + "credit_map": [ + 0.02929443120956421, + 0.019586920738220215, + 0.01813686266541481, + 0.019499175250530243, + 0.10633480548858643, + 0.10285148024559021, + 0.026552749797701836, + 0.021033674478530884, + 0.006698543671518564, + 0.009149757213890553, + 0.007020134944468737, + 0.004812194034457207, + 0.014926370233297348, + 0.010026478208601475, + 0.009965053759515285, + 0.007182912901043892, + 0.013818416744470596, + 0.016578052192926407, + 0.015734275802969933, + 0.01093204040080309, + 0.016633426770567894, + 0.014191613532602787, + 0.01280976366251707, + 0.014439114369452, + 0.008147068321704865, + 0.009629535488784313, + 0.010818451642990112, + 0.021649785339832306, + 0.007609534077346325, + 0.010546022094786167, + 0.012614186853170395, + 0.015603137202560902, + 0.030870012938976288, + 0.01808580569922924, + 0.011117431335151196, + 0.012433191761374474, + 0.020787382498383522, + 0.03747817128896713, + 0.018453367054462433, + 0.013067346066236496, + 0.007435498759150505, + 0.009677759371697903, + 0.012620196677744389, + 0.014042149297893047, + 0.013013997115194798, + 0.009098938666284084, + 0.009947913698852062, + 0.01373837236315012, + 0.006005191709846258, + 0.014989050105214119, + 0.00620993971824646, + 0.006865371484309435, + 0.006220007315278053, + 0.005257106386125088, + 0.007332183420658112, + 0.008064132183790207, + 0.0021569482050836086, + 0.004060251638293266, + 0.003005139296874404, + 0.0025548022240400314, + 0.0028147955890744925, + 0.0022214779164642096, + 0.002311720047146082, + 0.001858612522482872, + 0.0020387701224535704, + 0.0026678917929530144, + 0.00559533154591918, + 0.0029560725670307875, + 0.0029672684613615274, + 0.0029112878255546093, + 0.0018066171323880553, + 0.0024606219958513975, + 0.002961087506264448, + 0.005175107158720493, + 0.0020045340061187744, + 0.0012523774057626724, + 0.0020212510135024786, + 0.0014584665186703205, + 0.001317195245064795, + 0.001785970525816083 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 2 + ], + "selected_reward": -0.016075992956757545 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a cinematic long shot featuring a flock of geese in a V-formation?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera shot from a low angle and smoothly panning right to follow the geese?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the geese fly low over a mirror-smooth lake surface with wing tips almost touching the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the sky painted with a sunset gradient from fiery orange to soft pink-purple, casting golden light on the geese?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are deep silhouette outlines of rolling mountains visible on the distant horizon contrasting with the sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0013377059949561954 + ], + "opposite_logprob_tokens": [ + -6.62633752822876 + ], + "logprob_reward": -0.0013377059949561954, + "yes_no_margin_reward": 6.625, + "token_credit_reward": -0.0013377059949561954, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.34280747175216675, + 0.209811270236969, + 0.09716887027025223, + 0.044095229357481, + 0.0658646747469902, + 0.04811980202794075, + 0.1097254827618599, + 0.030090415850281715, + 0.02118794433772564, + 0.03112884983420372 + ], + "credit_map": [ + 0.03310169652104378, + 0.02076275832951069, + 0.0257241390645504, + 0.025636738166213036, + 0.07617301493883133, + 0.07036585360765457, + 0.030525773763656616, + 0.06051747873425484, + 0.01203045342117548, + 0.015578682534396648, + 0.012268457561731339, + 0.01055072620511055, + 0.0738566517829895, + 0.04153459891676903, + 0.03180176392197609, + 0.012189934030175209, + 0.009268097579479218, + 0.012880384922027588, + 0.012141727842390537, + 0.012075583450496197, + 0.01593559980392456, + 0.008597149513661861, + 0.014549619518220425, + 0.011720702983438969, + 0.005637640133500099, + 0.005103098228573799, + 0.004302064422518015, + 0.0033480231650173664, + 0.005799029488116503, + 0.00919415708631277, + 0.005019233096390963, + 0.005691985134035349, + 0.012719682417809963, + 0.007890206761658192, + 0.007563652470707893, + 0.006833708379417658, + 0.006175231654196978, + 0.007492403499782085, + 0.010002376511693, + 0.007187414914369583, + 0.005026310682296753, + 0.009196192026138306, + 0.004662677180022001, + 0.004737290553748608, + 0.008106313645839691, + 0.00654525775462389, + 0.0048364996910095215, + 0.005009259097278118, + 0.009715527296066284, + 0.022641556337475777, + 0.009377621114253998, + 0.008890599943697453, + 0.019096845760941505, + 0.015186985023319721, + 0.011794581077992916, + 0.013021760620176792, + 0.003809264861047268, + 0.006695658899843693, + 0.0034540421329438686, + 0.002877976279705763, + 0.004522155970335007, + 0.0032844217494130135, + 0.0027396006044000387, + 0.0027072960510849953, + 0.002234600018709898, + 0.0025933494325727224, + 0.00405041454359889, + 0.0031005912460386753, + 0.0033611899707466364, + 0.0013756421394646168, + 0.0025913442950695753, + 0.0018808128079399467, + 0.003614310175180435, + 0.005447968374937773, + 0.003973696380853653, + 0.002456346061080694, + 0.004459898453205824, + 0.004585681948810816, + 0.002529468620195985, + 0.004061481449753046 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 6 + ], + "selected_reward": -0.0013377059949561954 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a cinematic long shot featuring a flock of geese in a V-formation?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera shot from a low angle and smoothly panning right to follow the geese?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the geese fly low over a mirror-smooth lake surface with wing tips almost touching the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the sky painted with a sunset gradient from fiery orange to soft pink-purple, casting golden light on the geese?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are deep silhouette outlines of rolling mountains visible on the distant horizon contrasting with the sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0036118291318416595 + ], + "opposite_logprob_tokens": [ + -5.628612041473389 + ], + "logprob_reward": -0.0036118291318416595, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.0036118291318416595, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.530376672744751, + 0.03620593622326851, + 0.012952044606208801, + 0.18410871922969818, + 0.015395051799714565, + 0.08067275583744049, + 0.07102366536855698, + 0.03323625773191452, + 0.019847428426146507, + 0.016181476414203644 + ], + "credit_map": [ + 0.03323165327310562, + 0.022713689133524895, + 0.020383724942803383, + 0.023454848676919937, + 0.1976783573627472, + 0.1679021120071411, + 0.02456694282591343, + 0.040445324033498764, + 0.0033772948663681746, + 0.005609048996120691, + 0.00396682508289814, + 0.00348946126177907, + 0.005890856496989727, + 0.0029002593364566565, + 0.004893249366432428, + 0.006078940350562334, + 0.0013287161709740758, + 0.0015535567654296756, + 0.0015772635815665126, + 0.001975898863747716, + 0.002036744263023138, + 0.0013450027909129858, + 0.001884728786535561, + 0.0012501315213739872, + 0.026292601600289345, + 0.03539564087986946, + 0.00801472831517458, + 0.009516035206615925, + 0.026723138988018036, + 0.054383646696805954, + 0.009120811708271503, + 0.01466211024671793, + 0.0024935367982834578, + 0.0024279311764985323, + 0.0014203665778040886, + 0.001369814039207995, + 0.001757035730406642, + 0.0028174088802188635, + 0.0016039767069742084, + 0.0015049806097522378, + 0.011046348139643669, + 0.010632198303937912, + 0.005077012348920107, + 0.005384951829910278, + 0.015796013176441193, + 0.020087504759430885, + 0.005809579975903034, + 0.0068391431123018265, + 0.007096780929714441, + 0.016594547778367996, + 0.0062265414744615555, + 0.004202557727694511, + 0.01113892626017332, + 0.011232432909309864, + 0.006374263670295477, + 0.008157609961926937, + 0.003017729613929987, + 0.007576546631753445, + 0.004326547496020794, + 0.0036401269026100636, + 0.004804417956620455, + 0.0033123516477644444, + 0.003653795225545764, + 0.002904739696532488, + 0.001857610186561942, + 0.0024531655944883823, + 0.0036607012152671814, + 0.0029936512000858784, + 0.0023736092261970043, + 0.002410507993772626, + 0.0022298223339021206, + 0.0018683593953028321, + 0.002372909337282181, + 0.0020214561372995377, + 0.0023625900503247976, + 0.0013984792167320848, + 0.002132941037416458, + 0.002244990086182952, + 0.0015691292937844992, + 0.002078981138765812 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 5 + ], + "selected_reward": -0.0036118291318416595 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "cinematic long shot, shot from a low angle, the camera smoothly pans right, following the flight path, to capture a flock of geese in a meticulously ordered V-formation, passing low over the mirror-smooth surface of a lake. The last rays of sunset paint the sky a gradient of hues from fiery orange to soft pink-purple. This warm light precisely outlines the golden edges of each goose's wings and casts a warm glow on the white feathers on their bellies. In exquisitely realistic detail, their powerful wings beat gracefully in sync, each gray-brown and white feather trembling slightly under the influence of aerodynamic forces. The tips of the wings almost touch the water, and each beat creates tiny, rapidly spreading ripples on the surface. On the distant horizon, the outlines of the rolling mountains are dyed into deep silhouettes, forming a sharp contrast with the colorful sky. The whole picture is full of a tranquil, magnificent and slightly poetic atmosphere, perfectly presenting the grandeur of nature and the dynamic beauty of life.", + "seed": 31891, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000021/group_5/index_3252_seed_31891_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": -0.026957518234848977, + "vlm_logprob": -0.026957518234848977, + "vlm_yes_no_margin": 4.55, + "vlm_token_credit": -0.026957518234848977, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a cinematic long shot featuring a flock of geese in a V-formation?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera shot from a low angle and smoothly panning right to follow the geese?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the geese fly low over a mirror-smooth lake surface with wing tips almost touching the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the sky painted with a sunset gradient from fiery orange to soft pink-purple, casting golden light on the geese?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are deep silhouette outlines of rolling mountains visible on the distant horizon contrasting with the sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.060205813497304916 + ], + "opposite_logprob_tokens": [ + -2.9352056980133057 + ], + "logprob_reward": -0.060205813497304916, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.060205813497304916, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17779472470283508, + 0.05690690502524376, + 0.3588193953037262, + 0.02558816224336624, + 0.036119844764471054, + 0.03982589766383171, + 0.04104768857359886, + 0.06166661158204079, + 0.11104106158018112, + 0.09118971228599548 + ], + "credit_map": [ + 0.016551395878195763, + 0.019757980480790138, + 0.023079337552189827, + 0.03277188912034035, + 0.028887171298265457, + 0.019786128774285316, + 0.019313393160700798, + 0.017647432163357735, + 0.005424719303846359, + 0.006863559130579233, + 0.005501336883753538, + 0.006424203515052795, + 0.006367613561451435, + 0.01074665691703558, + 0.009633558802306652, + 0.005945263430476189, + 0.02745160274207592, + 0.02997477911412716, + 0.04123058542609215, + 0.08698192983865738, + 0.07349856942892075, + 0.0365578830242157, + 0.03107902966439724, + 0.032045021653175354, + 0.004081912338733673, + 0.003558746073395014, + 0.0021813411731272936, + 0.0019123858073726296, + 0.0030896528623998165, + 0.00246960180811584, + 0.00396761717274785, + 0.004326904658228159, + 0.0047543770633637905, + 0.0049690078012645245, + 0.004132636822760105, + 0.0052756667137146, + 0.003877219744026661, + 0.004717939533293247, + 0.0035386616364121437, + 0.004854337312281132, + 0.0033278160262852907, + 0.004310503602027893, + 0.005317916162312031, + 0.005147502291947603, + 0.004223598167300224, + 0.004647174850106239, + 0.00836097076535225, + 0.0044904169626533985, + 0.003247577929869294, + 0.002646938432008028, + 0.004565380048006773, + 0.006038818042725325, + 0.0037355271633714437, + 0.00564114423468709, + 0.010400274768471718, + 0.00477202981710434, + 0.005725070834159851, + 0.004946880973875523, + 0.008062930777668953, + 0.008450275287032127, + 0.009167814627289772, + 0.00496972631663084, + 0.013561687432229519, + 0.006782232318073511, + 0.014885141514241695, + 0.021624408662319183, + 0.016779744997620583, + 0.013124959543347359, + 0.015447976998984814, + 0.008636846207082272, + 0.009299395605921745, + 0.01124260202050209, + 0.005846378393471241, + 0.009132745675742626, + 0.013747629709541798, + 0.00854003056883812, + 0.01862558163702488, + 0.014191281981766224, + 0.012956212274730206, + 0.00814985390752554 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 8 + ], + "selected_reward": -0.060205813497304916 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a cinematic long shot featuring a flock of geese in a V-formation?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera shot from a low angle and smoothly panning right to follow the geese?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the geese fly low over a mirror-smooth lake surface with wing tips almost touching the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the sky painted with a sunset gradient from fiery orange to soft pink-purple, casting golden light on the geese?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are deep silhouette outlines of rolling mountains visible on the distant horizon contrasting with the sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03811107203364372 + ], + "opposite_logprob_tokens": [ + -3.2881109714508057 + ], + "logprob_reward": -0.03811107203364372, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03811107203364372, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24277977645397186, + 0.03473234549164772, + 0.23051562905311584, + 0.0416770838201046, + 0.036796100437641144, + 0.0360063761472702, + 0.05570903420448303, + 0.0711139366030693, + 0.13372838497161865, + 0.11694133281707764 + ], + "credit_map": [ + 0.021319150924682617, + 0.022570336237549782, + 0.046393610537052155, + 0.06857666373252869, + 0.027213620021939278, + 0.017841095104813576, + 0.022181138396263123, + 0.016684141010046005, + 0.004597816616296768, + 0.00620455015450716, + 0.0032785660587251186, + 0.002715061418712139, + 0.004307004157453775, + 0.005858451593667269, + 0.003944105003029108, + 0.003826791187748313, + 0.026172993704676628, + 0.01707022823393345, + 0.030244309455156326, + 0.03864774852991104, + 0.02267325110733509, + 0.02732400968670845, + 0.02339772693812847, + 0.04498535767197609, + 0.005787323694676161, + 0.004467609338462353, + 0.004093226045370102, + 0.0048624733462929726, + 0.006045416463166475, + 0.004263309296220541, + 0.005777361337095499, + 0.006380363833159208, + 0.004348299466073513, + 0.0044967359863221645, + 0.003987795207649469, + 0.004802628420293331, + 0.004145881626754999, + 0.0049403090961277485, + 0.004170289728790522, + 0.005904159042984247, + 0.003803195431828499, + 0.0034833604004234076, + 0.0038482979871332645, + 0.003947901539504528, + 0.004420917015522718, + 0.0038940301164984703, + 0.008322379551827908, + 0.004286292474716902, + 0.004687695298343897, + 0.005390860140323639, + 0.007280339021235704, + 0.008885114453732967, + 0.003259944962337613, + 0.0036080284044146538, + 0.012806462123990059, + 0.00979058537632227, + 0.008360045962035656, + 0.006971381139010191, + 0.007040392141789198, + 0.00878172367811203, + 0.01246089767664671, + 0.008166979067027569, + 0.012347476556897163, + 0.006985045038163662, + 0.01143508218228817, + 0.01751992478966713, + 0.014998034574091434, + 0.010677404701709747, + 0.017453869804739952, + 0.017243746668100357, + 0.011941126547753811, + 0.03245919570326805, + 0.015130706131458282, + 0.011620724573731422, + 0.019590431824326515, + 0.013422874733805656, + 0.018982496112585068, + 0.013554347679018974, + 0.014248639345169067, + 0.010391109623014927 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 8 + ], + "selected_reward": -0.03811107203364372 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a cinematic long shot featuring a flock of geese in a V-formation?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera shot from a low angle and smoothly panning right to follow the geese?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the geese fly low over a mirror-smooth lake surface with wing tips almost touching the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the sky painted with a sunset gradient from fiery orange to soft pink-purple, casting golden light on the geese?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are deep silhouette outlines of rolling mountains visible on the distant horizon contrasting with the sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.033714812248945236 + ], + "opposite_logprob_tokens": [ + -3.408714771270752 + ], + "logprob_reward": -0.033714812248945236, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.033714812248945236, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.28206348419189453, + 0.02843073382973671, + 0.2560005187988281, + 0.020495211705565453, + 0.027059406042099, + 0.06013959273695946, + 0.013827729038894176, + 0.012412885203957558, + 0.19676192104816437, + 0.10280856490135193 + ], + "credit_map": [ + 0.02274421788752079, + 0.030330412089824677, + 0.046946194022893906, + 0.0680931806564331, + 0.046537838876247406, + 0.02061084657907486, + 0.022278273478150368, + 0.024522505700588226, + 0.0028434856794774532, + 0.00397959491237998, + 0.002873969729989767, + 0.003310569329187274, + 0.0031662853434681892, + 0.005164747592061758, + 0.004004305228590965, + 0.0030877748504281044, + 0.016404904425144196, + 0.014209278859198093, + 0.029536399990320206, + 0.09292922914028168, + 0.02332807146012783, + 0.01505086850374937, + 0.018568260595202446, + 0.045973509550094604, + 0.0031846959609538317, + 0.005609963089227676, + 0.0016720760613679886, + 0.0013757185079157352, + 0.0036736566107720137, + 0.0021111974492669106, + 0.0014808052219450474, + 0.0013870978727936745, + 0.003139832988381386, + 0.004456887021660805, + 0.0032995296642184258, + 0.0035354099236428738, + 0.002116333693265915, + 0.0023465584963560104, + 0.003062739735469222, + 0.005102114286273718, + 0.006036360748112202, + 0.005824403837323189, + 0.009033456444740295, + 0.005111090838909149, + 0.008681753650307655, + 0.011222312226891518, + 0.007640658877789974, + 0.006589555647224188, + 0.0018085422925651073, + 0.0016343725146725774, + 0.0021542306058108807, + 0.0018172881100326777, + 0.0012501283781602979, + 0.001092501450330019, + 0.0023580121342092752, + 0.0017126542516052723, + 0.0010603232076391578, + 0.00130713009275496, + 0.0014842376112937927, + 0.002416857285425067, + 0.0010928973788395524, + 0.0013491078279912472, + 0.0018739966908469796, + 0.0018283335957676172, + 0.03478160873055458, + 0.0458955243229866, + 0.02556488662958145, + 0.014212965033948421, + 0.027334820479154587, + 0.014849591068923473, + 0.013218027539551258, + 0.02090447209775448, + 0.009975910186767578, + 0.010180669836699963, + 0.013312332332134247, + 0.012698277831077576, + 0.01850392110645771, + 0.013806194067001343, + 0.01563551276922226, + 0.008695745840668678 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 8 + ], + "selected_reward": -0.033714812248945236 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a cinematic long shot featuring a flock of geese in a V-formation?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera shot from a low angle and smoothly panning right to follow the geese?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the geese fly low over a mirror-smooth lake surface with wing tips almost touching the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the sky painted with a sunset gradient from fiery orange to soft pink-purple, casting golden light on the geese?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are deep silhouette outlines of rolling mountains visible on the distant horizon contrasting with the sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0017151420470327139 + ], + "opposite_logprob_tokens": [ + -6.376715183258057 + ], + "logprob_reward": -0.0017151420470327139, + "yes_no_margin_reward": 6.375, + "token_credit_reward": -0.0017151420470327139, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2763426601886749, + 0.0442240945994854, + 0.31940722465515137, + 0.05757443979382515, + 0.023109054192900658, + 0.03637675195932388, + 0.03758156672120094, + 0.0255117230117321, + 0.12604990601539612, + 0.053822506219148636 + ], + "credit_map": [ + 0.02450166456401348, + 0.03169397637248039, + 0.04576990753412247, + 0.06777478754520416, + 0.04095302149653435, + 0.018659338355064392, + 0.023451384156942368, + 0.023538587614893913, + 0.005539943464100361, + 0.005826312117278576, + 0.006716008763760328, + 0.003968711011111736, + 0.005838381592184305, + 0.005248808767646551, + 0.005709872115403414, + 0.005376055371016264, + 0.019090691581368446, + 0.04221869632601738, + 0.02833865024149418, + 0.060848090797662735, + 0.059650570154190063, + 0.01509816013276577, + 0.02166406624019146, + 0.07249830663204193, + 0.011341805569827557, + 0.021814832463860512, + 0.0033394102938473225, + 0.0025748361367732286, + 0.008242262527346611, + 0.003267020685598254, + 0.00325561361387372, + 0.0037386564072221518, + 0.0029836404137313366, + 0.003118161577731371, + 0.0031559038907289505, + 0.0027654492296278477, + 0.0028725028969347477, + 0.002413538284599781, + 0.002796825719997287, + 0.003003031015396118, + 0.004856393206864595, + 0.004959210753440857, + 0.0033503170125186443, + 0.007989674806594849, + 0.003518263576552272, + 0.0031437936704605818, + 0.0058812634088099, + 0.0026778371538966894, + 0.0033830308821052313, + 0.003325593424960971, + 0.004193268250674009, + 0.006518085021525621, + 0.0020911172032356262, + 0.0029740668833255768, + 0.008971793577075005, + 0.006124608684331179, + 0.0018198578618466854, + 0.0021967231296002865, + 0.0037025162018835545, + 0.0034221915993839502, + 0.00355848902836442, + 0.0021789341699332, + 0.005622577853500843, + 0.0030104334000498056, + 0.014791124500334263, + 0.03992052376270294, + 0.009189808741211891, + 0.006452742498368025, + 0.025411754846572876, + 0.01660686731338501, + 0.006619763560593128, + 0.007057329639792442, + 0.0062944223172962666, + 0.006545999553054571, + 0.005335152614861727, + 0.005165277514606714, + 0.010900161229074001, + 0.008595322258770466, + 0.00640480313450098, + 0.00458136573433876 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 8 + ], + "selected_reward": -0.0017151420470327139 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a cinematic long shot featuring a flock of geese in a V-formation?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera shot from a low angle and smoothly panning right to follow the geese?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the geese fly low over a mirror-smooth lake surface with wing tips almost touching the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the sky painted with a sunset gradient from fiery orange to soft pink-purple, casting golden light on the geese?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are deep silhouette outlines of rolling mountains visible on the distant horizon contrasting with the sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0010407513473182917 + ], + "opposite_logprob_tokens": [ + -6.876040935516357 + ], + "logprob_reward": -0.0010407513473182917, + "yes_no_margin_reward": 6.875, + "token_credit_reward": -0.0010407513473182917, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19303271174430847, + 0.04236210882663727, + 0.29100659489631653, + 0.12865345180034637, + 0.04040398821234703, + 0.02530897594988346, + 0.04063975065946579, + 0.1086667850613594, + 0.06591112911701202, + 0.06401447206735611 + ], + "credit_map": [ + 0.017557965591549873, + 0.01998383179306984, + 0.03051602654159069, + 0.04987405985593796, + 0.02516643889248371, + 0.016480712220072746, + 0.018071375787258148, + 0.015382308512926102, + 0.003387123579159379, + 0.005454897414892912, + 0.004883022513240576, + 0.00666913902387023, + 0.006870040204375982, + 0.007235864643007517, + 0.00429017748683691, + 0.003571844194084406, + 0.020005803555250168, + 0.014593164436519146, + 0.03512636572122574, + 0.0986720621585846, + 0.02192755974829197, + 0.01163394283503294, + 0.0208290982991457, + 0.06821858882904053, + 0.01962399110198021, + 0.03938906267285347, + 0.007263295818120241, + 0.0068210638128221035, + 0.01766977645456791, + 0.009712579660117626, + 0.012102349661290646, + 0.016071341931819916, + 0.005330032203346491, + 0.008680528029799461, + 0.003750554518774152, + 0.004619188606739044, + 0.003691180841997266, + 0.004271005280315876, + 0.004473433364182711, + 0.005588064435869455, + 0.0033940307330340147, + 0.0031848468352109194, + 0.0032807025127112865, + 0.0029773504938930273, + 0.0030292076990008354, + 0.003002830082550645, + 0.0036265396047383547, + 0.00281346682459116, + 0.004236242268234491, + 0.004651978146284819, + 0.005438255146145821, + 0.005498891230672598, + 0.002904970897361636, + 0.004031321499496698, + 0.006716892588883638, + 0.007161199115216732, + 0.008252788335084915, + 0.008469239808619022, + 0.015688421204686165, + 0.015104847028851509, + 0.014406641945242882, + 0.00821384135633707, + 0.025389784947037697, + 0.013141218572854996, + 0.010454469360411167, + 0.0077463495545089245, + 0.008507870137691498, + 0.007877814583480358, + 0.010264884680509567, + 0.005281307268887758, + 0.0059647709131240845, + 0.00981366541236639, + 0.007767789997160435, + 0.007766170427203178, + 0.006584952585399151, + 0.0074093942530453205, + 0.012974168173968792, + 0.009482001885771751, + 0.007809868082404137, + 0.004220122005790472 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 3 + ], + "selected_reward": -0.0010407513473182917 + } + ], + "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/0000021/group_6/index_tensor([7112, 7112])_group_6_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000021/group_6/index_tensor([7112, 7112])_group_6_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..d5583c0436b7ec03d61aa88c7fdb867f176985ab --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000021/group_6/index_tensor([7112, 7112])_group_6_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "smooth, low-angle tracking shot captures a professional man in his 30s with neatly coiffed dark hair as he effortlessly glides through a sun-drenched, modern urban business district on a stylish electric scooter. Dressed in an impeccably tailored charcoal gray suit over a crisp white shirt, his posture is confident and relaxed. The scooter itself is a design marvel, its frame constructed of matte black carbon fiber with a subtle, woven texture that draws attention. As he approaches a wide, clearly marked crosswalk, he deliberately and smoothly applies the brakes. The scooter decelerates gracefully, its advanced braking system underscored by the vibrant red LED strip on the rear fender. He comes to a perfect, steady stop just before the white line, placing one polished black leather Oxford shoe firmly on the clean sidewalk. He calmly turns his head, yielding to the sleek, high-end car passing before him. The entire scene is bathed in the warm golden glow of the afternoon sun, casting long shadows on the glass and steel skyscrapers in the background and producing beautiful lens flares, creating a sophisticated and aspirational atmosphere that showcases the city's sophisticated transportation and cutting-edge security.", + "seed": 14916, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000021/group_6/index_7112_seed_14916_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": -0.23118540979921817, + "vlm_logprob": -0.23118540979921817, + "vlm_yes_no_margin": 1.75, + "vlm_token_credit": -0.23118540979921817, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 main subject a man in his 30s wearing a charcoal gray suit and white shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the electric scooter's frame made of matte black carbon fiber with a woven texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scooter have a vibrant red LED strip on the rear fender?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene include long shadows and lens flares caused by afternoon sun?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.24399708211421967 + ], + "opposite_logprob_tokens": [ + -1.8689970970153809 + ], + "logprob_reward": -0.24399708211421967, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.24399708211421967, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.42597725987434387, + 0.02450997196137905, + 0.029019268229603767, + 0.032069336622953415, + 0.013779891654849052, + 0.03233836963772774, + 0.04125014320015907, + 0.028000859543681145, + 0.3510788083076477, + 0.021976128220558167 + ], + "credit_map": [ + 0.07049188017845154, + 0.09000512957572937, + 0.04227349907159805, + 0.04346393421292305, + 0.03964689373970032, + 0.05831998959183693, + 0.04536764323711395, + 0.03640829026699066, + 0.0034105221275240183, + 0.003158463863655925, + 0.0025144817773252726, + 0.002974758855998516, + 0.0029647608753293753, + 0.004050380550324917, + 0.0031272151973098516, + 0.0023093887139111757, + 0.0027247201651334763, + 0.0036393869668245316, + 0.003128394251689315, + 0.004752465523779392, + 0.00507698068395257, + 0.002995833521708846, + 0.003915674518793821, + 0.0027858130633831024, + 0.0029763937927782536, + 0.002663594437763095, + 0.0047585973516106606, + 0.005516414996236563, + 0.005460645072162151, + 0.0035824584774672985, + 0.00412000622600317, + 0.0029912260361015797, + 0.0014866332057863474, + 0.0018021556315943599, + 0.0018210782436653972, + 0.0024027263280004263, + 0.0021076882258057594, + 0.0016957724001258612, + 0.0014463806292042136, + 0.001017456641420722, + 0.003089423757046461, + 0.005789898801594973, + 0.002912084572017193, + 0.004267207346856594, + 0.005509562324732542, + 0.003994937054812908, + 0.004702214617282152, + 0.002073042094707489, + 0.004499450325965881, + 0.005219118669629097, + 0.004286203533411026, + 0.0036845167633146048, + 0.005767835304141045, + 0.008534063585102558, + 0.004762514028698206, + 0.004496441688388586, + 0.006223320495337248, + 0.0028835826087743044, + 0.0057708630338311195, + 0.0022848527878522873, + 0.0019072751747444272, + 0.0035724579356610775, + 0.002777670742943883, + 0.0025808382779359818, + 0.0684409812092781, + 0.02128220908343792, + 0.01853778585791588, + 0.01751111075282097, + 0.015932420268654823, + 0.14561304450035095, + 0.021197326481342316, + 0.04256391152739525, + 0.002416303614154458, + 0.002871844684705138, + 0.0025150789879262447, + 0.004135241266340017, + 0.002319429535418749, + 0.0018095694249495864, + 0.002690986031666398, + 0.0032176736276596785 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 6 + ], + "selected_reward": -0.24399708211421967 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 main subject a man in his 30s wearing a charcoal gray suit and white shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the electric scooter's frame made of matte black carbon fiber with a woven texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scooter have a vibrant red LED strip on the rear fender?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene include long shadows and lens flares caused by afternoon sun?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2521471381187439 + ], + "opposite_logprob_tokens": [ + -1.5021471977233887 + ], + "logprob_reward": -0.2521471381187439, + "yes_no_margin_reward": 1.25, + "token_credit_reward": -0.2521471381187439, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15800732374191284, + 0.05666262283921242, + 0.056486859917640686, + 0.10390935093164444, + 0.035441748797893524, + 0.021083975210785866, + 0.05020786449313164, + 0.0450461246073246, + 0.34570929408073425, + 0.12744484841823578 + ], + "credit_map": [ + 0.02188216708600521, + 0.0275206808000803, + 0.013283008709549904, + 0.019213562831282616, + 0.023158064112067223, + 0.020164869725704193, + 0.019024141132831573, + 0.013760821893811226, + 0.005825759377330542, + 0.008113575167953968, + 0.0028373359236866236, + 0.004418950527906418, + 0.010024409741163254, + 0.007449834141880274, + 0.011085446923971176, + 0.006907311733812094, + 0.005409909877926111, + 0.006268527824431658, + 0.0048736464232206345, + 0.005155662540346384, + 0.010742519050836563, + 0.005951696541160345, + 0.010245128534734249, + 0.007839767262339592, + 0.011221182532608509, + 0.01768571510910988, + 0.00817827694118023, + 0.01241784542798996, + 0.024081263691186905, + 0.012010731734335423, + 0.01168073620647192, + 0.006633603014051914, + 0.003040981013327837, + 0.0055662039667367935, + 0.003585339291021228, + 0.005501340143382549, + 0.006025902461260557, + 0.005349492188543081, + 0.004376670345664024, + 0.0019958168268203735, + 0.0030277289915829897, + 0.0031146807596087456, + 0.002169734798371792, + 0.002030705101788044, + 0.003851589746773243, + 0.002825544448569417, + 0.0019271040800958872, + 0.0021368868183344603, + 0.004316345788538456, + 0.0066123781725764275, + 0.005823557265102863, + 0.004978442098945379, + 0.007598763331770897, + 0.00658723758533597, + 0.008913747034966946, + 0.005377395078539848, + 0.006999483332037926, + 0.00471079908311367, + 0.006555072497576475, + 0.005422856658697128, + 0.0047971634194254875, + 0.008634364232420921, + 0.004464453551918268, + 0.003461932996287942, + 0.08428493142127991, + 0.02948835864663124, + 0.01990552619099617, + 0.02274259179830551, + 0.019693318754434586, + 0.040253691375255585, + 0.025723949074745178, + 0.10361693054437637, + 0.026237331330776215, + 0.009570702910423279, + 0.007248586043715477, + 0.008087247610092163, + 0.008350602351129055, + 0.018492447212338448, + 0.009395983070135117, + 0.04006195068359375 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 9 + ], + "selected_reward": -0.2521471381187439 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 main subject a man in his 30s wearing a charcoal gray suit and white shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the electric scooter's frame made of matte black carbon fiber with a woven texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scooter have a vibrant red LED strip on the rear fender?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene include long shadows and lens flares caused by afternoon sun?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.38711249828338623 + ], + "opposite_logprob_tokens": [ + -1.1371124982833862 + ], + "logprob_reward": -0.38711249828338623, + "yes_no_margin_reward": 0.75, + "token_credit_reward": -0.38711249828338623, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.36001914739608765, + 0.020196013152599335, + 0.032365553081035614, + 0.026742566376924515, + 0.022132616490125656, + 0.03152340650558472, + 0.10709314793348312, + 0.042997755110263824, + 0.33627498149871826, + 0.020654739812016487 + ], + "credit_map": [ + 0.057932741940021515, + 0.07832224667072296, + 0.03237854689359665, + 0.03820709139108658, + 0.033548105508089066, + 0.04711006209254265, + 0.04088164120912552, + 0.031638745218515396, + 0.0026166539173573256, + 0.002502732677385211, + 0.0010523417731747031, + 0.002339435974135995, + 0.0030325872357934713, + 0.002765933284536004, + 0.004010915756225586, + 0.0018754132324829698, + 0.004718419164419174, + 0.004483467899262905, + 0.0024233225267380476, + 0.003270515240728855, + 0.006110329646617174, + 0.004040537867695093, + 0.004275752697139978, + 0.003043205477297306, + 0.002437862567603588, + 0.0034941169433295727, + 0.0037289480678737164, + 0.0034867902286350727, + 0.004380456171929836, + 0.0028163890819996595, + 0.003857458708807826, + 0.0025405443739145994, + 0.0024403403513133526, + 0.002431746805086732, + 0.0021992505062371492, + 0.004300213418900967, + 0.003485730616375804, + 0.003327915444970131, + 0.002197839319705963, + 0.001749579794704914, + 0.003774406621232629, + 0.006168667692691088, + 0.004546263255178928, + 0.00279042380861938, + 0.005159810185432434, + 0.003722918452695012, + 0.003370814025402069, + 0.001990100834518671, + 0.01367910485714674, + 0.010871690697968006, + 0.009895659051835537, + 0.01073888223618269, + 0.012223806232213974, + 0.023829272016882896, + 0.014193424955010414, + 0.011661304160952568, + 0.009164424613118172, + 0.004736966919153929, + 0.00838392786681652, + 0.004066574852913618, + 0.0037455500569194555, + 0.0044355206191539764, + 0.004409092478454113, + 0.004055699333548546, + 0.036189090460538864, + 0.01805238611996174, + 0.01242383848875761, + 0.011334867216646671, + 0.01080422755330801, + 0.12603527307510376, + 0.01438430231064558, + 0.10705098509788513, + 0.0024795765057206154, + 0.0028947414830327034, + 0.003108762437477708, + 0.002766096731647849, + 0.002514554187655449, + 0.0022513284347951412, + 0.002681450918316841, + 0.001958228647708893 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 6 + ], + "selected_reward": -0.38711249828338623 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 main subject a man in his 30s wearing a charcoal gray suit and white shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the electric scooter's frame made of matte black carbon fiber with a woven texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scooter have a vibrant red LED strip on the rear fender?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene include long shadows and lens flares caused by afternoon sun?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.25208720564842224 + ], + "opposite_logprob_tokens": [ + -1.5020872354507446 + ], + "logprob_reward": -0.25208720564842224, + "yes_no_margin_reward": 1.25, + "token_credit_reward": -0.25208720564842224, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13926774263381958, + 0.03291352838277817, + 0.03520672768354416, + 0.03713340684771538, + 0.021969210356473923, + 0.03374217078089714, + 0.0433255173265934, + 0.1296294778585434, + 0.26871970295906067, + 0.2580925226211548 + ], + "credit_map": [ + 0.01842651329934597, + 0.01635439321398735, + 0.01135531160980463, + 0.009671494364738464, + 0.02031748555600643, + 0.020952410995960236, + 0.02940724790096283, + 0.012782878242433071, + 0.004117558244615793, + 0.0033876500092446804, + 0.0023007895797491074, + 0.004443979822099209, + 0.007516735699027777, + 0.0038245769683271646, + 0.00436055613681674, + 0.002961677499115467, + 0.0031312531791627407, + 0.003688289551064372, + 0.004534287843853235, + 0.00435056583955884, + 0.0056702676229178905, + 0.0034483089111745358, + 0.006533301901072264, + 0.0038504551630467176, + 0.003761306870728731, + 0.006904683541506529, + 0.004478430841118097, + 0.004258422181010246, + 0.006689861416816711, + 0.004901345353573561, + 0.003317854832857847, + 0.0028214987833052874, + 0.0031278564129024744, + 0.0027593376580625772, + 0.001975103048607707, + 0.003272769972681999, + 0.0033660358749330044, + 0.002632318064570427, + 0.0026762678753584623, + 0.0021595205180346966, + 0.004104234743863344, + 0.004674973431974649, + 0.004196492489427328, + 0.00384607445448637, + 0.006054075434803963, + 0.004736227914690971, + 0.0030641965568065643, + 0.003065891098231077, + 0.00461586331948638, + 0.004313118755817413, + 0.006707505322992802, + 0.0053072418086230755, + 0.004734460264444351, + 0.005942140705883503, + 0.00855710357427597, + 0.003148077754303813, + 0.011980731971561909, + 0.011739935725927353, + 0.014211820438504219, + 0.019178491085767746, + 0.010017412714660168, + 0.01467198133468628, + 0.036251187324523926, + 0.011577917262911797, + 0.03524022549390793, + 0.016266267746686935, + 0.010982736945152283, + 0.012154221534729004, + 0.012173010967671871, + 0.07209070026874542, + 0.01421995460987091, + 0.09559257328510284, + 0.05021929740905762, + 0.013448617421090603, + 0.028754182159900665, + 0.017102405428886414, + 0.01315827202051878, + 0.03891462832689285, + 0.019715899601578712, + 0.07677922397851944 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 0 + ], + "selected_reward": -0.25208720564842224 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 main subject a man in his 30s wearing a charcoal gray suit and white shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the electric scooter's frame made of matte black carbon fiber with a woven texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scooter have a vibrant red LED strip on the rear fender?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene include long shadows and lens flares caused by afternoon sun?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020583124831318855 + ], + "opposite_logprob_tokens": [ + -3.895583152770996 + ], + "logprob_reward": -0.020583124831318855, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020583124831318855, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17551736533641815, + 0.06543584167957306, + 0.0303469430655241, + 0.028098050504922867, + 0.028358321636915207, + 0.04175354167819023, + 0.12164861708879471, + 0.2686143219470978, + 0.05660314857959747, + 0.183623805642128 + ], + "credit_map": [ + 0.0261178407818079, + 0.031933605670928955, + 0.02079092711210251, + 0.021892890334129333, + 0.016186533495783806, + 0.026392458006739616, + 0.016816161572933197, + 0.01538695115596056, + 0.011283710598945618, + 0.008968573994934559, + 0.005765467416495085, + 0.008986413478851318, + 0.004964066669344902, + 0.01514971163123846, + 0.004766227211803198, + 0.005551675334572792, + 0.0035513821057975292, + 0.0036594620905816555, + 0.0029246318154037, + 0.004352492745965719, + 0.004437617491930723, + 0.0043662847019732, + 0.003774995217099786, + 0.0032800764311105013, + 0.0026254800613969564, + 0.0034020491875708103, + 0.003987849224358797, + 0.0033402335830032825, + 0.0033784268889576197, + 0.00436399457976222, + 0.004072089679539204, + 0.00292792497202754, + 0.00513943936675787, + 0.0031678241211920977, + 0.0033674004953354597, + 0.005008202977478504, + 0.0031846624333411455, + 0.0034006605856120586, + 0.002933276118710637, + 0.002156855072826147, + 0.004256875719875097, + 0.007572159171104431, + 0.00494254007935524, + 0.004239258356392384, + 0.008211917243897915, + 0.005365089979022741, + 0.003272880567237735, + 0.003892818931490183, + 0.012261783704161644, + 0.019203025847673416, + 0.011635146103799343, + 0.012410666793584824, + 0.02000836655497551, + 0.01591308042407036, + 0.012197664938867092, + 0.018018875271081924, + 0.05553139001131058, + 0.026750730350613594, + 0.0542757622897625, + 0.02813478745520115, + 0.01895708404481411, + 0.028448183089494705, + 0.031377121806144714, + 0.02513926289975643, + 0.005959629081189632, + 0.003219663631170988, + 0.0034912594128400087, + 0.002690677996724844, + 0.0020190016366541386, + 0.01979164034128189, + 0.0029449956491589546, + 0.016486283391714096, + 0.02570495940744877, + 0.019115453585982323, + 0.019734257832169533, + 0.0209732074290514, + 0.027456222102046013, + 0.023557553067803383, + 0.0192390289157629, + 0.027843136340379715 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 9, + 0 + ], + "selected_reward": -0.020583124831318855 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "smooth, low-angle tracking shot captures a professional man in his 30s with neatly coiffed dark hair as he effortlessly glides through a sun-drenched, modern urban business district on a stylish electric scooter. Dressed in an impeccably tailored charcoal gray suit over a crisp white shirt, his posture is confident and relaxed. The scooter itself is a design marvel, its frame constructed of matte black carbon fiber with a subtle, woven texture that draws attention. As he approaches a wide, clearly marked crosswalk, he deliberately and smoothly applies the brakes. The scooter decelerates gracefully, its advanced braking system underscored by the vibrant red LED strip on the rear fender. He comes to a perfect, steady stop just before the white line, placing one polished black leather Oxford shoe firmly on the clean sidewalk. He calmly turns his head, yielding to the sleek, high-end car passing before him. The entire scene is bathed in the warm golden glow of the afternoon sun, casting long shadows on the glass and steel skyscrapers in the background and producing beautiful lens flares, creating a sophisticated and aspirational atmosphere that showcases the city's sophisticated transportation and cutting-edge security.", + "seed": 14917, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000021/group_6/index_7112_seed_14917_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": -0.14792497120797635, + "vlm_logprob": -0.14792497120797635, + "vlm_yes_no_margin": 2.349999952316284, + "vlm_token_credit": -0.14792497120797635, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 main subject a man in his 30s wearing a charcoal gray suit and white shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the electric scooter's frame made of matte black carbon fiber with a woven texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scooter have a vibrant red LED strip on the rear fender?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene include long shadows and lens flares caused by afternoon sun?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08540015667676926 + ], + "opposite_logprob_tokens": [ + -3.210400104522705 + ], + "logprob_reward": -0.08540015667676926, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.08540015667676926, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.27435925602912903, + 0.03634896129369736, + 0.02236327715218067, + 0.05049313232302666, + 0.063356414437294, + 0.10662829130887985, + 0.06694991141557693, + 0.0966772735118866, + 0.2328798919916153, + 0.04994351789355278 + ], + "credit_map": [ + 0.059832118451595306, + 0.02194233238697052, + 0.03768255189061165, + 0.030860045924782753, + 0.020081009715795517, + 0.03533394634723663, + 0.02127169258892536, + 0.04735557734966278, + 0.003603540826588869, + 0.0037204374093562365, + 0.005136226303875446, + 0.005209252238273621, + 0.0051278104074299335, + 0.004859862849116325, + 0.00574903329834342, + 0.002942798426374793, + 0.0026002831291407347, + 0.0023547757882624865, + 0.0029951126780360937, + 0.0033561550080776215, + 0.003366749966517091, + 0.0018148048548027873, + 0.0037929227109998465, + 0.002082476392388344, + 0.004680974874645472, + 0.006712870206683874, + 0.005272517446428537, + 0.004927092231810093, + 0.0052245124243199825, + 0.0041809892281889915, + 0.011899258941411972, + 0.007594924885779619, + 0.007398073095828295, + 0.009831959381699562, + 0.00570288160815835, + 0.007793165277689695, + 0.012399764731526375, + 0.006852025631815195, + 0.007116137538105249, + 0.006262414623051882, + 0.019781850278377533, + 0.009350502863526344, + 0.015487759374082088, + 0.006983977276831865, + 0.010299519635736942, + 0.014856758527457714, + 0.012061253190040588, + 0.017806673422455788, + 0.008182302117347717, + 0.008763473480939865, + 0.014373238198459148, + 0.00663547171279788, + 0.005965749267488718, + 0.007444226182997227, + 0.009996487759053707, + 0.005588965956121683, + 0.011639258824288845, + 0.019580526277422905, + 0.008448081091046333, + 0.009629420004785061, + 0.011642233468592167, + 0.013677969574928284, + 0.011181147769093513, + 0.010878641158342361, + 0.027344835922122, + 0.051789991557598114, + 0.025107944384217262, + 0.021335389465093613, + 0.02153616026043892, + 0.037031140178442, + 0.021961510181427002, + 0.026772962883114815, + 0.006992473267018795, + 0.004129363689571619, + 0.007779452484101057, + 0.006734216120094061, + 0.0035542957484722137, + 0.004249948542565107, + 0.008183327503502369, + 0.00832044705748558 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 5 + ], + "selected_reward": -0.08540015667676926 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 main subject a man in his 30s wearing a charcoal gray suit and white shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the electric scooter's frame made of matte black carbon fiber with a woven texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scooter have a vibrant red LED strip on the rear fender?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene include long shadows and lens flares caused by afternoon sun?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.12705540657043457 + ], + "opposite_logprob_tokens": [ + -2.1270554065704346 + ], + "logprob_reward": -0.12705540657043457, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.12705540657043457, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.39509502053260803, + 0.04682418704032898, + 0.050054971128702164, + 0.02902863919734955, + 0.08692332357168198, + 0.03261639550328255, + 0.17683382332324982, + 0.07382914423942566, + 0.08008165657520294, + 0.028712859377264977 + ], + "credit_map": [ + 0.07076114416122437, + 0.029223576188087463, + 0.0503048375248909, + 0.041139695793390274, + 0.06352293491363525, + 0.03785082697868347, + 0.053541794419288635, + 0.04875018820166588, + 0.005476532503962517, + 0.005098042543977499, + 0.006206532008945942, + 0.005072103347629309, + 0.0032778470776975155, + 0.008491325192153454, + 0.006199028342962265, + 0.00700277928262949, + 0.003866918385028839, + 0.004719654563814402, + 0.012718536891043186, + 0.006207798607647419, + 0.004709774162620306, + 0.004810297396034002, + 0.005012734793126583, + 0.008009258657693863, + 0.00272714183665812, + 0.0054645235650241375, + 0.003834324888885021, + 0.004757333546876907, + 0.0029062731191515923, + 0.0030904854647815228, + 0.0038878910709172487, + 0.0023606643080711365, + 0.010155931115150452, + 0.010139921680092812, + 0.009412341751158237, + 0.010153261013329029, + 0.011199397966265678, + 0.00899591390043497, + 0.016671959310770035, + 0.010194595903158188, + 0.006215546280145645, + 0.003319517243653536, + 0.004811129532754421, + 0.00341348210349679, + 0.004727092571556568, + 0.0029069248121231794, + 0.003201533341780305, + 0.004021167289465666, + 0.017904359847307205, + 0.019291561096906662, + 0.027445578947663307, + 0.019892342388629913, + 0.016010073944926262, + 0.014207517728209496, + 0.04652648791670799, + 0.015555905178189278, + 0.008321505971252918, + 0.01414257287979126, + 0.008548365905880928, + 0.011021227575838566, + 0.0056830826215445995, + 0.007953069172799587, + 0.010555734857916832, + 0.007603587117046118, + 0.009302781894803047, + 0.017827654257416725, + 0.0067312708124518394, + 0.0061233448795974255, + 0.00916958786547184, + 0.007932024076581001, + 0.013793227262794971, + 0.009201758541166782, + 0.0030671143904328346, + 0.0028072441928088665, + 0.004406232386827469, + 0.004081010352820158, + 0.0023452755995094776, + 0.002307343063876033, + 0.005737475119531155, + 0.003961163107305765 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 4 + ], + "selected_reward": -0.12705540657043457 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 main subject a man in his 30s wearing a charcoal gray suit and white shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the electric scooter's frame made of matte black carbon fiber with a woven texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scooter have a vibrant red LED strip on the rear fender?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene include long shadows and lens flares caused by afternoon sun?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.3486485779285431 + ], + "opposite_logprob_tokens": [ + -1.2236485481262207 + ], + "logprob_reward": -0.3486485779285431, + "yes_no_margin_reward": 0.875, + "token_credit_reward": -0.3486485779285431, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15280497074127197, + 0.04939766228199005, + 0.23399260640144348, + 0.05811171978712082, + 0.043660953640937805, + 0.11817719787359238, + 0.12071668356657028, + 0.052075985819101334, + 0.10783745348453522, + 0.06322475522756577 + ], + "credit_map": [ + 0.020688995718955994, + 0.009607742540538311, + 0.01763823628425598, + 0.01030677743256092, + 0.014198699034750462, + 0.02711872011423111, + 0.019521299749612808, + 0.03372450917959213, + 0.005799889098852873, + 0.007535939104855061, + 0.006355921737849712, + 0.004754961002618074, + 0.008215775713324547, + 0.005659322254359722, + 0.007154688239097595, + 0.003921169321984053, + 0.024196682497859, + 0.04061627388000488, + 0.029438622295856476, + 0.026417916640639305, + 0.03659570589661598, + 0.026132524013519287, + 0.0265144445002079, + 0.024080434814095497, + 0.005358678754419088, + 0.006215222179889679, + 0.007901064120233059, + 0.008761663921177387, + 0.00802670419216156, + 0.00808822363615036, + 0.007140396162867546, + 0.006619768682867289, + 0.00578267639502883, + 0.004935741424560547, + 0.0073734307661652565, + 0.004756242968142033, + 0.004280779976397753, + 0.004087591078132391, + 0.005943786818534136, + 0.006500707473605871, + 0.025527670979499817, + 0.011198786087334156, + 0.01943536475300789, + 0.007288729306310415, + 0.010264822281897068, + 0.013293850235641003, + 0.00910088513046503, + 0.022067086771130562, + 0.011793527752161026, + 0.01361435279250145, + 0.01765439473092556, + 0.014036317355930805, + 0.009483926929533482, + 0.013202072121202946, + 0.031080404296517372, + 0.009851690381765366, + 0.00503840297460556, + 0.012584667652845383, + 0.006224270444363356, + 0.006402573082596064, + 0.004334624391049147, + 0.004995678085833788, + 0.00730933528393507, + 0.005186437629163265, + 0.012895189225673676, + 0.016404690220952034, + 0.010824520140886307, + 0.009509635157883167, + 0.012036986649036407, + 0.01671482063829899, + 0.01596282608807087, + 0.013488797470927238, + 0.008739665150642395, + 0.006237954366952181, + 0.007016485091298819, + 0.005734063684940338, + 0.005041734781116247, + 0.006712316535413265, + 0.012357285246253014, + 0.011385255493223667 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 6 + ], + "selected_reward": -0.3486485779285431 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 main subject a man in his 30s wearing a charcoal gray suit and white shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the electric scooter's frame made of matte black carbon fiber with a woven texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scooter have a vibrant red LED strip on the rear fender?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene include long shadows and lens flares caused by afternoon sun?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.16034160554409027 + ], + "opposite_logprob_tokens": [ + -1.9103416204452515 + ], + "logprob_reward": -0.16034160554409027, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.16034160554409027, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1379452496767044, + 0.053010959178209305, + 0.07712328433990479, + 0.027277454733848572, + 0.06539870798587799, + 0.034369949251413345, + 0.3450542092323303, + 0.1302817463874817, + 0.07164080440998077, + 0.05789763480424881 + ], + "credit_map": [ + 0.021512769162654877, + 0.008004612289369106, + 0.015213212929666042, + 0.011586475186049938, + 0.010422434657812119, + 0.02835102193057537, + 0.009856210090219975, + 0.03299851343035698, + 0.007952283136546612, + 0.007739530876278877, + 0.006059396546334028, + 0.007783076725900173, + 0.0068503739312291145, + 0.005282691679894924, + 0.0072549995966255665, + 0.004088611342012882, + 0.008129050955176353, + 0.013441457413136959, + 0.00796906091272831, + 0.009830191731452942, + 0.014255826361477375, + 0.007390399929136038, + 0.008989042602479458, + 0.007118257228285074, + 0.0027277530170977116, + 0.005268807057291269, + 0.0036051564384251833, + 0.0032558029051870108, + 0.0029359927866607904, + 0.00424149539321661, + 0.0031863332260400057, + 0.0020561153069138527, + 0.007076021283864975, + 0.011784570291638374, + 0.004796517081558704, + 0.008502097800374031, + 0.009419639594852924, + 0.005798981059342623, + 0.010722970589995384, + 0.007297911215573549, + 0.004857191350311041, + 0.004495913162827492, + 0.004175599664449692, + 0.0031636939384043217, + 0.0035852217115461826, + 0.004100253339856863, + 0.005279724486172199, + 0.004712354391813278, + 0.029614925384521484, + 0.036700859665870667, + 0.05070706084370613, + 0.036873046308755875, + 0.02183622680604458, + 0.031043024733662605, + 0.10787249356508255, + 0.030406612902879715, + 0.01560606062412262, + 0.024863887578248978, + 0.012657509185373783, + 0.012186188250780106, + 0.012779993005096912, + 0.01688998192548752, + 0.01784055307507515, + 0.01745758391916752, + 0.006544911302626133, + 0.011058875359594822, + 0.004624552093446255, + 0.006967473775148392, + 0.008844197727739811, + 0.008160991594195366, + 0.016899483278393745, + 0.008540324866771698, + 0.007039163261651993, + 0.004808129742741585, + 0.006458093412220478, + 0.005435774102807045, + 0.004772217012941837, + 0.0070041632279753685, + 0.012247459962964058, + 0.010132640600204468 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 7 + ], + "selected_reward": -0.16034160554409027 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 main subject a man in his 30s wearing a charcoal gray suit and white shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the electric scooter's frame made of matte black carbon fiber with a woven texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scooter have a vibrant red LED strip on the rear fender?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene include long shadows and lens flares caused by afternoon sun?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018179109320044518 + ], + "opposite_logprob_tokens": [ + -4.018178939819336 + ], + "logprob_reward": -0.018179109320044518, + "yes_no_margin_reward": 3.999999761581421, + "token_credit_reward": -0.018179109320044518, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14385199546813965, + 0.08346535265445709, + 0.10284930467605591, + 0.026480022817850113, + 0.06646819412708282, + 0.202071413397789, + 0.047634124755859375, + 0.047911182045936584, + 0.19482895731925964, + 0.0844394862651825 + ], + "credit_map": [ + 0.02073253132402897, + 0.00927547924220562, + 0.02759309485554695, + 0.01241021603345871, + 0.00869250949472189, + 0.021027013659477234, + 0.01367155835032463, + 0.03044959343969822, + 0.007987100630998611, + 0.0073173376731574535, + 0.01028800942003727, + 0.009430443868041039, + 0.012861130759119987, + 0.011936676688492298, + 0.01710715889930725, + 0.006537489127367735, + 0.008726004511117935, + 0.014566205441951752, + 0.024614859372377396, + 0.012078769505023956, + 0.007678569294512272, + 0.01000219490379095, + 0.009083472192287445, + 0.016099225729703903, + 0.002652899594977498, + 0.003157060593366623, + 0.0035268019419163465, + 0.0035869970452040434, + 0.004244911950081587, + 0.0033863915596157312, + 0.0038201662246137857, + 0.0021047950722277164, + 0.005856772884726524, + 0.006863221991807222, + 0.008267778903245926, + 0.007230525370687246, + 0.010645365342497826, + 0.007763794157654047, + 0.011260155588388443, + 0.008580582216382027, + 0.03219085559248924, + 0.020984606817364693, + 0.028816621750593185, + 0.020439419895410538, + 0.01900383085012436, + 0.0249103382229805, + 0.02486255392432213, + 0.0308632031083107, + 0.004851185716688633, + 0.006197735667228699, + 0.005158546846359968, + 0.004879505839198828, + 0.0034713090863078833, + 0.005658361129462719, + 0.012108072638511658, + 0.005309406202286482, + 0.005567940883338451, + 0.006337172817438841, + 0.004828608129173517, + 0.0027822526171803474, + 0.008512457832694054, + 0.008028385229408741, + 0.0054817767813801765, + 0.0063725863583385944, + 0.028017964214086533, + 0.024837642908096313, + 0.01426614634692669, + 0.010928249917924404, + 0.03376906365156174, + 0.04538126662373543, + 0.015215568244457245, + 0.022413058206439018, + 0.011771325953304768, + 0.008620870299637318, + 0.012622509151697159, + 0.009652446024119854, + 0.007909157313406467, + 0.007961975410580635, + 0.01347234658896923, + 0.012428855523467064 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 8, + 0 + ], + "selected_reward": -0.018179109320044518 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "smooth, low-angle tracking shot captures a professional man in his 30s with neatly coiffed dark hair as he effortlessly glides through a sun-drenched, modern urban business district on a stylish electric scooter. Dressed in an impeccably tailored charcoal gray suit over a crisp white shirt, his posture is confident and relaxed. The scooter itself is a design marvel, its frame constructed of matte black carbon fiber with a subtle, woven texture that draws attention. As he approaches a wide, clearly marked crosswalk, he deliberately and smoothly applies the brakes. The scooter decelerates gracefully, its advanced braking system underscored by the vibrant red LED strip on the rear fender. He comes to a perfect, steady stop just before the white line, placing one polished black leather Oxford shoe firmly on the clean sidewalk. He calmly turns his head, yielding to the sleek, high-end car passing before him. The entire scene is bathed in the warm golden glow of the afternoon sun, casting long shadows on the glass and steel skyscrapers in the background and producing beautiful lens flares, creating a sophisticated and aspirational atmosphere that showcases the city's sophisticated transportation and cutting-edge security.", + "seed": 14918, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000021/group_6/index_7112_seed_14918_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": -0.14922301918268205, + "vlm_logprob": -0.14922301918268205, + "vlm_yes_no_margin": 2.475, + "vlm_token_credit": -0.14922301918268205, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 main subject a man in his 30s wearing a charcoal gray suit and white shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the electric scooter's frame made of matte black carbon fiber with a woven texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scooter have a vibrant red LED strip on the rear fender?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene include long shadows and lens flares caused by afternoon sun?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11573424935340881 + ], + "opposite_logprob_tokens": [ + -2.615734338760376 + ], + "logprob_reward": -0.11573424935340881, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.11573424935340881, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1774577796459198, + 0.18671198189258575, + 0.08899883180856705, + 0.2012619823217392, + 0.03144402056932449, + 0.07010549306869507, + 0.08388256281614304, + 0.062190115451812744, + 0.042512599378824234, + 0.055434614419937134 + ], + "credit_map": [ + 0.021018892526626587, + 0.019901426509022713, + 0.027867037802934647, + 0.01978319324553013, + 0.01993398554623127, + 0.02950538508594036, + 0.02062559500336647, + 0.01882226951420307, + 0.01126494538038969, + 0.03435641527175903, + 0.013780025765299797, + 0.010843412950634956, + 0.022999849170446396, + 0.06334396451711655, + 0.011671320535242558, + 0.01845204457640648, + 0.006295554339885712, + 0.007378436159342527, + 0.008957171812653542, + 0.007052985951304436, + 0.020460402593016624, + 0.0076880124397575855, + 0.019913412630558014, + 0.011252854019403458, + 0.07080481201410294, + 0.0219871923327446, + 0.011740405112504959, + 0.009115074761211872, + 0.031484127044677734, + 0.032341934740543365, + 0.01523318886756897, + 0.008555267006158829, + 0.0028605398256331682, + 0.003907822072505951, + 0.004288116469979286, + 0.005117406602948904, + 0.003834853647276759, + 0.0047592888586223125, + 0.004231463652104139, + 0.0024445296730846167, + 0.008509744890034199, + 0.009297751821577549, + 0.008057348430156708, + 0.011044306680560112, + 0.01344003900885582, + 0.008068090304732323, + 0.005107835400849581, + 0.006580373737961054, + 0.008592863567173481, + 0.010126465931534767, + 0.008426099084317684, + 0.012691800482571125, + 0.0163014754652977, + 0.0077693872153759, + 0.011022815480828285, + 0.00895165279507637, + 0.006547314114868641, + 0.007835249416530132, + 0.009920435026288033, + 0.006733006332069635, + 0.010392038151621819, + 0.006904002279043198, + 0.0077799842692911625, + 0.006078087724745274, + 0.0043547372333705425, + 0.005959145724773407, + 0.0037555797025561333, + 0.0036169211380183697, + 0.010396194644272327, + 0.006190852262079716, + 0.004603293724358082, + 0.003635875415056944, + 0.00811760500073433, + 0.008667727932333946, + 0.0055692302994430065, + 0.005548485554754734, + 0.009599283337593079, + 0.008212930522859097, + 0.004307440482079983, + 0.005411913618445396 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 1, + 0 + ], + "selected_reward": -0.11573424935340881 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 main subject a man in his 30s wearing a charcoal gray suit and white shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the electric scooter's frame made of matte black carbon fiber with a woven texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scooter have a vibrant red LED strip on the rear fender?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene include long shadows and lens flares caused by afternoon sun?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.20159879326820374 + ], + "opposite_logprob_tokens": [ + -1.7015987634658813 + ], + "logprob_reward": -0.20159879326820374, + "yes_no_margin_reward": 1.5, + "token_credit_reward": -0.20159879326820374, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14788861572742462, + 0.365409791469574, + 0.06014923378825188, + 0.09213333576917648, + 0.04700919985771179, + 0.023798344656825066, + 0.05083906650543213, + 0.1407812237739563, + 0.02568114548921585, + 0.04631008580327034 + ], + "credit_map": [ + 0.016584662720561028, + 0.01725287176668644, + 0.021230963990092278, + 0.025800015777349472, + 0.021939866244792938, + 0.013395286165177822, + 0.013981327414512634, + 0.017703616991639137, + 0.027875268831849098, + 0.05784282088279724, + 0.04138626530766487, + 0.030863292515277863, + 0.04552435129880905, + 0.09413476288318634, + 0.037653129547834396, + 0.030129874125123024, + 0.005357372108846903, + 0.006613972131162882, + 0.009173602797091007, + 0.005187740549445152, + 0.010694678872823715, + 0.005993744358420372, + 0.009457922540605068, + 0.007670195773243904, + 0.02980772964656353, + 0.007542522624135017, + 0.006739053409546614, + 0.004894729238003492, + 0.014378953725099564, + 0.016844693571329117, + 0.00692402059212327, + 0.005001624580472708, + 0.003766313660889864, + 0.00483359768986702, + 0.005609332118183374, + 0.011071879416704178, + 0.005719367880374193, + 0.003988546784967184, + 0.00643902737647295, + 0.0055811298079788685, + 0.0030788234435021877, + 0.002575014950707555, + 0.0033688100520521402, + 0.0029014041647315025, + 0.006079872604459524, + 0.001976324710994959, + 0.0019448427483439445, + 0.0018732473254203796, + 0.004678776487708092, + 0.005958484951406717, + 0.005588850472122431, + 0.005994495470076799, + 0.008565887808799744, + 0.006454709451645613, + 0.008661503903567791, + 0.004936352837830782, + 0.012529538944363594, + 0.019325312227010727, + 0.014672777615487576, + 0.013720255345106125, + 0.015108325518667698, + 0.014199777506291866, + 0.03238765150308609, + 0.018837573006749153, + 0.0043431115336716175, + 0.0027337472420185804, + 0.003875704715028405, + 0.0025850259698927402, + 0.003547178115695715, + 0.0036289000418037176, + 0.0019438479794189334, + 0.003023626282811165, + 0.005541435908526182, + 0.005453169811517, + 0.004916834644973278, + 0.00609120586887002, + 0.006982075050473213, + 0.005611209198832512, + 0.004682541359215975, + 0.0070316060446202755 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 7 + ], + "selected_reward": -0.20159879326820374 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 main subject a man in his 30s wearing a charcoal gray suit and white shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the electric scooter's frame made of matte black carbon fiber with a woven texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scooter have a vibrant red LED strip on the rear fender?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene include long shadows and lens flares caused by afternoon sun?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.28135427832603455 + ], + "opposite_logprob_tokens": [ + -1.406354308128357 + ], + "logprob_reward": -0.28135427832603455, + "yes_no_margin_reward": 1.125, + "token_credit_reward": -0.28135427832603455, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.34207648038864136, + 0.17936411499977112, + 0.046314992010593414, + 0.1677163988351822, + 0.03433125093579292, + 0.03507257625460625, + 0.061157260090112686, + 0.05816825479269028, + 0.03632504492998123, + 0.03947363793849945 + ], + "credit_map": [ + 0.02543778344988823, + 0.053973522037267685, + 0.042263664305210114, + 0.03776109963655472, + 0.06400679796934128, + 0.05674339085817337, + 0.03195486217737198, + 0.02993534877896309, + 0.008232680149376392, + 0.03234139457345009, + 0.0139384800568223, + 0.009458967484533787, + 0.02335287444293499, + 0.06107252463698387, + 0.014036738313734531, + 0.01693044789135456, + 0.004879381973296404, + 0.0037729209288954735, + 0.009969634935259819, + 0.00452513387426734, + 0.00844535231590271, + 0.0032524976413697004, + 0.00870466884225607, + 0.0027653942815959454, + 0.06132814288139343, + 0.012870801612734795, + 0.009987950325012207, + 0.005481621250510216, + 0.02590165287256241, + 0.03553733602166176, + 0.008328056894242764, + 0.008280819281935692, + 0.0035853195004165173, + 0.00525134801864624, + 0.004467378370463848, + 0.005569308530539274, + 0.003627587342634797, + 0.005090703256428242, + 0.0037137630861252546, + 0.003025840036571026, + 0.005801813676953316, + 0.005557936616241932, + 0.0039433762431144714, + 0.004191241227090359, + 0.005633240099996328, + 0.0047508832067251205, + 0.0024551046080887318, + 0.002738978248089552, + 0.012396649457514286, + 0.006031534168869257, + 0.006443404592573643, + 0.006284546572715044, + 0.010062627494335175, + 0.006523194257169962, + 0.007157558109611273, + 0.006257742643356323, + 0.006670869421213865, + 0.0052556367591023445, + 0.005423611495643854, + 0.006367363967001438, + 0.0051079257391393185, + 0.006164726801216602, + 0.015283294022083282, + 0.007894818671047688, + 0.003073920262977481, + 0.0046084788627922535, + 0.003420683089643717, + 0.0030481000430881977, + 0.008012826554477215, + 0.005394690670073032, + 0.005306142847985029, + 0.003460197476670146, + 0.00505397142842412, + 0.004655170254409313, + 0.004267560318112373, + 0.006504627410322428, + 0.004867083393037319, + 0.005155453458428383, + 0.0032494503539055586, + 0.005720316432416439 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 3 + ], + "selected_reward": -0.28135427832603455 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 main subject a man in his 30s wearing a charcoal gray suit and white shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the electric scooter's frame made of matte black carbon fiber with a woven texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scooter have a vibrant red LED strip on the rear fender?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene include long shadows and lens flares caused by afternoon sun?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1427910476922989 + ], + "opposite_logprob_tokens": [ + -2.0177910327911377 + ], + "logprob_reward": -0.1427910476922989, + "yes_no_margin_reward": 1.875, + "token_credit_reward": -0.1427910476922989, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1460951566696167, + 0.06421539187431335, + 0.06833650916814804, + 0.10986689478158951, + 0.06596393138170242, + 0.046353213489055634, + 0.06727492064237595, + 0.2609812319278717, + 0.06588739901781082, + 0.10502538084983826 + ], + "credit_map": [ + 0.014280331321060658, + 0.015309471637010574, + 0.02718433365225792, + 0.020224353298544884, + 0.022105704993009567, + 0.009846480563282967, + 0.01765718124806881, + 0.01948729157447815, + 0.004328736569732428, + 0.009913953021168709, + 0.00512872776016593, + 0.005106193479150534, + 0.009570921771228313, + 0.016873545944690704, + 0.008397472091019154, + 0.004895838443189859, + 0.007185847964137793, + 0.006350470706820488, + 0.008946800604462624, + 0.004784503951668739, + 0.018090544268488884, + 0.006067688576877117, + 0.011300474405288696, + 0.005610178224742413, + 0.01294419914484024, + 0.013314038515090942, + 0.012371276505291462, + 0.011186409741640091, + 0.015264001674950123, + 0.019651558250188828, + 0.015079698525369167, + 0.010055704042315483, + 0.0058450475335121155, + 0.00883568450808525, + 0.007754182908684015, + 0.011807008646428585, + 0.007968421094119549, + 0.007969380356371403, + 0.01027696579694748, + 0.005507238674908876, + 0.005766755435615778, + 0.00709871482104063, + 0.004645481705665588, + 0.004797148052603006, + 0.007890033535659313, + 0.008692150935530663, + 0.0033639760222285986, + 0.004098952282220125, + 0.005416868254542351, + 0.005604881793260574, + 0.007290767040103674, + 0.014639991335570812, + 0.01115715317428112, + 0.006253266707062721, + 0.011014388874173164, + 0.005897603929042816, + 0.020278124138712883, + 0.033094339072704315, + 0.029272535815835, + 0.03465647995471954, + 0.020170865580439568, + 0.025143994018435478, + 0.0651906356215477, + 0.03317425400018692, + 0.006437170784920454, + 0.006346480920910835, + 0.007215507794171572, + 0.00978883821517229, + 0.008141331374645233, + 0.00904110912233591, + 0.01105035562068224, + 0.007866603322327137, + 0.014593039639294147, + 0.00855270680040121, + 0.011792770586907864, + 0.01826445385813713, + 0.010741684585809708, + 0.013019412755966187, + 0.014790457673370838, + 0.013270841911435127 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 3 + ], + "selected_reward": -0.1427910476922989 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 main subject a man in his 30s wearing a charcoal gray suit and white shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the electric scooter's frame made of matte black carbon fiber with a woven texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scooter have a vibrant red LED strip on the rear fender?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene include long shadows and lens flares caused by afternoon sun?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"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": [ + 0.12918676435947418, + 0.07996315509080887, + 0.055118806660175323, + 0.09626928716897964, + 0.27083635330200195, + 0.05492838844656944, + 0.056292422115802765, + 0.11269186437129974, + 0.08876103907823563, + 0.05595189705491066 + ], + "credit_map": [ + 0.014296385459601879, + 0.02103240415453911, + 0.009318639524281025, + 0.01204642467200756, + 0.037613462656736374, + 0.01570240966975689, + 0.009494236670434475, + 0.009682809002697468, + 0.008978532627224922, + 0.011003399267792702, + 0.015585212968289852, + 0.0057834177277982235, + 0.008332700468599796, + 0.01823447085916996, + 0.005615714937448502, + 0.006429703906178474, + 0.006203615106642246, + 0.005653435364365578, + 0.009074840694665909, + 0.004974749870598316, + 0.009113711304962635, + 0.004102442879229784, + 0.01191775780171156, + 0.0040782541036605835, + 0.03608434647321701, + 0.01042220089584589, + 0.006053087301552296, + 0.003510784124955535, + 0.014460356906056404, + 0.01420863438397646, + 0.0065280599519610405, + 0.005001818761229515, + 0.028545383363962173, + 0.02844047173857689, + 0.0279863104224205, + 0.03186547011137009, + 0.04798175394535065, + 0.05732773244380951, + 0.027496421709656715, + 0.02119278535246849, + 0.0082706268876791, + 0.006267638877034187, + 0.005096345208585262, + 0.005998035427182913, + 0.012295442633330822, + 0.008508730679750443, + 0.003929788246750832, + 0.004561777226626873, + 0.007112354505807161, + 0.005650684237480164, + 0.004242728929966688, + 0.008137935772538185, + 0.009778057225048542, + 0.0062995352782309055, + 0.010280266404151917, + 0.004790854640305042, + 0.010351558215916157, + 0.013099484145641327, + 0.02292119897902012, + 0.013853100128471851, + 0.007671821396797895, + 0.013795009814202785, + 0.018276620656251907, + 0.01272306777536869, + 0.00784245878458023, + 0.008527019061148167, + 0.009698785841464996, + 0.012730367481708527, + 0.011223254725337029, + 0.011196840554475784, + 0.014527690596878529, + 0.013014617376029491, + 0.006920390762388706, + 0.0063353111036121845, + 0.007310632616281509, + 0.006968955975025892, + 0.007133477367460728, + 0.006560469511896372, + 0.005304915364831686, + 0.009417744353413582 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 7 + ], + "selected_reward": -0.004636727273464203 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "smooth, low-angle tracking shot captures a professional man in his 30s with neatly coiffed dark hair as he effortlessly glides through a sun-drenched, modern urban business district on a stylish electric scooter. Dressed in an impeccably tailored charcoal gray suit over a crisp white shirt, his posture is confident and relaxed. The scooter itself is a design marvel, its frame constructed of matte black carbon fiber with a subtle, woven texture that draws attention. As he approaches a wide, clearly marked crosswalk, he deliberately and smoothly applies the brakes. The scooter decelerates gracefully, its advanced braking system underscored by the vibrant red LED strip on the rear fender. He comes to a perfect, steady stop just before the white line, placing one polished black leather Oxford shoe firmly on the clean sidewalk. He calmly turns his head, yielding to the sleek, high-end car passing before him. The entire scene is bathed in the warm golden glow of the afternoon sun, casting long shadows on the glass and steel skyscrapers in the background and producing beautiful lens flares, creating a sophisticated and aspirational atmosphere that showcases the city's sophisticated transportation and cutting-edge security.", + "seed": 14919, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000021/group_6/index_7112_seed_14919_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": -0.2630868196487427, + "vlm_logprob": -0.2630868196487427, + "vlm_yes_no_margin": 1.724999988079071, + "vlm_token_credit": -0.2630868196487427, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 main subject a man in his 30s wearing a charcoal gray suit and white shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the electric scooter's frame made of matte black carbon fiber with a woven texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scooter have a vibrant red LED strip on the rear fender?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene include long shadows and lens flares caused by afternoon sun?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1857433021068573 + ], + "opposite_logprob_tokens": [ + -2.1857433319091797 + ], + "logprob_reward": -0.1857433021068573, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.1857433021068573, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07270926982164383, + 0.0695231705904007, + 0.12701216340065002, + 0.2099277824163437, + 0.06155994534492493, + 0.048804398626089096, + 0.03490613400936127, + 0.30188071727752686, + 0.05088897421956062, + 0.022787457332015038 + ], + "credit_map": [ + 0.011294114403426647, + 0.01032655406743288, + 0.007453165017068386, + 0.008810766041278839, + 0.008336065337061882, + 0.006663131527602673, + 0.00786400493234396, + 0.011961471289396286, + 0.020779220387339592, + 0.021149279549717903, + 0.0029251794330775738, + 0.004949429538100958, + 0.005301771219819784, + 0.00448231678456068, + 0.006133741233497858, + 0.003802236635237932, + 0.018479952588677406, + 0.02846689149737358, + 0.011551711708307266, + 0.016741471365094185, + 0.011668610386550426, + 0.013975719921290874, + 0.015509339980781078, + 0.010618465952575207, + 0.030465401709079742, + 0.022152861580252647, + 0.026565568521618843, + 0.03444753959774971, + 0.021719686686992645, + 0.021330522373318672, + 0.03048868104815483, + 0.022757507860660553, + 0.004189311061054468, + 0.003016672097146511, + 0.009790727868676186, + 0.010033491998910904, + 0.005890174303203821, + 0.011750461533665657, + 0.008655382320284843, + 0.008233725093305111, + 0.006193944253027439, + 0.0048621282912790775, + 0.005173428915441036, + 0.005246064160019159, + 0.0061490098014473915, + 0.008959341794252396, + 0.005890927743166685, + 0.006329551804810762, + 0.002248835051432252, + 0.0051813507452607155, + 0.0034873259719461203, + 0.005366906989365816, + 0.003593707922846079, + 0.0049918219447135925, + 0.004703272134065628, + 0.005332914646714926, + 0.007978703826665878, + 0.008560067042708397, + 0.1250682920217514, + 0.1289919912815094, + 0.005000999663025141, + 0.005293064285069704, + 0.013654472306370735, + 0.007333134301006794, + 0.006499209441244602, + 0.008337819948792458, + 0.006748826242983341, + 0.008351225405931473, + 0.00584886921569705, + 0.0052974591962993145, + 0.004156786482781172, + 0.005648776423186064, + 0.002091670874506235, + 0.002189078601077199, + 0.0029949089512228966, + 0.002603766741231084, + 0.0019504375522956252, + 0.0031403691973537207, + 0.003465972375124693, + 0.004351250361651182 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 3, + 2 + ], + "selected_reward": -0.1857433021068573 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 main subject a man in his 30s wearing a charcoal gray suit and white shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the electric scooter's frame made of matte black carbon fiber with a woven texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scooter have a vibrant red LED strip on the rear fender?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene include long shadows and lens flares caused by afternoon sun?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2521747350692749 + ], + "opposite_logprob_tokens": [ + -1.502174735069275 + ], + "logprob_reward": -0.2521747350692749, + "yes_no_margin_reward": 1.25, + "token_credit_reward": -0.2521747350692749, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18319076299667358, + 0.1919628083705902, + 0.0833505392074585, + 0.0743674486875534, + 0.04604623466730118, + 0.05238233879208565, + 0.10381923615932465, + 0.09548455476760864, + 0.11062067747116089, + 0.058775395154953 + ], + "credit_map": [ + 0.018601777032017708, + 0.0329340323805809, + 0.013422997668385506, + 0.020934956148266792, + 0.0225080493837595, + 0.02536557987332344, + 0.03005523979663849, + 0.0193681288510561, + 0.05519919842481613, + 0.04693296179175377, + 0.009988993406295776, + 0.015637461096048355, + 0.0166020467877388, + 0.011869705282151699, + 0.021175023168325424, + 0.0145574314519763, + 0.010153399780392647, + 0.010780565440654755, + 0.007585461717098951, + 0.015919048339128494, + 0.007748593110591173, + 0.00867024902254343, + 0.015411312691867352, + 0.007081907242536545, + 0.012458287179470062, + 0.005709564778953791, + 0.013555475510656834, + 0.007142553571611643, + 0.009997956454753876, + 0.009093532338738441, + 0.008899619802832603, + 0.007510458584874868, + 0.004885742906481028, + 0.0027492011431604624, + 0.005871998146176338, + 0.00630793534219265, + 0.004724706523120403, + 0.004859240725636482, + 0.008340185508131981, + 0.008307226002216339, + 0.009263120591640472, + 0.00575623894110322, + 0.006554636172950268, + 0.006628155708312988, + 0.004176819231361151, + 0.0038042492233216763, + 0.008134271018207073, + 0.008064845576882362, + 0.0047553288750350475, + 0.005758349783718586, + 0.007727631367743015, + 0.03609698638319969, + 0.004720501601696014, + 0.005652792286127806, + 0.03065355494618416, + 0.008454094640910625, + 0.007501054555177689, + 0.013742025941610336, + 0.015097598545253277, + 0.011608900502324104, + 0.016391806304454803, + 0.00755067216232419, + 0.01602042280137539, + 0.007572071626782417, + 0.00913702417165041, + 0.011782074347138405, + 0.015812676399946213, + 0.015614277683198452, + 0.009766356088221073, + 0.016029343008995056, + 0.009288482367992401, + 0.023190442472696304, + 0.003865422448143363, + 0.005399995017796755, + 0.013452546671032906, + 0.013895712792873383, + 0.004376124124974012, + 0.004220173228532076, + 0.006137421820312738, + 0.0074280002154409885 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 8 + ], + "selected_reward": -0.2521747350692749 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 main subject a man in his 30s wearing a charcoal gray suit and white shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the electric scooter's frame made of matte black carbon fiber with a woven texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scooter have a vibrant red LED strip on the rear fender?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene include long shadows and lens flares caused by afternoon sun?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.4743916988372803 + ], + "opposite_logprob_tokens": [ + -0.9743916988372803 + ], + "logprob_reward": -0.4743916988372803, + "yes_no_margin_reward": 0.5, + "token_credit_reward": -0.4743916988372803, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03453199937939644, + 0.09739167243242264, + 0.020932579413056374, + 0.03963281586766243, + 0.01702352613210678, + 0.04974422603845596, + 0.14073382318019867, + 0.5699920058250427, + 0.02168557420372963, + 0.00833177287131548 + ], + "credit_map": [ + 0.0064642601646482944, + 0.005092037376016378, + 0.00339893763884902, + 0.0036719157360494137, + 0.004119346849620342, + 0.003565473249182105, + 0.004632579628378153, + 0.0035874468740075827, + 0.03377314284443855, + 0.032806724309921265, + 0.004040004685521126, + 0.005524132400751114, + 0.00519177969545126, + 0.006286893039941788, + 0.004149621818214655, + 0.005619358737021685, + 0.00228780135512352, + 0.0023811813443899155, + 0.0023818768095225096, + 0.00253389123827219, + 0.0021955803968012333, + 0.002260008594021201, + 0.0035205399617552757, + 0.0033716976176947355, + 0.004461576696485281, + 0.004516857676208019, + 0.00484935799613595, + 0.003661123802885413, + 0.0077764424495399, + 0.0055124107748270035, + 0.005422323010861874, + 0.0034327171742916107, + 0.0017073472263291478, + 0.0011387113481760025, + 0.002217786153778434, + 0.002254430204629898, + 0.0018472728552296758, + 0.002099208766594529, + 0.0026885385159403086, + 0.0030702282674610615, + 0.007034715265035629, + 0.006706808228045702, + 0.007204775232821703, + 0.007571784313768148, + 0.005442979279905558, + 0.004354297649115324, + 0.005294709466397762, + 0.0061341519467532635, + 0.0052018966525793076, + 0.00735981110483408, + 0.0074098724871873856, + 0.051111314445734024, + 0.006826695986092091, + 0.008386077359318733, + 0.04418785870075226, + 0.010250282473862171, + 0.020164767280220985, + 0.015741340816020966, + 0.2323969453573227, + 0.24136675894260406, + 0.009283998981118202, + 0.01026275847107172, + 0.025900866836309433, + 0.014874565415084362, + 0.001560504431836307, + 0.0019855147693306208, + 0.003269470063969493, + 0.0024813353084027767, + 0.002509647747501731, + 0.004202625248581171, + 0.001559182070195675, + 0.004117291420698166, + 0.0008376832702197134, + 0.00084958306979388, + 0.0010589971207082272, + 0.00075046147685498, + 0.0007482673972845078, + 0.0007660128176212311, + 0.0011594401439651847, + 0.0021613265853375196 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 6, + 1 + ], + "selected_reward": -0.4743916988372803 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 main subject a man in his 30s wearing a charcoal gray suit and white shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the electric scooter's frame made of matte black carbon fiber with a woven texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scooter have a vibrant red LED strip on the rear fender?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene include long shadows and lens flares caused by afternoon sun?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.38705235719680786 + ], + "opposite_logprob_tokens": [ + -1.137052297592163 + ], + "logprob_reward": -0.38705235719680786, + "yes_no_margin_reward": 0.7499999403953552, + "token_credit_reward": -0.38705235719680786, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03609871491789818, + 0.1683247685432434, + 0.07763895392417908, + 0.03985915705561638, + 0.03499981015920639, + 0.08646143227815628, + 0.08497606217861176, + 0.4102121591567993, + 0.028405409306287766, + 0.03302351385354996 + ], + "credit_map": [ + 0.00573760038241744, + 0.004262865055352449, + 0.0027679193299263716, + 0.0033452364150434732, + 0.0038663893938064575, + 0.007005702704191208, + 0.004631811287254095, + 0.004481183364987373, + 0.05464475229382515, + 0.0566241629421711, + 0.007058315444737673, + 0.013648198917508125, + 0.015179237350821495, + 0.007128888741135597, + 0.007384040858596563, + 0.0066571589559316635, + 0.012826608493924141, + 0.010003292001783848, + 0.007965072989463806, + 0.015354051254689693, + 0.005975858308374882, + 0.00962646771222353, + 0.009045328944921494, + 0.006842268165200949, + 0.004656107630580664, + 0.0048831854946911335, + 0.005378847476094961, + 0.005496463272720575, + 0.005590956192463636, + 0.004782160744071007, + 0.004793855827301741, + 0.004277581814676523, + 0.003996433690190315, + 0.0029014332685619593, + 0.004907422699034214, + 0.0034906836226582527, + 0.004663684871047735, + 0.0043052672408521175, + 0.004221650306135416, + 0.006513231433928013, + 0.010281023569405079, + 0.006714366376399994, + 0.008661014027893543, + 0.012255063280463219, + 0.006345878355205059, + 0.005085911136120558, + 0.013807780109345913, + 0.023310387507081032, + 0.0025209065061062574, + 0.004703725688159466, + 0.005880543030798435, + 0.030252056196331978, + 0.0037397039122879505, + 0.005098625551909208, + 0.0250086709856987, + 0.007771824486553669, + 0.01034537609666586, + 0.010583193972706795, + 0.17405176162719727, + 0.17852520942687988, + 0.0037775749806314707, + 0.0056619662791490555, + 0.01733490265905857, + 0.009932169690728188, + 0.0017844375688582659, + 0.003740526968613267, + 0.0039778719656169415, + 0.004980645142495632, + 0.0027896780520677567, + 0.004664772655814886, + 0.001885032164864242, + 0.004582442343235016, + 0.00238408031873405, + 0.002983206883072853, + 0.003959632478654385, + 0.004197539296001196, + 0.0028177150525152683, + 0.002920081838965416, + 0.0037061336915940046, + 0.010055122897028923 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 1, + 5 + ], + "selected_reward": -0.38705235719680786 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 main subject a man in his 30s wearing a charcoal gray suit and white shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the electric scooter's frame made of matte black carbon fiber with a woven texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scooter have a vibrant red LED strip on the rear fender?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene include long shadows and lens flares caused by afternoon sun?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"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": [ + 0.07199298590421677, + 0.04829295352101326, + 0.1543440967798233, + 0.06165436655282974, + 0.05374332517385483, + 0.12891308963298798, + 0.14950601756572723, + 0.21391630172729492, + 0.038254909217357635, + 0.07938194274902344 + ], + "credit_map": [ + 0.010729699395596981, + 0.008215776644647121, + 0.008954547345638275, + 0.005245829466730356, + 0.01039650198072195, + 0.013926100917160511, + 0.006840880960226059, + 0.007683648262172937, + 0.004568786360323429, + 0.00884114671498537, + 0.0028619759250432253, + 0.008803064934909344, + 0.0052463882602751255, + 0.0038647230248898268, + 0.011015762574970722, + 0.003091105492785573, + 0.020932890474796295, + 0.038131844252347946, + 0.012450705282390118, + 0.01825563609600067, + 0.01656891591846943, + 0.018544357270002365, + 0.017038969323039055, + 0.012420780956745148, + 0.011738037690520287, + 0.006377754267305136, + 0.016271313652396202, + 0.007916945964097977, + 0.004574487917125225, + 0.005043049342930317, + 0.005881506484001875, + 0.0038512723986059427, + 0.0062649487517774105, + 0.004349296446889639, + 0.01117259357124567, + 0.011437158100306988, + 0.0052980100736021996, + 0.004774231929332018, + 0.005025767721235752, + 0.005421316251158714, + 0.016142738983035088, + 0.012598377652466297, + 0.013739090412855148, + 0.017257748171687126, + 0.014285040087997913, + 0.02130751498043537, + 0.017609978094697, + 0.015972603112459183, + 0.005726896226406097, + 0.00811468344181776, + 0.006551825907081366, + 0.05493693798780441, + 0.006696953438222408, + 0.00987898837774992, + 0.046740636229515076, + 0.010859088972210884, + 0.00648984219878912, + 0.006316892337054014, + 0.0878869816660881, + 0.09121058881282806, + 0.003622246440500021, + 0.0035369987599551678, + 0.009471610188484192, + 0.005381127819418907, + 0.0032531628385186195, + 0.005291655194014311, + 0.0057402742095291615, + 0.006826095748692751, + 0.0057259732857346535, + 0.004889221396297216, + 0.003044604090973735, + 0.003483922453597188, + 0.0064702811650931835, + 0.006945709697902203, + 0.02144775167107582, + 0.022487124428153038, + 0.0050063468515872955, + 0.003987369127571583, + 0.0081779221072793, + 0.004859443288296461 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 2, + 6 + ], + "selected_reward": -0.016072005033493042 + } + ], + "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/0000021/group_7/index_tensor([9748, 9748])_group_7_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000021/group_7/index_tensor([9748, 9748])_group_7_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..c39d3dc368d60f204ca36fdbc41f8c72510e32e0 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000021/group_7/index_tensor([9748, 9748])_group_7_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "close-up overhead shot, the camera held still, captures a moment brimming with tension in incredibly slow motion: a rich, viscous drop of pure black ink, its perfectly spherical surface reflecting soft, focused light, silently descends onto a sheet of aged rice paper spread flat on a dark wooden tabletop. The paper is a natural beige, its fibrous texture clearly visible in the sidelight, its edges slightly curled with age. Just as the droplet is about to touch the paper, it doesn't soak in or disperse as expected. Instead, it miraculously evaporates upon contact, its liquid form instantly vanishing. Instead, a dense, tangible plume of black smoke curls up from the paper. Unlike ordinary smoke, it doesn't dissipate. Instead, it swirls, twists, and stretches in the air with an elegant and controlled gesture, like a living spirit performing a silent dance, forming complex, abstract forms resembling cursive calligraphy. It then slowly ascends, finally dissipating into the still air. The entire scene is illuminated by a warm side light, casting soft shadows on the ancient paper and smoke, creating an atmosphere of tranquility, concentration and mysterious oriental aesthetics.", + "seed": 6100, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000021/group_7/index_9748_seed_6100_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": -0.6140347750857472, + "vlm_logprob": -0.6140347750857472, + "vlm_yes_no_margin": 2.575, + "vlm_token_credit": -0.6140347750857472, + "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 overhead view of a black ink drop descending onto aged rice paper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the video filmed in incredibly slow motion with the camera held still?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the black ink drop evaporate instantly upon contact with the paper instead of soaking in?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a dense plume of black smoke rise from the paper and form abstract shapes resembling cursive calligraphy?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by warm side light casting soft shadows on the paper and 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": [ + -0.028599431738257408 + ], + "opposite_logprob_tokens": [ + -4.403599262237549 + ], + "logprob_reward": -0.028599431738257408, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.028599431738257408, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12717223167419434, + 0.05690758675336838, + 0.05923302844166756, + 0.11453613638877869, + 0.12545441091060638, + 0.09060972183942795, + 0.08029469847679138, + 0.06456247717142105, + 0.15132874250411987, + 0.12990088760852814 + ], + "credit_map": [ + 0.0097613874822855, + 0.017635749652981758, + 0.023552296683192253, + 0.014817004092037678, + 0.02138005569577217, + 0.011832775548100471, + 0.017051734030246735, + 0.011141236871480942, + 0.0038302596658468246, + 0.007841355167329311, + 0.007331265136599541, + 0.0062684654258191586, + 0.006937569938600063, + 0.008917396888136864, + 0.007152607198804617, + 0.008628666400909424, + 0.004724419675767422, + 0.005789283663034439, + 0.010331769473850727, + 0.0069752042181789875, + 0.0051375883631408215, + 0.011044136248528957, + 0.008298303931951523, + 0.0069323210045695305, + 0.008585990406572819, + 0.011368636041879654, + 0.021675458177924156, + 0.007333752233535051, + 0.013627903535962105, + 0.022705448791384697, + 0.01759122870862484, + 0.011647704057395458, + 0.011354846879839897, + 0.009381357580423355, + 0.012487152591347694, + 0.02006036974489689, + 0.010346841998398304, + 0.023859180510044098, + 0.026497816666960716, + 0.011466842144727707, + 0.0072683291509747505, + 0.00894579105079174, + 0.0096680773422122, + 0.011471250094473362, + 0.007792985532432795, + 0.02388269267976284, + 0.012646622955799103, + 0.008933975361287594, + 0.006413232069462538, + 0.010105070658028126, + 0.009787364862859249, + 0.010916233994066715, + 0.008490736596286297, + 0.014439736492931843, + 0.011025922372937202, + 0.009116399101912975, + 0.008545909076929092, + 0.0074570439755916595, + 0.013867679052054882, + 0.007028896827250719, + 0.007470953743904829, + 0.00723833916708827, + 0.005385800264775753, + 0.0075678578577935696, + 0.01260961964726448, + 0.01347107719630003, + 0.02708096243441105, + 0.03198525682091713, + 0.015950581058859825, + 0.019326483830809593, + 0.01615224778652191, + 0.014752520248293877, + 0.009032933041453362, + 0.008930263109505177, + 0.02897586114704609, + 0.015740787610411644, + 0.012236327864229679, + 0.021557915955781937, + 0.0185259897261858, + 0.0149008147418499 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 0 + ], + "selected_reward": -0.028599431738257408 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 overhead view of a black ink drop descending onto aged rice paper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the video filmed in incredibly slow motion with the camera held still?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the black ink drop evaporate instantly upon contact with the paper instead of soaking in?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a dense plume of black smoke rise from the paper and form abstract shapes resembling cursive calligraphy?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by warm side light casting soft shadows on the paper and 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": [ + -0.01257055252790451 + ], + "opposite_logprob_tokens": [ + -4.387570381164551 + ], + "logprob_reward": -0.01257055252790451, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.01257055252790451, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09650857746601105, + 0.08862663805484772, + 0.07636348158121109, + 0.12490099668502808, + 0.0456792488694191, + 0.06725765019655228, + 0.12692543864250183, + 0.1138119176030159, + 0.1339542418718338, + 0.12597179412841797 + ], + "credit_map": [ + 0.006618992891162634, + 0.0093660494312644, + 0.015506249852478504, + 0.010729188099503517, + 0.015525450929999352, + 0.012727429158985615, + 0.015467042103409767, + 0.010568171739578247, + 0.006574265658855438, + 0.0101369209587574, + 0.01430672500282526, + 0.010274012573063374, + 0.012588059529662132, + 0.01223020814359188, + 0.008670941926538944, + 0.013845506124198437, + 0.007402656599879265, + 0.009153594262897968, + 0.006258884444832802, + 0.008628529496490955, + 0.00906673539429903, + 0.024774739518761635, + 0.00554090179502964, + 0.005537445656955242, + 0.009759400971233845, + 0.011671219952404499, + 0.020826373249292374, + 0.007534489966928959, + 0.01433231495320797, + 0.03179885447025299, + 0.018759023398160934, + 0.010219328105449677, + 0.004116713535040617, + 0.004077784717082977, + 0.003954460844397545, + 0.007656774949282408, + 0.004835041705518961, + 0.006686235778033733, + 0.008767062798142433, + 0.0055851717479527, + 0.006467565428465605, + 0.007372145541012287, + 0.009017796255648136, + 0.009980759583413601, + 0.007428520824760199, + 0.008137324824929237, + 0.009316111914813519, + 0.009537428617477417, + 0.010975449346005917, + 0.013343599624931812, + 0.013097718358039856, + 0.015159121714532375, + 0.01568329520523548, + 0.028341762721538544, + 0.016661446541547775, + 0.01366304513067007, + 0.0149998115375638, + 0.014783616177737713, + 0.02726287953555584, + 0.009729596786201, + 0.012109312228858471, + 0.01400996558368206, + 0.009447023272514343, + 0.011469717137515545, + 0.011392530053853989, + 0.01244408916682005, + 0.03212960809469223, + 0.024801872670650482, + 0.013615947216749191, + 0.015105968341231346, + 0.012933487072587013, + 0.011530733667314053, + 0.028977081179618835, + 0.008518393151462078, + 0.04424737021327019, + 0.0069369785487651825, + 0.010329640470445156, + 0.012732408940792084, + 0.007216428406536579, + 0.007013488560914993 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 6, + 9 + ], + "selected_reward": -0.01257055252790451 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 overhead view of a black ink drop descending onto aged rice paper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the video filmed in incredibly slow motion with the camera held still?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the black ink drop evaporate instantly upon contact with the paper instead of soaking in?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a dense plume of black smoke rise from the paper and form abstract shapes resembling cursive calligraphy?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by warm side light casting soft shadows on the paper and 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": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -2.9299497604370117 + ], + "opposite_logprob_tokens": [ + -0.054949671030044556 + ], + "logprob_reward": -2.9299497604370117, + "yes_no_margin_reward": -2.875, + "token_credit_reward": -2.9299497604370117, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.12144254148006439, + 0.05591033399105072, + 0.03682556375861168, + 0.08606177568435669, + 0.1278892457485199, + 0.04370693489909172, + 0.11453232169151306, + 0.05332000181078911, + 0.10799329727888107, + 0.2523179054260254 + ], + "credit_map": [ + 0.008586786687374115, + 0.017028771340847015, + 0.013977128081023693, + 0.012809647247195244, + 0.0253866296261549, + 0.01644045300781727, + 0.01695295237004757, + 0.010260196402668953, + 0.0036082176957279444, + 0.00544621841982007, + 0.009593039751052856, + 0.007510723080486059, + 0.004184779245406389, + 0.008831590414047241, + 0.00643629627302289, + 0.010299477726221085, + 0.0031928399112075567, + 0.0038229115307331085, + 0.005813658703118563, + 0.0053842454217374325, + 0.0043020714074373245, + 0.007143502589315176, + 0.0038633206859230995, + 0.003303017234429717, + 0.007369926664978266, + 0.008144116960465908, + 0.017086060717701912, + 0.005630902945995331, + 0.008137216791510582, + 0.021701186895370483, + 0.012013163417577744, + 0.0059792110696434975, + 0.01243982370942831, + 0.010791896842420101, + 0.009717202745378017, + 0.017922572791576385, + 0.012470130808651447, + 0.031200388446450233, + 0.02265036292374134, + 0.010696894489228725, + 0.006068193353712559, + 0.0038855536840856075, + 0.007990037091076374, + 0.005286813713610172, + 0.004334267694503069, + 0.00787148717790842, + 0.004630234558135271, + 0.0036403548438102007, + 0.010132658295333385, + 0.009510613046586514, + 0.022536495700478554, + 0.008895989507436752, + 0.00681697903200984, + 0.04200929030776024, + 0.009119033813476562, + 0.005511275492608547, + 0.005781467072665691, + 0.005274360999464989, + 0.007937504909932613, + 0.006549745798110962, + 0.006566817406564951, + 0.009307005442678928, + 0.006714438553899527, + 0.0051886653527617455, + 0.011554772965610027, + 0.011755428276956081, + 0.022171933203935623, + 0.013832706026732922, + 0.013804112561047077, + 0.011327260173857212, + 0.010937107726931572, + 0.012609992176294327, + 0.04185573011636734, + 0.01998010277748108, + 0.016579238697886467, + 0.024851122871041298, + 0.02270698733627796, + 0.08016631007194519, + 0.026629790663719177, + 0.019548682495951653 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 4, + 0 + ], + "selected_reward": -2.9299497604370117 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 overhead view of a black ink drop descending onto aged rice paper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the video filmed in incredibly slow motion with the camera held still?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the black ink drop evaporate instantly upon contact with the paper instead of soaking in?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a dense plume of black smoke rise from the paper and form abstract shapes resembling cursive calligraphy?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by warm side light casting soft shadows on the paper and 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": [ + -0.08921535313129425 + ], + "opposite_logprob_tokens": [ + -2.4642152786254883 + ], + "logprob_reward": -0.08921535313129425, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08921535313129425, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06579902023077011, + 0.04333851858973503, + 0.05637115240097046, + 0.12194016575813293, + 0.24583645164966583, + 0.018658407032489777, + 0.054657526314258575, + 0.06215891242027283, + 0.05694929137825966, + 0.2742905914783478 + ], + "credit_map": [ + 0.003988682758063078, + 0.006187804043292999, + 0.004913893528282642, + 0.006972768809646368, + 0.018378810957074165, + 0.007690088357776403, + 0.009807192720472813, + 0.007859776727855206, + 0.003415454179048538, + 0.006415997631847858, + 0.005911723710596561, + 0.004842066206037998, + 0.005065709352493286, + 0.007450108416378498, + 0.0034918177407234907, + 0.00674564391374588, + 0.004841129761189222, + 0.0062857987359166145, + 0.007243400439620018, + 0.004392941482365131, + 0.007231306284666061, + 0.016044750809669495, + 0.006096377037465572, + 0.004235444590449333, + 0.008726072497665882, + 0.011602677404880524, + 0.020693408325314522, + 0.007707723416388035, + 0.014240690506994724, + 0.025524061173200607, + 0.022813957184553146, + 0.01063158456236124, + 0.022016119211912155, + 0.019385341554880142, + 0.019712934270501137, + 0.03879541531205177, + 0.020062129944562912, + 0.05377867445349693, + 0.0478927344083786, + 0.024193091318011284, + 0.0015790697652846575, + 0.0017076184740290046, + 0.002303030574694276, + 0.0030724862590432167, + 0.001919585163705051, + 0.0028331757057458162, + 0.0032447841949760914, + 0.0019986575935035944, + 0.005509689915925264, + 0.006142312660813332, + 0.008994970470666885, + 0.00974210724234581, + 0.0046910676173865795, + 0.0057981894351542, + 0.008007119409739971, + 0.005772068165242672, + 0.0040174201130867004, + 0.004708034917712212, + 0.01656421460211277, + 0.005891871638596058, + 0.004500857554376125, + 0.018477067351341248, + 0.0043253167532384396, + 0.003674129955470562, + 0.006556577980518341, + 0.006308781448751688, + 0.007330885156989098, + 0.008609366603195667, + 0.004865226801484823, + 0.01283235289156437, + 0.005239163059741259, + 0.005206932779401541, + 0.047017812728881836, + 0.0238010473549366, + 0.028116095811128616, + 0.01336120255291462, + 0.018492238596081734, + 0.11382737755775452, + 0.017580274492502213, + 0.012094552628695965 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 4, + 3 + ], + "selected_reward": -0.08921535313129425 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 overhead view of a black ink drop descending onto aged rice paper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the video filmed in incredibly slow motion with the camera held still?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the black ink drop evaporate instantly upon contact with the paper instead of soaking in?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a dense plume of black smoke rise from the paper and form abstract shapes resembling cursive calligraphy?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by warm side light casting soft shadows on the paper and 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": [ + -0.009838777594268322 + ], + "opposite_logprob_tokens": [ + -4.634838581085205 + ], + "logprob_reward": -0.009838777594268322, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009838777594268322, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10127925127744675, + 0.09866943955421448, + 0.07937297224998474, + 0.2352800816297531, + 0.1023191586136818, + 0.025257235392928123, + 0.05825743079185486, + 0.05445656552910805, + 0.06096296012401581, + 0.18414494395256042 + ], + "credit_map": [ + 0.008398836478590965, + 0.017122138291597366, + 0.00976715236902237, + 0.010221114382147789, + 0.022942066192626953, + 0.010720335878431797, + 0.01256529800593853, + 0.009542315267026424, + 0.007543930783867836, + 0.00991984736174345, + 0.0187224131077528, + 0.01250580046325922, + 0.010757717303931713, + 0.013805457390844822, + 0.012120721861720085, + 0.013293554075062275, + 0.008903200738132, + 0.009210003539919853, + 0.0050628697499632835, + 0.00815874245017767, + 0.008695135824382305, + 0.02442413941025734, + 0.0075496709905564785, + 0.007369214668869972, + 0.020440097898244858, + 0.03141022473573685, + 0.04107378050684929, + 0.014451120980083942, + 0.015914229676127434, + 0.07879545539617538, + 0.014875958673655987, + 0.018319204449653625, + 0.007400178350508213, + 0.009669759310781956, + 0.007818046025931835, + 0.018849937245249748, + 0.01104777678847313, + 0.013092514127492905, + 0.02205914817750454, + 0.012381798587739468, + 0.0031815932597965, + 0.002155665308237076, + 0.0030439558904618025, + 0.004112868569791317, + 0.002425925573334098, + 0.004634609911590815, + 0.003238516626879573, + 0.0024641002528369427, + 0.004144658800214529, + 0.004589327145367861, + 0.012317643500864506, + 0.005558271426707506, + 0.004566938616335392, + 0.01786251924932003, + 0.0056151109747588634, + 0.003602957120165229, + 0.005135927349328995, + 0.006798350717872381, + 0.00980313029140234, + 0.008502001874148846, + 0.005227683577686548, + 0.007787758484482765, + 0.005443928297609091, + 0.005757786333560944, + 0.006046764086931944, + 0.006376252975314856, + 0.0091686537489295, + 0.009359331801533699, + 0.007958203554153442, + 0.005842673592269421, + 0.009163995273411274, + 0.007047088351100683, + 0.022762471809983253, + 0.013581265695393085, + 0.021156098693609238, + 0.01932324841618538, + 0.015104386024177074, + 0.05521409586071968, + 0.021788422018289566, + 0.015214959159493446 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 9, + 4 + ], + "selected_reward": -0.009838777594268322 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "No", + "Yes", + "Yes" + ] + }, + { + "prompt": "close-up overhead shot, the camera held still, captures a moment brimming with tension in incredibly slow motion: a rich, viscous drop of pure black ink, its perfectly spherical surface reflecting soft, focused light, silently descends onto a sheet of aged rice paper spread flat on a dark wooden tabletop. The paper is a natural beige, its fibrous texture clearly visible in the sidelight, its edges slightly curled with age. Just as the droplet is about to touch the paper, it doesn't soak in or disperse as expected. Instead, it miraculously evaporates upon contact, its liquid form instantly vanishing. Instead, a dense, tangible plume of black smoke curls up from the paper. Unlike ordinary smoke, it doesn't dissipate. Instead, it swirls, twists, and stretches in the air with an elegant and controlled gesture, like a living spirit performing a silent dance, forming complex, abstract forms resembling cursive calligraphy. It then slowly ascends, finally dissipating into the still air. The entire scene is illuminated by a warm side light, casting soft shadows on the ancient paper and smoke, creating an atmosphere of tranquility, concentration and mysterious oriental aesthetics.", + "seed": 6101, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000021/group_7/index_9748_seed_6101_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": -0.5775562196969986, + "vlm_logprob": -0.5775562196969986, + "vlm_yes_no_margin": 2.275, + "vlm_token_credit": -0.5775562196969986, + "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 overhead view of a black ink drop descending onto aged rice paper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the video filmed in incredibly slow motion with the camera held still?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the black ink drop evaporate instantly upon contact with the paper instead of soaking in?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a dense plume of black smoke rise from the paper and form abstract shapes resembling cursive calligraphy?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by warm side light casting soft shadows on the paper and 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": [ + -0.022827232256531715 + ], + "opposite_logprob_tokens": [ + -5.0228271484375 + ], + "logprob_reward": -0.022827232256531715, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.022827232256531715, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.396217405796051, + 0.03717104718089104, + 0.06509619206190109, + 0.043228477239608765, + 0.07977452874183655, + 0.04822508990764618, + 0.14744502305984497, + 0.07724446803331375, + 0.07869862020015717, + 0.026899181306362152 + ], + "credit_map": [ + 0.04899633303284645, + 0.016192331910133362, + 0.017912743613123894, + 0.10851766914129257, + 0.025063225999474525, + 0.02113562636077404, + 0.06876145303249359, + 0.08963800966739655, + 0.004995714407414198, + 0.0035845397505909204, + 0.004399114288389683, + 0.0043912953697144985, + 0.007316336501389742, + 0.003722962224856019, + 0.00578935444355011, + 0.002971731359139085, + 0.021687055006623268, + 0.005185092333704233, + 0.005214186850935221, + 0.007087701931595802, + 0.0030240246560424566, + 0.004112586844712496, + 0.005259335972368717, + 0.013526204973459244, + 0.014793826267123222, + 0.003549229120835662, + 0.0032404446974396706, + 0.0030367402359843254, + 0.0027794039342552423, + 0.0024535434786230326, + 0.0030133859254419804, + 0.010361905209720135, + 0.024221675470471382, + 0.004190405365079641, + 0.004502878524363041, + 0.005671378690749407, + 0.005178721621632576, + 0.005732348188757896, + 0.004291914869099855, + 0.025985201820731163, + 0.015652960166335106, + 0.002875964855775237, + 0.003246342297643423, + 0.005289030261337757, + 0.006025900598615408, + 0.002717619063332677, + 0.00399685837328434, + 0.008420413359999657, + 0.041703417897224426, + 0.010753733105957508, + 0.01147210132330656, + 0.018770501017570496, + 0.012643065303564072, + 0.008349049836397171, + 0.008933528326451778, + 0.034819625318050385, + 0.018071865662932396, + 0.006608583964407444, + 0.004738869145512581, + 0.015087315812706947, + 0.01244871225208044, + 0.005067445803433657, + 0.008669772185385227, + 0.006551895756274462, + 0.013240277767181396, + 0.008291058242321014, + 0.0069027431309223175, + 0.005058969371020794, + 0.019664207473397255, + 0.004744286183267832, + 0.009390131570398808, + 0.011406941339373589, + 0.004883382003754377, + 0.0027118646539747715, + 0.003228750079870224, + 0.0019306192407384515, + 0.0044648596085608006, + 0.0026190620847046375, + 0.00449272571131587, + 0.002567917574197054 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 4 + ], + "selected_reward": -0.022827232256531715 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 overhead view of a black ink drop descending onto aged rice paper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the video filmed in incredibly slow motion with the camera held still?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the black ink drop evaporate instantly upon contact with the paper instead of soaking in?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a dense plume of black smoke rise from the paper and form abstract shapes resembling cursive calligraphy?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by warm side light casting soft shadows on the paper and 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": [ + -0.02063835971057415 + ], + "opposite_logprob_tokens": [ + -3.8956384658813477 + ], + "logprob_reward": -0.02063835971057415, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.02063835971057415, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13455058634281158, + 0.06410595029592514, + 0.04166954755783081, + 0.21579718589782715, + 0.2684248089790344, + 0.07711664587259293, + 0.06352736055850983, + 0.04248054325580597, + 0.05695488303899765, + 0.03537248075008392 + ], + "credit_map": [ + 0.019354872405529022, + 0.010606592521071434, + 0.011113489046692848, + 0.012635590508580208, + 0.03878500685095787, + 0.012166233733296394, + 0.015547954477369785, + 0.014340856112539768, + 0.01047713402658701, + 0.005052024498581886, + 0.005528242792934179, + 0.012545811012387276, + 0.01170154009014368, + 0.0042094276286661625, + 0.00927381869405508, + 0.005317952483892441, + 0.009592673741281033, + 0.002847827272489667, + 0.0042221094481647015, + 0.006174934562295675, + 0.003759714076295495, + 0.0034176334738731384, + 0.006094491574913263, + 0.005560167599469423, + 0.06380455940961838, + 0.014626660384237766, + 0.008583847433328629, + 0.02725485898554325, + 0.006834542378783226, + 0.014396590180695057, + 0.009053375571966171, + 0.07124276459217072, + 0.10993002355098724, + 0.016745999455451965, + 0.012009896337985992, + 0.006881438661366701, + 0.010396333411335945, + 0.013968474231660366, + 0.010002024471759796, + 0.08849062770605087, + 0.029496129602193832, + 0.004670255817472935, + 0.004142508376389742, + 0.007101827766746283, + 0.007003077305853367, + 0.0032823653891682625, + 0.003908523358404636, + 0.017511971294879913, + 0.009230994619429111, + 0.004896875470876694, + 0.006681900005787611, + 0.008533173240721226, + 0.010194960050284863, + 0.006089279893785715, + 0.007937685586512089, + 0.009962494485080242, + 0.009537740610539913, + 0.004138480871915817, + 0.0041378578171133995, + 0.0046934764832258224, + 0.008207009173929691, + 0.0028591689188033342, + 0.004871487617492676, + 0.00403532525524497, + 0.012663455680012703, + 0.004799878690391779, + 0.007254030089825392, + 0.004534230101853609, + 0.008001673966646194, + 0.003939252346754074, + 0.00745476083829999, + 0.008307605050504208, + 0.007875055074691772, + 0.004093335475772619, + 0.002696388866752386, + 0.001956674735993147, + 0.006585827562958002, + 0.0020681798923760653, + 0.0037085062358528376, + 0.006388518959283829 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 3, + 0 + ], + "selected_reward": -0.02063835971057415 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 overhead view of a black ink drop descending onto aged rice paper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the video filmed in incredibly slow motion with the camera held still?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the black ink drop evaporate instantly upon contact with the paper instead of soaking in?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a dense plume of black smoke rise from the paper and form abstract shapes resembling cursive calligraphy?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by warm side light casting soft shadows on the paper and 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": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -2.6950204372406006 + ], + "opposite_logprob_tokens": [ + -0.07002048194408417 + ], + "logprob_reward": -2.6950204372406006, + "yes_no_margin_reward": -2.625, + "token_credit_reward": -2.6950204372406006, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.07475113123655319, + 0.05436472222208977, + 0.07453225553035736, + 0.22352860867977142, + 0.132355198264122, + 0.08828601241111755, + 0.18690000474452972, + 0.05079539492726326, + 0.06956858932971954, + 0.044918086379766464 + ], + "credit_map": [ + 0.00803230982273817, + 0.006492075975984335, + 0.004241961985826492, + 0.014273066073656082, + 0.018187925219535828, + 0.0048667616210877895, + 0.013156898319721222, + 0.00550013268366456, + 0.00817288551479578, + 0.005885283928364515, + 0.007449100259691477, + 0.004504386335611343, + 0.00913999043405056, + 0.005438479594886303, + 0.008226960897445679, + 0.0055476343259215355, + 0.02281780354678631, + 0.005614790599793196, + 0.004350839648395777, + 0.006590487435460091, + 0.004980358295142651, + 0.003747740760445595, + 0.009074746631085873, + 0.017355486750602722, + 0.06373856961727142, + 0.014407150447368622, + 0.010296273045241833, + 0.031489863991737366, + 0.007572761736810207, + 0.014563895761966705, + 0.010949157178401947, + 0.07051094621419907, + 0.054903171956539154, + 0.009107060730457306, + 0.0065652974881231785, + 0.006499748211354017, + 0.005834359209984541, + 0.0057333726435899734, + 0.005936546251177788, + 0.03777565062046051, + 0.033348944038152695, + 0.00491446303203702, + 0.0041686492040753365, + 0.00708717480301857, + 0.008744003251194954, + 0.003588731400668621, + 0.005032593850046396, + 0.02140144817531109, + 0.06113087385892868, + 0.015304978005588055, + 0.010598874650895596, + 0.01204150915145874, + 0.02336614951491356, + 0.009271882474422455, + 0.01830988936126232, + 0.03687584772706032, + 0.008217529393732548, + 0.005116209853440523, + 0.00418959790840745, + 0.0038435577880591154, + 0.014760186895728111, + 0.0035150160547345877, + 0.006170061882585287, + 0.004983236081898212, + 0.013067169114947319, + 0.00756889209151268, + 0.00718969339504838, + 0.0044938805513083935, + 0.013748467899858952, + 0.004797648172825575, + 0.007305466569960117, + 0.011397370137274265, + 0.013127629645168781, + 0.004650890361517668, + 0.00327152363024652, + 0.006457061972469091, + 0.00680432328954339, + 0.0029848620761185884, + 0.004518083296716213, + 0.003103713970631361 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 6, + 4 + ], + "selected_reward": -2.6950204372406006 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 overhead view of a black ink drop descending onto aged rice paper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the video filmed in incredibly slow motion with the camera held still?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the black ink drop evaporate instantly upon contact with the paper instead of soaking in?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a dense plume of black smoke rise from the paper and form abstract shapes resembling cursive calligraphy?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by warm side light casting soft shadows on the paper and 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": [ + -0.07015318423509598 + ], + "opposite_logprob_tokens": [ + -2.69515323638916 + ], + "logprob_reward": -0.07015318423509598, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07015318423509598, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1812528669834137, + 0.07236175984144211, + 0.047542884945869446, + 0.16562199592590332, + 0.08210281282663345, + 0.047849442809820175, + 0.20686037838459015, + 0.0706469789147377, + 0.05969274789094925, + 0.06606818735599518 + ], + "credit_map": [ + 0.023192070424556732, + 0.015757804736495018, + 0.010451826266944408, + 0.01565779186785221, + 0.06023045629262924, + 0.015330924652516842, + 0.023207729682326317, + 0.017424268648028374, + 0.008738072589039803, + 0.006684125401079655, + 0.008130722679197788, + 0.010310822166502476, + 0.011494201608002186, + 0.005486791487783194, + 0.011795531958341599, + 0.009721488691866398, + 0.010572022758424282, + 0.0032084633130580187, + 0.0036506804171949625, + 0.0036664195358753204, + 0.00756417540833354, + 0.0029253163374960423, + 0.006356651429086924, + 0.009599153883755207, + 0.05894886702299118, + 0.011120970360934734, + 0.007937164977192879, + 0.012473542243242264, + 0.0060823336243629456, + 0.009588157758116722, + 0.007139511406421661, + 0.05233145132660866, + 0.032771408557891846, + 0.006085307337343693, + 0.004680931568145752, + 0.006439754739403725, + 0.004652834963053465, + 0.003307910170406103, + 0.004505368415266275, + 0.019659293815493584, + 0.008148176595568657, + 0.004336860496550798, + 0.004080117680132389, + 0.0030083025339990854, + 0.012225279584527016, + 0.002760763745754957, + 0.007719077635556459, + 0.005570867098867893, + 0.02989766374230385, + 0.01980576105415821, + 0.010129103437066078, + 0.021410498768091202, + 0.04431072250008583, + 0.012466373853385448, + 0.02749335579574108, + 0.041346896439790726, + 0.013850534334778786, + 0.005472420249134302, + 0.005967376753687859, + 0.00739996787160635, + 0.018258675932884216, + 0.006158542353659868, + 0.008719795383512974, + 0.004819669760763645, + 0.010832243598997593, + 0.0037981572095304728, + 0.003529581241309643, + 0.004329940769821405, + 0.010207301005721092, + 0.0057485513389110565, + 0.010399716906249523, + 0.010847260244190693, + 0.016228903084993362, + 0.005894005764275789, + 0.004245017655193806, + 0.0043182699009776115, + 0.01164118479937315, + 0.004075981210917234, + 0.012140666134655476, + 0.007524159736931324 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 3 + ], + "selected_reward": -0.07015318423509598 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 overhead view of a black ink drop descending onto aged rice paper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the video filmed in incredibly slow motion with the camera held still?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the black ink drop evaporate instantly upon contact with the paper instead of soaking in?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a dense plume of black smoke rise from the paper and form abstract shapes resembling cursive calligraphy?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by warm side light casting soft shadows on the paper and 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": [ + -0.07914188504219055 + ], + "opposite_logprob_tokens": [ + -2.579141855239868 + ], + "logprob_reward": -0.07914188504219055, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07914188504219055, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18206787109375, + 0.05522412806749344, + 0.05550150200724602, + 0.14877451956272125, + 0.06553368270397186, + 0.20819169282913208, + 0.12195262312889099, + 0.0511210635304451, + 0.06608549505472183, + 0.045547422021627426 + ], + "credit_map": [ + 0.026746263727545738, + 0.014733932912349701, + 0.01367308758199215, + 0.021265724673867226, + 0.057906754314899445, + 0.01435117144137621, + 0.022060517221689224, + 0.01133042760193348, + 0.007331874221563339, + 0.005178759340196848, + 0.00679031852632761, + 0.006337146274745464, + 0.008174863643944263, + 0.0049402364529669285, + 0.007596001960337162, + 0.008874926716089249, + 0.01582260988652706, + 0.003264104714617133, + 0.003455040045082569, + 0.005257039796561003, + 0.0033623753115534782, + 0.0031036240980029106, + 0.007852902635931969, + 0.013383808545768261, + 0.025967655703425407, + 0.009250825271010399, + 0.0077966079115867615, + 0.03599023073911667, + 0.00505814841017127, + 0.010952786542475224, + 0.010079753585159779, + 0.04367850348353386, + 0.023059139028191566, + 0.004286612384021282, + 0.005821430590003729, + 0.008904938586056232, + 0.003654573345556855, + 0.002607614966109395, + 0.0038564251735806465, + 0.013342943042516708, + 0.08548854291439056, + 0.011019174940884113, + 0.010769805870950222, + 0.02232147380709648, + 0.009469547308981419, + 0.008152100257575512, + 0.00910062063485384, + 0.0518704354763031, + 0.03664035722613335, + 0.011841199360787868, + 0.008450539782643318, + 0.0065208966843783855, + 0.020976701751351357, + 0.009798584505915642, + 0.011517481878399849, + 0.016206862404942513, + 0.008343330584466457, + 0.0040787807665765285, + 0.004232111386954784, + 0.003418724285438657, + 0.016495581716299057, + 0.004643668420612812, + 0.005380914080888033, + 0.0045279511250555515, + 0.011875385418534279, + 0.0064306012354791164, + 0.0062683988362550735, + 0.0052855126559734344, + 0.012588686309754848, + 0.004359254613518715, + 0.009694756008684635, + 0.009582898579537868, + 0.009750175289809704, + 0.004548718221485615, + 0.004271495155990124, + 0.008209913969039917, + 0.006747753359377384, + 0.003411879064515233, + 0.004702962003648281, + 0.00390452379360795 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 3 + ], + "selected_reward": -0.07914188504219055 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "No", + "Yes", + "Yes" + ] + }, + { + "prompt": "close-up overhead shot, the camera held still, captures a moment brimming with tension in incredibly slow motion: a rich, viscous drop of pure black ink, its perfectly spherical surface reflecting soft, focused light, silently descends onto a sheet of aged rice paper spread flat on a dark wooden tabletop. The paper is a natural beige, its fibrous texture clearly visible in the sidelight, its edges slightly curled with age. Just as the droplet is about to touch the paper, it doesn't soak in or disperse as expected. Instead, it miraculously evaporates upon contact, its liquid form instantly vanishing. Instead, a dense, tangible plume of black smoke curls up from the paper. Unlike ordinary smoke, it doesn't dissipate. Instead, it swirls, twists, and stretches in the air with an elegant and controlled gesture, like a living spirit performing a silent dance, forming complex, abstract forms resembling cursive calligraphy. It then slowly ascends, finally dissipating into the still air. The entire scene is illuminated by a warm side light, casting soft shadows on the ancient paper and smoke, creating an atmosphere of tranquility, concentration and mysterious oriental aesthetics.", + "seed": 6102, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000021/group_7/index_9748_seed_6102_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": -0.5889535926282405, + "vlm_logprob": -0.5889535926282405, + "vlm_yes_no_margin": 2.525, + "vlm_token_credit": -0.5889535926282405, + "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 overhead view of a black ink drop descending onto aged rice paper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the video filmed in incredibly slow motion with the camera held still?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the black ink drop evaporate instantly upon contact with the paper instead of soaking in?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a dense plume of black smoke rise from the paper and form abstract shapes resembling cursive calligraphy?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by warm side light casting soft shadows on the paper and 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": [ + -0.020950565114617348 + ], + "opposite_logprob_tokens": [ + -5.020950794219971 + ], + "logprob_reward": -0.020950565114617348, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.020950565114617348, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13823392987251282, + 0.06053721159696579, + 0.056347429752349854, + 0.16475357115268707, + 0.025716036558151245, + 0.05904756113886833, + 0.07861847430467606, + 0.10110490769147873, + 0.22387869656085968, + 0.09176208823919296 + ], + "credit_map": [ + 0.029696259647607803, + 0.006457626353949308, + 0.006856026127934456, + 0.010723351500928402, + 0.043916720896959305, + 0.017814770340919495, + 0.011974383145570755, + 0.010794807225465775, + 0.008972471579909325, + 0.00439114635810256, + 0.0077531165443360806, + 0.005337072070688009, + 0.013185164891183376, + 0.009224393405020237, + 0.0063295792788267136, + 0.005344267934560776, + 0.009052793495357037, + 0.007065258454531431, + 0.006098407786339521, + 0.0055288164876401424, + 0.011336633935570717, + 0.006967931520193815, + 0.004336419049650431, + 0.005961169954389334, + 0.02629215270280838, + 0.02759142778813839, + 0.013760015368461609, + 0.013525653630495071, + 0.02136303298175335, + 0.030287308618426323, + 0.013104328885674477, + 0.018829643726348877, + 0.003241270314902067, + 0.0021717241033911705, + 0.002313996898010373, + 0.002149885054677725, + 0.00391876045614481, + 0.005464017391204834, + 0.0027411559130996466, + 0.003715226659551263, + 0.0071969651617109776, + 0.017860883846879005, + 0.0031848002690821886, + 0.0038548798765987158, + 0.005791935138404369, + 0.01245412789285183, + 0.0037635366898030043, + 0.004940434359014034, + 0.008921687491238117, + 0.016634352505207062, + 0.007263560313731432, + 0.008644883520901203, + 0.006797680631279945, + 0.010016883723437786, + 0.008375069126486778, + 0.01196435559540987, + 0.011905355378985405, + 0.015905223786830902, + 0.005466351751238108, + 0.013882803730666637, + 0.019304126501083374, + 0.017859455198049545, + 0.01095180120319128, + 0.005829792004078627, + 0.028576012700796127, + 0.020559897646307945, + 0.009569494053721428, + 0.008933927863836288, + 0.08135708421468735, + 0.04043501242995262, + 0.02179686166346073, + 0.012650403194129467, + 0.011912671849131584, + 0.0074196793138980865, + 0.0062993490137159824, + 0.008502522483468056, + 0.017917467281222343, + 0.016898471862077713, + 0.010580869391560555, + 0.012231056578457355 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 3, + 0 + ], + "selected_reward": -0.020950565114617348 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 overhead view of a black ink drop descending onto aged rice paper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the video filmed in incredibly slow motion with the camera held still?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the black ink drop evaporate instantly upon contact with the paper instead of soaking in?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a dense plume of black smoke rise from the paper and form abstract shapes resembling cursive calligraphy?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by warm side light casting soft shadows on the paper and 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": [ + -0.023333145305514336 + ], + "opposite_logprob_tokens": [ + -3.7733330726623535 + ], + "logprob_reward": -0.023333145305514336, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023333145305514336, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1358151137828827, + 0.04683384299278259, + 0.06553623825311661, + 0.045171987265348434, + 0.028966883197426796, + 0.0539245530962944, + 0.062045078724622726, + 0.08859526365995407, + 0.351265013217926, + 0.12184599041938782 + ], + "credit_map": [ + 0.02303915284574032, + 0.007072266191244125, + 0.014589149504899979, + 0.008819674141705036, + 0.042930539697408676, + 0.018805621191859245, + 0.012445067055523396, + 0.00811365619301796, + 0.007473235949873924, + 0.003771815914660692, + 0.007112670224159956, + 0.0026537333615124226, + 0.010006032884120941, + 0.006036096718162298, + 0.004564311355352402, + 0.005215949844568968, + 0.012008611112833023, + 0.008409607224166393, + 0.005914735607802868, + 0.007534315809607506, + 0.013067352585494518, + 0.00701932655647397, + 0.005546786822378635, + 0.006035508122295141, + 0.009275109507143497, + 0.00447138212621212, + 0.003734487807378173, + 0.0035492146853357553, + 0.004742149729281664, + 0.011970563791692257, + 0.0031998315826058388, + 0.004229250829666853, + 0.006899620406329632, + 0.0042626867070794106, + 0.0015985905192792416, + 0.0020579330157488585, + 0.0030471873469650745, + 0.005895860027521849, + 0.001992533216252923, + 0.003212475683540106, + 0.006980102974921465, + 0.016602981835603714, + 0.002936830511316657, + 0.002864087000489235, + 0.0056804330088198185, + 0.012408534064888954, + 0.0028526855167001486, + 0.0035989051684737206, + 0.01081748865544796, + 0.010649221017956734, + 0.004867567680776119, + 0.004295926075428724, + 0.005580632947385311, + 0.01499621570110321, + 0.0047165569849312305, + 0.006121477112174034, + 0.019376616925001144, + 0.00924742128700018, + 0.004375753458589315, + 0.006537133362144232, + 0.0124949486926198, + 0.026137391105294228, + 0.00562775693833828, + 0.004798253532499075, + 0.03804158791899681, + 0.03977763652801514, + 0.013302025385200977, + 0.01627851463854313, + 0.11852870881557465, + 0.07149688154459, + 0.03304996341466904, + 0.020789721980690956, + 0.021126262843608856, + 0.007981078699231148, + 0.005384495016187429, + 0.009260650724172592, + 0.0409560427069664, + 0.016373950988054276, + 0.010266212746500969, + 0.010497311130166054 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 9 + ], + "selected_reward": -0.023333145305514336 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 overhead view of a black ink drop descending onto aged rice paper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the video filmed in incredibly slow motion with the camera held still?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the black ink drop evaporate instantly upon contact with the paper instead of soaking in?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a dense plume of black smoke rise from the paper and form abstract shapes resembling cursive calligraphy?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by warm side light casting soft shadows on the paper and 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": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -2.812039613723755 + ], + "opposite_logprob_tokens": [ + -0.062039658427238464 + ], + "logprob_reward": -2.812039613723755, + "yes_no_margin_reward": -2.75, + "token_credit_reward": -2.812039613723755, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.1565149873495102, + 0.057562075555324554, + 0.04810512065887451, + 0.05430086702108383, + 0.07316836714744568, + 0.05050235614180565, + 0.07773592323064804, + 0.1416131556034088, + 0.274800568819046, + 0.06569661945104599 + ], + "credit_map": [ + 0.0392424650490284, + 0.006566145457327366, + 0.008933772332966328, + 0.004791624378412962, + 0.061542488634586334, + 0.021302904933691025, + 0.008895142003893852, + 0.005240431986749172, + 0.008367695845663548, + 0.00461930176243186, + 0.00995638594031334, + 0.005036736372858286, + 0.01129116304218769, + 0.0079280324280262, + 0.004578470252454281, + 0.005784282926470041, + 0.00858406163752079, + 0.0065081436187028885, + 0.004792654421180487, + 0.00526328943669796, + 0.008934476412832737, + 0.0060035353526473045, + 0.0038895297329872847, + 0.004129426088184118, + 0.006970415357500315, + 0.009806079789996147, + 0.005924289580434561, + 0.005873850081115961, + 0.0063713728450238705, + 0.009911845438182354, + 0.004094919189810753, + 0.005348091945052147, + 0.00925394892692566, + 0.012081188149750233, + 0.004783837124705315, + 0.006760630756616592, + 0.00988670065999031, + 0.01610497012734413, + 0.007628186140209436, + 0.006668899208307266, + 0.008406588807702065, + 0.004637124948203564, + 0.0025897060986608267, + 0.002076311968266964, + 0.007868744432926178, + 0.017333492636680603, + 0.0028800242580473423, + 0.0047103604301810265, + 0.019115621224045753, + 0.009224891662597656, + 0.002930823713541031, + 0.006938665173947811, + 0.005210707429796457, + 0.024661274626851082, + 0.004834847990423441, + 0.004819083958864212, + 0.03071175515651703, + 0.01312309131026268, + 0.0057465070858597755, + 0.009802182205021381, + 0.017196130007505417, + 0.04937739670276642, + 0.008497331291437149, + 0.007158752530813217, + 0.05227236822247505, + 0.020002758130431175, + 0.009490188211202621, + 0.011399996466934681, + 0.0647159144282341, + 0.08450217545032501, + 0.020264938473701477, + 0.012152224779129028, + 0.009932594373822212, + 0.009673791006207466, + 0.004580256063491106, + 0.009068948216736317, + 0.010334404185414314, + 0.007720524910837412, + 0.0073442659340798855, + 0.007041833363473415 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 7 + ], + "selected_reward": -2.812039613723755 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 overhead view of a black ink drop descending onto aged rice paper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the video filmed in incredibly slow motion with the camera held still?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the black ink drop evaporate instantly upon contact with the paper instead of soaking in?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a dense plume of black smoke rise from the paper and form abstract shapes resembling cursive calligraphy?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by warm side light casting soft shadows on the paper and 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": [ + -0.07018230110406876 + ], + "opposite_logprob_tokens": [ + -2.6951823234558105 + ], + "logprob_reward": -0.07018230110406876, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07018230110406876, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03697541728615761, + 0.025979211553931236, + 0.02661253698170185, + 0.04853057861328125, + 0.0115587767213583, + 0.041189827024936676, + 0.04341384395956993, + 0.07094896584749222, + 0.6467544436454773, + 0.048036348074674606 + ], + "credit_map": [ + 0.006403310690075159, + 0.002721545286476612, + 0.0037988373078405857, + 0.003756212070584297, + 0.008984788320958614, + 0.005034445319324732, + 0.0035770339891314507, + 0.002699249889701605, + 0.003226633183658123, + 0.002441226039081812, + 0.0032973175402730703, + 0.0024701780639588833, + 0.004644697532057762, + 0.004084690008312464, + 0.00277179223485291, + 0.00304267811588943, + 0.004785139113664627, + 0.0034790027420967817, + 0.0022533524315804243, + 0.0025076651945710182, + 0.00475152675062418, + 0.0037777882535010576, + 0.002194324741140008, + 0.0028637410141527653, + 0.006147565320134163, + 0.008211487904191017, + 0.0046929484233260155, + 0.0046397666446864605, + 0.006765434518456459, + 0.00791110284626484, + 0.004001897294074297, + 0.0061603751964867115, + 0.0018459578277543187, + 0.0015042789746075869, + 0.0008802367374300957, + 0.0007499617640860379, + 0.00193830206990242, + 0.0017272660043090582, + 0.0013610542519018054, + 0.0015517196152359247, + 0.008346280083060265, + 0.008304517716169357, + 0.0017979878466576338, + 0.0017647455679252744, + 0.00432517658919096, + 0.012712654657661915, + 0.001518586534075439, + 0.0024198798928409815, + 0.004294941201806068, + 0.007430644240230322, + 0.002635245444253087, + 0.006955109536647797, + 0.0052475049160420895, + 0.009003564715385437, + 0.0041253529489040375, + 0.0037214818876236677, + 0.009733680635690689, + 0.01099466159939766, + 0.003835038049146533, + 0.003979801200330257, + 0.012311174534261227, + 0.018136853352189064, + 0.0077207861468195915, + 0.004236976616084576, + 0.09089983254671097, + 0.052063196897506714, + 0.021738583222031593, + 0.0257036741822958, + 0.21721050143241882, + 0.15616121888160706, + 0.055976517498493195, + 0.027000999078154564, + 0.008209776133298874, + 0.003881253767758608, + 0.0021283638197928667, + 0.0033918628469109535, + 0.015591342933475971, + 0.006007916759699583, + 0.004621685482561588, + 0.004204150754958391 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 3 + ], + "selected_reward": -0.07018230110406876 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 overhead view of a black ink drop descending onto aged rice paper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the video filmed in incredibly slow motion with the camera held still?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the black ink drop evaporate instantly upon contact with the paper instead of soaking in?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a dense plume of black smoke rise from the paper and form abstract shapes resembling cursive calligraphy?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by warm side light casting soft shadows on the paper and 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": [ + -0.018262337893247604 + ], + "opposite_logprob_tokens": [ + -4.0182623863220215 + ], + "logprob_reward": -0.018262337893247604, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018262337893247604, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2091655135154724, + 0.06850333511829376, + 0.015929535031318665, + 0.023402433842420578, + 0.018473664298653603, + 0.04240844026207924, + 0.05305441468954086, + 0.13896414637565613, + 0.1183047741651535, + 0.3117937445640564 + ], + "credit_map": [ + 0.042469125241041183, + 0.010317159816622734, + 0.011545895598828793, + 0.012813371606171131, + 0.07183751463890076, + 0.02857244946062565, + 0.017491936683654785, + 0.014118055813014507, + 0.011435839347541332, + 0.004509213380515575, + 0.007828987203538418, + 0.004401105921715498, + 0.01103507075458765, + 0.009495981968939304, + 0.009192700497806072, + 0.010604443028569221, + 0.0017829489661380649, + 0.0017227052012458444, + 0.0017053920309990644, + 0.0012882831506431103, + 0.002364876214414835, + 0.0029643664602190256, + 0.0016525228274986148, + 0.0024484407622367144, + 0.0038319271989166737, + 0.003826374188065529, + 0.0018005993915721774, + 0.0022174501791596413, + 0.0026337469462305307, + 0.0053228032775223255, + 0.0015952797839418054, + 0.002174252411350608, + 0.002178210997954011, + 0.0019737426191568375, + 0.0017219135770574212, + 0.0013569107977673411, + 0.0026281606405973434, + 0.003715845290571451, + 0.0020854277536273003, + 0.0028134530875831842, + 0.007484854198992252, + 0.007837471552193165, + 0.002109366934746504, + 0.0020395717583596706, + 0.004827774129807949, + 0.01166443806141615, + 0.0026419786736369133, + 0.0038029872812330723, + 0.010999448597431183, + 0.006400527432560921, + 0.0027147794608026743, + 0.00519781606271863, + 0.0051461076363921165, + 0.016794146969914436, + 0.002571303630247712, + 0.003230290487408638, + 0.031621772795915604, + 0.011195105500519276, + 0.004377208650112152, + 0.009503810666501522, + 0.016894835978746414, + 0.050907500088214874, + 0.007530202157795429, + 0.006933706346899271, + 0.0156373493373394, + 0.0119398208335042, + 0.00647145276889205, + 0.007358412258327007, + 0.03271341696381569, + 0.02201755717396736, + 0.013639416545629501, + 0.008527352474629879, + 0.05307777598500252, + 0.021220475435256958, + 0.013664258643984795, + 0.02036508545279503, + 0.08173348009586334, + 0.04775625467300415, + 0.026582913473248482, + 0.047393497079610825 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 7 + ], + "selected_reward": -0.018262337893247604 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "No", + "Yes", + "Yes" + ] + }, + { + "prompt": "close-up overhead shot, the camera held still, captures a moment brimming with tension in incredibly slow motion: a rich, viscous drop of pure black ink, its perfectly spherical surface reflecting soft, focused light, silently descends onto a sheet of aged rice paper spread flat on a dark wooden tabletop. The paper is a natural beige, its fibrous texture clearly visible in the sidelight, its edges slightly curled with age. Just as the droplet is about to touch the paper, it doesn't soak in or disperse as expected. Instead, it miraculously evaporates upon contact, its liquid form instantly vanishing. Instead, a dense, tangible plume of black smoke curls up from the paper. Unlike ordinary smoke, it doesn't dissipate. Instead, it swirls, twists, and stretches in the air with an elegant and controlled gesture, like a living spirit performing a silent dance, forming complex, abstract forms resembling cursive calligraphy. It then slowly ascends, finally dissipating into the still air. The entire scene is illuminated by a warm side light, casting soft shadows on the ancient paper and smoke, creating an atmosphere of tranquility, concentration and mysterious oriental aesthetics.", + "seed": 6103, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000021/group_7/index_9748_seed_6103_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": -0.5357150167226792, + "vlm_logprob": -0.5357150167226792, + "vlm_yes_no_margin": 2.75, + "vlm_token_credit": -0.5357150167226792, + "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 overhead view of a black ink drop descending onto aged rice paper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the video filmed in incredibly slow motion with the camera held still?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the black ink drop evaporate instantly upon contact with the paper instead of soaking in?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a dense plume of black smoke rise from the paper and form abstract shapes resembling cursive calligraphy?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by warm side light casting soft shadows on the paper and 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": [ + -0.027961483225226402 + ], + "opposite_logprob_tokens": [ + -4.652961254119873 + ], + "logprob_reward": -0.027961483225226402, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.027961483225226402, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.36612817645072937, + 0.17636580765247345, + 0.026049619540572166, + 0.06815563887357712, + 0.04597292095422745, + 0.016882218420505524, + 0.07264477014541626, + 0.09361648559570312, + 0.06840761005878448, + 0.0657767578959465 + ], + "credit_map": [ + 0.12283682823181152, + 0.03448055312037468, + 0.02123427949845791, + 0.14139245450496674, + 0.012742339633405209, + 0.005527566187083721, + 0.01593019999563694, + 0.01198397297412157, + 0.06567379087209702, + 0.059089627116918564, + 0.018006261438131332, + 0.0069437045603990555, + 0.008897405117750168, + 0.005366159602999687, + 0.006077789701521397, + 0.006311069708317518, + 0.004743535071611404, + 0.005272388458251953, + 0.0041600000113248825, + 0.002006220631301403, + 0.004305338021367788, + 0.0017366571119055152, + 0.001884138211607933, + 0.001941341906785965, + 0.01105688326060772, + 0.010156363248825073, + 0.009222684428095818, + 0.012595598585903645, + 0.008467482402920723, + 0.004061414394527674, + 0.0068239690735936165, + 0.005771243944764137, + 0.005710928700864315, + 0.00735874380916357, + 0.005755861289799213, + 0.008044850081205368, + 0.006720284465700388, + 0.00401969812810421, + 0.0033983488101512194, + 0.004964204039424658, + 0.0026276474818587303, + 0.002000365639105439, + 0.0024291900917887688, + 0.00200192560441792, + 0.0015354672214016318, + 0.00135716306976974, + 0.0013180627720430493, + 0.0036123963072896004, + 0.006722562480717897, + 0.011641062796115875, + 0.011859873309731483, + 0.014027518220245838, + 0.011157026514410973, + 0.005446924362331629, + 0.006577975116670132, + 0.005211822688579559, + 0.010779484175145626, + 0.01990436390042305, + 0.012069177813827991, + 0.011533074080944061, + 0.020259197801351547, + 0.005768168717622757, + 0.007215630263090134, + 0.006087391637265682, + 0.0066504767164587975, + 0.01212148554623127, + 0.008780369535088539, + 0.009024584665894508, + 0.017470616847276688, + 0.006203500088304281, + 0.004399890545755625, + 0.0037566812243312597, + 0.008126704953610897, + 0.013623353093862534, + 0.007899017073214054, + 0.007493302691727877, + 0.011922899633646011, + 0.005136124324053526, + 0.0050951577723026276, + 0.0064801969565451145 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 7 + ], + "selected_reward": -0.027961483225226402 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 overhead view of a black ink drop descending onto aged rice paper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the video filmed in incredibly slow motion with the camera held still?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the black ink drop evaporate instantly upon contact with the paper instead of soaking in?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a dense plume of black smoke rise from the paper and form abstract shapes resembling cursive calligraphy?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by warm side light casting soft shadows on the paper and 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": [ + -0.014231733046472073 + ], + "opposite_logprob_tokens": [ + -4.2642316818237305 + ], + "logprob_reward": -0.014231733046472073, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014231733046472073, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.468172162771225, + 0.06342688202857971, + 0.011738816276192665, + 0.043358925729990005, + 0.14166714251041412, + 0.044471096247434616, + 0.04051662236452103, + 0.057816408574581146, + 0.046730298548936844, + 0.08210165798664093 + ], + "credit_map": [ + 0.16419877111911774, + 0.04240565747022629, + 0.02511834353208542, + 0.1849316507577896, + 0.01431171502918005, + 0.00664525618776679, + 0.018646234646439552, + 0.01191448513418436, + 0.0161568783223629, + 0.018924174830317497, + 0.0054468256421387196, + 0.005202762316912413, + 0.006507929414510727, + 0.0021583412308245897, + 0.004898973274976015, + 0.00413099117577076, + 0.0018115552375093102, + 0.001948906690813601, + 0.00228087417781353, + 0.0013204762944951653, + 0.0014095603255555034, + 0.0006914944970048964, + 0.0011016125790774822, + 0.001174336182884872, + 0.005636402405798435, + 0.0071424804627895355, + 0.005363339092582464, + 0.008253918029367924, + 0.005018592346459627, + 0.0026182453148066998, + 0.004780289717018604, + 0.004545654635876417, + 0.020038703456521034, + 0.038381319493055344, + 0.017138095572590828, + 0.02192111127078533, + 0.018237518146634102, + 0.0060894424095749855, + 0.010011356323957443, + 0.009849585592746735, + 0.006393645890057087, + 0.005044534802436829, + 0.006273355800658464, + 0.007016248069703579, + 0.0045622289180755615, + 0.003308108076453209, + 0.004251732956618071, + 0.0076212394051253796, + 0.004801723174750805, + 0.006078843493014574, + 0.005107329227030277, + 0.00505779217928648, + 0.006898828782141209, + 0.0029730868991464376, + 0.0031362909357994795, + 0.006462725345045328, + 0.007990655489265919, + 0.008844689466059208, + 0.0074914125725626945, + 0.005214195232838392, + 0.007693026680499315, + 0.00465071015059948, + 0.0036976858973503113, + 0.012234030291438103, + 0.005878285504877567, + 0.011345570906996727, + 0.006684785708785057, + 0.004711443558335304, + 0.009385106153786182, + 0.0031048317905515432, + 0.002753693610429764, + 0.0028665773570537567, + 0.010027647018432617, + 0.011847849935293198, + 0.010992218740284443, + 0.011464090086519718, + 0.013960521668195724, + 0.006739513482898474, + 0.006431640591472387, + 0.010638177394866943 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 9 + ], + "selected_reward": -0.014231733046472073 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 overhead view of a black ink drop descending onto aged rice paper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the video filmed in incredibly slow motion with the camera held still?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the black ink drop evaporate instantly upon contact with the paper instead of soaking in?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a dense plume of black smoke rise from the paper and form abstract shapes resembling cursive calligraphy?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by warm side light casting soft shadows on the paper and 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": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -2.578974723815918 + ], + "opposite_logprob_tokens": [ + -0.07897467911243439 + ], + "logprob_reward": -2.578974723815918, + "yes_no_margin_reward": -2.5, + "token_credit_reward": -2.578974723815918, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.3563973903656006, + 0.030481934547424316, + 0.02192114293575287, + 0.03657349571585655, + 0.07495228946208954, + 0.01378388050943613, + 0.05495835840702057, + 0.07419455051422119, + 0.12349415570497513, + 0.21324275434017181 + ], + "credit_map": [ + 0.12715518474578857, + 0.028251158073544502, + 0.017464542761445045, + 0.14236997067928314, + 0.008664298802614212, + 0.006414404604583979, + 0.015562336891889572, + 0.010515525005757809, + 0.007414339575916529, + 0.007215183228254318, + 0.0029968293383717537, + 0.003401437308639288, + 0.0032661575824022293, + 0.0014362926594913006, + 0.002585635520517826, + 0.002166059333831072, + 0.002727748826146126, + 0.0033038719557225704, + 0.002615405013784766, + 0.0013807258801534772, + 0.002587668364867568, + 0.0020993477664887905, + 0.0017836993793025613, + 0.005422675050795078, + 0.007826589979231358, + 0.008051582612097263, + 0.004654728341847658, + 0.005267643835395575, + 0.003840650198981166, + 0.0014389174757525325, + 0.0028390707448124886, + 0.0026543138083070517, + 0.011652421206235886, + 0.015069076791405678, + 0.008690410293638706, + 0.01504161674529314, + 0.008991280570626259, + 0.003655498381704092, + 0.007644344121217728, + 0.004207638092339039, + 0.0021349722519516945, + 0.0019068849505856633, + 0.0027547702193260193, + 0.0018925040494650602, + 0.0021037498954683542, + 0.0008358824998140335, + 0.0008999238489195704, + 0.0012551936088129878, + 0.006813573185354471, + 0.010491358116269112, + 0.007808465510606766, + 0.00954382959753275, + 0.007218961138278246, + 0.0029592816717922688, + 0.004585358779877424, + 0.005537534132599831, + 0.009370601736009121, + 0.008506784215569496, + 0.010651099495589733, + 0.007785792928189039, + 0.007893115282058716, + 0.007457382511347532, + 0.005100829992443323, + 0.017428942024707794, + 0.013747104443609715, + 0.027726678177714348, + 0.015192003920674324, + 0.011969550512731075, + 0.03068159706890583, + 0.010821144096553326, + 0.006483828648924828, + 0.0068722437135875225, + 0.02121216617524624, + 0.04446236044168472, + 0.024645371362566948, + 0.020091254264116287, + 0.040451936423778534, + 0.018346892669796944, + 0.01095995120704174, + 0.033072832971811295 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 8 + ], + "selected_reward": -2.578974723815918 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 overhead view of a black ink drop descending onto aged rice paper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the video filmed in incredibly slow motion with the camera held still?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the black ink drop evaporate instantly upon contact with the paper instead of soaking in?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a dense plume of black smoke rise from the paper and form abstract shapes resembling cursive calligraphy?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by warm side light casting soft shadows on the paper and 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": [ + -0.04315930977463722 + ], + "opposite_logprob_tokens": [ + -3.168159246444702 + ], + "logprob_reward": -0.04315930977463722, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04315930977463722, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11926944553852081, + 0.05068933218717575, + 0.022250274196267128, + 0.028667984530329704, + 0.10971137136220932, + 0.01663651131093502, + 0.13302046060562134, + 0.10209334641695023, + 0.15124866366386414, + 0.2664125859737396 + ], + "credit_map": [ + 0.023279421031475067, + 0.011845373548567295, + 0.014029761776328087, + 0.031678203493356705, + 0.010144788771867752, + 0.00520375557243824, + 0.01070895791053772, + 0.012379192747175694, + 0.01042864192277193, + 0.012577002868056297, + 0.005226136650890112, + 0.0048992689698934555, + 0.006617262028157711, + 0.0024010457564145327, + 0.004603528883308172, + 0.00393644068390131, + 0.0041587515734136105, + 0.00318469968624413, + 0.003658905392512679, + 0.0014086728915572166, + 0.0020364299416542053, + 0.0018210392445325851, + 0.0016013552667573094, + 0.004380419850349426, + 0.004638953600078821, + 0.004261679016053677, + 0.004605747759342194, + 0.00410392927005887, + 0.003795660799369216, + 0.001694437232799828, + 0.0026986906304955482, + 0.0028688875027000904, + 0.01638106256723404, + 0.01880466565489769, + 0.01244628056883812, + 0.022164592519402504, + 0.01428922824561596, + 0.006373063661158085, + 0.01129618939012289, + 0.007956290617585182, + 0.0024265446700155735, + 0.0020073289051651955, + 0.0025680901017040014, + 0.002607791917398572, + 0.0015520770102739334, + 0.0013383233454078436, + 0.0015348618617281318, + 0.002601493149995804, + 0.014549903571605682, + 0.025892578065395355, + 0.014679819345474243, + 0.019031520932912827, + 0.022439146414399147, + 0.008040676824748516, + 0.01279676053673029, + 0.015590060502290726, + 0.012127785012125969, + 0.02060827426612377, + 0.010668091475963593, + 0.008222103118896484, + 0.017892777919769287, + 0.007270646281540394, + 0.007618510629981756, + 0.017685161903500557, + 0.01725274883210659, + 0.03273480758070946, + 0.017579710111021996, + 0.013839741237461567, + 0.04001486673951149, + 0.013744594529271126, + 0.00828003790229559, + 0.00780215859413147, + 0.030766472220420837, + 0.053434476256370544, + 0.03271043300628662, + 0.02514669857919216, + 0.04456436634063721, + 0.023379046469926834, + 0.012928434647619724, + 0.043482668697834015 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 6 + ], + "selected_reward": -0.04315930977463722 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 overhead view of a black ink drop descending onto aged rice paper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the video filmed in incredibly slow motion with the camera held still?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the black ink drop evaporate instantly upon contact with the paper instead of soaking in?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a dense plume of black smoke rise from the paper and form abstract shapes resembling cursive calligraphy?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by warm side light casting soft shadows on the paper and 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": [ + -0.014247833751142025 + ], + "opposite_logprob_tokens": [ + -4.264247894287109 + ], + "logprob_reward": -0.014247833751142025, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014247833751142025, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17911529541015625, + 0.06648333370685577, + 0.032948773354291916, + 0.07760928571224213, + 0.3026493489742279, + 0.017356975004076958, + 0.0358881801366806, + 0.05715017765760422, + 0.0852428451180458, + 0.14555571973323822 + ], + "credit_map": [ + 0.03789319843053818, + 0.02739529125392437, + 0.017437027767300606, + 0.04407390207052231, + 0.018181316554546356, + 0.005203830543905497, + 0.019121766090393066, + 0.009808980859816074, + 0.021589607000350952, + 0.019098762422800064, + 0.007226184476166964, + 0.0037353606894612312, + 0.005054611712694168, + 0.0024009891785681248, + 0.0037792236544191837, + 0.0035985985305160284, + 0.005310489796102047, + 0.0038615912199020386, + 0.00432988116517663, + 0.0020877092611044645, + 0.0031449999660253525, + 0.0028421240858733654, + 0.0028532801661640406, + 0.008518701419234276, + 0.011085344478487968, + 0.011489706113934517, + 0.008817619644105434, + 0.011938830837607384, + 0.013367634266614914, + 0.004565307404845953, + 0.009053343906998634, + 0.0072915018536150455, + 0.04226522892713547, + 0.09115980565547943, + 0.03239401429891586, + 0.04337962716817856, + 0.03830548748373985, + 0.01362544484436512, + 0.020082613453269005, + 0.021437160670757294, + 0.0022090652491897345, + 0.002627723850309849, + 0.0027031810022890568, + 0.00261695752851665, + 0.002644944004714489, + 0.0012144158827140927, + 0.0017423259560018778, + 0.0015983630437403917, + 0.0034668368753045797, + 0.005855329800397158, + 0.004547186195850372, + 0.007419008295983076, + 0.005638837814331055, + 0.0026297857984900475, + 0.0032538368832319975, + 0.0030773659236729145, + 0.007191849872469902, + 0.006947642657905817, + 0.007484006695449352, + 0.004130411893129349, + 0.006286391522735357, + 0.005981322377920151, + 0.003500450635328889, + 0.015628106892108917, + 0.009139505214989185, + 0.01773993857204914, + 0.010189627297222614, + 0.009212596341967583, + 0.020485790446400642, + 0.007861253805458546, + 0.004741812124848366, + 0.005872329697012901, + 0.014160508289933205, + 0.030729256570339203, + 0.01719602383673191, + 0.014369381591677666, + 0.026901189237833023, + 0.012236224487423897, + 0.0074259815737605095, + 0.022537169978022575 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 9 + ], + "selected_reward": -0.014247833751142025 + } + ], + "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/0000021/group_8/index_tensor([8546, 8546])_group_8_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000021/group_8/index_tensor([8546, 8546])_group_8_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..217875bc522fe37f7b4fc3d517ba280fc1df53ed --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000021/group_8/index_tensor([8546, 8546])_group_8_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "An extreme close-up, shot at eye level, captures the dramatic moment of life's birth. At the center of the frame is a newborn tree frog, its disproportionately large eyes a genuine, shimmering, jewel-red hue, the steady, intense focal point of the vibrant scene. It struggles with all its might to free itself from a translucent, shimmering, gelatinous eggshell. Its tiny body, half curled up, its moist, emerald green skin dotted with tiny bumps. A slender hind leg has already pierced the membrane, and its head and forelimbs are slowly but surely pushing through the breach. Everything around it, from the broad, veined green leaves in the background to the water droplets reflecting light on them, is not realistic but rather flows feverishly in Van Gogh's signature swirling brushstrokes: intense oils of deep green, yellow-green, and cobalt blue create the dancing leaves, while the water light is stretched and twisted into bright, trembling strokes of white and lemon yellow. A vibrant ray of light penetrates this swirl of color, coating the frog's body and the remnants of its eggshell with a tangible, impasto-like highlight. The entire scene is imbued with the texture of thick oil paint, as if pulsating with a primal, frenetic energy.", + "seed": 29908, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000021/group_8/index_8546_seed_29908_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": -0.013296889420598746, + "vlm_logprob": -0.013296889420598746, + "vlm_yes_no_margin": 4.8, + "vlm_token_credit": -0.013296889420598746, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a newborn tree frog struggling to free itself from a translucent, gelatinous eggshell?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the frog have disproportionately large eyes with a shimmering jewel-red hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the frog's skin moist, emerald green, and dotted with tiny bumps?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background composed of broad, veined green leaves and water droplets rendered in Van Gogh's swirling brushstroke style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the entire scene rendered with the texture of thick oil paint featuring impasto-like highlights and vibrant colors?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014500241726636887 + ], + "opposite_logprob_tokens": [ + -4.389500141143799 + ], + "logprob_reward": -0.014500241726636887, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.014500241726636887, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19395987689495087, + 0.10415489971637726, + 0.053310371935367584, + 0.10423536598682404, + 0.07635664939880371, + 0.09626074135303497, + 0.042645130306482315, + 0.15999671816825867, + 0.09455849230289459, + 0.0745217502117157 + ], + "credit_map": [ + 0.020241040736436844, + 0.023321326822042465, + 0.020596006885170937, + 0.02894458919763565, + 0.020283641293644905, + 0.018694447353482246, + 0.03551042824983597, + 0.026368379592895508, + 0.014569036662578583, + 0.012939889915287495, + 0.008001589216291904, + 0.011832330375909805, + 0.01690402440726757, + 0.012765388004481792, + 0.013507612980902195, + 0.013635030016303062, + 0.005040368065237999, + 0.005616451613605022, + 0.0068345749750733376, + 0.00992960762232542, + 0.004093617666512728, + 0.0062932539731264114, + 0.009668159298598766, + 0.0058343373239040375, + 0.01280110701918602, + 0.01290835440158844, + 0.010786529630422592, + 0.01599113829433918, + 0.013118062168359756, + 0.01309350598603487, + 0.0160219706594944, + 0.009514698758721352, + 0.007552384864538908, + 0.005362526047974825, + 0.007931588217616081, + 0.011246908456087112, + 0.006979906931519508, + 0.007590653840452433, + 0.021049268543720245, + 0.00864341203123331, + 0.01226872205734253, + 0.012956126593053341, + 0.008277887478470802, + 0.010465388186275959, + 0.016469739377498627, + 0.012226182036101818, + 0.009693999774754047, + 0.013902697712182999, + 0.0056773764081299305, + 0.006356327794492245, + 0.0035597882233560085, + 0.00662347674369812, + 0.004588953219354153, + 0.004989842418581247, + 0.005820600315928459, + 0.00502876378595829, + 0.02198433130979538, + 0.018756601959466934, + 0.01870197057723999, + 0.021026138216257095, + 0.01779100112617016, + 0.02031325176358223, + 0.023498833179473877, + 0.0179245974868536, + 0.013643072918057442, + 0.009525461122393608, + 0.01304912380874157, + 0.01302146352827549, + 0.011678572744131088, + 0.011600903235375881, + 0.01216237060725689, + 0.00987752340734005, + 0.009691904298961163, + 0.006697871256619692, + 0.02278590388596058, + 0.004639011342078447, + 0.006458847317844629, + 0.010777979157865047, + 0.0056319707073271275, + 0.00783825945109129 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 3 + ], + "selected_reward": -0.014500241726636887 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a newborn tree frog struggling to free itself from a translucent, gelatinous eggshell?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the frog have disproportionately large eyes with a shimmering jewel-red hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the frog's skin moist, emerald green, and dotted with tiny bumps?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background composed of broad, veined green leaves and water droplets rendered in Van Gogh's swirling brushstroke style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the entire scene rendered with the texture of thick oil paint featuring impasto-like highlights and vibrant colors?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03811497241258621 + ], + "opposite_logprob_tokens": [ + -3.2881150245666504 + ], + "logprob_reward": -0.03811497241258621, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03811497241258621, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21512502431869507, + 0.10257157683372498, + 0.0801548883318901, + 0.14873233437538147, + 0.04305977001786232, + 0.14277926087379456, + 0.060839101672172546, + 0.10098184645175934, + 0.031081004068255424, + 0.07467519491910934 + ], + "credit_map": [ + 0.024103369563817978, + 0.0240167286247015, + 0.025858059525489807, + 0.029252804815769196, + 0.01971173845231533, + 0.022881904616951942, + 0.036385975778102875, + 0.03291444852948189, + 0.014444833621382713, + 0.014075317420065403, + 0.014279095456004143, + 0.008713309653103352, + 0.019872022792696953, + 0.01321661751717329, + 0.00935528427362442, + 0.008615100756287575, + 0.008684257045388222, + 0.0071597620844841, + 0.013054770417511463, + 0.01238766685128212, + 0.007088604383170605, + 0.008942323736846447, + 0.012669318355619907, + 0.010168183594942093, + 0.013497347943484783, + 0.008934229612350464, + 0.017187001183629036, + 0.024647537618875504, + 0.022633228451013565, + 0.016262749210000038, + 0.02704705484211445, + 0.01852317713201046, + 0.005548762157559395, + 0.0045541999861598015, + 0.005067944526672363, + 0.0051473272033035755, + 0.003928798716515303, + 0.004895676858723164, + 0.006519646849483252, + 0.0073974160477519035, + 0.021298488602042198, + 0.023705247789621353, + 0.010717318393290043, + 0.013668407686054707, + 0.02890479750931263, + 0.01947677880525589, + 0.011892986483871937, + 0.01311523001641035, + 0.00864228792488575, + 0.006302692927420139, + 0.006558316759765148, + 0.009001358412206173, + 0.007345984689891338, + 0.007137918379157782, + 0.008677161298692226, + 0.007173381745815277, + 0.014492700807750225, + 0.012537140399217606, + 0.012769348919391632, + 0.013777868822216988, + 0.009366138838231564, + 0.013085673563182354, + 0.016296982765197754, + 0.008655991405248642, + 0.004691917449235916, + 0.0033340754453092813, + 0.0027897909749299288, + 0.0051740724593400955, + 0.0028467283118516207, + 0.0037198769859969616, + 0.0048776837065815926, + 0.003646859899163246, + 0.010180788114666939, + 0.007109423633664846, + 0.015905462205410004, + 0.0072610024362802505, + 0.008523340336978436, + 0.00961040798574686, + 0.008119089528918266, + 0.007965678349137306 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 5 + ], + "selected_reward": -0.03811497241258621 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a newborn tree frog struggling to free itself from a translucent, gelatinous eggshell?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the frog have disproportionately large eyes with a shimmering jewel-red hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the frog's skin moist, emerald green, and dotted with tiny bumps?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background composed of broad, veined green leaves and water droplets rendered in Van Gogh's swirling brushstroke style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the entire scene rendered with the texture of thick oil paint featuring impasto-like highlights and vibrant colors?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0067427014000713825 + ], + "opposite_logprob_tokens": [ + -5.006742477416992 + ], + "logprob_reward": -0.0067427014000713825, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.0067427014000713825, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1996106654405594, + 0.13036863505840302, + 0.057425905019044876, + 0.08141591399908066, + 0.05197659134864807, + 0.17179739475250244, + 0.05274534970521927, + 0.07133545726537704, + 0.07867688685655594, + 0.10464716702699661 + ], + "credit_map": [ + 0.02385369874536991, + 0.02458941750228405, + 0.025777526199817657, + 0.024176131933927536, + 0.01641147956252098, + 0.023334771394729614, + 0.02986225299537182, + 0.031605400145053864, + 0.012631907127797604, + 0.014281358569860458, + 0.012517444789409637, + 0.01624748483300209, + 0.0289468877017498, + 0.015555447898805141, + 0.01538375299423933, + 0.014804362319409847, + 0.0059816776774823666, + 0.005328698083758354, + 0.00868960004299879, + 0.009189722128212452, + 0.007494187913835049, + 0.006754761561751366, + 0.00877973809838295, + 0.005207519046962261, + 0.008217149414122105, + 0.0067404997535049915, + 0.00805396493524313, + 0.015699023380875587, + 0.010138122364878654, + 0.008524266071617603, + 0.014876602217555046, + 0.00916629284620285, + 0.006699915509670973, + 0.007518871687352657, + 0.004657513462007046, + 0.007170452270656824, + 0.00454040989279747, + 0.005448178853839636, + 0.011304609477519989, + 0.004636647645384073, + 0.025847408920526505, + 0.023350898176431656, + 0.013744252733886242, + 0.02031364105641842, + 0.031282488256692886, + 0.022668415680527687, + 0.018389102071523666, + 0.016201192513108253, + 0.0076302471570670605, + 0.004999884404242039, + 0.005687839351594448, + 0.006270798854529858, + 0.00635048933327198, + 0.00694372458383441, + 0.008570212870836258, + 0.0062921601347625256, + 0.01026720181107521, + 0.009588900953531265, + 0.008300683461129665, + 0.008558747358620167, + 0.0064600263722240925, + 0.008895516395568848, + 0.011558371596038342, + 0.00770601537078619, + 0.011642750352621078, + 0.015820367261767387, + 0.006114979740232229, + 0.01094367541372776, + 0.006321524735540152, + 0.011259014718234539, + 0.007369718514382839, + 0.009204861707985401, + 0.01396141480654478, + 0.01154663972556591, + 0.020179079845547676, + 0.012647001072764397, + 0.009917599149048328, + 0.013267211616039276, + 0.009734570048749447, + 0.013393656350672245 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 1 + ], + "selected_reward": -0.0067427014000713825 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a newborn tree frog struggling to free itself from a translucent, gelatinous eggshell?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the frog have disproportionately large eyes with a shimmering jewel-red hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the frog's skin moist, emerald green, and dotted with tiny bumps?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background composed of broad, veined green leaves and water droplets rendered in Van Gogh's swirling brushstroke style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the entire scene rendered with the texture of thick oil paint featuring impasto-like highlights and vibrant colors?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004635896533727646 + ], + "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": [ + 0.1625804454088211, + 0.11100523173809052, + 0.13747556507587433, + 0.07168612629175186, + 0.06352456659078598, + 0.15629954636096954, + 0.026200145483016968, + 0.1670903116464615, + 0.05988959223031998, + 0.044248469173908234 + ], + "credit_map": [ + 0.027873950079083443, + 0.025645194575190544, + 0.017594240605831146, + 0.015011939220130444, + 0.013158434070646763, + 0.022581545636057854, + 0.018781539052724838, + 0.021933602169156075, + 0.018569549545645714, + 0.01718785986304283, + 0.014495314098894596, + 0.009820112027227879, + 0.013180631212890148, + 0.015430879779160023, + 0.01187625341117382, + 0.01044463086873293, + 0.030957136303186417, + 0.02431652881205082, + 0.013106614351272583, + 0.010961408726871014, + 0.01584957167506218, + 0.019909588620066643, + 0.009941513650119305, + 0.012433194555342197, + 0.007025381084531546, + 0.006042092107236385, + 0.00837832409888506, + 0.010483324527740479, + 0.009420267306268215, + 0.007825438864529133, + 0.013931636698544025, + 0.00857966486364603, + 0.011285105720162392, + 0.009388230741024017, + 0.006637325510382652, + 0.00614959979429841, + 0.007093502674251795, + 0.007192917168140411, + 0.008531983941793442, + 0.007245905697345734, + 0.021876275539398193, + 0.01831119693815708, + 0.015007796697318554, + 0.019206998869776726, + 0.02354518137872219, + 0.019640589132905006, + 0.018606344237923622, + 0.020105155184864998, + 0.003760672640055418, + 0.0029296278953552246, + 0.0026006782427430153, + 0.003042062744498253, + 0.0029813835863023996, + 0.0031542403157800436, + 0.002713926136493683, + 0.005017551127821207, + 0.024211188778281212, + 0.023184241726994514, + 0.01604926399886608, + 0.024120384827256203, + 0.0174118559807539, + 0.022223424166440964, + 0.02432907372713089, + 0.015560881234705448, + 0.009369004517793655, + 0.008823349140584469, + 0.006023353897035122, + 0.008137673139572144, + 0.005086322780698538, + 0.008268659934401512, + 0.006250664591789246, + 0.007930567488074303, + 0.007025004830211401, + 0.00461521465331316, + 0.011941883713006973, + 0.003579080803319812, + 0.004295441787689924, + 0.006286544259637594, + 0.002798382192850113, + 0.003706918563693762 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 5 + ], + "selected_reward": -0.004635896533727646 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a newborn tree frog struggling to free itself from a translucent, gelatinous eggshell?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the frog have disproportionately large eyes with a shimmering jewel-red hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the frog's skin moist, emerald green, and dotted with tiny bumps?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background composed of broad, veined green leaves and water droplets rendered in Van Gogh's swirling brushstroke style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the entire scene rendered with the texture of thick oil paint featuring impasto-like highlights and vibrant colors?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0024906350299715996 + ], + "opposite_logprob_tokens": [ + -6.002490520477295 + ], + "logprob_reward": -0.0024906350299715996, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.0024906350299715996, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18136653304100037, + 0.09675505757331848, + 0.046181246638298035, + 0.05365973711013794, + 0.04048634693026543, + 0.2836052179336548, + 0.04301995038986206, + 0.07100563496351242, + 0.05582201108336449, + 0.12809820473194122 + ], + "credit_map": [ + 0.029838619753718376, + 0.026343997567892075, + 0.021025201305747032, + 0.017294492572546005, + 0.015332033857703209, + 0.02505331300199032, + 0.022504888474941254, + 0.02397400327026844, + 0.013826238922774792, + 0.011603566817939281, + 0.01609249971807003, + 0.013114411383867264, + 0.009894215501844883, + 0.01275374460965395, + 0.009726227261126041, + 0.00974416546523571, + 0.005875782575458288, + 0.005530972965061665, + 0.004245970398187637, + 0.008699789643287659, + 0.005154056940227747, + 0.005218783859163523, + 0.0074662319384515285, + 0.00398966483771801, + 0.0082927281036973, + 0.008248679339885712, + 0.007396580185741186, + 0.004423921927809715, + 0.008876650594174862, + 0.006740633398294449, + 0.0052803391590714455, + 0.004400209058076143, + 0.0059427362866699696, + 0.0048385318368673325, + 0.0030620284378528595, + 0.005984071642160416, + 0.0038096210919320583, + 0.004817066248506308, + 0.007330149412155151, + 0.004702148027718067, + 0.04052937403321266, + 0.03811343014240265, + 0.024452267214655876, + 0.03317699953913689, + 0.04805338755249977, + 0.036799993366003036, + 0.03021066077053547, + 0.03226913884282112, + 0.0055460697039961815, + 0.004584426525980234, + 0.0030465975869446993, + 0.00758529594168067, + 0.005116076208651066, + 0.0047904085367918015, + 0.008416221477091312, + 0.0039348602294921875, + 0.012696533463895321, + 0.012217783369123936, + 0.00747096911072731, + 0.00961555726826191, + 0.006320382934063673, + 0.00910940207540989, + 0.008241409435868263, + 0.005333606619387865, + 0.010880539193749428, + 0.007610442116856575, + 0.004671457223594189, + 0.007034460082650185, + 0.006051001138985157, + 0.0067911213263869286, + 0.007391341961920261, + 0.005391653627157211, + 0.02011040784418583, + 0.015192737802863121, + 0.04175180569291115, + 0.009199448861181736, + 0.007767192553728819, + 0.018167315050959587, + 0.005699459929019213, + 0.010209844447672367 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 9 + ], + "selected_reward": -0.0024906350299715996 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An extreme close-up, shot at eye level, captures the dramatic moment of life's birth. At the center of the frame is a newborn tree frog, its disproportionately large eyes a genuine, shimmering, jewel-red hue, the steady, intense focal point of the vibrant scene. It struggles with all its might to free itself from a translucent, shimmering, gelatinous eggshell. Its tiny body, half curled up, its moist, emerald green skin dotted with tiny bumps. A slender hind leg has already pierced the membrane, and its head and forelimbs are slowly but surely pushing through the breach. Everything around it, from the broad, veined green leaves in the background to the water droplets reflecting light on them, is not realistic but rather flows feverishly in Van Gogh's signature swirling brushstrokes: intense oils of deep green, yellow-green, and cobalt blue create the dancing leaves, while the water light is stretched and twisted into bright, trembling strokes of white and lemon yellow. A vibrant ray of light penetrates this swirl of color, coating the frog's body and the remnants of its eggshell with a tangible, impasto-like highlight. The entire scene is imbued with the texture of thick oil paint, as if pulsating with a primal, frenetic energy.", + "seed": 29909, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000021/group_8/index_8546_seed_29909_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": -0.011348706646822393, + "vlm_logprob": -0.011348706646822393, + "vlm_yes_no_margin": 4.975, + "vlm_token_credit": -0.011348706646822393, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a newborn tree frog struggling to free itself from a translucent, gelatinous eggshell?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the frog have disproportionately large eyes with a shimmering jewel-red hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the frog's skin moist, emerald green, and dotted with tiny bumps?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background composed of broad, veined green leaves and water droplets rendered in Van Gogh's swirling brushstroke style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the entire scene rendered with the texture of thick oil paint featuring impasto-like highlights and vibrant colors?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009357285685837269 + ], + "opposite_logprob_tokens": [ + -4.884357452392578 + ], + "logprob_reward": -0.009357285685837269, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.009357285685837269, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21073172986507416, + 0.15559981763362885, + 0.07395724952220917, + 0.0706278458237648, + 0.09480499476194382, + 0.07226710021495819, + 0.051954396069049835, + 0.047114934772253036, + 0.11521563678979874, + 0.10772630572319031 + ], + "credit_map": [ + 0.031343474984169006, + 0.02371179684996605, + 0.027450663968920708, + 0.030564574524760246, + 0.02456318773329258, + 0.015925347805023193, + 0.034497082233428955, + 0.022675611078739166, + 0.024975795298814774, + 0.0433863066136837, + 0.019054673612117767, + 0.02081933431327343, + 0.010070203803479671, + 0.01086905412375927, + 0.016179624944925308, + 0.010244825854897499, + 0.009620235301554203, + 0.008922568522393703, + 0.008525979705154896, + 0.014604835771024227, + 0.006518075242638588, + 0.0074331508949398994, + 0.011680809780955315, + 0.006651606876403093, + 0.012323997914791107, + 0.007454912178218365, + 0.008405201137065887, + 0.008438545279204845, + 0.007177406921982765, + 0.011052270419895649, + 0.006763576064258814, + 0.009011936374008656, + 0.012192140333354473, + 0.00814813282340765, + 0.01150776632130146, + 0.017486657947301865, + 0.011964773759245872, + 0.0077885775826871395, + 0.01671769842505455, + 0.008999249897897243, + 0.009517138823866844, + 0.007032562047243118, + 0.01017492450773716, + 0.01261915173381567, + 0.007473423611372709, + 0.008660933002829552, + 0.0071591273881495, + 0.009629840962588787, + 0.007636161986738443, + 0.005260134115815163, + 0.007085970137268305, + 0.01078615803271532, + 0.0032796249724924564, + 0.005351274739950895, + 0.007281928788870573, + 0.005273144692182541, + 0.00610480597242713, + 0.0036814201157540083, + 0.006841265596449375, + 0.008920581080019474, + 0.0059489100240170956, + 0.00469399755820632, + 0.0069936844520270824, + 0.0039302692748606205, + 0.016539165750145912, + 0.013528247363865376, + 0.013286220841109753, + 0.018896378576755524, + 0.013145911507308483, + 0.009566555730998516, + 0.018973413854837418, + 0.01127974409610033, + 0.015149986371397972, + 0.012371107004582882, + 0.016133366152644157, + 0.014349216595292091, + 0.010753489099442959, + 0.012153265066444874, + 0.008168301545083523, + 0.018647575750947 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 8 + ], + "selected_reward": -0.009357285685837269 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a newborn tree frog struggling to free itself from a translucent, gelatinous eggshell?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the frog have disproportionately large eyes with a shimmering jewel-red hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the frog's skin moist, emerald green, and dotted with tiny bumps?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background composed of broad, veined green leaves and water droplets rendered in Van Gogh's swirling brushstroke style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the entire scene rendered with the texture of thick oil paint featuring impasto-like highlights and vibrant colors?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03371596708893776 + ], + "opposite_logprob_tokens": [ + -3.4087159633636475 + ], + "logprob_reward": -0.03371596708893776, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03371596708893776, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24040354788303375, + 0.13669873774051666, + 0.04366346448659897, + 0.04652535170316696, + 0.15351131558418274, + 0.18558001518249512, + 0.06948918849229813, + 0.025993527844548225, + 0.04283543676137924, + 0.055299416184425354 + ], + "credit_map": [ + 0.03497191518545151, + 0.02260538563132286, + 0.03874204680323601, + 0.02907111868262291, + 0.0307710450142622, + 0.022265303879976273, + 0.03934908285737038, + 0.022627651691436768, + 0.019401736557483673, + 0.02424115687608719, + 0.02290898561477661, + 0.02217330038547516, + 0.010051296092569828, + 0.010063881054520607, + 0.01679045706987381, + 0.011067934334278107, + 0.00558315496891737, + 0.005117790307849646, + 0.004846708383411169, + 0.009670809842646122, + 0.0034993456210941076, + 0.0040857079438865185, + 0.007332649547606707, + 0.0035273004323244095, + 0.006592731922864914, + 0.009254313074052334, + 0.005474886391311884, + 0.005880729760974646, + 0.005173270590603352, + 0.0038623821455985308, + 0.00644473684951663, + 0.003842300735414028, + 0.019670970737934113, + 0.013901046477258205, + 0.01706692762672901, + 0.030465569347143173, + 0.014505945146083832, + 0.012560579925775528, + 0.031378284096717834, + 0.013961990363895893, + 0.027562296018004417, + 0.019597716629505157, + 0.024301500990986824, + 0.03975397348403931, + 0.018414262682199478, + 0.016225822269916534, + 0.021276723593473434, + 0.01844772882759571, + 0.009381592273712158, + 0.00855683721601963, + 0.009192814119160175, + 0.012624965980648994, + 0.005836407653987408, + 0.00860450230538845, + 0.007258288562297821, + 0.008033779449760914, + 0.003230242058634758, + 0.002265451243147254, + 0.00510141858831048, + 0.004772539250552654, + 0.0033687662798911333, + 0.0025334067177027464, + 0.002929891226813197, + 0.001791813294403255, + 0.005064394325017929, + 0.0048455484211444855, + 0.0058534531854093075, + 0.008287186734378338, + 0.004497972782701254, + 0.004290688782930374, + 0.004948807414621115, + 0.005047383718192577, + 0.007922890596091747, + 0.0076582226902246475, + 0.007717088796198368, + 0.00564843975007534, + 0.005822340026497841, + 0.007839514873921871, + 0.004150724504143, + 0.008540195412933826 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 4 + ], + "selected_reward": -0.03371596708893776 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a newborn tree frog struggling to free itself from a translucent, gelatinous eggshell?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the frog have disproportionately large eyes with a shimmering jewel-red hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the frog's skin moist, emerald green, and dotted with tiny bumps?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background composed of broad, veined green leaves and water droplets rendered in Van Gogh's swirling brushstroke style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the entire scene rendered with the texture of thick oil paint featuring impasto-like highlights and vibrant colors?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00763652054592967 + ], + "opposite_logprob_tokens": [ + -4.882636547088623 + ], + "logprob_reward": -0.00763652054592967, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.00763652054592967, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22792643308639526, + 0.10160458832979202, + 0.06523388624191284, + 0.11493682861328125, + 0.06874813884496689, + 0.15296947956085205, + 0.07387525588274002, + 0.05482376366853714, + 0.08890856057405472, + 0.050973061472177505 + ], + "credit_map": [ + 0.031786900013685226, + 0.02543000318109989, + 0.022617578506469727, + 0.037231169641017914, + 0.02657264657318592, + 0.022882897406816483, + 0.0386839397251606, + 0.0227213017642498, + 0.01448989287018776, + 0.013701754622161388, + 0.016078831627964973, + 0.017934808507561684, + 0.007408441975712776, + 0.009042680263519287, + 0.014834697358310223, + 0.008113473653793335, + 0.008971866220235825, + 0.008219190873205662, + 0.006125196814537048, + 0.01277163252234459, + 0.005951509345322847, + 0.007168257609009743, + 0.009686472825706005, + 0.006339757703244686, + 0.018441805616021156, + 0.014429209753870964, + 0.01761683262884617, + 0.019304389134049416, + 0.00872446596622467, + 0.011789198964834213, + 0.01412940677255392, + 0.010501516051590443, + 0.007680834736675024, + 0.005232300143688917, + 0.01025561522692442, + 0.015912683680653572, + 0.006450147833675146, + 0.005011347122490406, + 0.012339934706687927, + 0.005865280982106924, + 0.026897011324763298, + 0.014415065757930279, + 0.022720403969287872, + 0.025253798812627792, + 0.01449340209364891, + 0.013341948390007019, + 0.016310501843690872, + 0.01953735388815403, + 0.0112669812515378, + 0.006663859821856022, + 0.010617516934871674, + 0.01608520746231079, + 0.004745566286146641, + 0.006632231175899506, + 0.010967807844281197, + 0.006896079983562231, + 0.007716410327702761, + 0.006123166996985674, + 0.00931174773722887, + 0.008004112169146538, + 0.0049101198092103004, + 0.00681379996240139, + 0.006189573556184769, + 0.005754831247031689, + 0.011866804212331772, + 0.012063047848641872, + 0.013812010176479816, + 0.009343599900603294, + 0.01158085186034441, + 0.01062261313199997, + 0.009740573354065418, + 0.00987906102091074, + 0.006300490349531174, + 0.004468859639018774, + 0.005557193420827389, + 0.0073432656936347485, + 0.007770072668790817, + 0.006712458096444607, + 0.0060401554219424725, + 0.006780562456697226 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 3 + ], + "selected_reward": -0.00763652054592967 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a newborn tree frog struggling to free itself from a translucent, gelatinous eggshell?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the frog have disproportionately large eyes with a shimmering jewel-red hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the frog's skin moist, emerald green, and dotted with tiny bumps?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background composed of broad, veined green leaves and water droplets rendered in Van Gogh's swirling brushstroke style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the entire scene rendered with the texture of thick oil paint featuring impasto-like highlights and vibrant colors?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004092176910489798 + ], + "opposite_logprob_tokens": [ + -5.504092216491699 + ], + "logprob_reward": -0.004092176910489798, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004092176910489798, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23248030245304108, + 0.08383672684431076, + 0.032226067036390305, + 0.15104372799396515, + 0.08629036694765091, + 0.08074444532394409, + 0.059822164475917816, + 0.020920606330037117, + 0.1032918244600296, + 0.14934378862380981 + ], + "credit_map": [ + 0.036092620342969894, + 0.02642986737191677, + 0.028694206848740578, + 0.027307190001010895, + 0.029821250587701797, + 0.023781921714544296, + 0.032470740377902985, + 0.027882495895028114, + 0.011926616542041302, + 0.018170880153775215, + 0.012984735891222954, + 0.008843856863677502, + 0.009471099823713303, + 0.008858062326908112, + 0.0069779036566615105, + 0.006603566464036703, + 0.004490599501878023, + 0.005120426882058382, + 0.0037987150717526674, + 0.0037443581968545914, + 0.003468627342954278, + 0.005131774581968784, + 0.003719120752066374, + 0.002752444241195917, + 0.022177960723638535, + 0.016780827194452286, + 0.013015181757509708, + 0.017433950677514076, + 0.023537179455161095, + 0.02457609958946705, + 0.015764283016324043, + 0.017758233472704887, + 0.012118786573410034, + 0.009863420389592648, + 0.011172334663569927, + 0.009841891005635262, + 0.01322445459663868, + 0.010747784748673439, + 0.007794053293764591, + 0.011527635157108307, + 0.012865342199802399, + 0.009253251366317272, + 0.011528109200298786, + 0.011526515707373619, + 0.00883540790528059, + 0.008243322372436523, + 0.00786538515239954, + 0.010627109557390213, + 0.008389847353100777, + 0.006934232544153929, + 0.009248641319572926, + 0.011689849197864532, + 0.004560815170407295, + 0.007485460489988327, + 0.005323898978531361, + 0.00618941942229867, + 0.0028391308151185513, + 0.002184509066864848, + 0.0029277692083269358, + 0.0025527195539325476, + 0.003023726399987936, + 0.0030476516112685204, + 0.0024284818209707737, + 0.001916615990921855, + 0.014078848995268345, + 0.011454909108579159, + 0.011947780847549438, + 0.018204450607299805, + 0.008343466557562351, + 0.007462349254637957, + 0.020986588671803474, + 0.010813425295054913, + 0.020368222147226334, + 0.02083921618759632, + 0.021187977865338326, + 0.010399487800896168, + 0.018463779240846634, + 0.017613649368286133, + 0.012020993046462536, + 0.028450453653931618 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 9 + ], + "selected_reward": -0.004092176910489798 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a newborn tree frog struggling to free itself from a translucent, gelatinous eggshell?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the frog have disproportionately large eyes with a shimmering jewel-red hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the frog's skin moist, emerald green, and dotted with tiny bumps?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background composed of broad, veined green leaves and water droplets rendered in Van Gogh's swirling brushstroke style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the entire scene rendered with the texture of thick oil paint featuring impasto-like highlights and vibrant colors?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0019415830029174685 + ], + "opposite_logprob_tokens": [ + -6.251941680908203 + ], + "logprob_reward": -0.0019415830029174685, + "yes_no_margin_reward": 6.25, + "token_credit_reward": -0.0019415830029174685, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23093779385089874, + 0.08000071346759796, + 0.029592573642730713, + 0.05181322991847992, + 0.1076396107673645, + 0.18090710043907166, + 0.06068255007266998, + 0.028921527788043022, + 0.07561299204826355, + 0.15389196574687958 + ], + "credit_map": [ + 0.0321861207485199, + 0.025280434638261795, + 0.0398409441113472, + 0.026770731434226036, + 0.025737086310982704, + 0.019426941871643066, + 0.03885924443602562, + 0.02283628284931183, + 0.010316281579434872, + 0.010881900787353516, + 0.011806534603238106, + 0.009570187889039516, + 0.008405160158872604, + 0.010637406259775162, + 0.00971486046910286, + 0.00866838451474905, + 0.003793593728914857, + 0.004339213017374277, + 0.0030251729767769575, + 0.0035961971152573824, + 0.002861360553652048, + 0.004795924294739962, + 0.004604363348335028, + 0.0025767493061721325, + 0.008348225615918636, + 0.005915021989494562, + 0.006980317644774914, + 0.006436144933104515, + 0.006071509327739477, + 0.005607529543340206, + 0.006716402713209391, + 0.005738077685236931, + 0.015258034691214561, + 0.010071934200823307, + 0.01689928211271763, + 0.012768533080816269, + 0.011550873517990112, + 0.009823753498494625, + 0.02062373422086239, + 0.010643462650477886, + 0.030676638707518578, + 0.020730197429656982, + 0.02889177016913891, + 0.024182971566915512, + 0.019943322986364365, + 0.015819745138287544, + 0.020765993744134903, + 0.019896458834409714, + 0.008134054951369762, + 0.0064448462799191475, + 0.008895847015082836, + 0.013747293502092361, + 0.004628640599548817, + 0.007249909918755293, + 0.005785189103335142, + 0.005796767305582762, + 0.0037869566585868597, + 0.0026957441587001085, + 0.004481614101678133, + 0.003876527538523078, + 0.0026491633616387844, + 0.004134288989007473, + 0.004052247852087021, + 0.003244982333853841, + 0.0096746189519763, + 0.008733765222132206, + 0.010101612657308578, + 0.009549234062433243, + 0.010092194192111492, + 0.008656240068376064, + 0.009127230383455753, + 0.009678094647824764, + 0.02114204317331314, + 0.021163370460271835, + 0.019227420911192894, + 0.01067317184060812, + 0.02058270014822483, + 0.020462965592741966, + 0.012830142863094807, + 0.027810147032141685 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 9 + ], + "selected_reward": -0.0019415830029174685 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An extreme close-up, shot at eye level, captures the dramatic moment of life's birth. At the center of the frame is a newborn tree frog, its disproportionately large eyes a genuine, shimmering, jewel-red hue, the steady, intense focal point of the vibrant scene. It struggles with all its might to free itself from a translucent, shimmering, gelatinous eggshell. Its tiny body, half curled up, its moist, emerald green skin dotted with tiny bumps. A slender hind leg has already pierced the membrane, and its head and forelimbs are slowly but surely pushing through the breach. Everything around it, from the broad, veined green leaves in the background to the water droplets reflecting light on them, is not realistic but rather flows feverishly in Van Gogh's signature swirling brushstrokes: intense oils of deep green, yellow-green, and cobalt blue create the dancing leaves, while the water light is stretched and twisted into bright, trembling strokes of white and lemon yellow. A vibrant ray of light penetrates this swirl of color, coating the frog's body and the remnants of its eggshell with a tangible, impasto-like highlight. The entire scene is imbued with the texture of thick oil paint, as if pulsating with a primal, frenetic energy.", + "seed": 29910, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000021/group_8/index_8546_seed_29910_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": -0.00986581842880696, + "vlm_logprob": -0.00986581842880696, + "vlm_yes_no_margin": 5.225, + "vlm_token_credit": -0.00986581842880696, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a newborn tree frog struggling to free itself from a translucent, gelatinous eggshell?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the frog have disproportionately large eyes with a shimmering jewel-red hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the frog's skin moist, emerald green, and dotted with tiny bumps?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background composed of broad, veined green leaves and water droplets rendered in Van Gogh's swirling brushstroke style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the entire scene rendered with the texture of thick oil paint featuring impasto-like highlights and vibrant colors?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.010714031755924225 + ], + "opposite_logprob_tokens": [ + -4.635714054107666 + ], + "logprob_reward": -0.010714031755924225, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.010714031755924225, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07725082337856293, + 0.19337016344070435, + 0.23351076245307922, + 0.10152552276849747, + 0.07738038897514343, + 0.15695615112781525, + 0.05640527606010437, + 0.03270972892642021, + 0.04953964799642563, + 0.021351579576730728 + ], + "credit_map": [ + 0.011081000789999962, + 0.011490259319543839, + 0.009076624177396297, + 0.009522633627057076, + 0.006429185159504414, + 0.008803236298263073, + 0.009701088070869446, + 0.01114678755402565, + 0.07373933494091034, + 0.015386516228318214, + 0.010485241189599037, + 0.05295536667108536, + 0.011736047454178333, + 0.009299044497311115, + 0.011473850347101688, + 0.008294755592942238, + 0.06631176918745041, + 0.01909375749528408, + 0.01670212857425213, + 0.04823852702975273, + 0.03230740502476692, + 0.014566057361662388, + 0.02417229861021042, + 0.012118810787796974, + 0.020055046305060387, + 0.0072234501130878925, + 0.007283991202712059, + 0.013245304115116596, + 0.0195467509329319, + 0.0071582309901714325, + 0.01894518919289112, + 0.008067554794251919, + 0.008499222807586193, + 0.010920843109488487, + 0.007672804407775402, + 0.008136225864291191, + 0.013496424071490765, + 0.007743553724139929, + 0.015908341854810715, + 0.005002971738576889, + 0.02589411847293377, + 0.013068441301584244, + 0.01615598425269127, + 0.03648895025253296, + 0.01488808449357748, + 0.012567187659442425, + 0.027256634086370468, + 0.010636735707521439, + 0.006158856209367514, + 0.007991465739905834, + 0.006849987432360649, + 0.007951481267809868, + 0.006015739403665066, + 0.005141907837241888, + 0.011766878888010979, + 0.004528955090790987, + 0.003375693690031767, + 0.0036440303083509207, + 0.003485135966911912, + 0.005327890161424875, + 0.005332268308848143, + 0.002514633582904935, + 0.006101129110902548, + 0.002928946865722537, + 0.006503959186375141, + 0.00759296165779233, + 0.004973447881639004, + 0.007365251891314983, + 0.0053399852477014065, + 0.004003320354968309, + 0.009262247942388058, + 0.00449847336858511, + 0.0034239620435982943, + 0.0024851406924426556, + 0.0023836526088416576, + 0.004190024454146624, + 0.002017224906012416, + 0.0012827467871829867, + 0.0037571412976831198, + 0.0018116876017302275 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 1, + 5 + ], + "selected_reward": -0.010714031755924225 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a newborn tree frog struggling to free itself from a translucent, gelatinous eggshell?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the frog have disproportionately large eyes with a shimmering jewel-red hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the frog's skin moist, emerald green, and dotted with tiny bumps?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background composed of broad, veined green leaves and water droplets rendered in Van Gogh's swirling brushstroke style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the entire scene rendered with the texture of thick oil paint featuring impasto-like highlights and vibrant colors?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02981683425605297 + ], + "opposite_logprob_tokens": [ + -3.5298168659210205 + ], + "logprob_reward": -0.02981683425605297, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.02981683425605297, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12055204063653946, + 0.1638806313276291, + 0.5152702927589417, + 0.07568828016519547, + 0.03829549625515938, + 0.013147336430847645, + 0.016443615779280663, + 0.02874886617064476, + 0.015830937772989273, + 0.012142523191869259 + ], + "credit_map": [ + 0.020353753119707108, + 0.018596122041344643, + 0.013713867403566837, + 0.014703390188515186, + 0.009861962869763374, + 0.013744216412305832, + 0.013612236827611923, + 0.015966491773724556, + 0.030538080260157585, + 0.013179978355765343, + 0.0076699769124388695, + 0.03819403424859047, + 0.023743081837892532, + 0.00960997212678194, + 0.030645718798041344, + 0.010299787856638432, + 0.06291677802801132, + 0.06664453446865082, + 0.04199117422103882, + 0.08147673308849335, + 0.11755956709384918, + 0.04857337102293968, + 0.060865264385938644, + 0.03524281829595566, + 0.015162888914346695, + 0.004789629951119423, + 0.004012943711131811, + 0.011585631407797337, + 0.013013484887778759, + 0.0047174966894090176, + 0.016230953857302666, + 0.00617525028064847, + 0.007583657279610634, + 0.003239880781620741, + 0.002803277689963579, + 0.009439711458981037, + 0.005087338853627443, + 0.002582537941634655, + 0.005033273249864578, + 0.002525818534195423, + 0.0015896601835265756, + 0.001182837295345962, + 0.0010993188479915261, + 0.0014165635220706463, + 0.0016698786057531834, + 0.0010950823780149221, + 0.0037952414713799953, + 0.0012987545924261212, + 0.0026977085508406162, + 0.0030152604449540377, + 0.0016200931277126074, + 0.0017938428791239858, + 0.002118912059813738, + 0.0014012529281899333, + 0.0026367369573563337, + 0.0011598077835515141, + 0.004730000626295805, + 0.0028111282736063004, + 0.0028151143342256546, + 0.006897086277604103, + 0.003161970293149352, + 0.0019492959836497903, + 0.004410781431943178, + 0.0019734883680939674, + 0.00227695656940341, + 0.001847037230618298, + 0.0015595590230077505, + 0.0029376985039561987, + 0.0017560033593326807, + 0.001065594144165516, + 0.003206613939255476, + 0.0011814753524959087, + 0.002654800657182932, + 0.0010145206470042467, + 0.00118508271407336, + 0.0032651845831424, + 0.0010368910152465105, + 0.0008746988605707884, + 0.0012448406778275967, + 0.0008665031637065113 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 1, + 0 + ], + "selected_reward": -0.02981683425605297 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a newborn tree frog struggling to free itself from a translucent, gelatinous eggshell?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the frog have disproportionately large eyes with a shimmering jewel-red hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the frog's skin moist, emerald green, and dotted with tiny bumps?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background composed of broad, veined green leaves and water droplets rendered in Van Gogh's swirling brushstroke style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the entire scene rendered with the texture of thick oil paint featuring impasto-like highlights and vibrant colors?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005262095481157303 + ], + "opposite_logprob_tokens": [ + -5.2552618980407715 + ], + "logprob_reward": -0.005262095481157303, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005262095481157303, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1133832111954689, + 0.2370377779006958, + 0.25344282388687134, + 0.10254549980163574, + 0.09444840252399445, + 0.058739516884088516, + 0.06161286681890488, + 0.04922031983733177, + 0.01261215191334486, + 0.016957497224211693 + ], + "credit_map": [ + 0.017598995938897133, + 0.019124465063214302, + 0.013434682041406631, + 0.01350560411810875, + 0.009068642742931843, + 0.013489463366568089, + 0.014165572822093964, + 0.012995782308280468, + 0.08858756721019745, + 0.018357427790760994, + 0.01166482176631689, + 0.07437262684106827, + 0.008212430402636528, + 0.010637441650032997, + 0.01747879385948181, + 0.007726668380200863, + 0.034211158752441406, + 0.02866818569600582, + 0.01592942327260971, + 0.06870588660240173, + 0.05179951712489128, + 0.013870303519070148, + 0.027082037180662155, + 0.013176308013498783, + 0.009903178550302982, + 0.006650757510215044, + 0.003329782048240304, + 0.01707451418042183, + 0.04362471029162407, + 0.006218865979462862, + 0.009311829693615437, + 0.006431862246245146, + 0.019504301249980927, + 0.0104881776496768, + 0.0073527991771698, + 0.02441057749092579, + 0.005914031527936459, + 0.007105433847755194, + 0.01252648327499628, + 0.007146598305553198, + 0.010157523676753044, + 0.00516946567222476, + 0.0061884974129498005, + 0.013677755370736122, + 0.004926441702991724, + 0.004954291507601738, + 0.009212045930325985, + 0.004453490953892469, + 0.006824364420026541, + 0.009884419851005077, + 0.006943258456885815, + 0.00862952508032322, + 0.006794011685997248, + 0.006106812506914139, + 0.01220614742487669, + 0.004224327392876148, + 0.009861456230282784, + 0.004348930902779102, + 0.004221781622618437, + 0.012634193524718285, + 0.003884420497342944, + 0.0032433071173727512, + 0.007357587572187185, + 0.0036686426028609276, + 0.0015388785395771265, + 0.002213916275650263, + 0.0014668434159830213, + 0.0013512435834854841, + 0.002212012652307749, + 0.001147401169873774, + 0.0017643626779317856, + 0.0009174938313663006, + 0.003321858122944832, + 0.0015122457407414913, + 0.0015138298040255904, + 0.0039426651783287525, + 0.0020609849598258734, + 0.0009264214313589036, + 0.002515621017664671, + 0.0011638703290373087 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 1, + 0 + ], + "selected_reward": -0.005262095481157303 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a newborn tree frog struggling to free itself from a translucent, gelatinous eggshell?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the frog have disproportionately large eyes with a shimmering jewel-red hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the frog's skin moist, emerald green, and dotted with tiny bumps?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background composed of broad, veined green leaves and water droplets rendered in Van Gogh's swirling brushstroke style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the entire scene rendered with the texture of thick oil paint featuring impasto-like highlights and vibrant colors?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002197353169322014 + ], + "opposite_logprob_tokens": [ + -6.127197265625 + ], + "logprob_reward": -0.002197353169322014, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.002197353169322014, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13501381874084473, + 0.20470532774925232, + 0.28761106729507446, + 0.050984710454940796, + 0.0596877746284008, + 0.05189063772559166, + 0.07385914772748947, + 0.02066998928785324, + 0.09801826626062393, + 0.017559248954057693 + ], + "credit_map": [ + 0.024818474426865578, + 0.02122916467487812, + 0.01554248295724392, + 0.015993865206837654, + 0.009994388557970524, + 0.015440640971064568, + 0.013463344424962997, + 0.018531465902924538, + 0.07349283993244171, + 0.01651429571211338, + 0.010166199877858162, + 0.04067891463637352, + 0.012375163845717907, + 0.030354894697666168, + 0.01132733840495348, + 0.009795683436095715, + 0.05920391157269478, + 0.02926819585263729, + 0.015408358536660671, + 0.08103539794683456, + 0.047233354300260544, + 0.01741516776382923, + 0.025226155295968056, + 0.012820567935705185, + 0.005480729974806309, + 0.0034196439664810896, + 0.0045677609741687775, + 0.008542899042367935, + 0.017157914116978645, + 0.0028352318331599236, + 0.005556041374802589, + 0.003424491034820676, + 0.009863309562206268, + 0.006339084357023239, + 0.005775329191237688, + 0.009949522092938423, + 0.007329365238547325, + 0.0053329733200371265, + 0.010434982366859913, + 0.004663209896534681, + 0.00823328085243702, + 0.006823423784226179, + 0.006414470728486776, + 0.009658433496952057, + 0.0055227139964699745, + 0.004280314780771732, + 0.007129155099391937, + 0.0038288459181785583, + 0.010198739357292652, + 0.010294269770383835, + 0.007349347695708275, + 0.014764616265892982, + 0.006958432029932737, + 0.0067366366274654865, + 0.012969430536031723, + 0.0045876773074269295, + 0.0026905068662017584, + 0.0019190422026440501, + 0.0024596068542450666, + 0.0028400816954672337, + 0.003176473779603839, + 0.0021017121616750956, + 0.0031411098316311836, + 0.002341456711292267, + 0.027443982660770416, + 0.00986209511756897, + 0.006021232344210148, + 0.027990732342004776, + 0.005454213824123144, + 0.006150173023343086, + 0.006297104526311159, + 0.008798729628324509, + 0.003742899978533387, + 0.00202595884911716, + 0.001718171639367938, + 0.004582597874104977, + 0.001359941903501749, + 0.001024896977469325, + 0.001778668025508523, + 0.0013261144049465656 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 1, + 0 + ], + "selected_reward": -0.002197353169322014 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a newborn tree frog struggling to free itself from a translucent, gelatinous eggshell?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the frog have disproportionately large eyes with a shimmering jewel-red hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the frog's skin moist, emerald green, and dotted with tiny bumps?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background composed of broad, veined green leaves and water droplets rendered in Van Gogh's swirling brushstroke style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the entire scene rendered with the texture of thick oil paint featuring impasto-like highlights and vibrant colors?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0013387774815782905 + ], + "opposite_logprob_tokens": [ + -6.626338958740234 + ], + "logprob_reward": -0.0013387774815782905, + "yes_no_margin_reward": 6.625, + "token_credit_reward": -0.0013387774815782905, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2609265446662903, + 0.07102765887975693, + 0.1515616774559021, + 0.03630645573139191, + 0.04344106465578079, + 0.14652635157108307, + 0.10989398509263992, + 0.04126045107841492, + 0.10000457614660263, + 0.039051126688718796 + ], + "credit_map": [ + 0.06324461102485657, + 0.039359815418720245, + 0.027074813842773438, + 0.027721986174583435, + 0.01857249066233635, + 0.02477082423865795, + 0.026779532432556152, + 0.033402517437934875, + 0.012422301806509495, + 0.00789455696940422, + 0.004527002573013306, + 0.023241685703396797, + 0.00483108963817358, + 0.007661144249141216, + 0.005611739587038755, + 0.004838148131966591, + 0.02093101665377617, + 0.01873229630291462, + 0.010807354934513569, + 0.041317954659461975, + 0.03095068968832493, + 0.007844550535082817, + 0.014203650876879692, + 0.006774189881980419, + 0.005318695213645697, + 0.004406491294503212, + 0.0035994623322039843, + 0.006903499830514193, + 0.0042154621332883835, + 0.0030155193526297808, + 0.006335858721286058, + 0.002511471277102828, + 0.006687003653496504, + 0.0038573541678488255, + 0.00440210709348321, + 0.009191677905619144, + 0.004603378474712372, + 0.004088486544787884, + 0.007262247148901224, + 0.003348813159391284, + 0.03088243119418621, + 0.011929154396057129, + 0.012068016454577446, + 0.041641831398010254, + 0.009961077943444252, + 0.010634937323629856, + 0.019571596756577492, + 0.009837337769567966, + 0.01202322356402874, + 0.015301886945962906, + 0.01274812500923872, + 0.016897745430469513, + 0.010894009843468666, + 0.010060813277959824, + 0.023916829377412796, + 0.008051375858485699, + 0.005013398826122284, + 0.004157077055424452, + 0.004395825322717428, + 0.008350754156708717, + 0.004837582353502512, + 0.003069535596296191, + 0.008255361579358578, + 0.0031809222418814898, + 0.02125106006860733, + 0.012984754517674446, + 0.008174160495400429, + 0.02577045187354088, + 0.007038629148155451, + 0.006658558733761311, + 0.011424240656197071, + 0.006702729035168886, + 0.01026149932295084, + 0.004000924993306398, + 0.003491306910291314, + 0.010138874873518944, + 0.002907133661210537, + 0.002213558880612254, + 0.0033616237342357635, + 0.0026762105990201235 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 5 + ], + "selected_reward": -0.0013387774815782905 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An extreme close-up, shot at eye level, captures the dramatic moment of life's birth. At the center of the frame is a newborn tree frog, its disproportionately large eyes a genuine, shimmering, jewel-red hue, the steady, intense focal point of the vibrant scene. It struggles with all its might to free itself from a translucent, shimmering, gelatinous eggshell. Its tiny body, half curled up, its moist, emerald green skin dotted with tiny bumps. A slender hind leg has already pierced the membrane, and its head and forelimbs are slowly but surely pushing through the breach. Everything around it, from the broad, veined green leaves in the background to the water droplets reflecting light on them, is not realistic but rather flows feverishly in Van Gogh's signature swirling brushstrokes: intense oils of deep green, yellow-green, and cobalt blue create the dancing leaves, while the water light is stretched and twisted into bright, trembling strokes of white and lemon yellow. A vibrant ray of light penetrates this swirl of color, coating the frog's body and the remnants of its eggshell with a tangible, impasto-like highlight. The entire scene is imbued with the texture of thick oil paint, as if pulsating with a primal, frenetic energy.", + "seed": 29911, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000021/group_8/index_8546_seed_29911_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": -0.011589890765026211, + "vlm_logprob": -0.011589890765026211, + "vlm_yes_no_margin": 4.95, + "vlm_token_credit": -0.011589890765026211, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a newborn tree frog struggling to free itself from a translucent, gelatinous eggshell?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the frog have disproportionately large eyes with a shimmering jewel-red hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the frog's skin moist, emerald green, and dotted with tiny bumps?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background composed of broad, veined green leaves and water droplets rendered in Van Gogh's swirling brushstroke style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the entire scene rendered with the texture of thick oil paint featuring impasto-like highlights and vibrant colors?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009837242774665356 + ], + "opposite_logprob_tokens": [ + -4.8848371505737305 + ], + "logprob_reward": -0.009837242774665356, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.009837242774665356, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16379930078983307, + 0.0640212744474411, + 0.05008821189403534, + 0.06953507661819458, + 0.3533337116241455, + 0.049126312136650085, + 0.04207373782992363, + 0.029851125553250313, + 0.04813149571418762, + 0.13003972172737122 + ], + "credit_map": [ + 0.018054313957691193, + 0.018825005739927292, + 0.014198612421751022, + 0.019402405247092247, + 0.028497567400336266, + 0.01753491722047329, + 0.025914901867508888, + 0.021371610462665558, + 0.007111771963536739, + 0.006387449335306883, + 0.009515040554106236, + 0.006193313281983137, + 0.010780044831335545, + 0.009177855215966702, + 0.009124008007347584, + 0.005731790326535702, + 0.00412182928994298, + 0.004935265053063631, + 0.005048531573265791, + 0.00586533360183239, + 0.007171287201344967, + 0.005528775043785572, + 0.011016612872481346, + 0.006400584243237972, + 0.005346531979739666, + 0.006709631532430649, + 0.005252808332443237, + 0.008822380565106869, + 0.014060295186936855, + 0.01210611779242754, + 0.011170954443514347, + 0.006066369358450174, + 0.015350022353231907, + 0.01933392509818077, + 0.01654650829732418, + 0.011663218960165977, + 0.1497749388217926, + 0.1007017195224762, + 0.016647405922412872, + 0.023316001519560814, + 0.005287715699523687, + 0.007683844305574894, + 0.005855262745171785, + 0.007077020592987537, + 0.006135137751698494, + 0.0061838966794312, + 0.00528303487226367, + 0.0056204041466116905, + 0.005448669660836458, + 0.004269447177648544, + 0.004049475770443678, + 0.004805488046258688, + 0.006902205292135477, + 0.00683211162686348, + 0.005077107809484005, + 0.004689237102866173, + 0.0033785849809646606, + 0.0033970451913774014, + 0.0028767730109393597, + 0.003737084334716201, + 0.0035729696974158287, + 0.0049696327187120914, + 0.00392401497811079, + 0.003995025996118784, + 0.00413668854162097, + 0.005713248159736395, + 0.003824046580120921, + 0.006507112644612789, + 0.009145122021436691, + 0.0072061894461512566, + 0.006254119332879782, + 0.005344972480088472, + 0.011178679764270782, + 0.01732771098613739, + 0.012888322584331036, + 0.021728523075580597, + 0.013876447454094887, + 0.01751076616346836, + 0.021572789177298546, + 0.013956492766737938 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 9 + ], + "selected_reward": -0.009837242774665356 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a newborn tree frog struggling to free itself from a translucent, gelatinous eggshell?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the frog have disproportionately large eyes with a shimmering jewel-red hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the frog's skin moist, emerald green, and dotted with tiny bumps?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background composed of broad, veined green leaves and water droplets rendered in Van Gogh's swirling brushstroke style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the entire scene rendered with the texture of thick oil paint featuring impasto-like highlights and vibrant colors?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03371100872755051 + ], + "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": [ + 0.1264275461435318, + 0.07140449434518814, + 0.10397519171237946, + 0.16917426884174347, + 0.2755797803401947, + 0.042443204671144485, + 0.022837363183498383, + 0.022161606699228287, + 0.03881185129284859, + 0.12718477845191956 + ], + "credit_map": [ + 0.011129514314234257, + 0.01142751146107912, + 0.011650748550891876, + 0.013983283191919327, + 0.021225688979029655, + 0.017126895487308502, + 0.020162897184491158, + 0.01972099021077156, + 0.008347122929990292, + 0.0063362629152834415, + 0.007523400709033012, + 0.00880315713584423, + 0.01258982066065073, + 0.009603319689631462, + 0.011652741581201553, + 0.006548659410327673, + 0.0069503504782915115, + 0.009206892922520638, + 0.009465205483138561, + 0.011936874128878117, + 0.018576819449663162, + 0.0116100599989295, + 0.02599654719233513, + 0.01023243460804224, + 0.014762847684323788, + 0.014177457429468632, + 0.010835121385753155, + 0.031140169128775597, + 0.01966940239071846, + 0.020861875265836716, + 0.03875952586531639, + 0.018967855721712112, + 0.011318840086460114, + 0.01381644420325756, + 0.013471384532749653, + 0.013022803701460361, + 0.1164533719420433, + 0.07392996549606323, + 0.015157490037381649, + 0.018409451469779015, + 0.0033453076612204313, + 0.006613370031118393, + 0.005146168638020754, + 0.007674751337617636, + 0.006231169681996107, + 0.0031665670685470104, + 0.005815396551042795, + 0.004450470209121704, + 0.003662245348095894, + 0.0025878765154629946, + 0.002179250819608569, + 0.003079741494730115, + 0.0030083211604505777, + 0.0031762400176376104, + 0.002706443425267935, + 0.0024372441694140434, + 0.0023825429379940033, + 0.003062834497541189, + 0.0018010534113273025, + 0.0032168938778340816, + 0.002190450904890895, + 0.0024865216109901667, + 0.003326834412291646, + 0.003694473300129175, + 0.0028079566545784473, + 0.005544841755181551, + 0.0032300606835633516, + 0.005556374788284302, + 0.00741451745852828, + 0.004341308493167162, + 0.00541286077350378, + 0.004503926262259483, + 0.01271138247102499, + 0.01426535751670599, + 0.010457046329975128, + 0.02109207957983017, + 0.014520308002829552, + 0.013873135671019554, + 0.028287013992667198, + 0.0119784502312541 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 3, + 9 + ], + "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 main subject a newborn tree frog struggling to free itself from a translucent, gelatinous eggshell?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the frog have disproportionately large eyes with a shimmering jewel-red hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the frog's skin moist, emerald green, and dotted with tiny bumps?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background composed of broad, veined green leaves and water droplets rendered in Van Gogh's swirling brushstroke style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the entire scene rendered with the texture of thick oil paint featuring impasto-like highlights and vibrant colors?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0067422278225421906 + ], + "opposite_logprob_tokens": [ + -5.006742000579834 + ], + "logprob_reward": -0.0067422278225421906, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.0067422278225421906, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21758563816547394, + 0.13912947475910187, + 0.07789589464664459, + 0.10707148164510727, + 0.13585729897022247, + 0.04875415191054344, + 0.04587123915553093, + 0.02770281210541725, + 0.03829828277230263, + 0.1618337482213974 + ], + "credit_map": [ + 0.021405376493930817, + 0.02251051552593708, + 0.027984892949461937, + 0.024191370233893394, + 0.03269364684820175, + 0.027686581015586853, + 0.030580634251236916, + 0.03053261525928974, + 0.015512133948504925, + 0.016083667054772377, + 0.01554498914629221, + 0.01305985264480114, + 0.027832364663481712, + 0.019490474835038185, + 0.018140386790037155, + 0.013465595431625843, + 0.0063817379996180534, + 0.008130679838359356, + 0.00668381666764617, + 0.009307574480772018, + 0.017355632036924362, + 0.009870932437479496, + 0.010813669301569462, + 0.00935184396803379, + 0.006930870935320854, + 0.012142913416028023, + 0.005992588121443987, + 0.012151948176324368, + 0.02584843896329403, + 0.018853606656193733, + 0.015022923238575459, + 0.010128198191523552, + 0.010809419676661491, + 0.01403429638594389, + 0.010433347895741463, + 0.019792109727859497, + 0.03031403198838234, + 0.020263126119971275, + 0.018725281581282616, + 0.011485690250992775, + 0.007002276834100485, + 0.006487265229225159, + 0.005236064549535513, + 0.007584530860185623, + 0.00616418058052659, + 0.0061015537939965725, + 0.005304552149027586, + 0.004873728379607201, + 0.00714503088966012, + 0.0065408730879426, + 0.004331797361373901, + 0.007207997143268585, + 0.004307985305786133, + 0.004755340050905943, + 0.006652472540736198, + 0.004929741378873587, + 0.0030524141620844603, + 0.002814760897308588, + 0.0035494836047291756, + 0.004195565823465586, + 0.0034530425909906626, + 0.0034116588067263365, + 0.003784622298553586, + 0.003441261127591133, + 0.003338582580909133, + 0.005669471807777882, + 0.0055931671522557735, + 0.0054963696748018265, + 0.005755981896072626, + 0.003493237541988492, + 0.004912652540951967, + 0.004038819577544928, + 0.013836761936545372, + 0.02333458513021469, + 0.019940972328186035, + 0.0226975716650486, + 0.019818011671304703, + 0.019737601280212402, + 0.022550417110323906, + 0.019917819648981094 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 1 + ], + "selected_reward": -0.0067422278225421906 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a newborn tree frog struggling to free itself from a translucent, gelatinous eggshell?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the frog have disproportionately large eyes with a shimmering jewel-red hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the frog's skin moist, emerald green, and dotted with tiny bumps?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background composed of broad, veined green leaves and water droplets rendered in Van Gogh's swirling brushstroke style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the entire scene rendered with the texture of thick oil paint featuring impasto-like highlights and vibrant colors?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005945141427218914 + ], + "opposite_logprob_tokens": [ + -5.130945205688477 + ], + "logprob_reward": -0.005945141427218914, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005945141427218914, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0610300675034523, + 0.04603828489780426, + 0.03162730857729912, + 0.03323183208703995, + 0.37317728996276855, + 0.08749634772539139, + 0.07586918026208878, + 0.041879910975694656, + 0.05978978052735329, + 0.1898600161075592 + ], + "credit_map": [ + 0.0037015427369624376, + 0.006933092605322599, + 0.006742329802364111, + 0.007865245454013348, + 0.01164026465266943, + 0.006852520629763603, + 0.01056734286248684, + 0.006727725267410278, + 0.005868909880518913, + 0.00573006784543395, + 0.0054105049930512905, + 0.004901766777038574, + 0.0072881365194916725, + 0.006137712858617306, + 0.005510910879820585, + 0.005190275609493256, + 0.003100482514128089, + 0.0036421893164515495, + 0.0030267247930169106, + 0.0033130545634776354, + 0.005666005425155163, + 0.0037166241090744734, + 0.005689234472811222, + 0.0034729933831840754, + 0.0040796962566673756, + 0.0029695609118789434, + 0.0028441615868359804, + 0.006213249638676643, + 0.0037655283231288195, + 0.003088263561949134, + 0.006667952984571457, + 0.003603419754654169, + 0.01490701548755169, + 0.01843148097395897, + 0.017581718042492867, + 0.015942761674523354, + 0.15974867343902588, + 0.10426570475101471, + 0.016438735648989677, + 0.0258612260222435, + 0.009377649985253811, + 0.00974056962877512, + 0.007789236959069967, + 0.011730971746146679, + 0.014364711008965969, + 0.01733706332743168, + 0.007212151773273945, + 0.00994399469345808, + 0.011036049574613571, + 0.01024642027914524, + 0.007334994617849588, + 0.011559130623936653, + 0.00812546070665121, + 0.009417790919542313, + 0.0087399547919631, + 0.009409378282725811, + 0.0060295648872852325, + 0.004931008443236351, + 0.0031139650382101536, + 0.005643761716783047, + 0.006822341121733189, + 0.0053700730204582214, + 0.004660173784941435, + 0.005309025291353464, + 0.004513410851359367, + 0.007031143177300692, + 0.005759069230407476, + 0.008583045564591885, + 0.010737068951129913, + 0.008464163169264793, + 0.007650428917258978, + 0.007051452528685331, + 0.01715506613254547, + 0.02398771420121193, + 0.015609937720000744, + 0.02327502705156803, + 0.02806960791349411, + 0.026900550350546837, + 0.02857006900012493, + 0.026292039081454277 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 9, + 5 + ], + "selected_reward": -0.005945141427218914 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a newborn tree frog struggling to free itself from a translucent, gelatinous eggshell?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the frog have disproportionately large eyes with a shimmering jewel-red hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the frog's skin moist, emerald green, and dotted with tiny bumps?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background composed of broad, veined green leaves and water droplets rendered in Van Gogh's swirling brushstroke style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the entire scene rendered with the texture of thick oil paint featuring impasto-like highlights and vibrant colors?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0017138330731540918 + ], + "opposite_logprob_tokens": [ + -6.376713752746582 + ], + "logprob_reward": -0.0017138330731540918, + "yes_no_margin_reward": 6.375, + "token_credit_reward": -0.0017138330731540918, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12442287802696228, + 0.0662078931927681, + 0.06026739254593849, + 0.11460932344198227, + 0.0756772831082344, + 0.15203320980072021, + 0.16228927671909332, + 0.07216239720582962, + 0.08764757961034775, + 0.08468277007341385 + ], + "credit_map": [ + 0.01679733581840992, + 0.012032260186970234, + 0.016980605199933052, + 0.014491190202534199, + 0.01230678241699934, + 0.01363025140017271, + 0.01804407872259617, + 0.020140383392572403, + 0.0115731880068779, + 0.005894824396818876, + 0.012199905700981617, + 0.006674469448626041, + 0.006875098217278719, + 0.007733379490673542, + 0.008629526942968369, + 0.006627494469285011, + 0.005316128022968769, + 0.005839591845870018, + 0.004875715356320143, + 0.00663668941706419, + 0.011074780486524105, + 0.006999643053859472, + 0.013863357715308666, + 0.00566148292273283, + 0.010904285125434399, + 0.009110677987337112, + 0.01018040906637907, + 0.023653429001569748, + 0.013810776174068451, + 0.013226984068751335, + 0.020233912393450737, + 0.013488846831023693, + 0.008329366333782673, + 0.005901881027966738, + 0.006181749049574137, + 0.014383646659553051, + 0.01044513937085867, + 0.01022571325302124, + 0.01216339785605669, + 0.00804639421403408, + 0.0191322211176157, + 0.012632295489311218, + 0.011555464938282967, + 0.013763075694441795, + 0.03033655323088169, + 0.036943066865205765, + 0.013741561211645603, + 0.013928965665400028, + 0.019778674468398094, + 0.022295689210295677, + 0.015054155141115189, + 0.028137942776083946, + 0.01603369601070881, + 0.015496808104217052, + 0.023731475695967674, + 0.0217608455568552, + 0.00971950776875019, + 0.008497153408825397, + 0.0052665164694190025, + 0.009216605685651302, + 0.009610502980649471, + 0.010024510324001312, + 0.009717689827084541, + 0.0101099144667387, + 0.005817625671625137, + 0.012637518346309662, + 0.0068733347579836845, + 0.012668483890593052, + 0.0152513412758708, + 0.011681647039949894, + 0.012401456944644451, + 0.010316172614693642, + 0.008376563899219036, + 0.009956834837794304, + 0.00784423016011715, + 0.012648108415305614, + 0.010077791288495064, + 0.011825820431113243, + 0.013864538632333279, + 0.010088881477713585 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 5, + 0 + ], + "selected_reward": -0.0017138330731540918 + } + ], + "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/0000021/group_9/index_tensor([5732, 5732])_group_9_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000021/group_9/index_tensor([5732, 5732])_group_9_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..e4676ceb5807ab3b140a46813727290cd74d0904 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000021/group_9/index_tensor([5732, 5732])_group_9_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "smooth tracking shot sweeps slowly from left to right, revealing a chaotic laboratory imbued with a heavy, industrial feel. The camera first passes over dusty concrete walls and exposed metal ducts in the ceiling, the low hum of rusted vents. A tangle of wires, like vines, entangles various delicate machines—an old oscilloscope screen flickers an erratic green waveform, while a centrifuge spins steadily in the corner. As the camera pans to the center of the laboratory, it finally focuses on a metal workbench cluttered with tools, beakers, and scribbled blueprints. Bent over it is a scientist with graying hair, as unkempt as a clump of grass. He wears a white lab coat stained with oil and an unidentified substance, his sleeves rolled up high, revealing his muscular, tense forearms. His bloodshot eyes, bloodshot from chronic sleep deprivation, gleamed with a feverish glint as he wielded a heavy brass mortar, exerting all his strength to crush, one by one, an irregular lump of otherworldly crystal matter pulsing with an eerie purple light. With each impact, tiny luminous particles and tiny sparks of energy flew forth, briefly illuminating his face, which was slightly distorted by intense concentration. The flat overhead fluorescent lights cast sharp, cold shadows, intertwining with the eerie purple light emanating from the crystal itself. The realistic yet rough details created an atmosphere of tension, danger, and profound creativity.", + "seed": 28188, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000021/group_9/index_5732_seed_28188_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": -0.12560108713805676, + "vlm_logprob": -0.12560108713805676, + "vlm_yes_no_margin": 2.65, + "vlm_token_credit": -0.12560108713805676, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" 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 perform a smooth tracking shot sweeping slowly from left to right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a chaotic laboratory with dusty concrete walls, exposed metal ducts, and rusted vents?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a scientist with graying hair wearing a white lab coat stained with oil visible at a metal workbench?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scientist wielding a heavy brass mortar to crush an irregular lump of crystal matter pulsing with eerie purple light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do flat overhead fluorescent lights cast sharp, cold shadows in 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": [ + -0.34229394793510437 + ], + "opposite_logprob_tokens": [ + -1.3422939777374268 + ], + "logprob_reward": -0.34229394793510437, + "yes_no_margin_reward": 1.0, + "token_credit_reward": -0.34229394793510437, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14568881690502167, + 0.24137471616268158, + 0.2685210704803467, + 0.0457046702504158, + 0.03575972467660904, + 0.03944765776395798, + 0.04406062141060829, + 0.030107365921139717, + 0.035905130207538605, + 0.1134302169084549 + ], + "credit_map": [ + 0.010933667421340942, + 0.014521030709147453, + 0.02706729993224144, + 0.011879323981702328, + 0.020468026399612427, + 0.02065831422805786, + 0.017856676131486893, + 0.022304464131593704, + 0.009425726719200611, + 0.017508680000901222, + 0.0158547293394804, + 0.011637160554528236, + 0.07564626634120941, + 0.02007320523262024, + 0.01649007759988308, + 0.07473888248205185, + 0.01447366364300251, + 0.08123166114091873, + 0.009320101700723171, + 0.014766435138881207, + 0.09803348779678345, + 0.01682908460497856, + 0.02385709248483181, + 0.010009526275098324, + 0.0035833255387842655, + 0.005422435700893402, + 0.0051498739048838615, + 0.010806163772940636, + 0.00588487833738327, + 0.0056629846803843975, + 0.005684241186827421, + 0.003510769922286272, + 0.0036624656058847904, + 0.004164877813309431, + 0.0032788615208119154, + 0.00657019205391407, + 0.004243611358106136, + 0.005954172927886248, + 0.004347455222159624, + 0.003538087708875537, + 0.005500126164406538, + 0.004945649299770594, + 0.005123424343764782, + 0.0054291123524308205, + 0.007719443179666996, + 0.0029312565457075834, + 0.004339860752224922, + 0.003458782797679305, + 0.003991943784058094, + 0.014199073426425457, + 0.014330356381833553, + 0.0027062317822128534, + 0.0025703299324959517, + 0.0014648024225607514, + 0.001899439375847578, + 0.0028984432574361563, + 0.005424387753009796, + 0.0035360215697437525, + 0.004709136672317982, + 0.003853862639516592, + 0.004534921608865261, + 0.0017994155641645193, + 0.0030558109283447266, + 0.0031938094180077314, + 0.004958891309797764, + 0.0065384358167648315, + 0.0035163359716534615, + 0.004001350607722998, + 0.007561227772384882, + 0.00199087243527174, + 0.004214297980070114, + 0.0031237173825502396, + 0.014882814139127731, + 0.040888894349336624, + 0.006437573116272688, + 0.007914635352790356, + 0.006955265998840332, + 0.006226487923413515, + 0.0053707933984696865, + 0.0247537549585104 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 1, + 0 + ], + "selected_reward": -0.34229394793510437 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" 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 perform a smooth tracking shot sweeping slowly from left to right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a chaotic laboratory with dusty concrete walls, exposed metal ducts, and rusted vents?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a scientist with graying hair wearing a white lab coat stained with oil visible at a metal workbench?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scientist wielding a heavy brass mortar to crush an irregular lump of crystal matter pulsing with eerie purple light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do flat overhead fluorescent lights cast sharp, cold shadows in 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": [ + -0.03373014181852341 + ], + "opposite_logprob_tokens": [ + -3.4087300300598145 + ], + "logprob_reward": -0.03373014181852341, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03373014181852341, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20101240277290344, + 0.2328265905380249, + 0.2323639839887619, + 0.043848276138305664, + 0.04681748151779175, + 0.13939714431762695, + 0.01643918640911579, + 0.02515634149312973, + 0.03073873184621334, + 0.03139979764819145 + ], + "credit_map": [ + 0.015403815545141697, + 0.017560185864567757, + 0.0426822267472744, + 0.015886807814240456, + 0.028767608106136322, + 0.027560943737626076, + 0.019898608326911926, + 0.03325222060084343, + 0.011959883384406567, + 0.017731651663780212, + 0.014230894856154919, + 0.012482034042477608, + 0.0663817971944809, + 0.02813366800546646, + 0.040984392166137695, + 0.04092228412628174, + 0.011304868385195732, + 0.06884663552045822, + 0.0059629399329423904, + 0.011602031998336315, + 0.0878368467092514, + 0.0183681882917881, + 0.01876189187169075, + 0.009680598974227905, + 0.0029623007867485285, + 0.004555570427328348, + 0.0066653043031692505, + 0.009197898209095001, + 0.006586140021681786, + 0.006050785072147846, + 0.004186179954558611, + 0.0036441038828343153, + 0.009317491203546524, + 0.008632853627204895, + 0.005156744737178087, + 0.00543684558942914, + 0.0037093411665409803, + 0.006264046300202608, + 0.0042931027710437775, + 0.004007055889815092, + 0.02436183951795101, + 0.015004380606114864, + 0.01725633628666401, + 0.02544068731367588, + 0.019280271604657173, + 0.013342277146875858, + 0.0124864736571908, + 0.01222488097846508, + 0.0022379234433174133, + 0.0019146938575431705, + 0.003861246630549431, + 0.0019367781933397055, + 0.0021346062421798706, + 0.0007955524488352239, + 0.0013311674119904637, + 0.0022272206842899323, + 0.005011423025280237, + 0.0034639385994523764, + 0.0053541469387710094, + 0.00220357533544302, + 0.002631555078551173, + 0.0019727398175746202, + 0.0022430738899856806, + 0.002275891602039337, + 0.006354666780680418, + 0.004971024580299854, + 0.0037047683726996183, + 0.0027926636394113302, + 0.004295021761208773, + 0.002724813763052225, + 0.0031652231700718403, + 0.0027305509429425, + 0.005853655282407999, + 0.012027758173644543, + 0.00303363474085927, + 0.00210960628464818, + 0.0022442589979618788, + 0.0015410383930429816, + 0.0013657094677910209, + 0.0032241372391581535 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 2, + 0 + ], + "selected_reward": -0.03373014181852341 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" 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 perform a smooth tracking shot sweeping slowly from left to right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a chaotic laboratory with dusty concrete walls, exposed metal ducts, and rusted vents?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a scientist with graying hair wearing a white lab coat stained with oil visible at a metal workbench?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scientist wielding a heavy brass mortar to crush an irregular lump of crystal matter pulsing with eerie purple light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do flat overhead fluorescent lights cast sharp, cold shadows in 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": [ + -0.020592350512742996 + ], + "opposite_logprob_tokens": [ + -3.895592451095581 + ], + "logprob_reward": -0.020592350512742996, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020592350512742996, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09634668380022049, + 0.080137699842453, + 0.06576782464981079, + 0.12553828954696655, + 0.049698226153850555, + 0.040591418743133545, + 0.11971304565668106, + 0.08884033560752869, + 0.03335513174533844, + 0.3000113070011139 + ], + "credit_map": [ + 0.010109403170645237, + 0.010814376175403595, + 0.016021035611629486, + 0.009073342196643353, + 0.015879575163125992, + 0.015325985848903656, + 0.008390813134610653, + 0.010732155293226242, + 0.003422996960580349, + 0.0045797256752848625, + 0.0053140646778047085, + 0.004748343024402857, + 0.008053360506892204, + 0.012563930824398994, + 0.022740479558706284, + 0.018714793026447296, + 0.013195940293371677, + 0.010581336915493011, + 0.0040466319769620895, + 0.005027271807193756, + 0.01322863344103098, + 0.006587970536202192, + 0.008476763032376766, + 0.004623282700777054, + 0.008250663988292217, + 0.012612701393663883, + 0.022738616913557053, + 0.016625909134745598, + 0.014520974829792976, + 0.014821602031588554, + 0.015665225684642792, + 0.02030259557068348, + 0.004591737408190966, + 0.0064507173374295235, + 0.004371204413473606, + 0.005286424420773983, + 0.012566307559609413, + 0.004994119983166456, + 0.0073486375622451305, + 0.004089075606316328, + 0.004869633819907904, + 0.005160990636795759, + 0.003793097799643874, + 0.007862416096031666, + 0.0064040664583444595, + 0.0030346703715622425, + 0.00553751178085804, + 0.003929035272449255, + 0.010036258026957512, + 0.03562934324145317, + 0.03706132993102074, + 0.009208043105900288, + 0.009463243186473846, + 0.003219673177227378, + 0.007584144826978445, + 0.007511005271226168, + 0.009331239387392998, + 0.012525957077741623, + 0.009153896942734718, + 0.009691442362964153, + 0.016070229932665825, + 0.0054091946221888065, + 0.01541674043983221, + 0.01124163344502449, + 0.0038995600771158934, + 0.004196797497570515, + 0.003512290772050619, + 0.00505141681060195, + 0.005381885915994644, + 0.0021599363535642624, + 0.006404419895261526, + 0.002748822793364525, + 0.08862008899450302, + 0.09077198803424835, + 0.02053080126643181, + 0.032832249999046326, + 0.01911337859928608, + 0.01039782352745533, + 0.009560172446072102, + 0.028184816241264343 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 3, + 6 + ], + "selected_reward": -0.020592350512742996 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" 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 perform a smooth tracking shot sweeping slowly from left to right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a chaotic laboratory with dusty concrete walls, exposed metal ducts, and rusted vents?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a scientist with graying hair wearing a white lab coat stained with oil visible at a metal workbench?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scientist wielding a heavy brass mortar to crush an irregular lump of crystal matter pulsing with eerie purple light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do flat overhead fluorescent lights cast sharp, cold shadows in 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": [ + -0.029808619990944862 + ], + "opposite_logprob_tokens": [ + -3.529808521270752 + ], + "logprob_reward": -0.029808619990944862, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029808619990944862, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13872703909873962, + 0.09852392971515656, + 0.12295965105295181, + 0.09007518738508224, + 0.04573677107691765, + 0.08839526772499084, + 0.050497498363256454, + 0.03788543865084648, + 0.11644895374774933, + 0.2107502669095993 + ], + "credit_map": [ + 0.012583775445818901, + 0.01774144172668457, + 0.02048814855515957, + 0.010722531005740166, + 0.020336920395493507, + 0.022038504481315613, + 0.017163969576358795, + 0.017651742324233055, + 0.0044358596205711365, + 0.00804190058261156, + 0.00896511785686016, + 0.006730931811034679, + 0.01188686117529869, + 0.012190892361104488, + 0.028016837313771248, + 0.018255529925227165, + 0.007380837574601173, + 0.035968489944934845, + 0.0037245824933052063, + 0.00580420158803463, + 0.04617024585604668, + 0.007936430163681507, + 0.011371144093573093, + 0.00460370909422636, + 0.007787479553371668, + 0.009570061229169369, + 0.007479189895093441, + 0.016625212505459785, + 0.018550580367445946, + 0.011358410120010376, + 0.0114686768501997, + 0.007235568482428789, + 0.004543252754956484, + 0.00613982742652297, + 0.004111874848604202, + 0.007111425511538982, + 0.006634251214563847, + 0.006869054399430752, + 0.006331797689199448, + 0.003995288163423538, + 0.009630569256842136, + 0.009789065457880497, + 0.010951859876513481, + 0.013714909553527832, + 0.01772654429078102, + 0.006797092035412788, + 0.012365982867777348, + 0.00741924811154604, + 0.0085018090903759, + 0.01302262395620346, + 0.004525620955973864, + 0.004255420062690973, + 0.007615369278937578, + 0.0038588624447584152, + 0.0041373842395842075, + 0.004580408800393343, + 0.005963320843875408, + 0.00681333988904953, + 0.003718799911439419, + 0.0038005607202649117, + 0.006376712117344141, + 0.0018801892874762416, + 0.006014444399625063, + 0.003318071598187089, + 0.016224652528762817, + 0.01926223188638687, + 0.010412459261715412, + 0.01278094481676817, + 0.028438154608011246, + 0.005097790155559778, + 0.015811624005436897, + 0.008421089500188828, + 0.06677575409412384, + 0.06877361238002777, + 0.011363770812749863, + 0.01805001311004162, + 0.01573021337389946, + 0.00567271513864398, + 0.00789937749505043, + 0.016484804451465607 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 2 + ], + "selected_reward": -0.029808619990944862 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" 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 perform a smooth tracking shot sweeping slowly from left to right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a chaotic laboratory with dusty concrete walls, exposed metal ducts, and rusted vents?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a scientist with graying hair wearing a white lab coat stained with oil visible at a metal workbench?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scientist wielding a heavy brass mortar to crush an irregular lump of crystal matter pulsing with eerie purple light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do flat overhead fluorescent lights cast sharp, cold shadows in 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": [ + -0.20158037543296814 + ], + "opposite_logprob_tokens": [ + -1.7015804052352905 + ], + "logprob_reward": -0.20158037543296814, + "yes_no_margin_reward": 1.5, + "token_credit_reward": -0.20158037543296814, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4216857850551605, + 0.19759710133075714, + 0.056941673159599304, + 0.026215672492980957, + 0.025737162679433823, + 0.04349273443222046, + 0.03949384391307831, + 0.049320850521326065, + 0.12942668795585632, + 0.01008850708603859 + ], + "credit_map": [ + 0.03932884708046913, + 0.07050704956054688, + 0.056255172938108444, + 0.0365251749753952, + 0.04943690448999405, + 0.058032646775245667, + 0.049578193575143814, + 0.06202177330851555, + 0.012767389416694641, + 0.018956726416945457, + 0.01861770823597908, + 0.014511548914015293, + 0.028913894668221474, + 0.050474632531404495, + 0.030593298375606537, + 0.022761881351470947, + 0.005948144476860762, + 0.01315911766141653, + 0.0028998441994190216, + 0.003951605409383774, + 0.01684688962996006, + 0.005460120737552643, + 0.005800505168735981, + 0.002875440986827016, + 0.002425973303616047, + 0.0028023466002196074, + 0.0026779326144605875, + 0.004929732531309128, + 0.005098273511976004, + 0.004126067738980055, + 0.0020096078515052795, + 0.002145737875252962, + 0.0029666638001799583, + 0.0034838980063796043, + 0.002782918745651841, + 0.0038701104931533337, + 0.004568547010421753, + 0.0029007939156144857, + 0.0029145823791623116, + 0.0022496460005640984, + 0.004954211879521608, + 0.006366027053445578, + 0.006277401000261307, + 0.007069760467857122, + 0.0058284904807806015, + 0.005066987127065659, + 0.003625309793278575, + 0.004304545000195503, + 0.0066761053167283535, + 0.013265341520309448, + 0.003002563724294305, + 0.0025141967926174402, + 0.003754936857149005, + 0.0027012384962290525, + 0.0019699495751410723, + 0.005609509535133839, + 0.007910401560366154, + 0.005077444948256016, + 0.009806550107896328, + 0.006136721931397915, + 0.005520522594451904, + 0.005246590822935104, + 0.004611143376678228, + 0.0050114733166992664, + 0.02538786455988884, + 0.0148687195032835, + 0.02879684418439865, + 0.013090639375150204, + 0.011793261393904686, + 0.011997240595519543, + 0.010481813922524452, + 0.013010297901928425, + 0.001335343113169074, + 0.0012395564699545503, + 0.0014137242687866092, + 0.0012886828044429421, + 0.0018967895302921534, + 0.0009888889035210013, + 0.0009646622347645462, + 0.0009608588297851384 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 8 + ], + "selected_reward": -0.20158037543296814 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "smooth tracking shot sweeps slowly from left to right, revealing a chaotic laboratory imbued with a heavy, industrial feel. The camera first passes over dusty concrete walls and exposed metal ducts in the ceiling, the low hum of rusted vents. A tangle of wires, like vines, entangles various delicate machines—an old oscilloscope screen flickers an erratic green waveform, while a centrifuge spins steadily in the corner. As the camera pans to the center of the laboratory, it finally focuses on a metal workbench cluttered with tools, beakers, and scribbled blueprints. Bent over it is a scientist with graying hair, as unkempt as a clump of grass. He wears a white lab coat stained with oil and an unidentified substance, his sleeves rolled up high, revealing his muscular, tense forearms. His bloodshot eyes, bloodshot from chronic sleep deprivation, gleamed with a feverish glint as he wielded a heavy brass mortar, exerting all his strength to crush, one by one, an irregular lump of otherworldly crystal matter pulsing with an eerie purple light. With each impact, tiny luminous particles and tiny sparks of energy flew forth, briefly illuminating his face, which was slightly distorted by intense concentration. The flat overhead fluorescent lights cast sharp, cold shadows, intertwining with the eerie purple light emanating from the crystal itself. The realistic yet rough details created an atmosphere of tension, danger, and profound creativity.", + "seed": 28189, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000021/group_9/index_5732_seed_28189_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": -0.16611608415842055, + "vlm_logprob": -0.16611608415842055, + "vlm_yes_no_margin": 2.174999976158142, + "vlm_token_credit": -0.16611608415842055, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" 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 perform a smooth tracking shot sweeping slowly from left to right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a chaotic laboratory with dusty concrete walls, exposed metal ducts, and rusted vents?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a scientist with graying hair wearing a white lab coat stained with oil visible at a metal workbench?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scientist wielding a heavy brass mortar to crush an irregular lump of crystal matter pulsing with eerie purple light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do flat overhead fluorescent lights cast sharp, cold shadows in 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": [ + -0.17176401615142822 + ], + "opposite_logprob_tokens": [ + -2.0467638969421387 + ], + "logprob_reward": -0.17176401615142822, + "yes_no_margin_reward": 1.8749998807907104, + "token_credit_reward": -0.17176401615142822, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3218821585178375, + 0.07122854143381119, + 0.16090339422225952, + 0.1399931013584137, + 0.1093432679772377, + 0.04566844552755356, + 0.05446738377213478, + 0.06310970336198807, + 0.02050832100212574, + 0.012895653024315834 + ], + "credit_map": [ + 0.018053600564599037, + 0.022814080119132996, + 0.09162425249814987, + 0.053097914904356, + 0.02999100275337696, + 0.06524623185396194, + 0.023099923506379128, + 0.017955182120203972, + 0.00797212217003107, + 0.007562633603811264, + 0.008066252805292606, + 0.01437389012426138, + 0.005099286790937185, + 0.006200931500643492, + 0.011377902701497078, + 0.01057552918791771, + 0.019653743132948875, + 0.021074499934911728, + 0.022666163742542267, + 0.019260220229625702, + 0.01445508748292923, + 0.03086843155324459, + 0.0132558923214674, + 0.019669372588396072, + 0.01868821121752262, + 0.014385629445314407, + 0.018518105149269104, + 0.009694495238363743, + 0.01591375097632408, + 0.019898012280464172, + 0.014486439526081085, + 0.028408469632267952, + 0.00974431075155735, + 0.01319296658039093, + 0.00847109779715538, + 0.008279543370008469, + 0.01411673054099083, + 0.023577282205224037, + 0.013073191046714783, + 0.01888815313577652, + 0.002705545863136649, + 0.006516150198876858, + 0.003956527914851904, + 0.003597879083827138, + 0.007402114104479551, + 0.010273292660713196, + 0.006024316418915987, + 0.005192624405026436, + 0.005087398458272219, + 0.007444477640092373, + 0.0046644266694784164, + 0.006848737131804228, + 0.010886697098612785, + 0.005009073298424482, + 0.009511776268482208, + 0.005014804657548666, + 0.004947076551616192, + 0.010314050130546093, + 0.01744762435555458, + 0.006743124686181545, + 0.006239758804440498, + 0.0049791717901825905, + 0.004524300806224346, + 0.007914607413113117, + 0.003992006648331881, + 0.0031194528564810753, + 0.0017739005852490664, + 0.002384103601798415, + 0.002893771044909954, + 0.0018077364657074213, + 0.0021640099585056305, + 0.002373341005295515, + 0.0019281962886452675, + 0.001992159290239215, + 0.0008152278605848551, + 0.0013943189987912774, + 0.0020461988169699907, + 0.001235704985447228, + 0.0019140827935189009, + 0.001569764455780387 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 3 + ], + "selected_reward": -0.17176401615142822 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" 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 perform a smooth tracking shot sweeping slowly from left to right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a chaotic laboratory with dusty concrete walls, exposed metal ducts, and rusted vents?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a scientist with graying hair wearing a white lab coat stained with oil visible at a metal workbench?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scientist wielding a heavy brass mortar to crush an irregular lump of crystal matter pulsing with eerie purple light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do flat overhead fluorescent lights cast sharp, cold shadows in 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": [ + -0.252115935087204 + ], + "opposite_logprob_tokens": [ + -1.5021159648895264 + ], + "logprob_reward": -0.252115935087204, + "yes_no_margin_reward": 1.25, + "token_credit_reward": -0.252115935087204, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.27504321932792664, + 0.09406309574842453, + 0.09272229671478271, + 0.05532461032271385, + 0.14705140888690948, + 0.03171708062291145, + 0.05713896080851555, + 0.16735413670539856, + 0.04065028950572014, + 0.03893491253256798 + ], + "credit_map": [ + 0.01377434004098177, + 0.0282729659229517, + 0.038417309522628784, + 0.0631713941693306, + 0.0487372949719429, + 0.027063746005296707, + 0.03922061622142792, + 0.01638556271791458, + 0.014475171454250813, + 0.013535377569496632, + 0.009918026626110077, + 0.014998706988990307, + 0.008495855145156384, + 0.006195267662405968, + 0.012700000777840614, + 0.013744683936238289, + 0.013448324985802174, + 0.009011516347527504, + 0.01682751439511776, + 0.013959222473204136, + 0.005729353986680508, + 0.020877081900835037, + 0.006340199150145054, + 0.006529083475470543, + 0.009957574307918549, + 0.004842651542276144, + 0.008873970247805119, + 0.006666874513030052, + 0.0052687032148242, + 0.007622184231877327, + 0.004167148843407631, + 0.007925502024590969, + 0.012650592252612114, + 0.022112075239419937, + 0.011042867787182331, + 0.010447302833199501, + 0.019354287534952164, + 0.03298913314938545, + 0.014878918416798115, + 0.023576240986585617, + 0.0028485318180173635, + 0.005307839717715979, + 0.003226831089705229, + 0.0031423592008650303, + 0.0038180563133209944, + 0.004520860034972429, + 0.00525480043143034, + 0.0035978013183921576, + 0.01079821027815342, + 0.007019567769020796, + 0.004567250143736601, + 0.005681183189153671, + 0.008014954626560211, + 0.008968598209321499, + 0.006595488637685776, + 0.005493710283190012, + 0.01605193130671978, + 0.030408620834350586, + 0.0532924048602581, + 0.017770681530237198, + 0.011902544647455215, + 0.012904954142868519, + 0.011225459165871143, + 0.013797544874250889, + 0.003896035486832261, + 0.006747562903910875, + 0.00585382292047143, + 0.0046226331032812595, + 0.00451161153614521, + 0.00387222389690578, + 0.005705323535948992, + 0.005441074259579182, + 0.01125163584947586, + 0.004426066763699055, + 0.004434984643012285, + 0.0028123643714934587, + 0.0038193147629499435, + 0.003832452232018113, + 0.0036179721355438232, + 0.004740126430988312 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 4 + ], + "selected_reward": -0.252115935087204 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" 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 perform a smooth tracking shot sweeping slowly from left to right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a chaotic laboratory with dusty concrete walls, exposed metal ducts, and rusted vents?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a scientist with graying hair wearing a white lab coat stained with oil visible at a metal workbench?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scientist wielding a heavy brass mortar to crush an irregular lump of crystal matter pulsing with eerie purple light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do flat overhead fluorescent lights cast sharp, cold shadows in 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": [ + -0.014207758009433746 + ], + "opposite_logprob_tokens": [ + -4.26420783996582 + ], + "logprob_reward": -0.014207758009433746, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014207758009433746, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1828422248363495, + 0.07100539654493332, + 0.1148371547460556, + 0.0582776702940464, + 0.10260723531246185, + 0.06322390586137772, + 0.11761727929115295, + 0.10418199747800827, + 0.12073221802711487, + 0.06467488408088684 + ], + "credit_map": [ + 0.012312190607190132, + 0.012644007802009583, + 0.05441765487194061, + 0.024150675162672997, + 0.020029950886964798, + 0.0311898160725832, + 0.016900483518838882, + 0.011197463609278202, + 0.00999411754310131, + 0.009076970629394054, + 0.00976144801825285, + 0.016500217840075493, + 0.005564836785197258, + 0.004594788886606693, + 0.008804853074252605, + 0.006708162371069193, + 0.012401669286191463, + 0.01103186048567295, + 0.01816749759018421, + 0.0221339650452137, + 0.01003502868115902, + 0.014667908661067486, + 0.01369974110275507, + 0.012699490413069725, + 0.013097523711621761, + 0.006022556684911251, + 0.007940029725432396, + 0.004403636325150728, + 0.0056341649033129215, + 0.007351425942033529, + 0.004374115727841854, + 0.009454217739403248, + 0.011608687229454517, + 0.019639546051621437, + 0.010176186449825764, + 0.011630394496023655, + 0.013377023860812187, + 0.014981891959905624, + 0.011840060353279114, + 0.009353443048894405, + 0.004171262960880995, + 0.009683828800916672, + 0.004567382857203484, + 0.005028291139751673, + 0.010984404943883419, + 0.012333787977695465, + 0.01014777459204197, + 0.006307179573923349, + 0.013291583396494389, + 0.014331232756376266, + 0.00923144444823265, + 0.011511415243148804, + 0.025159737095236778, + 0.0169745571911335, + 0.013882460072636604, + 0.013234862126410007, + 0.01095153484493494, + 0.013784656301140785, + 0.01830252632498741, + 0.013725809752941132, + 0.013396594673395157, + 0.009989664889872074, + 0.013999386690557003, + 0.010031837038695812, + 0.0211293026804924, + 0.020299207419157028, + 0.010314024053514004, + 0.014944339171051979, + 0.014060115441679955, + 0.01179144624620676, + 0.016736093908548355, + 0.01145770214498043, + 0.014412753283977509, + 0.012216498143970966, + 0.005538941361010075, + 0.007102706469595432, + 0.006675756070762873, + 0.006247175391763449, + 0.005791353061795235, + 0.0066897086799144745 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 6 + ], + "selected_reward": -0.014207758009433746 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" 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 perform a smooth tracking shot sweeping slowly from left to right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a chaotic laboratory with dusty concrete walls, exposed metal ducts, and rusted vents?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a scientist with graying hair wearing a white lab coat stained with oil visible at a metal workbench?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scientist wielding a heavy brass mortar to crush an irregular lump of crystal matter pulsing with eerie purple light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do flat overhead fluorescent lights cast sharp, cold shadows in 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": [ + -0.07898899912834167 + ], + "opposite_logprob_tokens": [ + -2.578989028930664 + ], + "logprob_reward": -0.07898899912834167, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07898899912834167, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10811741650104523, + 0.06500597298145294, + 0.07860203832387924, + 0.052093129605054855, + 0.0836726501584053, + 0.13668327033519745, + 0.16545367240905762, + 0.13587813079357147, + 0.11050556600093842, + 0.0639880895614624 + ], + "credit_map": [ + 0.008597318083047867, + 0.0111478790640831, + 0.018756084144115448, + 0.01380415540188551, + 0.016666080802679062, + 0.018503496423363686, + 0.013002835214138031, + 0.007639585994184017, + 0.009336205199360847, + 0.008070620708167553, + 0.008365102112293243, + 0.012526116333901882, + 0.005766486749053001, + 0.004101507365703583, + 0.009774207137525082, + 0.00706573436036706, + 0.007033973466604948, + 0.009589139372110367, + 0.010898617096245289, + 0.011191267520189285, + 0.008458578959107399, + 0.015384985134005547, + 0.008131098933517933, + 0.00791438203305006, + 0.007421544753015041, + 0.006598062813282013, + 0.007293786853551865, + 0.005675086751580238, + 0.004835669416934252, + 0.006303756032139063, + 0.006472996436059475, + 0.00749223493039608, + 0.010230625048279762, + 0.01137377880513668, + 0.007047722116112709, + 0.00905187614262104, + 0.00998450256884098, + 0.013239697553217411, + 0.01098133996129036, + 0.011763113550841808, + 0.008263145573437214, + 0.020744139328598976, + 0.012086862698197365, + 0.012426034547388554, + 0.022720729932188988, + 0.028867486864328384, + 0.016449149698019028, + 0.015125717036426067, + 0.01702527515590191, + 0.023541951552033424, + 0.010586853139102459, + 0.020682118833065033, + 0.030018288642168045, + 0.019359532743692398, + 0.02689070999622345, + 0.017348965629935265, + 0.01186329685151577, + 0.019243361428380013, + 0.032892290502786636, + 0.013043048791587353, + 0.017416805028915405, + 0.010739585384726524, + 0.01293929759413004, + 0.017740460112690926, + 0.015589351765811443, + 0.022573264315724373, + 0.011770620010793209, + 0.008646378293633461, + 0.018623994663357735, + 0.01316991075873375, + 0.010530048981308937, + 0.009602002799510956, + 0.017364628612995148, + 0.006358729675412178, + 0.004719590302556753, + 0.003655265783891082, + 0.0061263213865458965, + 0.006120195146650076, + 0.005394408013671637, + 0.014248951338231564 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 5, + 7 + ], + "selected_reward": -0.07898899912834167 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" 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 perform a smooth tracking shot sweeping slowly from left to right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a chaotic laboratory with dusty concrete walls, exposed metal ducts, and rusted vents?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a scientist with graying hair wearing a white lab coat stained with oil visible at a metal workbench?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scientist wielding a heavy brass mortar to crush an irregular lump of crystal matter pulsing with eerie purple light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do flat overhead fluorescent lights cast sharp, cold shadows in 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": [ + -0.3135037124156952 + ], + "opposite_logprob_tokens": [ + -1.3135037422180176 + ], + "logprob_reward": -0.3135037124156952, + "yes_no_margin_reward": 1.0, + "token_credit_reward": -0.3135037124156952, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5861524343490601, + 0.05463109910488129, + 0.03350047767162323, + 0.03610723838210106, + 0.09758332371711731, + 0.027265295386314392, + 0.03233520686626434, + 0.06278686225414276, + 0.038873109966516495, + 0.030764885246753693 + ], + "credit_map": [ + 0.0375773124396801, + 0.03720541670918465, + 0.15802234411239624, + 0.07556416094303131, + 0.06705290824174881, + 0.12384536862373352, + 0.05262092500925064, + 0.034264031797647476, + 0.006349328905344009, + 0.004748135805130005, + 0.005970112048089504, + 0.0072056883946061134, + 0.007665323559194803, + 0.00631812633946538, + 0.007678345311433077, + 0.008696042001247406, + 0.00468128127977252, + 0.003874679561704397, + 0.005288324318826199, + 0.005459940992295742, + 0.002615397796034813, + 0.005394937936216593, + 0.0035183734726160765, + 0.002667544176802039, + 0.008397203870117664, + 0.004915696568787098, + 0.005568417254835367, + 0.0038515280466526747, + 0.003941090777516365, + 0.004036664497107267, + 0.002830187790095806, + 0.002566451206803322, + 0.010949214920401573, + 0.02004147134721279, + 0.010886811651289463, + 0.008695032447576523, + 0.01373862661421299, + 0.015284080058336258, + 0.007461495231837034, + 0.010526600293815136, + 0.003773381235077977, + 0.0038402299396693707, + 0.0030233438592404127, + 0.0023473650217056274, + 0.003803293453529477, + 0.004041306674480438, + 0.003927435725927353, + 0.002508939942345023, + 0.0028023854829370975, + 0.003645810531452298, + 0.0028040457982569933, + 0.004006013739854097, + 0.006374593358486891, + 0.003502302337437868, + 0.006522404029965401, + 0.002677651820704341, + 0.006902288179844618, + 0.011120526120066643, + 0.01175915077328682, + 0.006012033671140671, + 0.00808403268456459, + 0.007095085922628641, + 0.00608875323086977, + 0.005724999122321606, + 0.004790008068084717, + 0.006715061143040657, + 0.004410933703184128, + 0.0033931962680071592, + 0.0048790136352181435, + 0.0058143045753240585, + 0.004131908994168043, + 0.0047386870719492435, + 0.006252285558730364, + 0.004364539869129658, + 0.0027993638068437576, + 0.0021814575884491205, + 0.0024184752255678177, + 0.002298944164067507, + 0.003640643320977688, + 0.006809177342802286 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 7 + ], + "selected_reward": -0.3135037124156952 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "smooth tracking shot sweeps slowly from left to right, revealing a chaotic laboratory imbued with a heavy, industrial feel. The camera first passes over dusty concrete walls and exposed metal ducts in the ceiling, the low hum of rusted vents. A tangle of wires, like vines, entangles various delicate machines—an old oscilloscope screen flickers an erratic green waveform, while a centrifuge spins steadily in the corner. As the camera pans to the center of the laboratory, it finally focuses on a metal workbench cluttered with tools, beakers, and scribbled blueprints. Bent over it is a scientist with graying hair, as unkempt as a clump of grass. He wears a white lab coat stained with oil and an unidentified substance, his sleeves rolled up high, revealing his muscular, tense forearms. His bloodshot eyes, bloodshot from chronic sleep deprivation, gleamed with a feverish glint as he wielded a heavy brass mortar, exerting all his strength to crush, one by one, an irregular lump of otherworldly crystal matter pulsing with an eerie purple light. With each impact, tiny luminous particles and tiny sparks of energy flew forth, briefly illuminating his face, which was slightly distorted by intense concentration. The flat overhead fluorescent lights cast sharp, cold shadows, intertwining with the eerie purple light emanating from the crystal itself. The realistic yet rough details created an atmosphere of tension, danger, and profound creativity.", + "seed": 28190, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000021/group_9/index_5732_seed_28190_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": -0.21827764436602592, + "vlm_logprob": -0.21827764436602592, + "vlm_yes_no_margin": 2.15, + "vlm_token_credit": -0.21827764436602592, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" 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 perform a smooth tracking shot sweeping slowly from left to right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a chaotic laboratory with dusty concrete walls, exposed metal ducts, and rusted vents?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a scientist with graying hair wearing a white lab coat stained with oil visible at a metal workbench?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scientist wielding a heavy brass mortar to crush an irregular lump of crystal matter pulsing with eerie purple light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do flat overhead fluorescent lights cast sharp, cold shadows in 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": [ + -0.4956507086753845 + ], + "opposite_logprob_tokens": [ + -0.9956507086753845 + ], + "logprob_reward": -0.4956507086753845, + "yes_no_margin_reward": 0.5, + "token_credit_reward": -0.4956507086753845, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07847283035516739, + 0.08012210577726364, + 0.3961792588233948, + 0.06462334096431732, + 0.0792899951338768, + 0.07590358704328537, + 0.03455117717385292, + 0.03181023523211479, + 0.058331478387117386, + 0.10071597993373871 + ], + "credit_map": [ + 0.011292682029306889, + 0.011155773885548115, + 0.009808492846786976, + 0.008308542892336845, + 0.00978858396410942, + 0.00819967407733202, + 0.007515986915677786, + 0.012403097003698349, + 0.010673938319087029, + 0.007011585868895054, + 0.009811267256736755, + 0.009896256029605865, + 0.008237295784056187, + 0.012422885745763779, + 0.011532340198755264, + 0.010536536574363708, + 0.033053070306777954, + 0.014521989040076733, + 0.13793082535266876, + 0.05694753676652908, + 0.013484103605151176, + 0.015702057629823685, + 0.03149797022342682, + 0.09304173290729523, + 0.012703877873718739, + 0.007762271910905838, + 0.008801980875432491, + 0.006811031140387058, + 0.007085839286446571, + 0.007491156458854675, + 0.006796656642109156, + 0.007170530967414379, + 0.010921323671936989, + 0.014886504970490932, + 0.009945441037416458, + 0.01083142776042223, + 0.010107745416462421, + 0.007684777025133371, + 0.008722851984202862, + 0.0061899228021502495, + 0.006567341275513172, + 0.008699612691998482, + 0.012994795106351376, + 0.005930169951170683, + 0.01040714792907238, + 0.014367819763720036, + 0.00876505859196186, + 0.008171641267836094, + 0.003948016557842493, + 0.004953696858137846, + 0.009319586679339409, + 0.003582220757380128, + 0.00446260767057538, + 0.002615556353703141, + 0.0031493580900132656, + 0.0025201328098773956, + 0.004876980558037758, + 0.0018864694284275174, + 0.005269129294902086, + 0.0023852954618632793, + 0.003674132516607642, + 0.00839619804173708, + 0.002632871503010392, + 0.002689158311113715, + 0.008600093424320221, + 0.005353779997676611, + 0.013774513266980648, + 0.008507288061082363, + 0.008327987045049667, + 0.004208832513540983, + 0.005974352825433016, + 0.0035846359096467495, + 0.017714031040668488, + 0.0037376468535512686, + 0.04329376295208931, + 0.016279662027955055, + 0.004248111508786678, + 0.004655767697840929, + 0.005368195008486509, + 0.005418808199465275 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 9, + 1 + ], + "selected_reward": -0.4956507086753845 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" 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 perform a smooth tracking shot sweeping slowly from left to right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a chaotic laboratory with dusty concrete walls, exposed metal ducts, and rusted vents?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a scientist with graying hair wearing a white lab coat stained with oil visible at a metal workbench?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scientist wielding a heavy brass mortar to crush an irregular lump of crystal matter pulsing with eerie purple light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do flat overhead fluorescent lights cast sharp, cold shadows in 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": [ + -0.043099939823150635 + ], + "opposite_logprob_tokens": [ + -3.168099880218506 + ], + "logprob_reward": -0.043099939823150635, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.043099939823150635, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.26671159267425537, + 0.07706266641616821, + 0.04431198909878731, + 0.04103753715753555, + 0.04754720255732536, + 0.18702952563762665, + 0.06664986163377762, + 0.14377157390117645, + 0.09155530482530594, + 0.03432276099920273 + ], + "credit_map": [ + 0.02954610250890255, + 0.029839666560292244, + 0.027959151193499565, + 0.03228500857949257, + 0.0332813560962677, + 0.046032845973968506, + 0.032469742000103, + 0.035297691822052, + 0.008931511081755161, + 0.005452918354421854, + 0.015055551193654537, + 0.011721394956111908, + 0.00593835161998868, + 0.009592433460056782, + 0.01043770182877779, + 0.009932806715369225, + 0.003927870653569698, + 0.004272120073437691, + 0.00807910691946745, + 0.003918401896953583, + 0.0058974772691726685, + 0.0046706367284059525, + 0.005514147691428661, + 0.008032228797674179, + 0.006958688143640757, + 0.005588244646787643, + 0.005679262802004814, + 0.0048552039079368114, + 0.004523883573710918, + 0.005145673174411058, + 0.003496072720736265, + 0.004790507256984711, + 0.004497510381042957, + 0.009780886583030224, + 0.004818071611225605, + 0.004482688847929239, + 0.006541626062244177, + 0.004227927420288324, + 0.004802356939762831, + 0.008396132849156857, + 0.016719115898013115, + 0.01555532868951559, + 0.02032443881034851, + 0.016843605786561966, + 0.046219710260629654, + 0.03696748986840248, + 0.019399287179112434, + 0.015000540763139725, + 0.00731769809499383, + 0.007756717503070831, + 0.02079158090054989, + 0.006196143105626106, + 0.007118153851479292, + 0.005889293272048235, + 0.006279478780925274, + 0.005300796590745449, + 0.027162199839949608, + 0.004650560207664967, + 0.01004030741751194, + 0.004848754499107599, + 0.011167301796376705, + 0.06147237494587898, + 0.006488349288702011, + 0.017941728234291077, + 0.010308068245649338, + 0.010007599368691444, + 0.02400342747569084, + 0.015896128490567207, + 0.007840163074433804, + 0.010233967565000057, + 0.007067020516842604, + 0.006198932882398367, + 0.0050575826317071915, + 0.0032596972305327654, + 0.011197445914149284, + 0.0057100458070635796, + 0.0021769683808088303, + 0.002514155115932226, + 0.002338897669687867, + 0.0020679677836596966 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 7 + ], + "selected_reward": -0.043099939823150635 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" 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 perform a smooth tracking shot sweeping slowly from left to right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a chaotic laboratory with dusty concrete walls, exposed metal ducts, and rusted vents?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a scientist with graying hair wearing a white lab coat stained with oil visible at a metal workbench?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scientist wielding a heavy brass mortar to crush an irregular lump of crystal matter pulsing with eerie purple light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do flat overhead fluorescent lights cast sharp, cold shadows in 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": [ + -0.02330542728304863 + ], + "opposite_logprob_tokens": [ + -3.7733054161071777 + ], + "logprob_reward": -0.02330542728304863, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.02330542728304863, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06996253877878189, + 0.08530241996049881, + 0.32507291436195374, + 0.04040951281785965, + 0.028175804764032364, + 0.08074011653661728, + 0.06713781505823135, + 0.16743960976600647, + 0.0401882641017437, + 0.09557101875543594 + ], + "credit_map": [ + 0.006546665448695421, + 0.006271441467106342, + 0.008646667003631592, + 0.00801374763250351, + 0.007691667415201664, + 0.012139590457081795, + 0.009641182608902454, + 0.0110115772113204, + 0.00880307424813509, + 0.005996514577418566, + 0.02329874411225319, + 0.010046139359474182, + 0.006603766232728958, + 0.009970282204449177, + 0.00902959331870079, + 0.011554311960935593, + 0.027644136920571327, + 0.011785609647631645, + 0.13013499975204468, + 0.012873399071395397, + 0.015514248982071877, + 0.014650987461209297, + 0.024628587067127228, + 0.08784094452857971, + 0.0051598683930933475, + 0.005530975759029388, + 0.006236695684492588, + 0.004646775778383017, + 0.005624830722808838, + 0.006504879333078861, + 0.0021596834994852543, + 0.004545801784843206, + 0.003712767967954278, + 0.002558179432526231, + 0.0044390494003891945, + 0.004108556546270847, + 0.0035209336783736944, + 0.0025948146358132362, + 0.0030004470609128475, + 0.004241057671606541, + 0.0068216328509151936, + 0.01089922059327364, + 0.00785739254206419, + 0.0060777124017477036, + 0.017690006643533707, + 0.018892595544457436, + 0.007673345971852541, + 0.004828214645385742, + 0.006654517725110054, + 0.007555679883807898, + 0.009281834587454796, + 0.007720249705016613, + 0.012349634431302547, + 0.00681930361315608, + 0.011292667128145695, + 0.0054639289155602455, + 0.03440604731440544, + 0.006328103132545948, + 0.012602537870407104, + 0.006642812862992287, + 0.015656258910894394, + 0.07028405368328094, + 0.006924736779183149, + 0.014595059677958488, + 0.006232359446585178, + 0.0034098895266652107, + 0.012095453217625618, + 0.005179141648113728, + 0.003638782538473606, + 0.002954588970169425, + 0.0031881285831332207, + 0.0034899183083325624, + 0.01826881431043148, + 0.0031157394405454397, + 0.042437173426151276, + 0.012983135879039764, + 0.003274002578109503, + 0.004383328836411238, + 0.005722899921238422, + 0.0053859250620007515 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 7, + 9 + ], + "selected_reward": -0.02330542728304863 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" 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 perform a smooth tracking shot sweeping slowly from left to right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a chaotic laboratory with dusty concrete walls, exposed metal ducts, and rusted vents?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a scientist with graying hair wearing a white lab coat stained with oil visible at a metal workbench?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scientist wielding a heavy brass mortar to crush an irregular lump of crystal matter pulsing with eerie purple light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do flat overhead fluorescent lights cast sharp, cold shadows in 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": [ + -0.0549725741147995 + ], + "opposite_logprob_tokens": [ + -2.9299726486206055 + ], + "logprob_reward": -0.0549725741147995, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.0549725741147995, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23803338408470154, + 0.10022829473018646, + 0.05566669628024101, + 0.02928842417895794, + 0.03451303020119667, + 0.09973011165857315, + 0.13121472299098969, + 0.17287206649780273, + 0.028689458966255188, + 0.10976381599903107 + ], + "credit_map": [ + 0.0170485507696867, + 0.02677658386528492, + 0.028310570865869522, + 0.022425124421715736, + 0.037187039852142334, + 0.034134212881326675, + 0.030422337353229523, + 0.04172895848751068, + 0.01576012186706066, + 0.00982463639229536, + 0.016978485509753227, + 0.01235663890838623, + 0.008729812689125538, + 0.011170484125614166, + 0.009824847802519798, + 0.015583268366754055, + 0.0047842669300735, + 0.0053795939311385155, + 0.014362171292304993, + 0.007452741265296936, + 0.0044548301957547665, + 0.003745965426787734, + 0.005506481509655714, + 0.009980639442801476, + 0.004197224043309689, + 0.0035692693199962378, + 0.005259865429252386, + 0.0037629834841936827, + 0.0032087897416204214, + 0.004102982580661774, + 0.002534498693421483, + 0.0026528097223490477, + 0.004024517256766558, + 0.004997156094759703, + 0.00412584887817502, + 0.006470680236816406, + 0.00432196632027626, + 0.0026079427916556597, + 0.005071381106972694, + 0.002893533557653427, + 0.006782343611121178, + 0.013444090262055397, + 0.02462531439960003, + 0.013264517299830914, + 0.007958551868796349, + 0.0141534935683012, + 0.010879741981625557, + 0.008622055873274803, + 0.015036202035844326, + 0.016463739797472954, + 0.0327017642557621, + 0.013533743098378181, + 0.01909354142844677, + 0.008254698477685452, + 0.016328204423189163, + 0.00980281550437212, + 0.03108794242143631, + 0.00936802662909031, + 0.026427004486322403, + 0.009029584936797619, + 0.015962868928909302, + 0.05698724091053009, + 0.009709052741527557, + 0.0143003324046731, + 0.004286580253392458, + 0.0038197787944227457, + 0.005012572277337313, + 0.003396663349121809, + 0.0030673404689878225, + 0.0028518119361251593, + 0.003757509868592024, + 0.0024971989914774895, + 0.02219737134873867, + 0.004273927770555019, + 0.04799570515751839, + 0.01211115438491106, + 0.004632886033505201, + 0.005610049702227116, + 0.006619837135076523, + 0.006322877015918493 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 6 + ], + "selected_reward": -0.0549725741147995 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" 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 perform a smooth tracking shot sweeping slowly from left to right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a chaotic laboratory with dusty concrete walls, exposed metal ducts, and rusted vents?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a scientist with graying hair wearing a white lab coat stained with oil visible at a metal workbench?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scientist wielding a heavy brass mortar to crush an irregular lump of crystal matter pulsing with eerie purple light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do flat overhead fluorescent lights cast sharp, cold shadows in 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": [ + -0.47435957193374634 + ], + "opposite_logprob_tokens": [ + -0.9743595719337463 + ], + "logprob_reward": -0.47435957193374634, + "yes_no_margin_reward": 0.5, + "token_credit_reward": -0.47435957193374634, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11590421199798584, + 0.07481023669242859, + 0.22031766176223755, + 0.021180979907512665, + 0.03420603275299072, + 0.08225800842046738, + 0.07399668544530869, + 0.11507561057806015, + 0.06604940444231033, + 0.19620120525360107 + ], + "credit_map": [ + 0.011826863512396812, + 0.011028449982404709, + 0.01393614336848259, + 0.011402398347854614, + 0.01564188301563263, + 0.02348882332444191, + 0.014808163046836853, + 0.013771485537290573, + 0.009795072488486767, + 0.0048987651243805885, + 0.018718171864748, + 0.008448110893368721, + 0.0071076941676437855, + 0.008659953251481056, + 0.007630040403455496, + 0.009552424773573875, + 0.018357068300247192, + 0.007930572144687176, + 0.09157454967498779, + 0.009267497807741165, + 0.008959741331636906, + 0.008387022651731968, + 0.01557394303381443, + 0.06026726961135864, + 0.002253491198644042, + 0.002862235065549612, + 0.0021128531079739332, + 0.0028554864693433046, + 0.0024847672320902348, + 0.0038682539016008377, + 0.0022434971760958433, + 0.0025003959890455008, + 0.003946653567254543, + 0.0038208793848752975, + 0.0043382043950259686, + 0.0045875380747020245, + 0.0059549156576395035, + 0.003876244183629751, + 0.0032959189265966415, + 0.004385678097605705, + 0.0068352967500686646, + 0.007224529981613159, + 0.008102894760668278, + 0.01045326516032219, + 0.016148779541254044, + 0.01379956305027008, + 0.008963522501289845, + 0.010730152949690819, + 0.00832383707165718, + 0.009090485982596874, + 0.010360951535403728, + 0.009505677968263626, + 0.012940412387251854, + 0.006431919056922197, + 0.011272832751274109, + 0.0060705686919391155, + 0.02295064367353916, + 0.0044066342525184155, + 0.008827007375657558, + 0.004610888659954071, + 0.009860868565738201, + 0.04881822690367699, + 0.004441339522600174, + 0.01115999836474657, + 0.011145716533064842, + 0.0028090905398130417, + 0.026351293548941612, + 0.00971557293087244, + 0.0035318464506417513, + 0.004325571469962597, + 0.004138630349189043, + 0.004031681455671787, + 0.034514930099248886, + 0.005336306057870388, + 0.08572480082511902, + 0.035650722682476044, + 0.006428806576877832, + 0.008929316885769367, + 0.0098141273483634, + 0.009802187792956829 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 9, + 0 + ], + "selected_reward": -0.47435957193374634 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "smooth tracking shot sweeps slowly from left to right, revealing a chaotic laboratory imbued with a heavy, industrial feel. The camera first passes over dusty concrete walls and exposed metal ducts in the ceiling, the low hum of rusted vents. A tangle of wires, like vines, entangles various delicate machines—an old oscilloscope screen flickers an erratic green waveform, while a centrifuge spins steadily in the corner. As the camera pans to the center of the laboratory, it finally focuses on a metal workbench cluttered with tools, beakers, and scribbled blueprints. Bent over it is a scientist with graying hair, as unkempt as a clump of grass. He wears a white lab coat stained with oil and an unidentified substance, his sleeves rolled up high, revealing his muscular, tense forearms. His bloodshot eyes, bloodshot from chronic sleep deprivation, gleamed with a feverish glint as he wielded a heavy brass mortar, exerting all his strength to crush, one by one, an irregular lump of otherworldly crystal matter pulsing with an eerie purple light. With each impact, tiny luminous particles and tiny sparks of energy flew forth, briefly illuminating his face, which was slightly distorted by intense concentration. The flat overhead fluorescent lights cast sharp, cold shadows, intertwining with the eerie purple light emanating from the crystal itself. The realistic yet rough details created an atmosphere of tension, danger, and profound creativity.", + "seed": 28191, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000021/group_9/index_5732_seed_28191_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": -0.1428815893828869, + "vlm_logprob": -0.1428815893828869, + "vlm_yes_no_margin": 2.35, + "vlm_token_credit": -0.1428815893828869, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" 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 perform a smooth tracking shot sweeping slowly from left to right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a chaotic laboratory with dusty concrete walls, exposed metal ducts, and rusted vents?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a scientist with graying hair wearing a white lab coat stained with oil visible at a metal workbench?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scientist wielding a heavy brass mortar to crush an irregular lump of crystal matter pulsing with eerie purple light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do flat overhead fluorescent lights cast sharp, cold shadows in 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": [ + -0.30623897910118103 + ], + "opposite_logprob_tokens": [ + -1.4312390089035034 + ], + "logprob_reward": -0.30623897910118103, + "yes_no_margin_reward": 1.125, + "token_credit_reward": -0.30623897910118103, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1901800036430359, + 0.10540628433227539, + 0.13979582488536835, + 0.13125070929527283, + 0.06869196146726608, + 0.1506805419921875, + 0.04603012278676033, + 0.08582188189029694, + 0.05710211023688316, + 0.02504059113562107 + ], + "credit_map": [ + 0.029510602355003357, + 0.028023358434438705, + 0.03055248223245144, + 0.014778461307287216, + 0.02233738824725151, + 0.015496370382606983, + 0.01697109453380108, + 0.03251023590564728, + 0.009147299453616142, + 0.010362505912780762, + 0.021998317912220955, + 0.015545603819191456, + 0.009814361110329628, + 0.00919047649949789, + 0.013894141651690006, + 0.015453570522367954, + 0.015526700764894485, + 0.014893746934831142, + 0.018533596768975258, + 0.024154847487807274, + 0.015917222946882248, + 0.012971291318535805, + 0.019925115630030632, + 0.017873307690024376, + 0.012197194620966911, + 0.022422736510634422, + 0.016950825229287148, + 0.019689705222845078, + 0.018925530835986137, + 0.01933467760682106, + 0.012495432049036026, + 0.009234603494405746, + 0.009541252627968788, + 0.008371389470994473, + 0.008371211588382721, + 0.004533432889729738, + 0.005821953061968088, + 0.008679013699293137, + 0.007987764663994312, + 0.015385941602289677, + 0.022763965651392937, + 0.01543762069195509, + 0.024469740688800812, + 0.019115852192044258, + 0.011275511234998703, + 0.03141222149133682, + 0.012861971743404865, + 0.013343649916350842, + 0.006192338652908802, + 0.005678921937942505, + 0.005254009272903204, + 0.006962986662983894, + 0.004154330585151911, + 0.008690773509442806, + 0.0044072032906115055, + 0.004689555615186691, + 0.010255799628794193, + 0.023519549518823624, + 0.00840199738740921, + 0.00984850712120533, + 0.005373896565288305, + 0.0172289852052927, + 0.005187138449400663, + 0.0060060033574700356, + 0.014969626441597939, + 0.00836839247494936, + 0.007405908312648535, + 0.005755310412496328, + 0.004070667549967766, + 0.009081067517399788, + 0.0037038216833025217, + 0.003747316310182214, + 0.0037537766620516777, + 0.0043158396147191525, + 0.002659901976585388, + 0.004051429685205221, + 0.0020349924452602863, + 0.003591646673157811, + 0.0026120468974113464, + 0.0020209543872624636 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 2 + ], + "selected_reward": -0.30623897910118103 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" 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 perform a smooth tracking shot sweeping slowly from left to right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a chaotic laboratory with dusty concrete walls, exposed metal ducts, and rusted vents?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a scientist with graying hair wearing a white lab coat stained with oil visible at a metal workbench?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scientist wielding a heavy brass mortar to crush an irregular lump of crystal matter pulsing with eerie purple light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do flat overhead fluorescent lights cast sharp, cold shadows in 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": [ + -0.033720921725034714 + ], + "opposite_logprob_tokens": [ + -3.4087209701538086 + ], + "logprob_reward": -0.033720921725034714, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.033720921725034714, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.6398161053657532, + 0.06785029917955399, + 0.0419497974216938, + 0.022786961868405342, + 0.04797973111271858, + 0.05172530561685562, + 0.016524802893400192, + 0.08609003573656082, + 0.015867477282881737, + 0.009409439750015736 + ], + "credit_map": [ + 0.06068870797753334, + 0.03771238401532173, + 0.14693577587604523, + 0.03403829038143158, + 0.08631779998540878, + 0.029532847926020622, + 0.02443413995206356, + 0.2201562374830246, + 0.007088744547218084, + 0.0077058798633515835, + 0.014751844108104706, + 0.007751266937702894, + 0.007280416786670685, + 0.005830469075590372, + 0.010726866312325, + 0.006714812479913235, + 0.003552217036485672, + 0.004101487807929516, + 0.006532788276672363, + 0.005571845453232527, + 0.00601821206510067, + 0.005308903753757477, + 0.006908463779836893, + 0.003955879248678684, + 0.002024135086685419, + 0.0047058602795004845, + 0.0026280926540493965, + 0.0027436406817287207, + 0.0027963765896856785, + 0.002709857653826475, + 0.003243807004764676, + 0.001935190986841917, + 0.008559467270970345, + 0.0070952679961919785, + 0.006415828131139278, + 0.003097656648606062, + 0.00690828450024128, + 0.007719393819570541, + 0.0030396946240216494, + 0.005144137889146805, + 0.010091010481119156, + 0.003956667613238096, + 0.01658760756254196, + 0.004502505995333195, + 0.0030378098599612713, + 0.007910560816526413, + 0.0027663311921060085, + 0.002872810699045658, + 0.0019864162895828485, + 0.0017157165566459298, + 0.0017136818496510386, + 0.0025308034382760525, + 0.002095371950417757, + 0.0022970223799347878, + 0.002430733758956194, + 0.0017550578340888023, + 0.0111755495890975, + 0.022401541471481323, + 0.008627494797110558, + 0.008964036591351032, + 0.005379421636462212, + 0.017222296446561813, + 0.006019947584718466, + 0.006299749482423067, + 0.003212149953469634, + 0.0023996569216251373, + 0.0018228052649646997, + 0.0018730012234300375, + 0.0013936390168964863, + 0.002147287130355835, + 0.001257335883565247, + 0.0017616016557440162, + 0.0017921709222719073, + 0.001224595820531249, + 0.0011187995551154017, + 0.0014552309876307845, + 0.0008900760440155864, + 0.0013876133598387241, + 0.0008567029144614935, + 0.0006842498551122844 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 1 + ], + "selected_reward": -0.033720921725034714 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the camera perform a smooth tracking shot sweeping slowly from left to right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a chaotic laboratory with dusty concrete walls, exposed metal ducts, and rusted vents?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a scientist with graying hair wearing a white lab coat stained with oil visible at a metal workbench?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scientist wielding a heavy brass mortar to crush an irregular lump of crystal matter pulsing with eerie purple light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do flat overhead fluorescent lights cast sharp, cold shadows in 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": [ + -0.0381123311817646 + ], + "opposite_logprob_tokens": [ + -3.2881124019622803 + ], + "logprob_reward": -0.0381123311817646, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.0381123311817646, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2908402383327484, + 0.05375397205352783, + 0.028680451214313507, + 0.1373579204082489, + 0.0299247894436121, + 0.10613744705915451, + 0.040551476180553436, + 0.2675555944442749, + 0.026574797928333282, + 0.01862328313291073 + ], + "credit_map": [ + 0.02465825527906418, + 0.01917930319905281, + 0.06725945323705673, + 0.01707010343670845, + 0.036762986332178116, + 0.01838800311088562, + 0.010969455353915691, + 0.09655270725488663, + 0.005344490986317396, + 0.005674388725310564, + 0.009436622262001038, + 0.005366218741983175, + 0.005607516970485449, + 0.0061021908186376095, + 0.009490846656262875, + 0.006731705740094185, + 0.0036287319380789995, + 0.002688039094209671, + 0.004783933982253075, + 0.004196528345346451, + 0.003671718994155526, + 0.003094831947237253, + 0.003806525841355324, + 0.0028101408388465643, + 0.011655587702989578, + 0.02615690976381302, + 0.01739703118801117, + 0.02344323694705963, + 0.01842002011835575, + 0.014200208708643913, + 0.015892697498202324, + 0.010192242451012135, + 0.0040411269292235374, + 0.004795519169420004, + 0.00445463927462697, + 0.0023516814690083265, + 0.003995508886873722, + 0.004481873475015163, + 0.002610178431496024, + 0.0031942680943757296, + 0.016496701166033745, + 0.01679939590394497, + 0.014288481324911118, + 0.00815669260919094, + 0.010185868479311466, + 0.021793177351355553, + 0.00951992068439722, + 0.008897227235138416, + 0.005851763300597668, + 0.005395912565290928, + 0.0046513876877725124, + 0.006569256074726582, + 0.003268529660999775, + 0.007699647918343544, + 0.003635358763858676, + 0.0034796239342540503, + 0.03471339866518974, + 0.06675698608160019, + 0.027270587161183357, + 0.028097743168473244, + 0.022535841912031174, + 0.051796313375234604, + 0.015466798096895218, + 0.020917948335409164, + 0.0055317748337984085, + 0.003065074561163783, + 0.0031371775548905134, + 0.003467777743935585, + 0.0021489826031029224, + 0.003819170640781522, + 0.003557869466021657, + 0.0018469708738848567, + 0.0029618300031870604, + 0.002229216741397977, + 0.001903350348584354, + 0.0030696953181177378, + 0.0016333491075783968, + 0.0028363452292978764, + 0.002466073026880622, + 0.0015234240563586354 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 3 + ], + "selected_reward": -0.0381123311817646 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" 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 perform a smooth tracking shot sweeping slowly from left to right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a chaotic laboratory with dusty concrete walls, exposed metal ducts, and rusted vents?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a scientist with graying hair wearing a white lab coat stained with oil visible at a metal workbench?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scientist wielding a heavy brass mortar to crush an irregular lump of crystal matter pulsing with eerie purple light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do flat overhead fluorescent lights cast sharp, cold shadows in 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": [ + -0.05496874079108238 + ], + "opposite_logprob_tokens": [ + -2.92996883392334 + ], + "logprob_reward": -0.05496874079108238, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05496874079108238, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5774797201156616, + 0.03255671262741089, + 0.02488570474088192, + 0.03429911285638809, + 0.023220621049404144, + 0.042061422020196915, + 0.018006717786192894, + 0.20899227261543274, + 0.018515238538384438, + 0.019982509315013885 + ], + "credit_map": [ + 0.04697040095925331, + 0.028358690440654755, + 0.13937069475650787, + 0.028076941147446632, + 0.07992873340845108, + 0.025081396102905273, + 0.017708327621221542, + 0.21198450028896332, + 0.0042724208906292915, + 0.0035613509826362133, + 0.005237195175141096, + 0.004736820235848427, + 0.003356026951223612, + 0.0029769090469926596, + 0.0041302344761788845, + 0.004285759292542934, + 0.0018969322554767132, + 0.0031636550556868315, + 0.00351129868067801, + 0.003915121778845787, + 0.004579884000122547, + 0.002356954850256443, + 0.0032515868078917265, + 0.002210274338722229, + 0.00210567656904459, + 0.005590452812612057, + 0.003981269430369139, + 0.004109070170670748, + 0.003949672449380159, + 0.005522712599486113, + 0.004647229798138142, + 0.00439302995800972, + 0.0035459401551634073, + 0.004828058183193207, + 0.0030517796985805035, + 0.0014674377162009478, + 0.00213087210431695, + 0.0026445251423865557, + 0.0027016843669116497, + 0.002850324846804142, + 0.006729844957590103, + 0.00495015038177371, + 0.005235360469669104, + 0.004966704174876213, + 0.0038596841040998697, + 0.007351584732532501, + 0.004571439698338509, + 0.00439665699377656, + 0.0017389497952535748, + 0.0017897802172228694, + 0.0023679945152252913, + 0.002454133704304695, + 0.0019253287464380264, + 0.004257263615727425, + 0.0019527815748006105, + 0.0015204866649582982, + 0.023858943954110146, + 0.054263342171907425, + 0.018791094422340393, + 0.023034198209643364, + 0.014663580805063248, + 0.040689412504434586, + 0.01602870598435402, + 0.017663007602095604, + 0.0016989841824397445, + 0.0029854390304535627, + 0.001443653367459774, + 0.002768102800473571, + 0.002728759776800871, + 0.002739701885730028, + 0.002580539556220174, + 0.0015700594522058964, + 0.0015810142504051328, + 0.00225872197188437, + 0.0016095833852887154, + 0.0041169896721839905, + 0.0020094187930226326, + 0.0032354961149394512, + 0.0027622226625680923, + 0.0024090625811368227 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 5 + ], + "selected_reward": -0.05496874079108238 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" 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 perform a smooth tracking shot sweeping slowly from left to right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a chaotic laboratory with dusty concrete walls, exposed metal ducts, and rusted vents?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a scientist with graying hair wearing a white lab coat stained with oil visible at a metal workbench?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scientist wielding a heavy brass mortar to crush an irregular lump of crystal matter pulsing with eerie purple light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do flat overhead fluorescent lights cast sharp, cold shadows in 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": [ + -0.2813669741153717 + ], + "opposite_logprob_tokens": [ + -1.4063669443130493 + ], + "logprob_reward": -0.2813669741153717, + "yes_no_margin_reward": 1.125, + "token_credit_reward": -0.2813669741153717, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.6475120186805725, + 0.02933982014656067, + 0.04725702479481697, + 0.0230554286390543, + 0.052941396832466125, + 0.09069664776325226, + 0.020333142951130867, + 0.05528965964913368, + 0.01776573248207569, + 0.015809176489710808 + ], + "credit_map": [ + 0.07581748068332672, + 0.0410241074860096, + 0.14165662229061127, + 0.04245241731405258, + 0.06720402091741562, + 0.0722578689455986, + 0.06598501652479172, + 0.14111439883708954, + 0.0028297791723161936, + 0.003151859389618039, + 0.003624282544478774, + 0.003825177438557148, + 0.003950092010200024, + 0.0030912794172763824, + 0.004034444224089384, + 0.004832903854548931, + 0.002931961789727211, + 0.004132580012083054, + 0.00938885472714901, + 0.013585204258561134, + 0.005171529483050108, + 0.0021329671144485474, + 0.005679672583937645, + 0.004234247840940952, + 0.0021872082725167274, + 0.0026514504570513964, + 0.002674428978934884, + 0.004102572798728943, + 0.0033007876481860876, + 0.003373638028278947, + 0.002139338757842779, + 0.002626000437885523, + 0.014872362837195396, + 0.008761925622820854, + 0.008268623612821102, + 0.0033527591731399298, + 0.00369601184502244, + 0.006323406007140875, + 0.003345447825267911, + 0.004320852924138308, + 0.018561655655503273, + 0.010183733887970448, + 0.01501969713717699, + 0.009618012234568596, + 0.006237720604985952, + 0.017509911209344864, + 0.005987242329865694, + 0.00757866445928812, + 0.0022270355839282274, + 0.002824996132403612, + 0.0018994805868715048, + 0.002883452456444502, + 0.0028287621680647135, + 0.00326489913277328, + 0.002490547951310873, + 0.0019139668438583612, + 0.007381423842161894, + 0.014406370930373669, + 0.005509961396455765, + 0.006269688252359629, + 0.004327108152210712, + 0.010632839053869247, + 0.0033013608772307634, + 0.0034609015565365553, + 0.0025472668930888176, + 0.003267382038757205, + 0.0014745790977030993, + 0.002821794943884015, + 0.0017236450221389532, + 0.0022831903770565987, + 0.0018062650924548507, + 0.0018416063394397497, + 0.0038324175402522087, + 0.0017447079299017787, + 0.0019517128821462393, + 0.001215442200191319, + 0.0011099675903096795, + 0.003449466545134783, + 0.0011257384903728962, + 0.0013797221472486854 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 7 + ], + "selected_reward": -0.2813669741153717 + } + ], + "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/0000026/group_1/index_tensor([4563, 4563])_group_1_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000026/group_1/index_tensor([4563, 4563])_group_1_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..782f4bcd78c586bf45e037f3db9c713daa8fafeb --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000026/group_1/index_tensor([4563, 4563])_group_1_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "static, medium-length shot slowly, almost imperceptibly, begins to advance, capturing a deeply intimate moment within the cavernous nave of a magnificent, historic church. A brilliant, kaleidoscopic light pours through soaring, intricate stained-glass windows, painting the cold stone floor and ancient wooden pews with vibrant, shifting patterns of sapphire blue, ruby red, and emerald green. At the center of this ethereal illumination, a man and a woman embrace, creating a single, unified silhouette. Seen from behind, the man has short, dark, slightly disheveled hair and wears a thick, charcoal-gray wool coat; his pose is one of complete surrender, his face buried in her shoulder, his hands pressed firmly against her back. The woman, with her long, flowing auburn hair catching the colorful glow like a fiery halo, wears a simple, long-sleeved cream dress. Her eyes are closed, a tear trickling down her cheek, and as her arms wrap firmly around his neck, her expression is one of calm and overwhelming relief. The air is filled with dancing dust motes, each catching the light and enhancing the sacred and timeless quality of the scene, while slow, deliberate camera movements steadily draw the viewer into the overwhelming, silent, and deeply emotional atmosphere of their reunion.", + "seed": 21864, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000026/group_1/index_4563_seed_21864_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": -0.08227389650419355, + "vlm_logprob": -0.08227389650419355, + "vlm_yes_no_margin": 3.725, + "vlm_token_credit": -0.08227389650419355, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 the nave of a historic church with stone floors and wooden pews?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are there stained-glass windows casting sapphire blue, ruby red, and emerald green light patterns on the floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a man seen from behind wearing a thick charcoal-gray wool coat with short dark hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a woman seen from behind wearing a long-sleeved cream dress with long auburn hair and a tear on her cheek?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly advance forward while the couple embraces in the center 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": [ + -0.32751527428627014 + ], + "opposite_logprob_tokens": [ + -1.3275152444839478 + ], + "logprob_reward": -0.32751527428627014, + "yes_no_margin_reward": 1.0, + "token_credit_reward": -0.32751527428627014, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.6701473593711853, + 0.1122988611459732, + 0.057493217289447784, + 0.025153903290629387, + 0.04721563681960106, + 0.010759901255369186, + 0.012229888699948788, + 0.019890552386641502, + 0.023401103913784027, + 0.021409574896097183 + ], + "credit_map": [ + 0.023839645087718964, + 0.012053553946316242, + 0.011482237838208675, + 0.05475039780139923, + 0.023667098954319954, + 0.2922874391078949, + 0.21243347227573395, + 0.039633505046367645, + 0.012312846258282661, + 0.007044860161840916, + 0.006574331782758236, + 0.0328277088701725, + 0.008689073845744133, + 0.009334967471659184, + 0.023376738652586937, + 0.012138341553509235, + 0.008344672620296478, + 0.004574202466756105, + 0.005120071582496166, + 0.010881143622100353, + 0.005056838970631361, + 0.009121755138039589, + 0.005971308797597885, + 0.008423219434916973, + 0.0034478046000003815, + 0.0017434461042284966, + 0.0017579136183485389, + 0.004847928415983915, + 0.0019260392291471362, + 0.006346692331135273, + 0.0019906535744667053, + 0.003093424718827009, + 0.006341386120766401, + 0.003942470531910658, + 0.007721250876784325, + 0.010555712506175041, + 0.0022106727119535208, + 0.007067505735903978, + 0.004375775810331106, + 0.005000863689929247, + 0.0011452147737145424, + 0.0014500130200758576, + 0.0018677911721169949, + 0.0019265931332483888, + 0.0006608323892578483, + 0.0015721461968496442, + 0.0009952059481292963, + 0.0011421047383919358, + 0.0014187744818627834, + 0.0012472812086343765, + 0.0019848914816975594, + 0.0018468174384906888, + 0.0008740847697481513, + 0.001602657837793231, + 0.001903394004330039, + 0.0013519873609766364, + 0.0011933703208342195, + 0.002274158177897334, + 0.003046599216759205, + 0.003222381928935647, + 0.002597763668745756, + 0.0014489107998088002, + 0.003958487883210182, + 0.0021488796919584274, + 0.0028649477753788233, + 0.0038314592093229294, + 0.002948872512206435, + 0.0031685647554695606, + 0.003807316068559885, + 0.0015494778053835034, + 0.002842459362000227, + 0.0023880056105554104, + 0.002770522376522422, + 0.002604704350233078, + 0.0029487726278603077, + 0.0029755348805338144, + 0.002820574678480625, + 0.0029365953523665667, + 0.0017236998537555337, + 0.0026291694957762957 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.32751527428627014 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 the nave of a historic church with stone floors and wooden pews?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are there stained-glass windows casting sapphire blue, ruby red, and emerald green light patterns on the floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a man seen from behind wearing a thick charcoal-gray wool coat with short dark hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a woman seen from behind wearing a long-sleeved cream dress with long auburn hair and a tear on her cheek?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly advance forward while the couple embraces in the center 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": [ + -0.006779762916266918 + ], + "opposite_logprob_tokens": [ + -5.006779670715332 + ], + "logprob_reward": -0.006779762916266918, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006779762916266918, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.34653106331825256, + 0.08589943498373032, + 0.0544234924018383, + 0.036963511258363724, + 0.05145969241857529, + 0.05045751854777336, + 0.0309162475168705, + 0.023591909557580948, + 0.08938342332839966, + 0.23037369549274445 + ], + "credit_map": [ + 0.015184626914560795, + 0.007160366512835026, + 0.007143530994653702, + 0.02773543819785118, + 0.013627789914608002, + 0.14582104980945587, + 0.10907045751810074, + 0.020787816494703293, + 0.011149750091135502, + 0.005182015709578991, + 0.005163915921002626, + 0.024831051006913185, + 0.005056253168731928, + 0.008011572994291782, + 0.018773993477225304, + 0.00773088913410902, + 0.006046727765351534, + 0.004279902204871178, + 0.00425663311034441, + 0.008726881816983223, + 0.003682927694171667, + 0.010408610105514526, + 0.0074324654415249825, + 0.009589344263076782, + 0.007071367464959621, + 0.0024090909864753485, + 0.0028057631570845842, + 0.009010971523821354, + 0.0017886706627905369, + 0.007520180195569992, + 0.00250569568015635, + 0.0038517692591995, + 0.006830120459198952, + 0.00403138343244791, + 0.005358902271836996, + 0.008204195648431778, + 0.002565502654761076, + 0.012504752725362778, + 0.0053824554197490215, + 0.006582379341125488, + 0.005714336410164833, + 0.005843793973326683, + 0.008265122771263123, + 0.008344964124262333, + 0.0026879243087023497, + 0.009217720478773117, + 0.004128086846321821, + 0.006255570333451033, + 0.003324513789266348, + 0.004257069434970617, + 0.00447695842012763, + 0.005294464994221926, + 0.00257113971747458, + 0.0022966740652918816, + 0.006237743888050318, + 0.0024576818104833364, + 0.0019241905538365245, + 0.002810695208609104, + 0.003973385319113731, + 0.0036622777115553617, + 0.0024571139365434647, + 0.001386186690069735, + 0.005060575902462006, + 0.0023174830712378025, + 0.009472478181123734, + 0.013030712492763996, + 0.008455622009932995, + 0.011439300142228603, + 0.018406683579087257, + 0.007085314020514488, + 0.013004830107092857, + 0.008488484658300877, + 0.03209460526704788, + 0.02936873957514763, + 0.030672209337353706, + 0.035966936498880386, + 0.03277726098895073, + 0.027717843651771545, + 0.02005871944129467, + 0.021717391908168793 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 8 + ], + "selected_reward": -0.006779762916266918 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 the nave of a historic church with stone floors and wooden pews?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are there stained-glass windows casting sapphire blue, ruby red, and emerald green light patterns on the floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a man seen from behind wearing a thick charcoal-gray wool coat with short dark hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a woman seen from behind wearing a long-sleeved cream dress with long auburn hair and a tear on her cheek?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly advance forward while the couple embraces in the center 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": [ + -0.005258300807327032 + ], + "opposite_logprob_tokens": [ + -5.255258083343506 + ], + "logprob_reward": -0.005258300807327032, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005258300807327032, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.6534770727157593, + 0.04497202858328819, + 0.04272417351603508, + 0.031206311658024788, + 0.021259048953652382, + 0.02689262293279171, + 0.041739530861377716, + 0.03615262359380722, + 0.08203990757465363, + 0.019536711275577545 + ], + "credit_map": [ + 0.022213827818632126, + 0.011743091978132725, + 0.011530286632478237, + 0.05349878594279289, + 0.02354687638580799, + 0.2856904864311218, + 0.20683565735816956, + 0.038418062031269073, + 0.005524047650396824, + 0.004131403286010027, + 0.004031109157949686, + 0.00918971560895443, + 0.0036502464208751917, + 0.005543658509850502, + 0.008144126273691654, + 0.004757722374051809, + 0.005291901994496584, + 0.004402428399771452, + 0.00548507971689105, + 0.009147591888904572, + 0.003112224629148841, + 0.004676092881709337, + 0.004977887496352196, + 0.005630968604236841, + 0.0038660955615341663, + 0.003446621587499976, + 0.003932122141122818, + 0.00486652972176671, + 0.0031988786067813635, + 0.0037222637329250574, + 0.004913434386253357, + 0.0032603652216494083, + 0.002604598179459572, + 0.0020467573776841164, + 0.002535272855311632, + 0.004905374255031347, + 0.0015480093425139785, + 0.0020822028163820505, + 0.00330950110219419, + 0.002227334538474679, + 0.003452596254646778, + 0.00394356669858098, + 0.003236564341932535, + 0.004660354927182198, + 0.0018276451155543327, + 0.00392790324985981, + 0.0034669130109250546, + 0.002377079799771309, + 0.0027544826734811068, + 0.0115847522392869, + 0.004849622491747141, + 0.006754344794899225, + 0.004061422310769558, + 0.0026766308583319187, + 0.005894242785871029, + 0.003164030844345689, + 0.00488613685593009, + 0.004667445085942745, + 0.007600813172757626, + 0.005982779897749424, + 0.00269234087318182, + 0.002115569543093443, + 0.005227444693446159, + 0.002980094403028488, + 0.009063556790351868, + 0.013265855610370636, + 0.008274645544588566, + 0.010265011340379715, + 0.015169244259595871, + 0.007726934738457203, + 0.010899185203015804, + 0.007375472690910101, + 0.002822623588144779, + 0.0016664029099047184, + 0.0023159629199653864, + 0.003124599577859044, + 0.003064242424443364, + 0.0021566112991422415, + 0.0022854695562273264, + 0.0021007985342293978 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 1 + ], + "selected_reward": -0.005258300807327032 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 the nave of a historic church with stone floors and wooden pews?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are there stained-glass windows casting sapphire blue, ruby red, and emerald green light patterns on the floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a man seen from behind wearing a thick charcoal-gray wool coat with short dark hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a woman seen from behind wearing a long-sleeved cream dress with long auburn hair and a tear on her cheek?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly advance forward while the couple embraces in the center 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": [ + -0.009783769026398659 + ], + "opposite_logprob_tokens": [ + -4.634783744812012 + ], + "logprob_reward": -0.009783769026398659, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009783769026398659, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.31395119428634644, + 0.07695340365171432, + 0.04985556751489639, + 0.05337231606245041, + 0.06947002559900284, + 0.03465661779046059, + 0.04765205830335617, + 0.05088992789387703, + 0.25244516134262085, + 0.05075377970933914 + ], + "credit_map": [ + 0.013831053860485554, + 0.007519998122006655, + 0.007722131442278624, + 0.0255721528083086, + 0.014909341000020504, + 0.1297875940799713, + 0.09584414213895798, + 0.01876474730670452, + 0.009733766317367554, + 0.006128270179033279, + 0.004285808186978102, + 0.021162105724215508, + 0.005365830846130848, + 0.008452831767499447, + 0.0159151554107666, + 0.005909626837819815, + 0.004910068586468697, + 0.006745132151991129, + 0.004557592794299126, + 0.008435618132352829, + 0.005170359741896391, + 0.005930809304118156, + 0.006984076928347349, + 0.007121906150132418, + 0.011869613081216812, + 0.0032774158753454685, + 0.003480735933408141, + 0.012259752489626408, + 0.0028378136921674013, + 0.009056268259882927, + 0.006443399004638195, + 0.004147315863519907, + 0.00822390429675579, + 0.00782790221273899, + 0.007966536097228527, + 0.015628810971975327, + 0.0038594966754317284, + 0.008788284845650196, + 0.011789032258093357, + 0.005386053118854761, + 0.002797476015985012, + 0.005347920581698418, + 0.0057702879421412945, + 0.0071207815781235695, + 0.0026636288966983557, + 0.001949812052771449, + 0.006768228020519018, + 0.002238482004031539, + 0.0037062978371977806, + 0.010795198380947113, + 0.007073094602674246, + 0.008171219378709793, + 0.003780768718570471, + 0.002810940146446228, + 0.005687054246664047, + 0.005627484992146492, + 0.005660999100655317, + 0.010079212486743927, + 0.00714463135227561, + 0.008391477167606354, + 0.006787388119846582, + 0.002409926615655422, + 0.006774568464607, + 0.0036417206283658743, + 0.025218438357114792, + 0.04184505343437195, + 0.024637693539261818, + 0.03187112137675285, + 0.0523594506084919, + 0.022472882643342018, + 0.03280571475625038, + 0.021234780550003052, + 0.006574925035238266, + 0.005448210518807173, + 0.009032528847455978, + 0.007637360133230686, + 0.008299557492136955, + 0.0044930484145879745, + 0.005003774538636208, + 0.004264365881681442 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 1 + ], + "selected_reward": -0.009783769026398659 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 the nave of a historic church with stone floors and wooden pews?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are there stained-glass windows casting sapphire blue, ruby red, and emerald green light patterns on the floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a man seen from behind wearing a thick charcoal-gray wool coat with short dark hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a woman seen from behind wearing a long-sleeved cream dress with long auburn hair and a tear on her cheek?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly advance forward while the couple embraces in the center 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": [ + -0.06203237548470497 + ], + "opposite_logprob_tokens": [ + -2.812032461166382 + ], + "logprob_reward": -0.06203237548470497, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06203237548470497, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.41875159740448, + 0.13493677973747253, + 0.056871745735406876, + 0.05644718185067177, + 0.04603611305356026, + 0.06410030275583267, + 0.026792941614985466, + 0.10169391334056854, + 0.05775488540530205, + 0.036614540964365005 + ], + "credit_map": [ + 0.015925096347928047, + 0.011610545217990875, + 0.010253717191517353, + 0.03494402393698692, + 0.02148943580687046, + 0.17195837199687958, + 0.1270531862974167, + 0.02551724761724472, + 0.017525693401694298, + 0.009752125479280949, + 0.016130797564983368, + 0.03318120166659355, + 0.0075674839317798615, + 0.009866577573120594, + 0.025650961324572563, + 0.015261939726769924, + 0.008354823105037212, + 0.006716706790030003, + 0.003915687557309866, + 0.011040491983294487, + 0.004820576403290033, + 0.006313548423349857, + 0.009601416997611523, + 0.006108488887548447, + 0.012301396578550339, + 0.004452723544090986, + 0.005146029870957136, + 0.013601030223071575, + 0.004202881827950478, + 0.006112920586019754, + 0.006225875578820705, + 0.0044043236412107944, + 0.003717683954164386, + 0.005516830366104841, + 0.008576284162700176, + 0.008639154955744743, + 0.0031647563446313143, + 0.007148509845137596, + 0.005320661701261997, + 0.003952234052121639, + 0.0033201146870851517, + 0.011149178259074688, + 0.013217003084719181, + 0.012132572941482067, + 0.004897794686257839, + 0.0032379813492298126, + 0.011692172847688198, + 0.004453481640666723, + 0.001826211460866034, + 0.005410662852227688, + 0.0036580516025424004, + 0.004034898243844509, + 0.0025021072942763567, + 0.0026725768111646175, + 0.004132951144129038, + 0.0025554841849952936, + 0.011990715749561787, + 0.015609176829457283, + 0.013372776098549366, + 0.016461938619613647, + 0.010755804367363453, + 0.006872393190860748, + 0.01710786111652851, + 0.00952323991805315, + 0.005742743611335754, + 0.0069955564104020596, + 0.006514314562082291, + 0.008596318773925304, + 0.012439316138625145, + 0.0032637245021760464, + 0.009206150658428669, + 0.004996763542294502, + 0.004945059306919575, + 0.004404171835631132, + 0.004966504871845245, + 0.005112225189805031, + 0.004943687468767166, + 0.0030278137419372797, + 0.0060955192893743515, + 0.0031195594929158688 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 7 + ], + "selected_reward": -0.06203237548470497 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, medium-length shot slowly, almost imperceptibly, begins to advance, capturing a deeply intimate moment within the cavernous nave of a magnificent, historic church. A brilliant, kaleidoscopic light pours through soaring, intricate stained-glass windows, painting the cold stone floor and ancient wooden pews with vibrant, shifting patterns of sapphire blue, ruby red, and emerald green. At the center of this ethereal illumination, a man and a woman embrace, creating a single, unified silhouette. Seen from behind, the man has short, dark, slightly disheveled hair and wears a thick, charcoal-gray wool coat; his pose is one of complete surrender, his face buried in her shoulder, his hands pressed firmly against her back. The woman, with her long, flowing auburn hair catching the colorful glow like a fiery halo, wears a simple, long-sleeved cream dress. Her eyes are closed, a tear trickling down her cheek, and as her arms wrap firmly around his neck, her expression is one of calm and overwhelming relief. The air is filled with dancing dust motes, each catching the light and enhancing the sacred and timeless quality of the scene, while slow, deliberate camera movements steadily draw the viewer into the overwhelming, silent, and deeply emotional atmosphere of their reunion.", + "seed": 21865, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000026/group_1/index_4563_seed_21865_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": -0.05569229014217854, + "vlm_logprob": -0.05569229014217854, + "vlm_yes_no_margin": 3.475, + "vlm_token_credit": -0.05569229014217854, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 the nave of a historic church with stone floors and wooden pews?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are there stained-glass windows casting sapphire blue, ruby red, and emerald green light patterns on the floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a man seen from behind wearing a thick charcoal-gray wool coat with short dark hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a woman seen from behind wearing a long-sleeved cream dress with long auburn hair and a tear on her cheek?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly advance forward while the couple embraces in the center 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": [ + -0.1589922308921814 + ], + "opposite_logprob_tokens": [ + -2.158992290496826 + ], + "logprob_reward": -0.1589922308921814, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.1589922308921814, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03821168094873428, + 0.018562862649559975, + 0.3384624421596527, + 0.012509393505752087, + 0.017417358234524727, + 0.5205499529838562, + 0.00602450966835022, + 0.03691762685775757, + 0.00637809420004487, + 0.004966063424944878 + ], + "credit_map": [ + 0.005410721991211176, + 0.005807571578770876, + 0.005471083801239729, + 0.004429457243531942, + 0.0035973687190562487, + 0.006012436002492905, + 0.004748934879899025, + 0.0027341062668710947, + 0.003205652581527829, + 0.0023263522889465094, + 0.003194065298885107, + 0.0015755569329485297, + 0.001855212147347629, + 0.001879183342680335, + 0.0027628480456769466, + 0.001763991778716445, + 0.01703735813498497, + 0.011401613242924213, + 0.1297692358493805, + 0.01331198401749134, + 0.0166467297822237, + 0.0490078404545784, + 0.08956623077392578, + 0.011721428483724594, + 0.001921761897392571, + 0.0014830287545919418, + 0.0027392858173698187, + 0.0008934124489314854, + 0.0011838666396215558, + 0.0013588949805125594, + 0.0019536782056093216, + 0.0009754644124768674, + 0.002420770702883601, + 0.001905808225274086, + 0.002627237467095256, + 0.001709079253487289, + 0.002161038341000676, + 0.0013318810379132628, + 0.003757959231734276, + 0.001503583393059671, + 0.10507418215274811, + 0.04593285545706749, + 0.05025935173034668, + 0.06222853809595108, + 0.030135247856378555, + 0.11458339542150497, + 0.06113877892494202, + 0.051197599619627, + 0.0005816820776090026, + 0.0011042437981814146, + 0.001031465595588088, + 0.0008158850250765681, + 0.0005300903576426208, + 0.0005226171924732625, + 0.0009317826479673386, + 0.0005067427991889417, + 0.0085791926831007, + 0.002805819036439061, + 0.0022880607284605503, + 0.0026868057902902365, + 0.001424599438905716, + 0.005064650904387236, + 0.004944381769746542, + 0.00912411417812109, + 0.0009505445486865938, + 0.0010692948708310723, + 0.0010181963443756104, + 0.0006274966872297227, + 0.0006931473035365343, + 0.0004485302488319576, + 0.000731979263946414, + 0.0008389045833609998, + 0.0005302870413288474, + 0.00047338343574665487, + 0.0007001294288784266, + 0.0005954034859314561, + 0.0006034611142240465, + 0.0005855981144122779, + 0.00075454858597368, + 0.000723252072930336 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 2, + 0 + ], + "selected_reward": -0.1589922308921814 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 the nave of a historic church with stone floors and wooden pews?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are there stained-glass windows casting sapphire blue, ruby red, and emerald green light patterns on the floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a man seen from behind wearing a thick charcoal-gray wool coat with short dark hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a woman seen from behind wearing a long-sleeved cream dress with long auburn hair and a tear on her cheek?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly advance forward while the couple embraces in the center 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": [ + -0.009843381121754646 + ], + "opposite_logprob_tokens": [ + -4.634843349456787 + ], + "logprob_reward": -0.009843381121754646, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009843381121754646, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.037574268877506256, + 0.022685809060931206, + 0.8422030210494995, + 0.008683503605425358, + 0.01980852335691452, + 0.03577079623937607, + 0.004743799101561308, + 0.015033997595310211, + 0.005027701612561941, + 0.008468563668429852 + ], + "credit_map": [ + 0.005514177959412336, + 0.0042924825102090836, + 0.008998303674161434, + 0.0049675763584673405, + 0.0021099005825817585, + 0.003202454885467887, + 0.004668888635933399, + 0.0038204886950552464, + 0.004965739790350199, + 0.0026340384501963854, + 0.0033904570154845715, + 0.0026710082311183214, + 0.0012434418313205242, + 0.0027751289308071136, + 0.003690559882670641, + 0.001315438305027783, + 0.04266423359513283, + 0.03242874518036842, + 0.3142731487751007, + 0.032892610877752304, + 0.04219996556639671, + 0.14492681622505188, + 0.202945277094841, + 0.029872329905629158, + 0.0014946478186175227, + 0.00102138367947191, + 0.001783200423233211, + 0.0005288716056384146, + 0.0006986543885432184, + 0.0008469366002827883, + 0.001491036731749773, + 0.0008187727071344852, + 0.00256826588883996, + 0.002473806496709585, + 0.0016595956403762102, + 0.0014554570661857724, + 0.002215656219050288, + 0.0016537675401195884, + 0.00587988318875432, + 0.0019020943436771631, + 0.007714580278843641, + 0.003686219686642289, + 0.003066418692469597, + 0.004172704182565212, + 0.0022475060541182756, + 0.007897797040641308, + 0.0037829952780157328, + 0.003202581312507391, + 0.0005196319543756545, + 0.0008604133618064225, + 0.0006683237734250724, + 0.0007049958221614361, + 0.0004376497527118772, + 0.0004429593973327428, + 0.0007120808586478233, + 0.0003977451124228537, + 0.0013780423905700445, + 0.0024215634912252426, + 0.0012697638012468815, + 0.003088184632360935, + 0.0014165844768285751, + 0.0006859652348794043, + 0.0023034955374896526, + 0.0024703999515622854, + 0.0009364784345962107, + 0.0005871658213436604, + 0.0006448146887123585, + 0.0005915952497161925, + 0.0005415055784396827, + 0.0004096955817658454, + 0.0006759894895367324, + 0.0006404570303857327, + 0.0008697058074176311, + 0.000808567798230797, + 0.0016828262014314532, + 0.0013310626382008195, + 0.0009124961798079312, + 0.0006321262917481363, + 0.0010339656146243215, + 0.0011978138936683536 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 5 + ], + "selected_reward": -0.009843381121754646 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 the nave of a historic church with stone floors and wooden pews?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are there stained-glass windows casting sapphire blue, ruby red, and emerald green light patterns on the floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a man seen from behind wearing a thick charcoal-gray wool coat with short dark hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a woman seen from behind wearing a long-sleeved cream dress with long auburn hair and a tear on her cheek?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly advance forward while the couple embraces in the center 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": [ + -0.009794864803552628 + ], + "opposite_logprob_tokens": [ + -4.63479471206665 + ], + "logprob_reward": -0.009794864803552628, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009794864803552628, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09763028472661972, + 0.04144221916794777, + 0.4522421956062317, + 0.02733629383146763, + 0.036338865756988525, + 0.11632505059242249, + 0.026972291991114616, + 0.1331256777048111, + 0.03857804089784622, + 0.030009092763066292 + ], + "credit_map": [ + 0.0171648021787405, + 0.011136004701256752, + 0.012669001705944538, + 0.007194700185209513, + 0.014866541139781475, + 0.009503815323114395, + 0.015816569328308105, + 0.009278849698603153, + 0.0074604833498597145, + 0.004121650476008654, + 0.007674954365938902, + 0.0041099656373262405, + 0.004509382881224155, + 0.005273010116070509, + 0.006014970596879721, + 0.0022778043057769537, + 0.02319273352622986, + 0.01830337755382061, + 0.1667795479297638, + 0.01732090301811695, + 0.023966465145349503, + 0.0813162699341774, + 0.1046033576130867, + 0.016759512946009636, + 0.005098613910377026, + 0.0035158153623342514, + 0.003114477265626192, + 0.0018975419225171208, + 0.0026067739818245173, + 0.004845169372856617, + 0.0037252523470669985, + 0.0025326504837721586, + 0.00503958947956562, + 0.0038249443750828505, + 0.0033425043802708387, + 0.002782569732517004, + 0.004539881367236376, + 0.0022788667120039463, + 0.009820353239774704, + 0.004710158333182335, + 0.020235691219568253, + 0.01592489890754223, + 0.008701633661985397, + 0.012533605098724365, + 0.01236057747155428, + 0.021868277341127396, + 0.015044944360852242, + 0.009655426256358624, + 0.002017298946157098, + 0.00566558912396431, + 0.0022315627429634333, + 0.004278305917978287, + 0.002211904153227806, + 0.0026195868849754333, + 0.004691208712756634, + 0.0032568341121077538, + 0.015362489968538284, + 0.01742587424814701, + 0.01134082954376936, + 0.023783521726727486, + 0.010822051204741001, + 0.006256159860640764, + 0.02054649218916893, + 0.02758825570344925, + 0.005093343090265989, + 0.006706864107400179, + 0.005282594822347164, + 0.003468208946287632, + 0.006298408377915621, + 0.003423035843297839, + 0.004111210349947214, + 0.004194378387182951, + 0.0031704541761428118, + 0.002527110045775771, + 0.0042536756955087185, + 0.004057141486555338, + 0.003302808851003647, + 0.0037170303985476494, + 0.004175280220806599, + 0.004805592820048332 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 7, + 5 + ], + "selected_reward": -0.009794864803552628 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 the nave of a historic church with stone floors and wooden pews?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are there stained-glass windows casting sapphire blue, ruby red, and emerald green light patterns on the floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a man seen from behind wearing a thick charcoal-gray wool coat with short dark hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a woman seen from behind wearing a long-sleeved cream dress with long auburn hair and a tear on her cheek?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly advance forward while the couple embraces in the center 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": [ + -0.02981371060013771 + ], + "opposite_logprob_tokens": [ + -3.529813766479492 + ], + "logprob_reward": -0.02981371060013771, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.02981371060013771, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1268405169248581, + 0.07885981351137161, + 0.23056063055992126, + 0.06653433293104172, + 0.02204742655158043, + 0.12057501077651978, + 0.049608271569013596, + 0.17015644907951355, + 0.07443738728761673, + 0.06038021296262741 + ], + "credit_map": [ + 0.02639961987733841, + 0.013975844718515873, + 0.01832674629986286, + 0.009585517458617687, + 0.015489744022488594, + 0.01713757961988449, + 0.0166984461247921, + 0.00922701321542263, + 0.01837216503918171, + 0.0073566739447414875, + 0.012128849513828754, + 0.006247173063457012, + 0.00645226426422596, + 0.007970625534653664, + 0.013610483147203922, + 0.00672157621011138, + 0.012641300447285175, + 0.009950747713446617, + 0.08242001384496689, + 0.009143415838479996, + 0.014017213135957718, + 0.04604203253984451, + 0.046767834573984146, + 0.009578043594956398, + 0.009243283420801163, + 0.006889067590236664, + 0.006326610222458839, + 0.005574233829975128, + 0.006061295513063669, + 0.01811942644417286, + 0.00916038267314434, + 0.005160024855285883, + 0.0019987463019788265, + 0.0028814130928367376, + 0.0029718116857111454, + 0.0030752841848880053, + 0.004165048245340586, + 0.0014245498459786177, + 0.003733616089448333, + 0.0017969560576602817, + 0.00850705336779356, + 0.014475932344794273, + 0.02397020161151886, + 0.018958650529384613, + 0.008858897723257542, + 0.006383976899087429, + 0.022509686648845673, + 0.01691059209406376, + 0.00437861867249012, + 0.008599434047937393, + 0.006996121257543564, + 0.005152476951479912, + 0.004681019112467766, + 0.007433690596371889, + 0.008316557854413986, + 0.00405034888535738, + 0.013785437680780888, + 0.029650090262293816, + 0.01587068662047386, + 0.03898811712861061, + 0.015956968069076538, + 0.013491163030266762, + 0.026010405272245407, + 0.016403567045927048, + 0.008241205476224422, + 0.011954129673540592, + 0.008426922373473644, + 0.006690715439617634, + 0.011132228188216686, + 0.007089347578585148, + 0.008084069006145, + 0.012818761169910431, + 0.006310453172773123, + 0.00435014721006155, + 0.008588663302361965, + 0.009481707587838173, + 0.007013887166976929, + 0.007967133074998856, + 0.007590895518660545, + 0.00907731894403696 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 7, + 0 + ], + "selected_reward": -0.02981371060013771 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 the nave of a historic church with stone floors and wooden pews?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are there stained-glass windows casting sapphire blue, ruby red, and emerald green light patterns on the floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a man seen from behind wearing a thick charcoal-gray wool coat with short dark hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a woman seen from behind wearing a long-sleeved cream dress with long auburn hair and a tear on her cheek?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly advance forward while the couple embraces in the center 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": [ + -0.0700172632932663 + ], + "opposite_logprob_tokens": [ + -2.6950173377990723 + ], + "logprob_reward": -0.0700172632932663, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.0700172632932663, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09469246119260788, + 0.04482530802488327, + 0.13349901139736176, + 0.12397477775812149, + 0.031886689364910126, + 0.3058602511882782, + 0.05592328682541847, + 0.07722065597772598, + 0.07988477498292923, + 0.052232809364795685 + ], + "credit_map": [ + 0.014685285277664661, + 0.008622567169368267, + 0.014305511489510536, + 0.01206977665424347, + 0.013937688432633877, + 0.007439362816512585, + 0.014506048522889614, + 0.009126215241849422, + 0.009369431994855404, + 0.005138710606843233, + 0.005486426874995232, + 0.004697275348007679, + 0.0036036621313542128, + 0.006248712074011564, + 0.007135249674320221, + 0.003145839087665081, + 0.010693895630538464, + 0.008412553928792477, + 0.026322506368160248, + 0.005982282571494579, + 0.0098780682310462, + 0.0261016134172678, + 0.03964497148990631, + 0.006463115103542805, + 0.009807737544178963, + 0.01084175519645214, + 0.01102355308830738, + 0.011612878181040287, + 0.009745755232870579, + 0.036586858332157135, + 0.025831837207078934, + 0.008524400182068348, + 0.003273503854870796, + 0.003986777272075415, + 0.007074319291859865, + 0.002728434279561043, + 0.004052494652569294, + 0.0024865379091352224, + 0.005104123614728451, + 0.0031804977916181087, + 0.060093000531196594, + 0.026560509577393532, + 0.02854124829173088, + 0.03723813220858574, + 0.017999932169914246, + 0.06609420478343964, + 0.03672191500663757, + 0.032611310482025146, + 0.004455601330846548, + 0.010433553718030453, + 0.004336277488619089, + 0.010474850423634052, + 0.006983522325754166, + 0.00534014543518424, + 0.008781231939792633, + 0.005118102300912142, + 0.010034119710326195, + 0.011463267728686333, + 0.005223241169005632, + 0.01157389860600233, + 0.007023931015282869, + 0.0037979527842253447, + 0.011122097261250019, + 0.016982141882181168, + 0.008224294520914555, + 0.015286914072930813, + 0.00772031070664525, + 0.008609798736870289, + 0.015848025679588318, + 0.008457985706627369, + 0.008456499315798283, + 0.007280941586941481, + 0.0076128048822283745, + 0.006141584366559982, + 0.008049318566918373, + 0.0058801984414458275, + 0.005199864041060209, + 0.0060438066720962524, + 0.006522087845951319, + 0.006783138960599899 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 2, + 3 + ], + "selected_reward": -0.0700172632932663 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, medium-length shot slowly, almost imperceptibly, begins to advance, capturing a deeply intimate moment within the cavernous nave of a magnificent, historic church. A brilliant, kaleidoscopic light pours through soaring, intricate stained-glass windows, painting the cold stone floor and ancient wooden pews with vibrant, shifting patterns of sapphire blue, ruby red, and emerald green. At the center of this ethereal illumination, a man and a woman embrace, creating a single, unified silhouette. Seen from behind, the man has short, dark, slightly disheveled hair and wears a thick, charcoal-gray wool coat; his pose is one of complete surrender, his face buried in her shoulder, his hands pressed firmly against her back. The woman, with her long, flowing auburn hair catching the colorful glow like a fiery halo, wears a simple, long-sleeved cream dress. Her eyes are closed, a tear trickling down her cheek, and as her arms wrap firmly around his neck, her expression is one of calm and overwhelming relief. The air is filled with dancing dust motes, each catching the light and enhancing the sacred and timeless quality of the scene, while slow, deliberate camera movements steadily draw the viewer into the overwhelming, silent, and deeply emotional atmosphere of their reunion.", + "seed": 21866, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000026/group_1/index_4563_seed_21866_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": -0.0488251144066453, + "vlm_logprob": -0.0488251144066453, + "vlm_yes_no_margin": 3.575, + "vlm_token_credit": -0.0488251144066453, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 the nave of a historic church with stone floors and wooden pews?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are there stained-glass windows casting sapphire blue, ruby red, and emerald green light patterns on the floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a man seen from behind wearing a thick charcoal-gray wool coat with short dark hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a woman seen from behind wearing a long-sleeved cream dress with long auburn hair and a tear on her cheek?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly advance forward while the couple embraces in the center 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": [ + -0.10896891355514526 + ], + "opposite_logprob_tokens": [ + -2.48396897315979 + ], + "logprob_reward": -0.10896891355514526, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.10896891355514526, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1800243854522705, + 0.05931353196501732, + 0.032940108329057693, + 0.04751134291291237, + 0.30345094203948975, + 0.0303826704621315, + 0.02676536701619625, + 0.0766151174902916, + 0.0432569719851017, + 0.19973953068256378 + ], + "credit_map": [ + 0.02687676250934601, + 0.017626652494072914, + 0.027881426736712456, + 0.02162749506533146, + 0.015378380194306374, + 0.023910049349069595, + 0.01735541969537735, + 0.0293682012706995, + 0.005884194280952215, + 0.005741102620959282, + 0.011407389305531979, + 0.009823017753660679, + 0.003300238400697708, + 0.007757703308016062, + 0.007307359948754311, + 0.008092524483799934, + 0.006248203571885824, + 0.002622275147587061, + 0.004256298299878836, + 0.0037700633984059095, + 0.0032703066244721413, + 0.004108896944671869, + 0.004140160046517849, + 0.004523903131484985, + 0.005099078174680471, + 0.005360162816941738, + 0.0050961971282958984, + 0.004421515390276909, + 0.007376766763627529, + 0.008995302952826023, + 0.005337208043783903, + 0.00582510931417346, + 0.0438346303999424, + 0.026384813711047173, + 0.03543676435947418, + 0.04465143010020256, + 0.023493077605962753, + 0.027151107788085938, + 0.056391213089227676, + 0.046107906848192215, + 0.00453521404415369, + 0.002370962407439947, + 0.003430817509070039, + 0.0029702596366405487, + 0.003094426356256008, + 0.004464969504624605, + 0.001729347393848002, + 0.0077866725623607635, + 0.003931683488190174, + 0.0020357994362711906, + 0.0034957295283675194, + 0.003791533410549164, + 0.0026742173358798027, + 0.004256247077137232, + 0.0034068883396685123, + 0.0031732660718262196, + 0.007236178498715162, + 0.005475504789501429, + 0.00596466101706028, + 0.018285907804965973, + 0.004285979084670544, + 0.010465368628501892, + 0.009125507436692715, + 0.01577601581811905, + 0.006877325475215912, + 0.0037063441704958677, + 0.004753836430609226, + 0.006807003635913134, + 0.0028434661217033863, + 0.006900540553033352, + 0.006355506367981434, + 0.005012949928641319, + 0.02024270035326481, + 0.01592681184411049, + 0.03670743852853775, + 0.032797012478113174, + 0.010801007971167564, + 0.022484682500362396, + 0.037381500005722046, + 0.023398373275995255 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 9, + 0 + ], + "selected_reward": -0.10896891355514526 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 the nave of a historic church with stone floors and wooden pews?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are there stained-glass windows casting sapphire blue, ruby red, and emerald green light patterns on the floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a man seen from behind wearing a thick charcoal-gray wool coat with short dark hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a woman seen from behind wearing a long-sleeved cream dress with long auburn hair and a tear on her cheek?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly advance forward while the couple embraces in the center 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": [ + -0.007671656087040901 + ], + "opposite_logprob_tokens": [ + -4.88267183303833 + ], + "logprob_reward": -0.007671656087040901, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007671656087040901, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0893910676240921, + 0.06347338110208511, + 0.0695805698633194, + 0.22413401305675507, + 0.17325827479362488, + 0.18074919283390045, + 0.03578430041670799, + 0.09407418966293335, + 0.038520995527505875, + 0.031034007668495178 + ], + "credit_map": [ + 0.012907132506370544, + 0.01132531650364399, + 0.010194285772740841, + 0.012489113956689835, + 0.008124925196170807, + 0.014311645179986954, + 0.006998439319431782, + 0.013040213845670223, + 0.007047863211482763, + 0.0059797377325594425, + 0.007814638316631317, + 0.012865743599832058, + 0.0040464261546730995, + 0.007156936451792717, + 0.006766452919691801, + 0.0117955869063735, + 0.00808100588619709, + 0.007444391958415508, + 0.010638115927577019, + 0.006506878416985273, + 0.012060286477208138, + 0.008972900919616222, + 0.0054780966602265835, + 0.010398897342383862, + 0.03227991238236427, + 0.027380751445889473, + 0.01981767825782299, + 0.017579136416316032, + 0.03316403552889824, + 0.03161315992474556, + 0.017384933307766914, + 0.044914405792951584, + 0.022001950070261955, + 0.014562542550265789, + 0.018516946583986282, + 0.02138379029929638, + 0.01586422696709633, + 0.01972641795873642, + 0.025571051985025406, + 0.03563135489821434, + 0.05649173632264137, + 0.00724207004532218, + 0.013733012601733208, + 0.019773224368691444, + 0.006774701178073883, + 0.012532084248960018, + 0.006880749017000198, + 0.057321611791849136, + 0.004507681354880333, + 0.003284502075985074, + 0.004659757483750582, + 0.004872766323387623, + 0.00460823392495513, + 0.004674714058637619, + 0.004040652420371771, + 0.0051359920762479305, + 0.007465353701263666, + 0.007065558340400457, + 0.009348741732537746, + 0.019411882385611534, + 0.009142001159489155, + 0.011885766871273518, + 0.011665995232760906, + 0.018088888376951218, + 0.005206297617405653, + 0.0033465269953012466, + 0.003940637689083815, + 0.006289251148700714, + 0.0031320557463914156, + 0.007553219795227051, + 0.004107058979570866, + 0.004945945926010609, + 0.003684836672618985, + 0.002535872394219041, + 0.005437159910798073, + 0.004430311266332865, + 0.002802761970087886, + 0.003620073664933443, + 0.0038823732174932957, + 0.004640617407858372 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 5, + 4 + ], + "selected_reward": -0.007671656087040901 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 the nave of a historic church with stone floors and wooden pews?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are there stained-glass windows casting sapphire blue, ruby red, and emerald green light patterns on the floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a man seen from behind wearing a thick charcoal-gray wool coat with short dark hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a woman seen from behind wearing a long-sleeved cream dress with long auburn hair and a tear on her cheek?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly advance forward while the couple embraces in the center 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": [ + -0.008648204617202282 + ], + "opposite_logprob_tokens": [ + -4.75864839553833 + ], + "logprob_reward": -0.008648204617202282, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008648204617202282, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19356483221054077, + 0.12708210945129395, + 0.11050987988710403, + 0.21034245193004608, + 0.1337246596813202, + 0.07096917927265167, + 0.049217481166124344, + 0.04419882595539093, + 0.01801861822605133, + 0.042372025549411774 + ], + "credit_map": [ + 0.016160553321242332, + 0.019458554685115814, + 0.018620934337377548, + 0.029913445934653282, + 0.024423964321613312, + 0.02554486133158207, + 0.03550335764884949, + 0.023939136415719986, + 0.01391300093382597, + 0.010669550858438015, + 0.01532273180782795, + 0.027638258412480354, + 0.007386757060885429, + 0.014334498904645443, + 0.013724076561629772, + 0.024093221873044968, + 0.009309367276728153, + 0.014850444160401821, + 0.017157521098852158, + 0.014135573990643024, + 0.014640124514698982, + 0.010755348950624466, + 0.014289124868810177, + 0.01537236850708723, + 0.03485759347677231, + 0.019386937841773033, + 0.020304614678025246, + 0.02182604745030403, + 0.011669429950416088, + 0.0168790053576231, + 0.029206549748778343, + 0.05621226876974106, + 0.03400908038020134, + 0.00753653421998024, + 0.011290294118225574, + 0.014392335899174213, + 0.006041612941771746, + 0.010371459648013115, + 0.011452260427176952, + 0.03863107040524483, + 0.01468399167060852, + 0.0058503239415585995, + 0.007226971909403801, + 0.01161219459027052, + 0.004461895674467087, + 0.006144373677670956, + 0.009426277130842209, + 0.011563146486878395, + 0.005531069356948137, + 0.004323229193687439, + 0.005835510324686766, + 0.007670771796256304, + 0.005152672063559294, + 0.005652491468936205, + 0.0059270309284329414, + 0.009124701842665672, + 0.004209891892969608, + 0.0022419944871217012, + 0.0024377130903303623, + 0.012762411497533321, + 0.0019203074043616652, + 0.0038598054088652134, + 0.0051980167627334595, + 0.01156868040561676, + 0.0025705972220748663, + 0.0017357438337057829, + 0.0017758277244865894, + 0.003055426524952054, + 0.0011084495345130563, + 0.0027618040330708027, + 0.0032943598926067352, + 0.001716408645734191, + 0.006976158823817968, + 0.005572034046053886, + 0.004701729863882065, + 0.006675356533378363, + 0.0032510864548385143, + 0.005490120965987444, + 0.00526259234175086, + 0.004442943725734949 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 4 + ], + "selected_reward": -0.008648204617202282 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 the nave of a historic church with stone floors and wooden pews?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are there stained-glass windows casting sapphire blue, ruby red, and emerald green light patterns on the floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a man seen from behind wearing a thick charcoal-gray wool coat with short dark hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a woman seen from behind wearing a long-sleeved cream dress with long auburn hair and a tear on her cheek?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly advance forward while the couple embraces in the center 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": [ + -0.029807694256305695 + ], + "opposite_logprob_tokens": [ + -3.5298078060150146 + ], + "logprob_reward": -0.029807694256305695, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029807694256305695, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19658608734607697, + 0.23281021416187286, + 0.07587707042694092, + 0.09384781122207642, + 0.0536300428211689, + 0.07421760261058807, + 0.14799225330352783, + 0.06649921089410782, + 0.022727204486727715, + 0.03581247478723526 + ], + "credit_map": [ + 0.020633716136217117, + 0.02500125952064991, + 0.0165530014783144, + 0.024914810433983803, + 0.03240175172686577, + 0.028808020055294037, + 0.02167678065598011, + 0.02659674920141697, + 0.020395586267113686, + 0.025874482467770576, + 0.025463389232754707, + 0.051763229072093964, + 0.015688834711909294, + 0.027266373857855797, + 0.027731232345104218, + 0.038627080619335175, + 0.010082968510687351, + 0.00905875489115715, + 0.009100566618144512, + 0.009284427389502525, + 0.01029960811138153, + 0.006659583188593388, + 0.007909243926405907, + 0.01348192896693945, + 0.021559743210673332, + 0.003914370201528072, + 0.009886963292956352, + 0.006530735641717911, + 0.00352382636629045, + 0.00944541860371828, + 0.0042387135326862335, + 0.03474804386496544, + 0.008988846093416214, + 0.0034546602983027697, + 0.008544716984033585, + 0.006474320311099291, + 0.003627019701525569, + 0.00638666981831193, + 0.005568521562963724, + 0.010585291311144829, + 0.018280029296875, + 0.005182748194783926, + 0.00760557921603322, + 0.009014847688376904, + 0.0034580728970468044, + 0.0065682558342814445, + 0.007168704643845558, + 0.016939358785748482, + 0.015782181173563004, + 0.012901753187179565, + 0.01623145304620266, + 0.021918853744864464, + 0.01709078624844551, + 0.018007738515734673, + 0.015230139717459679, + 0.030829351395368576, + 0.007789929397404194, + 0.005899544805288315, + 0.005330033600330353, + 0.01092019397765398, + 0.004606747534126043, + 0.00841584149748087, + 0.014959569089114666, + 0.008577346801757812, + 0.0018965211929753423, + 0.0016465006629005075, + 0.002351499628275633, + 0.004039947409182787, + 0.0013555599143728614, + 0.0023923765402287245, + 0.006533617153763771, + 0.002511183498427272, + 0.00507690291851759, + 0.004105791915208101, + 0.004350975621491671, + 0.006418169476091862, + 0.0026392934378236532, + 0.004426266998052597, + 0.004895772319287062, + 0.0038993011694401503 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 6 + ], + "selected_reward": -0.029807694256305695 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 the nave of a historic church with stone floors and wooden pews?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are there stained-glass windows casting sapphire blue, ruby red, and emerald green light patterns on the floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a man seen from behind wearing a thick charcoal-gray wool coat with short dark hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a woman seen from behind wearing a long-sleeved cream dress with long auburn hair and a tear on her cheek?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly advance forward while the couple embraces in the center 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": [ + -0.08902910351753235 + ], + "opposite_logprob_tokens": [ + -2.46402907371521 + ], + "logprob_reward": -0.08902910351753235, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08902910351753235, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09800872951745987, + 0.08772402256727219, + 0.049614839255809784, + 0.07084634155035019, + 0.40148046612739563, + 0.06371067464351654, + 0.03952617570757866, + 0.07691412419080734, + 0.036035776138305664, + 0.07613882422447205 + ], + "credit_map": [ + 0.013568216934800148, + 0.00827829074114561, + 0.01767166703939438, + 0.015017888508737087, + 0.007727735210210085, + 0.010672435164451599, + 0.01110985316336155, + 0.013962646014988422, + 0.011008527129888535, + 0.006970670074224472, + 0.01386145781725645, + 0.011611515656113625, + 0.006609675008803606, + 0.009423498995602131, + 0.01607285812497139, + 0.012165823951363564, + 0.006469598039984703, + 0.007208556868135929, + 0.004787871614098549, + 0.007279050536453724, + 0.003827551379799843, + 0.006200876552611589, + 0.005379396490752697, + 0.008461940102279186, + 0.009973566979169846, + 0.008319205604493618, + 0.007952652871608734, + 0.005772704724222422, + 0.00721863703802228, + 0.00945089291781187, + 0.004878616891801357, + 0.017280064523220062, + 0.05164877325296402, + 0.036318812519311905, + 0.044245395809412, + 0.06124556437134743, + 0.03058079071342945, + 0.035929255187511444, + 0.07911377400159836, + 0.062398117035627365, + 0.01971358060836792, + 0.0029657555278390646, + 0.005426126532256603, + 0.007529051508754492, + 0.0023554887156933546, + 0.004666979890316725, + 0.0038771259132772684, + 0.017176568508148193, + 0.003982907626777887, + 0.003658503759652376, + 0.004972414113581181, + 0.005294699687510729, + 0.0048410785384476185, + 0.004756289999932051, + 0.004089352674782276, + 0.007930931635200977, + 0.00832157488912344, + 0.004512111656367779, + 0.0057985479943454266, + 0.023547954857349396, + 0.003946884535253048, + 0.005665722768753767, + 0.005167675670236349, + 0.01995365507900715, + 0.004250595346093178, + 0.002506810473278165, + 0.0035438844934105873, + 0.007838422432541847, + 0.003097259672358632, + 0.005731899756938219, + 0.004055970814079046, + 0.0050109331496059895, + 0.007477888837456703, + 0.008106768131256104, + 0.007649887818843126, + 0.012937773019075394, + 0.004942622035741806, + 0.006231650244444609, + 0.021442590281367302, + 0.007349645718932152 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 1 + ], + "selected_reward": -0.08902910351753235 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, medium-length shot slowly, almost imperceptibly, begins to advance, capturing a deeply intimate moment within the cavernous nave of a magnificent, historic church. A brilliant, kaleidoscopic light pours through soaring, intricate stained-glass windows, painting the cold stone floor and ancient wooden pews with vibrant, shifting patterns of sapphire blue, ruby red, and emerald green. At the center of this ethereal illumination, a man and a woman embrace, creating a single, unified silhouette. Seen from behind, the man has short, dark, slightly disheveled hair and wears a thick, charcoal-gray wool coat; his pose is one of complete surrender, his face buried in her shoulder, his hands pressed firmly against her back. The woman, with her long, flowing auburn hair catching the colorful glow like a fiery halo, wears a simple, long-sleeved cream dress. Her eyes are closed, a tear trickling down her cheek, and as her arms wrap firmly around his neck, her expression is one of calm and overwhelming relief. The air is filled with dancing dust motes, each catching the light and enhancing the sacred and timeless quality of the scene, while slow, deliberate camera movements steadily draw the viewer into the overwhelming, silent, and deeply emotional atmosphere of their reunion.", + "seed": 21867, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000026/group_1/index_4563_seed_21867_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": -0.21361594442278148, + "vlm_logprob": -0.21361594442278148, + "vlm_yes_no_margin": 2.8, + "vlm_token_credit": -0.21361594442278148, + "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 the nave of a historic church with stone floors and wooden pews?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are there stained-glass windows casting sapphire blue, ruby red, and emerald green light patterns on the floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a man seen from behind wearing a thick charcoal-gray wool coat with short dark hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a woman seen from behind wearing a long-sleeved cream dress with long auburn hair and a tear on her cheek?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly advance forward while the couple embraces in the center of the frame?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.9295135736465454 + ], + "opposite_logprob_tokens": [ + -0.5545135736465454 + ], + "logprob_reward": -0.9295135736465454, + "yes_no_margin_reward": -0.375, + "token_credit_reward": -0.9295135736465454, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.19200746715068817, + 0.10409290343523026, + 0.04674306884407997, + 0.03575926646590233, + 0.11285972595214844, + 0.05749375373125076, + 0.18080823123455048, + 0.1302507072687149, + 0.03773617371916771, + 0.10224869847297668 + ], + "credit_map": [ + 0.015354319475591183, + 0.013550166971981525, + 0.035467859357595444, + 0.03863198310136795, + 0.017882639542222023, + 0.015389830805361271, + 0.019188661128282547, + 0.036542002111673355, + 0.010259163565933704, + 0.0037331529892981052, + 0.023201704025268555, + 0.017897386103868484, + 0.0039503732696175575, + 0.004744661506265402, + 0.007045669015496969, + 0.03326079994440079, + 0.005449212621897459, + 0.004172804765403271, + 0.005067882593721151, + 0.009367656894028187, + 0.00469026155769825, + 0.00646078959107399, + 0.003180528525263071, + 0.00835393462330103, + 0.0066160643473267555, + 0.002181480871513486, + 0.004385295324027538, + 0.006331980228424072, + 0.0027095680125057697, + 0.002083134837448597, + 0.005526387598365545, + 0.005925356410443783, + 0.00855556596070528, + 0.011367956176400185, + 0.019695965573191643, + 0.03008219785988331, + 0.005743726156651974, + 0.0055177295580506325, + 0.009852415882050991, + 0.022044168785214424, + 0.0030185438226908445, + 0.0034455021377652884, + 0.013774128630757332, + 0.01681709662079811, + 0.003233223920688033, + 0.0030961192678660154, + 0.006594829726964235, + 0.0075143082067370415, + 0.014830570667982101, + 0.022971823811531067, + 0.02628989703953266, + 0.030072398483753204, + 0.015915997326374054, + 0.02984187752008438, + 0.017002584412693977, + 0.02388308383524418, + 0.020224524661898613, + 0.012846739031374454, + 0.021097412332892418, + 0.026926517486572266, + 0.006674981210380793, + 0.01302829384803772, + 0.017426099628210068, + 0.012026139535009861, + 0.004998259246349335, + 0.002494222018867731, + 0.007560605648905039, + 0.008957751095294952, + 0.0029494825284928083, + 0.00444689579308033, + 0.0036763667594641447, + 0.002652592957019806, + 0.010859689675271511, + 0.006934108678251505, + 0.015879997983574867, + 0.0178524199873209, + 0.010366017930209637, + 0.014078806154429913, + 0.016565414145588875, + 0.009712247177958488 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 7 + ], + "selected_reward": -0.9295135736465454 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 the nave of a historic church with stone floors and wooden pews?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are there stained-glass windows casting sapphire blue, ruby red, and emerald green light patterns on the floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a man seen from behind wearing a thick charcoal-gray wool coat with short dark hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a woman seen from behind wearing a long-sleeved cream dress with long auburn hair and a tear on her cheek?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly advance forward while the couple embraces in the center 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": [ + -0.02067105658352375 + ], + "opposite_logprob_tokens": [ + -3.8956711292266846 + ], + "logprob_reward": -0.02067105658352375, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.02067105658352375, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20285384356975555, + 0.2639647424221039, + 0.10055753588676453, + 0.06478867679834366, + 0.08589717000722885, + 0.06561215966939926, + 0.09253207594156265, + 0.046012330800294876, + 0.052447982132434845, + 0.025333460420370102 + ], + "credit_map": [ + 0.01847211830317974, + 0.02067197486758232, + 0.02700045518577099, + 0.03991164639592171, + 0.02725345455110073, + 0.017675481736660004, + 0.01740792766213417, + 0.034460779279470444, + 0.026179244741797447, + 0.009844504296779633, + 0.05689126253128052, + 0.053087569773197174, + 0.005590219981968403, + 0.010780817829072475, + 0.011962431482970715, + 0.0896286889910698, + 0.007387061603367329, + 0.012305734679102898, + 0.012404903769493103, + 0.021795788779854774, + 0.009570864029228687, + 0.011753112077713013, + 0.005264908540993929, + 0.020075157284736633, + 0.011618507094681263, + 0.003911707084625959, + 0.007681758608669043, + 0.012172868475317955, + 0.0034008892253041267, + 0.005511652212589979, + 0.007646285928785801, + 0.01284500677138567, + 0.007918154820799828, + 0.008371024392545223, + 0.012891370803117752, + 0.02196398936212063, + 0.003841300727799535, + 0.005743221379816532, + 0.007075504865497351, + 0.018092598766088486, + 0.00414583133533597, + 0.00454523554071784, + 0.013765659183263779, + 0.01774483174085617, + 0.0039433971978724, + 0.003502135630697012, + 0.009034430608153343, + 0.008930630050599575, + 0.008065003901720047, + 0.011868014931678772, + 0.015386044047772884, + 0.01780783012509346, + 0.00520229572430253, + 0.015109744854271412, + 0.008087271824479103, + 0.011005873791873455, + 0.004060704261064529, + 0.0031802647281438112, + 0.007719906512647867, + 0.007897610776126385, + 0.006037724204361439, + 0.0038991516921669245, + 0.00784087274223566, + 0.005376090295612812, + 0.004504121840000153, + 0.0038912701420485973, + 0.011868901550769806, + 0.011485404334962368, + 0.006681083235889673, + 0.004320579580962658, + 0.006022128742188215, + 0.003674489678815007, + 0.0018441331340000033, + 0.0016349090728908777, + 0.0037323865108191967, + 0.004389256238937378, + 0.0039798803627491, + 0.0020953372586518526, + 0.0048751202411949635, + 0.0027824356220662594 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 2 + ], + "selected_reward": -0.02067105658352375 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 the nave of a historic church with stone floors and wooden pews?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are there stained-glass windows casting sapphire blue, ruby red, and emerald green light patterns on the floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a man seen from behind wearing a thick charcoal-gray wool coat with short dark hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a woman seen from behind wearing a long-sleeved cream dress with long auburn hair and a tear on her cheek?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly advance forward while the couple embraces in the center 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": [ + -0.012553482316434383 + ], + "opposite_logprob_tokens": [ + -4.387553691864014 + ], + "logprob_reward": -0.012553482316434383, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012553482316434383, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08487570285797119, + 0.11460989713668823, + 0.1428094208240509, + 0.09216247498989105, + 0.18402570486068726, + 0.08019882440567017, + 0.06135459244251251, + 0.09254240244626999, + 0.02479313500225544, + 0.12262783944606781 + ], + "credit_map": [ + 0.0068954359740018845, + 0.0055762119591236115, + 0.016845431178808212, + 0.010551226325333118, + 0.008918283507227898, + 0.0064916471019387245, + 0.011686279438436031, + 0.017911188304424286, + 0.008561774156987667, + 0.007822815328836441, + 0.019841868430376053, + 0.021989023312926292, + 0.016550183296203613, + 0.007435594219714403, + 0.009593460708856583, + 0.022815166041254997, + 0.010923072695732117, + 0.007754607126116753, + 0.013727518729865551, + 0.03769281134009361, + 0.00624496303498745, + 0.010844218544661999, + 0.013663235120475292, + 0.041958998888731, + 0.012242598459124565, + 0.006560288369655609, + 0.009198884479701519, + 0.014852002263069153, + 0.0075432732701301575, + 0.01682421565055847, + 0.009220353327691555, + 0.015720859169960022, + 0.010896977037191391, + 0.014758630655705929, + 0.01690370962023735, + 0.05167878046631813, + 0.0052455877885222435, + 0.009614761918783188, + 0.007834065705537796, + 0.06709318608045578, + 0.0041474709287285805, + 0.005974438041448593, + 0.017471451312303543, + 0.014445330947637558, + 0.009288247674703598, + 0.003978160675615072, + 0.011305604130029678, + 0.01358811091631651, + 0.0033613652922213078, + 0.0031506449449807405, + 0.01488401833921671, + 0.01653730683028698, + 0.004423479549586773, + 0.0024915561079978943, + 0.009286579675972462, + 0.007219643797725439, + 0.013792349025607109, + 0.008584452793002129, + 0.01525855902582407, + 0.01931767351925373, + 0.005839794874191284, + 0.008564501069486141, + 0.012410757131874561, + 0.00877432245761156, + 0.0023771084379404783, + 0.0019592302851378918, + 0.004523188807070255, + 0.0049046920612454414, + 0.002260618144646287, + 0.0026704533956944942, + 0.0028612662572413683, + 0.00323657738044858, + 0.015716833993792534, + 0.00800083763897419, + 0.02117253839969635, + 0.017282728105783463, + 0.02236117422580719, + 0.01932618021965027, + 0.009750684723258018, + 0.009016874246299267 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 2, + 9 + ], + "selected_reward": -0.012553482316434383 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 the nave of a historic church with stone floors and wooden pews?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are there stained-glass windows casting sapphire blue, ruby red, and emerald green light patterns on the floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a man seen from behind wearing a thick charcoal-gray wool coat with short dark hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a woman seen from behind wearing a long-sleeved cream dress with long auburn hair and a tear on her cheek?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly advance forward while the couple embraces in the center 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": [ + -0.026359546929597855 + ], + "opposite_logprob_tokens": [ + -3.6513595581054688 + ], + "logprob_reward": -0.026359546929597855, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026359546929597855, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05627111345529556, + 0.07854855805635452, + 0.10736881196498871, + 0.08502642065286636, + 0.22033341228961945, + 0.06078450381755829, + 0.1960180103778839, + 0.05018973350524902, + 0.07215499877929688, + 0.0733044296503067 + ], + "credit_map": [ + 0.006546777673065662, + 0.006076854653656483, + 0.00724817207083106, + 0.0059152995236217976, + 0.009186283685266972, + 0.0071033877320587635, + 0.005937783047556877, + 0.00825655646622181, + 0.007525730412453413, + 0.00522951316088438, + 0.01734391413629055, + 0.010058974847197533, + 0.00941167026758194, + 0.006302995607256889, + 0.005749281961470842, + 0.016926482319831848, + 0.015519289299845695, + 0.007007866632193327, + 0.010013212449848652, + 0.027025308459997177, + 0.004911665804684162, + 0.005883791949599981, + 0.006116359028965235, + 0.03089131787419319, + 0.008867107331752777, + 0.004777030553668737, + 0.009665071032941341, + 0.01648932881653309, + 0.006241991184651852, + 0.012494970113039017, + 0.012652678415179253, + 0.013838251121342182, + 0.018089108169078827, + 0.020107237622141838, + 0.03446532040834427, + 0.06237005069851875, + 0.007913359440863132, + 0.011068297550082207, + 0.018225444480776787, + 0.04809459298849106, + 0.0033145856577903032, + 0.004701887723058462, + 0.011220147833228111, + 0.01114071998745203, + 0.008029752410948277, + 0.003615213092416525, + 0.011331156827509403, + 0.007431043777614832, + 0.013180471025407314, + 0.023875748738646507, + 0.039801884442567825, + 0.042205289006233215, + 0.010111929848790169, + 0.026600085198879242, + 0.01804003119468689, + 0.02220257930457592, + 0.005183383822441101, + 0.0034470614045858383, + 0.0075958603993058205, + 0.008828851394355297, + 0.0047728074714541435, + 0.005263129714876413, + 0.00992577988654375, + 0.0051728589460253716, + 0.008816595189273357, + 0.004659718833863735, + 0.016751617193222046, + 0.01674959436058998, + 0.006910504307597876, + 0.007342972792685032, + 0.00593090895563364, + 0.004993089009076357, + 0.00949904415756464, + 0.00448080338537693, + 0.012080423533916473, + 0.009417863562703133, + 0.013038148172199726, + 0.011774471029639244, + 0.007023751735687256, + 0.005989917553961277 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 6, + 2 + ], + "selected_reward": -0.026359546929597855 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 the nave of a historic church with stone floors and wooden pews?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are there stained-glass windows casting sapphire blue, ruby red, and emerald green light patterns on the floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a man seen from behind wearing a thick charcoal-gray wool coat with short dark hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a woman seen from behind wearing a long-sleeved cream dress with long auburn hair and a tear on her cheek?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly advance forward while the couple embraces in the center 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": [ + -0.07898206263780594 + ], + "opposite_logprob_tokens": [ + -2.57898211479187 + ], + "logprob_reward": -0.07898206263780594, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07898206263780594, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13193577527999878, + 0.05979476496577263, + 0.07433906197547913, + 0.03024383820593357, + 0.24007239937782288, + 0.07011032104492188, + 0.13352848589420319, + 0.18293344974517822, + 0.033935971558094025, + 0.043105896562337875 + ], + "credit_map": [ + 0.009986530058085918, + 0.006541901733726263, + 0.024235142394900322, + 0.022238701581954956, + 0.01175582967698574, + 0.009336997754871845, + 0.024385986849665642, + 0.023454688489437103, + 0.004971789661794901, + 0.002983412239700556, + 0.012882835231721401, + 0.011073296889662743, + 0.003421349683776498, + 0.003060360439121723, + 0.007403728552162647, + 0.013997990638017654, + 0.007342682220041752, + 0.005222052801400423, + 0.007575985975563526, + 0.020211651921272278, + 0.005247087217867374, + 0.004411341156810522, + 0.006360431667417288, + 0.017967823892831802, + 0.002585021313279867, + 0.0020909206941723824, + 0.0034333765506744385, + 0.008156053721904755, + 0.0016563785029575229, + 0.0019572945311665535, + 0.004236897453665733, + 0.006127893924713135, + 0.017440348863601685, + 0.02187535911798477, + 0.021226447075605392, + 0.058320216834545135, + 0.00785779021680355, + 0.021831505000591278, + 0.009639556519687176, + 0.08188117295503616, + 0.003431884106248617, + 0.003310754429548979, + 0.02257746458053589, + 0.01966513693332672, + 0.005395698826760054, + 0.0027256840839982033, + 0.006536911241710186, + 0.00646678963676095, + 0.00962299294769764, + 0.01689021661877632, + 0.023519277572631836, + 0.02423132210969925, + 0.012047822587192059, + 0.020786307752132416, + 0.011997637338936329, + 0.014432904310524464, + 0.027456043288111687, + 0.01787317730486393, + 0.029674746096134186, + 0.03861207887530327, + 0.011434298940002918, + 0.017735552042722702, + 0.02358822710812092, + 0.016559340059757233, + 0.005026284139603376, + 0.0025018525775521994, + 0.006510453298687935, + 0.006968297064304352, + 0.002888966118916869, + 0.0041717118583619595, + 0.0032100556418299675, + 0.00265835365280509, + 0.00545441173017025, + 0.0024965666234493256, + 0.006353475619107485, + 0.0066067129373550415, + 0.00553095294162631, + 0.006299327127635479, + 0.006588159129023552, + 0.0037762888241559267 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 7, + 6 + ], + "selected_reward": -0.07898206263780594 + } + ], + "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/0000026/group_10/index_tensor([8786, 8786])_group_10_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000026/group_10/index_tensor([8786, 8786])_group_10_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..e82f34a2fc310c0a0aa0ae365b56ce116586b3fb --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000026/group_10/index_tensor([8786, 8786])_group_10_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "striking low-angle tracking close-up follows a Siberian tiger at full speed across the snow. Its thick, shaggy orange and white fur is streaked with striking black stripes, and the taut muscles are clearly visible with every flex, radiating unparalleled strength. The beast's amber eyes, narrowed in intense concentration, lock onto its target, its scorching breath instantly condensing into a white mist in the sub-zero air. Its muscular body surges forward with breathtaking force, its limbs striking powerfully with each landing. Each landing sinks deeply into the untouched powdery snow below, sending a powerful, shimmering cloud of snow backward. This cloud gleams like diamonds in the cold winter sun, leaving a dynamic, diffuse white trail behind it. The entire scene, shimmering against a cold palette, is imbued with a sense of primal power and speed, creating a tense, breathtaking atmosphere of pursuit.", + "seed": 1974, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000026/group_10/index_8786_seed_1974_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": -0.5121204778552055, + "vlm_logprob": -0.5121204778552055, + "vlm_yes_no_margin": 0.975, + "vlm_token_credit": -0.5121204778552055, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle following the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject a Siberian tiger with orange, white, and black fur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the tiger running at full speed across snow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the tiger's movement cause deep snow displacement creating a shimmering cloud of snow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the tiger's breath visible as white mist in the cold 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": [ + -0.4811702072620392 + ], + "opposite_logprob_tokens": [ + -1.1061701774597168 + ], + "logprob_reward": -0.4811702072620392, + "yes_no_margin_reward": 0.625, + "token_credit_reward": -0.4811702072620392, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14861524105072021, + 0.07436518371105194, + 0.060776859521865845, + 0.05688279867172241, + 0.1555423140525818, + 0.05287232622504234, + 0.09976755082607269, + 0.043354734778404236, + 0.25944262742996216, + 0.048380348831415176 + ], + "credit_map": [ + 0.016831647604703903, + 0.009439131245017052, + 0.030093716457486153, + 0.010976618155837059, + 0.008958650752902031, + 0.028904113918542862, + 0.01283451821655035, + 0.03057684749364853, + 0.0036115599796175957, + 0.007990287616848946, + 0.0070571210235357285, + 0.006185689475387335, + 0.005103916395455599, + 0.018387559801340103, + 0.006140958052128553, + 0.019888095557689667, + 0.003909149207174778, + 0.003991839941591024, + 0.005929407197982073, + 0.006554019637405872, + 0.004278842359781265, + 0.01609974540770054, + 0.004128193948417902, + 0.015885664150118828, + 0.003902524709701538, + 0.005626371130347252, + 0.005886461585760117, + 0.012133671902120113, + 0.003874195273965597, + 0.008610395714640617, + 0.007280587684363127, + 0.009568591602146626, + 0.02256053499877453, + 0.010515779256820679, + 0.02823493629693985, + 0.013743432238698006, + 0.006734647322446108, + 0.01651165634393692, + 0.012746969237923622, + 0.04449436441063881, + 0.00874914601445198, + 0.006721365265548229, + 0.008158961310982704, + 0.005633152090013027, + 0.006251669954508543, + 0.004868344869464636, + 0.005331922322511673, + 0.007157768122851849, + 0.01646178588271141, + 0.00939987227320671, + 0.013407786376774311, + 0.010965977795422077, + 0.008592730388045311, + 0.009743216447532177, + 0.015473958104848862, + 0.015722230076789856, + 0.0044995988719165325, + 0.005976633168756962, + 0.00486353226006031, + 0.006650951690971851, + 0.005235765129327774, + 0.003106213640421629, + 0.006301071494817734, + 0.006720971781760454, + 0.0321430042386055, + 0.035585496574640274, + 0.023815756663680077, + 0.02718883939087391, + 0.03575730323791504, + 0.03510649874806404, + 0.0340251661837101, + 0.03582054749131203, + 0.008282926864922047, + 0.0031291977502405643, + 0.006205949932336807, + 0.004728452302515507, + 0.003845600876957178, + 0.004719999618828297, + 0.00918014720082283, + 0.008288074284791946 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 4, + 0 + ], + "selected_reward": -0.4811702072620392 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 following the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject a Siberian tiger with orange, white, and black fur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the tiger running at full speed across snow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the tiger's movement cause deep snow displacement creating a shimmering cloud of snow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the tiger's breath visible as white mist in the cold 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": [ + -0.04869883507490158 + ], + "opposite_logprob_tokens": [ + -3.048698902130127 + ], + "logprob_reward": -0.04869883507490158, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04869883507490158, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1678956151008606, + 0.030291927978396416, + 0.1132240816950798, + 0.023475952446460724, + 0.1568097621202469, + 0.02862325869500637, + 0.1446060687303543, + 0.0657799169421196, + 0.09226711094379425, + 0.17702622711658478 + ], + "credit_map": [ + 0.019447743892669678, + 0.017017671838402748, + 0.034555792808532715, + 0.017378073185682297, + 0.01024576835334301, + 0.031596194952726364, + 0.01181167084723711, + 0.025842707604169846, + 0.0020923425909131765, + 0.0027165126521140337, + 0.0024072034284472466, + 0.0028537153266370296, + 0.0028619468212127686, + 0.007458994630724192, + 0.002202914096415043, + 0.007698305416852236, + 0.007359487935900688, + 0.009776761755347252, + 0.010300411842763424, + 0.01433346513658762, + 0.00924853514879942, + 0.0257291067391634, + 0.011068462394177914, + 0.025407865643501282, + 0.0014068795135244727, + 0.0033158676233142614, + 0.002190826926380396, + 0.0028646294958889484, + 0.002722937148064375, + 0.0037738406099379063, + 0.00392917450517416, + 0.003271800000220537, + 0.025587378069758415, + 0.012273324653506279, + 0.022238489240407944, + 0.014898584224283695, + 0.014148206450045109, + 0.015963153913617134, + 0.014922482892870903, + 0.03677815943956375, + 0.0038888088893145323, + 0.003505094675347209, + 0.004380024038255215, + 0.0036367755383253098, + 0.004029462113976479, + 0.002784106880426407, + 0.0028794966638088226, + 0.0035194919910281897, + 0.01684751734137535, + 0.02648218721151352, + 0.01834987848997116, + 0.014953285455703735, + 0.01279712375253439, + 0.013945984654128551, + 0.026371289044618607, + 0.014858814887702465, + 0.0072715324349701405, + 0.009455992840230465, + 0.007705154363065958, + 0.008007766678929329, + 0.008118174970149994, + 0.0046872771345078945, + 0.010215303860604763, + 0.010318724438548088, + 0.008507156744599342, + 0.014301897957921028, + 0.009562651626765728, + 0.009922310709953308, + 0.011510868556797504, + 0.011291895061731339, + 0.013632065616548061, + 0.013538271188735962, + 0.025455661118030548, + 0.022212540730834007, + 0.019177569076418877, + 0.025105606764554977, + 0.018648743629455566, + 0.016363758593797684, + 0.027055833488702774, + 0.02300652302801609 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 4 + ], + "selected_reward": -0.04869883507490158 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 following the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject a Siberian tiger with orange, white, and black fur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the tiger running at full speed across snow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the tiger's movement cause deep snow displacement creating a shimmering cloud of snow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the tiger's breath visible as white mist in the cold 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": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.5020983219146729 + ], + "opposite_logprob_tokens": [ + -0.25209832191467285 + ], + "logprob_reward": -1.5020983219146729, + "yes_no_margin_reward": -1.25, + "token_credit_reward": -1.5020983219146729, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.09847745299339294, + 0.09657081961631775, + 0.09799658507108688, + 0.05663921684026718, + 0.15817560255527496, + 0.04343278706073761, + 0.1579252928495407, + 0.08979692310094833, + 0.14439353346824646, + 0.05659179762005806 + ], + "credit_map": [ + 0.008801272138953209, + 0.008637141436338425, + 0.020671427249908447, + 0.012730353511869907, + 0.007541967555880547, + 0.02126995474100113, + 0.008563772775232792, + 0.010261563584208488, + 0.004519140347838402, + 0.009748662821948528, + 0.009264513850212097, + 0.008511547930538654, + 0.005394706968218088, + 0.025362728163599968, + 0.006047513801604509, + 0.027722008526325226, + 0.0036275070160627365, + 0.004660241771489382, + 0.004437965806573629, + 0.007936818525195122, + 0.003520399332046509, + 0.03447870537638664, + 0.005968576762825251, + 0.0333663709461689, + 0.0035974362399429083, + 0.00595601461827755, + 0.004950724076479673, + 0.009914956986904144, + 0.004793431144207716, + 0.008671288378536701, + 0.008773071691393852, + 0.009982291609048843, + 0.029593346640467644, + 0.011992892250418663, + 0.026671525090932846, + 0.009443476796150208, + 0.006133037153631449, + 0.01750870794057846, + 0.010132018476724625, + 0.04670059308409691, + 0.008278660476207733, + 0.004234194289892912, + 0.005580200348049402, + 0.0039796046912670135, + 0.004483404103666544, + 0.003638937370851636, + 0.0047832452692091465, + 0.008454539813101292, + 0.014694228768348694, + 0.021092332899570465, + 0.016325807198882103, + 0.02014092728495598, + 0.01608937792479992, + 0.015202781185507774, + 0.0375371053814888, + 0.016842741519212723, + 0.01264965534210205, + 0.008717714808881283, + 0.011274381540715694, + 0.014754575677216053, + 0.008770810440182686, + 0.0067460504360497, + 0.01081453263759613, + 0.01606919802725315, + 0.015647754073143005, + 0.01993713155388832, + 0.016320958733558655, + 0.015660567209124565, + 0.018602972850203514, + 0.01931200549006462, + 0.01853085309267044, + 0.02038128860294819, + 0.008911057375371456, + 0.00380720105022192, + 0.005910825915634632, + 0.005708906799554825, + 0.004903487395495176, + 0.006613810546696186, + 0.01135115697979927, + 0.009385352954268456 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 6, + 8 + ], + "selected_reward": -1.5020983219146729 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 following the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject a Siberian tiger with orange, white, and black fur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the tiger running at full speed across snow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the tiger's movement cause deep snow displacement creating a shimmering cloud of snow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the tiger's breath visible as white mist in the cold 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": [ + -0.17995557188987732 + ], + "opposite_logprob_tokens": [ + -1.8049556016921997 + ], + "logprob_reward": -0.17995557188987732, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.17995557188987732, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17939087748527527, + 0.1033143699169159, + 0.04978041350841522, + 0.17906489968299866, + 0.1653667688369751, + 0.07768071442842484, + 0.08931485563516617, + 0.03214828670024872, + 0.049679145216941833, + 0.0742596760392189 + ], + "credit_map": [ + 0.019003624096512794, + 0.01501883752644062, + 0.034180767834186554, + 0.02070939727127552, + 0.01232024747878313, + 0.030238892883062363, + 0.018146051093935966, + 0.029773063957691193, + 0.008399516344070435, + 0.009899908676743507, + 0.008922751061618328, + 0.008766468614339828, + 0.006940321531146765, + 0.02556478977203369, + 0.008196878246963024, + 0.026623733341693878, + 0.0030724157113581896, + 0.004812285304069519, + 0.0051704938523471355, + 0.009479761123657227, + 0.004046380985528231, + 0.008574161678552628, + 0.0069001419469714165, + 0.007724774535745382, + 0.009731382131576538, + 0.01784217543900013, + 0.014749450609087944, + 0.029795000329613686, + 0.016236139461398125, + 0.02831834927201271, + 0.029919356107711792, + 0.03247303143143654, + 0.031512804329395294, + 0.014487004838883877, + 0.02353646233677864, + 0.014319399371743202, + 0.010391426272690296, + 0.01867697201669216, + 0.011642689816653728, + 0.040800005197525024, + 0.013319971039891243, + 0.007880988530814648, + 0.01154223270714283, + 0.010113619267940521, + 0.006793831940740347, + 0.007529694586992264, + 0.009872360154986382, + 0.010628020390868187, + 0.01009062398225069, + 0.010837094858288765, + 0.009891138412058353, + 0.011308164335787296, + 0.007765039335936308, + 0.009447569958865643, + 0.017749249935150146, + 0.012225974351167679, + 0.0030391414184123278, + 0.0044704219326376915, + 0.003994633909314871, + 0.004735398106276989, + 0.0032333508133888245, + 0.0027110716328024864, + 0.00427238130941987, + 0.005691885948181152, + 0.009775995276868343, + 0.007545184809714556, + 0.004765556193888187, + 0.005037850700318813, + 0.0058250646106898785, + 0.006159621756523848, + 0.004336425568908453, + 0.006233443971723318, + 0.014784622937440872, + 0.006085867527872324, + 0.00811524223536253, + 0.007337046321481466, + 0.006714410148561001, + 0.008223813958466053, + 0.010617407038807869, + 0.012381269596517086 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 4 + ], + "selected_reward": -0.17995557188987732 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 following the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject a Siberian tiger with orange, white, and black fur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the tiger running at full speed across snow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the tiger's movement cause deep snow displacement creating a shimmering cloud of snow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the tiger's breath visible as white mist in the cold 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": [ + -0.34867945313453674 + ], + "opposite_logprob_tokens": [ + -1.2236794233322144 + ], + "logprob_reward": -0.34867945313453674, + "yes_no_margin_reward": 0.875, + "token_credit_reward": -0.34867945313453674, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07698871195316315, + 0.09153321385383606, + 0.04064783826470375, + 0.09136059880256653, + 0.4203706979751587, + 0.041775111109018326, + 0.1204366385936737, + 0.020938191562891006, + 0.07593358308076859, + 0.02001546509563923 + ], + "credit_map": [ + 0.006858743727207184, + 0.0066703991033136845, + 0.013213049620389938, + 0.00916279200464487, + 0.007168974727392197, + 0.01662365347146988, + 0.008192515932023525, + 0.009098577313125134, + 0.0037199181970208883, + 0.003344404511153698, + 0.005673140287399292, + 0.004571378696709871, + 0.0033444520086050034, + 0.032339468598365784, + 0.004613555036485195, + 0.03392689302563667, + 0.002945044543594122, + 0.0033228208776563406, + 0.003433245699852705, + 0.0072418698109686375, + 0.003004655474796891, + 0.008387280628085136, + 0.004615125712007284, + 0.007697795517742634, + 0.005390350706875324, + 0.010389357805252075, + 0.008005339652299881, + 0.01649756357073784, + 0.007811269722878933, + 0.013715988025069237, + 0.013922955840826035, + 0.015627771615982056, + 0.08068740367889404, + 0.030495410785079002, + 0.07132329046726227, + 0.024057544767856598, + 0.017242269590497017, + 0.046848688274621964, + 0.02732171304523945, + 0.12239436060190201, + 0.007617866154760122, + 0.004580739885568619, + 0.006271689664572477, + 0.00495602423325181, + 0.004877655301243067, + 0.003907229285687208, + 0.004105756059288979, + 0.005458150524646044, + 0.01205410622060299, + 0.018381966277956963, + 0.013591405004262924, + 0.015389290638267994, + 0.009602490812540054, + 0.008504743687808514, + 0.02934030070900917, + 0.013572334311902523, + 0.002880838932469487, + 0.002564247464761138, + 0.002596950391307473, + 0.0027186654042452574, + 0.002051972085610032, + 0.0016418439336121082, + 0.0030422527343034744, + 0.003441422013565898, + 0.009233161807060242, + 0.011565182358026505, + 0.008401179686188698, + 0.009442036040127277, + 0.007799427956342697, + 0.008792766369879246, + 0.011587478220462799, + 0.009112347848713398, + 0.0034701619297266006, + 0.0015710884472355247, + 0.00228204857558012, + 0.0023507343139499426, + 0.0017480759415775537, + 0.002063983352854848, + 0.003761703846976161, + 0.0027676692698150873 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 6, + 1 + ], + "selected_reward": -0.34867945313453674 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "No", + "Yes", + "Yes" + ] + }, + { + "prompt": "striking low-angle tracking close-up follows a Siberian tiger at full speed across the snow. Its thick, shaggy orange and white fur is streaked with striking black stripes, and the taut muscles are clearly visible with every flex, radiating unparalleled strength. The beast's amber eyes, narrowed in intense concentration, lock onto its target, its scorching breath instantly condensing into a white mist in the sub-zero air. Its muscular body surges forward with breathtaking force, its limbs striking powerfully with each landing. Each landing sinks deeply into the untouched powdery snow below, sending a powerful, shimmering cloud of snow backward. This cloud gleams like diamonds in the cold winter sun, leaving a dynamic, diffuse white trail behind it. The entire scene, shimmering against a cold palette, is imbued with a sense of primal power and speed, creating a tense, breathtaking atmosphere of pursuit.", + "seed": 1975, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000026/group_10/index_8786_seed_1975_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": -0.27142059793695805, + "vlm_logprob": -0.27142059793695805, + "vlm_yes_no_margin": 2.4, + "vlm_token_credit": -0.27142059793695805, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle following the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject a Siberian tiger with orange, white, and black fur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the tiger running at full speed across snow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the tiger's movement cause deep snow displacement creating a shimmering cloud of snow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the tiger's breath visible as white mist in the cold 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": [ + -0.22589701414108276 + ], + "opposite_logprob_tokens": [ + -2.1008970737457275 + ], + "logprob_reward": -0.22589701414108276, + "yes_no_margin_reward": 1.875, + "token_credit_reward": -0.22589701414108276, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10048028826713562, + 0.12114586681127548, + 0.04323861747980118, + 0.08286925405263901, + 0.052260857075452805, + 0.2437278777360916, + 0.04144212231040001, + 0.09235246479511261, + 0.12023553997278214, + 0.10224707424640656 + ], + "credit_map": [ + 0.010050255805253983, + 0.017237108200788498, + 0.012485872022807598, + 0.018613962456583977, + 0.011761895380914211, + 0.008236102759838104, + 0.009196595288813114, + 0.012898508459329605, + 0.015420939773321152, + 0.01183291245251894, + 0.016798045486211777, + 0.01642715185880661, + 0.012051691301167011, + 0.020628012716770172, + 0.013498954474925995, + 0.0144881596788764, + 0.004303579218685627, + 0.004989597015082836, + 0.006473785266280174, + 0.00712204072624445, + 0.004813748877495527, + 0.00507358368486166, + 0.0070493705570697784, + 0.003412913763895631, + 0.008466556668281555, + 0.012622568756341934, + 0.009866305626928806, + 0.009253256022930145, + 0.008784334175288677, + 0.01827949285507202, + 0.006616593338549137, + 0.008980152197182178, + 0.004660762380808592, + 0.005698698107153177, + 0.007097475230693817, + 0.007364355493336916, + 0.006607747636735439, + 0.00839957594871521, + 0.0072653451934456825, + 0.005166903603821993, + 0.05149069055914879, + 0.03406863659620285, + 0.05299711227416992, + 0.030502093955874443, + 0.015718819573521614, + 0.01351878885179758, + 0.020624369382858276, + 0.02480737306177616, + 0.004393679555505514, + 0.00468391552567482, + 0.0057078637182712555, + 0.007836964912712574, + 0.005786441732198, + 0.004465377889573574, + 0.005021572113037109, + 0.003546312218531966, + 0.012639578431844711, + 0.010801132768392563, + 0.0110276248306036, + 0.015534854494035244, + 0.013050139881670475, + 0.008053326979279518, + 0.013847572728991508, + 0.007398237474262714, + 0.013408169150352478, + 0.014268883503973484, + 0.014505064114928246, + 0.016723953187465668, + 0.0197979174554348, + 0.010699467733502388, + 0.020247148349881172, + 0.010584944859147072, + 0.010517364367842674, + 0.013965483754873276, + 0.014704430475831032, + 0.013230212964117527, + 0.016562288627028465, + 0.013777053914964199, + 0.011577654629945755, + 0.007912594825029373 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 1, + 8 + ], + "selected_reward": -0.22589701414108276 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 following the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject a Siberian tiger with orange, white, and black fur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the tiger running at full speed across snow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the tiger's movement cause deep snow displacement creating a shimmering cloud of snow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the tiger's breath visible as white mist in the cold 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": [ + -0.004124231170862913 + ], + "opposite_logprob_tokens": [ + -5.504124164581299 + ], + "logprob_reward": -0.004124231170862913, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004124231170862913, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09864024817943573, + 0.11659206449985504, + 0.10904878377914429, + 0.1049252450466156, + 0.033323317766189575, + 0.1456238180398941, + 0.16259367763996124, + 0.08098400384187698, + 0.05845612660050392, + 0.089812733232975 + ], + "credit_map": [ + 0.007233699783682823, + 0.011445816606283188, + 0.016812028363347054, + 0.014607243239879608, + 0.010177026502788067, + 0.010270192287862301, + 0.01357996929436922, + 0.014514269307255745, + 0.01517628226429224, + 0.010715669952332973, + 0.01504219975322485, + 0.012790558859705925, + 0.008929645642638206, + 0.030248677358031273, + 0.011053226888179779, + 0.012635795399546623, + 0.01704123429954052, + 0.01385307777673006, + 0.016925936564803123, + 0.017055058851838112, + 0.013133318163454533, + 0.007170610595494509, + 0.015611108392477036, + 0.008258435875177383, + 0.01191712636500597, + 0.0181563887745142, + 0.014454093761742115, + 0.015387781895697117, + 0.009607535786926746, + 0.014493796974420547, + 0.010827431455254555, + 0.010081084445118904, + 0.0037733656354248524, + 0.00428043445572257, + 0.003976046107709408, + 0.005533706396818161, + 0.004922959953546524, + 0.0034040212631225586, + 0.00458826869726181, + 0.002844514325261116, + 0.031062504276633263, + 0.02084212936460972, + 0.01965436153113842, + 0.023238172754645348, + 0.012429456226527691, + 0.009184409864246845, + 0.01779256947338581, + 0.011420206166803837, + 0.01624525897204876, + 0.011863290332257748, + 0.01398223452270031, + 0.026294350624084473, + 0.021859202533960342, + 0.015507607720792294, + 0.03529861569404602, + 0.021543117240071297, + 0.01287230383604765, + 0.010031930170953274, + 0.010067357681691647, + 0.009225131943821907, + 0.015838488936424255, + 0.006660543382167816, + 0.009807989932596684, + 0.006480251904577017, + 0.008584653958678246, + 0.007214686367660761, + 0.004855478648096323, + 0.005081395152956247, + 0.011945562437176704, + 0.004936599172651768, + 0.010660052299499512, + 0.005177701823413372, + 0.006906654220074415, + 0.015631137415766716, + 0.007096131332218647, + 0.027204662561416626, + 0.009365654550492764, + 0.010857746005058289, + 0.008240229450166225, + 0.0045105209574103355 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 5, + 1 + ], + "selected_reward": -0.004124231170862913 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 following the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject a Siberian tiger with orange, white, and black fur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the tiger running at full speed across snow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the tiger's movement cause deep snow displacement creating a shimmering cloud of snow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the tiger's breath visible as white mist in the cold 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": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.7578449249267578 + ], + "opposite_logprob_tokens": [ + -0.6328449249267578 + ], + "logprob_reward": -0.7578449249267578, + "yes_no_margin_reward": -0.125, + "token_credit_reward": -0.7578449249267578, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.06949130445718765, + 0.20001102983951569, + 0.018543768674135208, + 0.04982854425907135, + 0.05645318701863289, + 0.0879468247294426, + 0.04421830549836159, + 0.05498199164867401, + 0.3696525990962982, + 0.04887254536151886 + ], + "credit_map": [ + 0.005143207963556051, + 0.009097306989133358, + 0.010406633839011192, + 0.008421818725764751, + 0.007017939817160368, + 0.005063211545348167, + 0.013270782306790352, + 0.01107039488852024, + 0.021790482103824615, + 0.022896528244018555, + 0.030624015256762505, + 0.03254315257072449, + 0.020894285291433334, + 0.02311682514846325, + 0.024942239746451378, + 0.023203495889902115, + 0.0025238285306841135, + 0.0018905291799455881, + 0.002759421244263649, + 0.002014386933296919, + 0.0022232928313314915, + 0.002081929938867688, + 0.0033079669810831547, + 0.0017424126854166389, + 0.004468139261007309, + 0.005835522897541523, + 0.010886686854064465, + 0.005744308233261108, + 0.005084588658064604, + 0.0066959550604224205, + 0.00557118421420455, + 0.005542156286537647, + 0.005150289740413427, + 0.005199949722737074, + 0.008188721723854542, + 0.008427813649177551, + 0.005517217796295881, + 0.00988524779677391, + 0.008493113331496716, + 0.005590828601270914, + 0.016301577910780907, + 0.00967294443398714, + 0.016241319477558136, + 0.013074951246380806, + 0.008425191976130009, + 0.006760534830391407, + 0.009455038234591484, + 0.008015265688300133, + 0.004291371908038855, + 0.002636657329276204, + 0.005561966449022293, + 0.009965814650058746, + 0.003735503414645791, + 0.003412653924897313, + 0.00887384731322527, + 0.005740485619753599, + 0.007334304042160511, + 0.006276754196733236, + 0.008094118908047676, + 0.008198020979762077, + 0.008825908415019512, + 0.006077868863940239, + 0.006311607081443071, + 0.003863404504954815, + 0.027061188593506813, + 0.047930557280778885, + 0.041060369461774826, + 0.04850267246365547, + 0.060612261295318604, + 0.03533240035176277, + 0.05081752687692642, + 0.058335576206445694, + 0.005502844229340553, + 0.00617957254871726, + 0.006603181827813387, + 0.00916646420955658, + 0.005215959623456001, + 0.0049598487094044685, + 0.006347510498017073, + 0.004897159989923239 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 1, + 5 + ], + "selected_reward": -0.7578449249267578 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 following the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject a Siberian tiger with orange, white, and black fur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the tiger running at full speed across snow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the tiger's movement cause deep snow displacement creating a shimmering cloud of snow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the tiger's breath visible as white mist in the cold 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": [ + -0.02058931440114975 + ], + "opposite_logprob_tokens": [ + -3.8955893516540527 + ], + "logprob_reward": -0.02058931440114975, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.02058931440114975, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09993354976177216, + 0.3669022023677826, + 0.04603194817900658, + 0.04030357301235199, + 0.025390496477484703, + 0.04495089873671532, + 0.036732301115989685, + 0.04540308937430382, + 0.07126309722661972, + 0.22308886051177979 + ], + "credit_map": [ + 0.0071489824913442135, + 0.010056170634925365, + 0.014630252495408058, + 0.01321377418935299, + 0.012373900040984154, + 0.018431348726153374, + 0.010858055204153061, + 0.013221057131886482, + 0.05638891085982323, + 0.03419801592826843, + 0.06808269768953323, + 0.05815105512738228, + 0.02415185049176216, + 0.0382971353828907, + 0.04205082729458809, + 0.045581694692373276, + 0.005499475635588169, + 0.005006498657166958, + 0.0069352248683571815, + 0.007120792288333178, + 0.005397305358201265, + 0.0046179029159247875, + 0.005891453009098768, + 0.005563293118029833, + 0.004682764410972595, + 0.006632886826992035, + 0.0051438696682453156, + 0.0062287477776408195, + 0.004566727671772242, + 0.00502766203135252, + 0.004339674953371286, + 0.003681236645206809, + 0.002837536158040166, + 0.002421563258394599, + 0.003159409621730447, + 0.0037815100513398647, + 0.003751302370801568, + 0.0025021331384778023, + 0.004530541133135557, + 0.002406500047072768, + 0.007969153113663197, + 0.004864983726292849, + 0.0070279547944664955, + 0.007375690154731274, + 0.005038462579250336, + 0.00327453576028347, + 0.005300442222505808, + 0.004099673125892878, + 0.004849211312830448, + 0.00424279784783721, + 0.00469714030623436, + 0.005426377058029175, + 0.00524761900305748, + 0.005132503341883421, + 0.0033916051033884287, + 0.0037450429517775774, + 0.007967263460159302, + 0.0051656244322657585, + 0.005509052891284227, + 0.005323270335793495, + 0.008047821931540966, + 0.0036963182501494884, + 0.006113763898611069, + 0.0035799702163785696, + 0.007652360945940018, + 0.009531361050903797, + 0.009826760739088058, + 0.008497203700244427, + 0.01241331733763218, + 0.006698779296129942, + 0.008728119544684887, + 0.007915188558399677, + 0.022486798465251923, + 0.03203914687037468, + 0.025821516290307045, + 0.04280105605721474, + 0.022054504603147507, + 0.029628708958625793, + 0.02567857876420021, + 0.022578544914722443 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 9, + 0 + ], + "selected_reward": -0.02058931440114975 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 following the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject a Siberian tiger with orange, white, and black fur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the tiger running at full speed across snow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the tiger's movement cause deep snow displacement creating a shimmering cloud of snow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the tiger's breath visible as white mist in the cold 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": [ + -0.34864750504493713 + ], + "opposite_logprob_tokens": [ + -1.2236474752426147 + ], + "logprob_reward": -0.34864750504493713, + "yes_no_margin_reward": 0.875, + "token_credit_reward": -0.34864750504493713, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16294096410274506, + 0.2525902986526489, + 0.08353700488805771, + 0.03526666387915611, + 0.04242158681154251, + 0.06442415714263916, + 0.08536577969789505, + 0.04569980502128601, + 0.1239272728562355, + 0.10382647812366486 + ], + "credit_map": [ + 0.016110053285956383, + 0.020827749744057655, + 0.02232595719397068, + 0.030479751527309418, + 0.01631942018866539, + 0.01800566166639328, + 0.019553225487470627, + 0.019319145008921623, + 0.03390650078654289, + 0.021160488948225975, + 0.03395235911011696, + 0.031250692903995514, + 0.018910350278019905, + 0.06436946243047714, + 0.02106303907930851, + 0.027977412566542625, + 0.012652418576180935, + 0.0079762302339077, + 0.013130762614309788, + 0.012194052338600159, + 0.009866548702120781, + 0.009143034927546978, + 0.01047473307698965, + 0.008099228143692017, + 0.0042732711881399155, + 0.005187641363590956, + 0.0050231823697686195, + 0.005324209108948708, + 0.0034431859385222197, + 0.005523917265236378, + 0.003311923937872052, + 0.003179332474246621, + 0.0036381734535098076, + 0.004256183747202158, + 0.005878569092601538, + 0.00700633367523551, + 0.0070021869614720345, + 0.0036454531364142895, + 0.0072676450945436954, + 0.003727039322257042, + 0.009466956369578838, + 0.006109698209911585, + 0.010789785534143448, + 0.009269637987017632, + 0.008857271634042263, + 0.005653218366205692, + 0.009355577640235424, + 0.004922011401504278, + 0.011823047883808613, + 0.009388153441250324, + 0.010599873960018158, + 0.01380446832627058, + 0.008339455351233482, + 0.008761248551309109, + 0.013734295032918453, + 0.008915234357118607, + 0.004446478094905615, + 0.005436614155769348, + 0.005996683146804571, + 0.007525720167905092, + 0.005808098707348108, + 0.006453755311667919, + 0.006008096970617771, + 0.004024357069283724, + 0.01830090954899788, + 0.015666089951992035, + 0.012761936523020267, + 0.012378535233438015, + 0.024455195292830467, + 0.011235308833420277, + 0.016545595601201057, + 0.012583700940012932, + 0.009491067379713058, + 0.017463218420743942, + 0.010814272798597813, + 0.020464159548282623, + 0.015354380942881107, + 0.011975432746112347, + 0.012203866615891457, + 0.0060600885190069675 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 8 + ], + "selected_reward": -0.34864750504493713 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "No", + "Yes", + "Yes" + ] + }, + { + "prompt": "striking low-angle tracking close-up follows a Siberian tiger at full speed across the snow. Its thick, shaggy orange and white fur is streaked with striking black stripes, and the taut muscles are clearly visible with every flex, radiating unparalleled strength. The beast's amber eyes, narrowed in intense concentration, lock onto its target, its scorching breath instantly condensing into a white mist in the sub-zero air. Its muscular body surges forward with breathtaking force, its limbs striking powerfully with each landing. Each landing sinks deeply into the untouched powdery snow below, sending a powerful, shimmering cloud of snow backward. This cloud gleams like diamonds in the cold winter sun, leaving a dynamic, diffuse white trail behind it. The entire scene, shimmering against a cold palette, is imbued with a sense of primal power and speed, creating a tense, breathtaking atmosphere of pursuit.", + "seed": 1976, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000026/group_10/index_8786_seed_1976_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": -0.34223737195134163, + "vlm_logprob": -0.34223737195134163, + "vlm_yes_no_margin": 2.1499999880790712, + "vlm_token_credit": -0.34223737195134163, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle following the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject a Siberian tiger with orange, white, and black fur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the tiger running at full speed across snow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the tiger's movement cause deep snow displacement creating a shimmering cloud of snow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the tiger's breath visible as white mist in the cold 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": [ + -0.0991191491484642 + ], + "opposite_logprob_tokens": [ + -3.349119186401367 + ], + "logprob_reward": -0.0991191491484642, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.0991191491484642, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15681900084018707, + 0.06360254436731339, + 0.04716029390692711, + 0.05410816892981529, + 0.0538385845720768, + 0.12736374139785767, + 0.17381726205348969, + 0.05130377411842346, + 0.09987694770097733, + 0.1721096783876419 + ], + "credit_map": [ + 0.01765533536672592, + 0.023363806307315826, + 0.02033540979027748, + 0.024956725537776947, + 0.018136421218514442, + 0.012730197980999947, + 0.021822314709424973, + 0.01781878061592579, + 0.0061564515344798565, + 0.008258186280727386, + 0.00912513118237257, + 0.006474157329648733, + 0.006693277508020401, + 0.0058458284474909306, + 0.015090225264430046, + 0.005959285888820887, + 0.008197708986699581, + 0.00732165714725852, + 0.005740133114159107, + 0.005259418860077858, + 0.004957476630806923, + 0.004013723693788052, + 0.006165973376482725, + 0.005504204425960779, + 0.007002856116741896, + 0.006371918600052595, + 0.007749389391392469, + 0.0062038288451731205, + 0.006058807484805584, + 0.006508634425699711, + 0.007894659414887428, + 0.006318077445030212, + 0.004351196810603142, + 0.007110480219125748, + 0.006200126837939024, + 0.009640219621360302, + 0.006505259312689304, + 0.006664137821644545, + 0.007332128472626209, + 0.0060350364074110985, + 0.030695591121912003, + 0.008873275481164455, + 0.02559409663081169, + 0.010792126879096031, + 0.010432206094264984, + 0.01283289585262537, + 0.015529577620327473, + 0.01261395774781704, + 0.024911964312195778, + 0.015332122333347797, + 0.036289334297180176, + 0.02592211216688156, + 0.01481469627469778, + 0.014875762164592743, + 0.018920129165053368, + 0.022751135751605034, + 0.005213466472923756, + 0.006845011841505766, + 0.005174548830837011, + 0.009260042570531368, + 0.004351137205958366, + 0.005684955045580864, + 0.008093760348856449, + 0.006680849473923445, + 0.008491731248795986, + 0.013740096241235733, + 0.014712036587297916, + 0.017518755048513412, + 0.011955180205404758, + 0.007067176513373852, + 0.011282750405371189, + 0.015109220519661903, + 0.02283254824578762, + 0.013579404912889004, + 0.02807379513978958, + 0.017985569313168526, + 0.01380967441946268, + 0.01922416500747204, + 0.015267746523022652, + 0.0413367860019207 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 9, + 0 + ], + "selected_reward": -0.0991191491484642 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 following the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject a Siberian tiger with orange, white, and black fur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the tiger running at full speed across snow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the tiger's movement cause deep snow displacement creating a shimmering cloud of snow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the tiger's breath visible as white mist in the cold 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": [ + -0.006756555289030075 + ], + "opposite_logprob_tokens": [ + -5.006756782531738 + ], + "logprob_reward": -0.006756555289030075, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006756555289030075, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12704338133335114, + 0.0457371287047863, + 0.10867788642644882, + 0.03678767383098602, + 0.09316547214984894, + 0.07252180576324463, + 0.20676939189434052, + 0.054188359528779984, + 0.14393532276153564, + 0.11117357760667801 + ], + "credit_map": [ + 0.01607563905417919, + 0.016934959217905998, + 0.017821326851844788, + 0.019017327576875687, + 0.009934168308973312, + 0.013454815372824669, + 0.017580870538949966, + 0.016224266961216927, + 0.00507403165102005, + 0.006743940059095621, + 0.006855284795165062, + 0.0049326298758387566, + 0.004772478714585304, + 0.00389685807749629, + 0.0074772522784769535, + 0.005984654184430838, + 0.012388993985950947, + 0.012253490276634693, + 0.014337420463562012, + 0.016133062541484833, + 0.007655440829694271, + 0.011968549340963364, + 0.019602037966251373, + 0.01433888915926218, + 0.0046205404214560986, + 0.0032974127680063248, + 0.0053175268694758415, + 0.005199606064707041, + 0.002605619840323925, + 0.004843676462769508, + 0.006348077207803726, + 0.00455521373078227, + 0.0071319593116641045, + 0.011810715310275555, + 0.012189882807433605, + 0.0188219603151083, + 0.009656158275902271, + 0.01092738937586546, + 0.013614407740533352, + 0.009013006463646889, + 0.017894914373755455, + 0.009287809953093529, + 0.012572704814374447, + 0.00758970994502306, + 0.005516566336154938, + 0.006042199209332466, + 0.006796091794967651, + 0.006821811199188232, + 0.03513624519109726, + 0.014645643532276154, + 0.027438096702098846, + 0.019351331517100334, + 0.01583251543343067, + 0.022169264033436775, + 0.013837289065122604, + 0.058358997106552124, + 0.006831306032836437, + 0.00824399758130312, + 0.006236738990992308, + 0.006898768711835146, + 0.006543928291648626, + 0.004507035482674837, + 0.007810512557625771, + 0.007116072811186314, + 0.010908600874245167, + 0.013302816078066826, + 0.022148072719573975, + 0.0236226674169302, + 0.019314199686050415, + 0.016031809151172638, + 0.02422015741467476, + 0.014386997558176517, + 0.024563120678067207, + 0.013002526015043259, + 0.014667578972876072, + 0.018527444452047348, + 0.014175374992191792, + 0.006954893935471773, + 0.010422361083328724, + 0.008860272355377674 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 8, + 0 + ], + "selected_reward": -0.006756555289030075 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 following the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject a Siberian tiger with orange, white, and black fur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the tiger running at full speed across snow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the tiger's movement cause deep snow displacement creating a shimmering cloud of snow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the tiger's breath visible as white mist in the cold 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": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.223628044128418 + ], + "opposite_logprob_tokens": [ + -0.3486279845237732 + ], + "logprob_reward": -1.223628044128418, + "yes_no_margin_reward": -0.8750000596046448, + "token_credit_reward": -1.223628044128418, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.04914569854736328, + 0.029481561854481697, + 0.052409734576940536, + 0.040416136384010315, + 0.047634996473789215, + 0.04883170872926712, + 0.19940567016601562, + 0.0997392013669014, + 0.07077650725841522, + 0.36215874552726746 + ], + "credit_map": [ + 0.005721197463572025, + 0.005658699665218592, + 0.007450429257005453, + 0.008396853692829609, + 0.005078468006104231, + 0.003431423334404826, + 0.006062462460249662, + 0.007346163969486952, + 0.00477593345567584, + 0.003385276300832629, + 0.004205997101962566, + 0.002928280970081687, + 0.002090533496811986, + 0.0026278900913894176, + 0.005620963405817747, + 0.0038466895930469036, + 0.00623659323900938, + 0.007469724398106337, + 0.006232344079762697, + 0.007980264723300934, + 0.0041559794917702675, + 0.007314383052289486, + 0.007099598180502653, + 0.00592084601521492, + 0.005271231755614281, + 0.004043450579047203, + 0.0062946234829723835, + 0.005004740320146084, + 0.0034742122516036034, + 0.003954468760639429, + 0.007018642500042915, + 0.005354768596589565, + 0.008458506315946579, + 0.007837604731321335, + 0.006437502335757017, + 0.0054925852455198765, + 0.0065068444237113, + 0.003251423127949238, + 0.005415493156760931, + 0.004235036205500364, + 0.011119358241558075, + 0.004859697073698044, + 0.009334255009889603, + 0.00513314688578248, + 0.00325383385643363, + 0.00484563410282135, + 0.005978678353130817, + 0.004307109862565994, + 0.02077440544962883, + 0.01947772689163685, + 0.030656984075903893, + 0.032925933599472046, + 0.016478389501571655, + 0.025318549945950508, + 0.024775870144367218, + 0.028997818008065224, + 0.013712799176573753, + 0.014630638062953949, + 0.01081080362200737, + 0.012901190668344498, + 0.012678809463977814, + 0.007852689363062382, + 0.016427986323833466, + 0.010724285617470741, + 0.0064794584177434444, + 0.011829806491732597, + 0.009885481558740139, + 0.010692042298614979, + 0.009531966410577297, + 0.004905014298856258, + 0.008908764459192753, + 0.008543968200683594, + 0.030002832412719727, + 0.02014288306236267, + 0.0733119398355484, + 0.04073910787701607, + 0.028228458017110825, + 0.0385819748044014, + 0.026856765151023865, + 0.10429476946592331 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 6, + 7 + ], + "selected_reward": -1.223628044128418 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 following the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject a Siberian tiger with orange, white, and black fur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the tiger running at full speed across snow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the tiger's movement cause deep snow displacement creating a shimmering cloud of snow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the tiger's breath visible as white mist in the cold 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": [ + -0.10033458471298218 + ], + "opposite_logprob_tokens": [ + -2.350334644317627 + ], + "logprob_reward": -0.10033458471298218, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10033458471298218, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07396001368761063, + 0.06466744840145111, + 0.05526590719819069, + 0.04457137733697891, + 0.16280949115753174, + 0.07501346617937088, + 0.10877549648284912, + 0.2140650600194931, + 0.04819153621792793, + 0.1526802033185959 + ], + "credit_map": [ + 0.009164894931018353, + 0.0087443757802248, + 0.00911037903279066, + 0.011996043846011162, + 0.00685559818521142, + 0.008201396092772484, + 0.010730770416557789, + 0.00915655866265297, + 0.007644874043762684, + 0.00885669607669115, + 0.008489097468554974, + 0.006864544004201889, + 0.006509751547127962, + 0.006026201881468296, + 0.012370838783681393, + 0.007905448786914349, + 0.00644939299672842, + 0.005989055149257183, + 0.007381292060017586, + 0.006371898576617241, + 0.0051046619191765785, + 0.005682863760739565, + 0.009788674302399158, + 0.008498069830238819, + 0.006135306321084499, + 0.004501502029597759, + 0.00683669513091445, + 0.004295777529478073, + 0.003009357023984194, + 0.004360899329185486, + 0.007870475761592388, + 0.007561365142464638, + 0.019243450835347176, + 0.026135096326470375, + 0.019963670521974564, + 0.02492183819413185, + 0.01795782335102558, + 0.01656181551516056, + 0.019721902906894684, + 0.018303895369172096, + 0.018041467294096947, + 0.0045527624897658825, + 0.01579422876238823, + 0.0061240722425282, + 0.005938762333244085, + 0.00841040350496769, + 0.009017782285809517, + 0.0071339840069413185, + 0.01635609194636345, + 0.008745496161282063, + 0.016605228185653687, + 0.01100117526948452, + 0.008551993407309055, + 0.013922358863055706, + 0.01166275143623352, + 0.021930404007434845, + 0.021445196121931076, + 0.014177368953824043, + 0.022712761536240578, + 0.03322596102952957, + 0.01247049868106842, + 0.02079659327864647, + 0.023943694308400154, + 0.06529299169778824, + 0.004863355774432421, + 0.00824266579002142, + 0.006673524621874094, + 0.007897012867033482, + 0.0060376692563295364, + 0.003136832034215331, + 0.0056848987005651, + 0.0056555746123194695, + 0.013194476254284382, + 0.008944042026996613, + 0.031380802392959595, + 0.016713302582502365, + 0.011958518996834755, + 0.016705568879842758, + 0.010665715672075748, + 0.04311778396368027 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 4, + 9 + ], + "selected_reward": -0.10033458471298218 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 following the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject a Siberian tiger with orange, white, and black fur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the tiger running at full speed across snow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the tiger's movement cause deep snow displacement creating a shimmering cloud of snow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the tiger's breath visible as white mist in the cold 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": [ + -0.2813485264778137 + ], + "opposite_logprob_tokens": [ + -1.406348466873169 + ], + "logprob_reward": -0.2813485264778137, + "yes_no_margin_reward": 1.125, + "token_credit_reward": -0.2813485264778137, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13262133300304413, + 0.05503765121102333, + 0.149619460105896, + 0.04654017463326454, + 0.37541669607162476, + 0.04181462526321411, + 0.06029560789465904, + 0.07841806858778, + 0.016968965530395508, + 0.04326748847961426 + ], + "credit_map": [ + 0.009498401544988155, + 0.014548864215612411, + 0.014115136116743088, + 0.025503037497401237, + 0.01312286127358675, + 0.018402541056275368, + 0.02723865956068039, + 0.01019181776791811, + 0.0074972789734601974, + 0.008070525713264942, + 0.00735322805121541, + 0.006823886651545763, + 0.003967875149101019, + 0.004901521373540163, + 0.011171388439834118, + 0.005251941736787558, + 0.02118990197777748, + 0.016612116247415543, + 0.019486207515001297, + 0.02000531367957592, + 0.013433165848255157, + 0.013668904080986977, + 0.023699987679719925, + 0.02152385003864765, + 0.0052785188890993595, + 0.0059271608479321, + 0.005682134535163641, + 0.006757376249879599, + 0.00455984752625227, + 0.004033295903354883, + 0.009587719105184078, + 0.004714117385447025, + 0.03743472695350647, + 0.04533350095152855, + 0.05052604526281357, + 0.07725139707326889, + 0.03508361056447029, + 0.03766358643770218, + 0.053092602640390396, + 0.03903120383620262, + 0.006219043862074614, + 0.003974230028688908, + 0.0061348513700068, + 0.004807981662452221, + 0.005027062259614468, + 0.0039358558133244514, + 0.006999533623456955, + 0.00471606245264411, + 0.011519639752805233, + 0.005476254038512707, + 0.010679562576115131, + 0.007512005046010017, + 0.004899453837424517, + 0.005457039922475815, + 0.00677125109359622, + 0.007980400696396828, + 0.007910741493105888, + 0.010346670635044575, + 0.007460078224539757, + 0.013191631063818932, + 0.004849649500101805, + 0.0077899363823235035, + 0.019118741154670715, + 0.007750607095658779, + 0.002112005138769746, + 0.0023282126057893038, + 0.002152204280719161, + 0.002215051557868719, + 0.0020113298669457436, + 0.0022241815458983183, + 0.0018760866951197386, + 0.002049892907962203, + 0.009147174656391144, + 0.0046962774358689785, + 0.005796860437840223, + 0.007028316147625446, + 0.003835064359009266, + 0.0029936779756098986, + 0.003820136422291398, + 0.0059499782510101795 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 2, + 0 + ], + "selected_reward": -0.2813485264778137 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "No", + "Yes", + "Yes" + ] + }, + { + "prompt": "striking low-angle tracking close-up follows a Siberian tiger at full speed across the snow. Its thick, shaggy orange and white fur is streaked with striking black stripes, and the taut muscles are clearly visible with every flex, radiating unparalleled strength. The beast's amber eyes, narrowed in intense concentration, lock onto its target, its scorching breath instantly condensing into a white mist in the sub-zero air. Its muscular body surges forward with breathtaking force, its limbs striking powerfully with each landing. Each landing sinks deeply into the untouched powdery snow below, sending a powerful, shimmering cloud of snow backward. This cloud gleams like diamonds in the cold winter sun, leaving a dynamic, diffuse white trail behind it. The entire scene, shimmering against a cold palette, is imbued with a sense of primal power and speed, creating a tense, breathtaking atmosphere of pursuit.", + "seed": 1977, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000026/group_10/index_8786_seed_1977_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": -0.19770315401256083, + "vlm_logprob": -0.19770315401256083, + "vlm_yes_no_margin": 2.7249999761581423, + "vlm_token_credit": -0.19770315401256083, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle following the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject a Siberian tiger with orange, white, and black fur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the tiger running at full speed across snow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the tiger's movement cause deep snow displacement creating a shimmering cloud of snow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the tiger's breath visible as white mist in the cold 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": [ + -0.13019981980323792 + ], + "opposite_logprob_tokens": [ + -2.880199909210205 + ], + "logprob_reward": -0.13019981980323792, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.13019981980323792, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07253559678792953, + 0.4799739420413971, + 0.039831940084695816, + 0.051495425403118134, + 0.03725830838084221, + 0.054100651293992996, + 0.08585385233163834, + 0.05745824798941612, + 0.06829883903265, + 0.05319323390722275 + ], + "credit_map": [ + 0.010291730985045433, + 0.007302477955818176, + 0.009969870559871197, + 0.007621282245963812, + 0.007551292423158884, + 0.006606645882129669, + 0.013383449986577034, + 0.009808849543333054, + 0.1282218098640442, + 0.03439247980713844, + 0.07989725470542908, + 0.04193760082125664, + 0.021790679544210434, + 0.04365386441349983, + 0.024348819628357887, + 0.10573145747184753, + 0.008363787084817886, + 0.004475600551813841, + 0.005781034007668495, + 0.004814884625375271, + 0.0039296625182032585, + 0.0034388157073408365, + 0.003772324649617076, + 0.005255831405520439, + 0.007092103827744722, + 0.004529634024947882, + 0.006756552029401064, + 0.0067721763625741005, + 0.00646594725549221, + 0.006163636222481728, + 0.006061045918613672, + 0.0076543292962014675, + 0.0027140723541378975, + 0.0060599143616855145, + 0.004280576016753912, + 0.0041546947322785854, + 0.006843233481049538, + 0.0037483209744095802, + 0.005703996401280165, + 0.0037535016890615225, + 0.004717950709164143, + 0.006091057322919369, + 0.008107216097414494, + 0.005855804309248924, + 0.005268128588795662, + 0.009048002772033215, + 0.006550062447786331, + 0.008462428115308285, + 0.011711442843079567, + 0.012448429130017757, + 0.010447095148265362, + 0.010501818731427193, + 0.0097427386790514, + 0.009432537481188774, + 0.011095749214291573, + 0.010474035516381264, + 0.007252177223563194, + 0.0051061660051345825, + 0.011744902469217777, + 0.005034615285694599, + 0.0047704423777759075, + 0.010434523224830627, + 0.0060727582313120365, + 0.007042666431516409, + 0.005258845165371895, + 0.006806506775319576, + 0.007381858769804239, + 0.012074815109372139, + 0.006652746815234423, + 0.010566453449428082, + 0.011596042662858963, + 0.007961574010550976, + 0.005332483910024166, + 0.005629201419651508, + 0.010216516442596912, + 0.0059145004488527775, + 0.004389024339616299, + 0.006759184412658215, + 0.006995395757257938, + 0.00795692391693592 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 6, + 0 + ], + "selected_reward": -0.13019981980323792 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 following the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject a Siberian tiger with orange, white, and black fur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the tiger running at full speed across snow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the tiger's movement cause deep snow displacement creating a shimmering cloud of snow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the tiger's breath visible as white mist in the cold 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": [ + -0.011093368753790855 + ], + "opposite_logprob_tokens": [ + -4.5110931396484375 + ], + "logprob_reward": -0.011093368753790855, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011093368753790855, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14071686565876007, + 0.2111351191997528, + 0.07584479451179504, + 0.10644616931676865, + 0.04371911287307739, + 0.09680923819541931, + 0.04605519399046898, + 0.13294005393981934, + 0.0705699697136879, + 0.07576359063386917 + ], + "credit_map": [ + 0.017076678574085236, + 0.01398066058754921, + 0.019288770854473114, + 0.017173048108816147, + 0.01667073369026184, + 0.01163096260279417, + 0.015058087185025215, + 0.02983791194856167, + 0.04218898341059685, + 0.02466638572514057, + 0.030717281624674797, + 0.017189864069223404, + 0.023698028177022934, + 0.02148539572954178, + 0.014912028796970844, + 0.036277126520872116, + 0.009064879268407822, + 0.007437815424054861, + 0.010458542965352535, + 0.00956460926681757, + 0.012406336143612862, + 0.006080264691263437, + 0.01096042338758707, + 0.00987191777676344, + 0.012888243421912193, + 0.008643010631203651, + 0.014119195751845837, + 0.020176440477371216, + 0.01091040950268507, + 0.008715060539543629, + 0.01792961172759533, + 0.013064189814031124, + 0.0038991791661828756, + 0.006407707929611206, + 0.005845692008733749, + 0.006539215333759785, + 0.007509078364819288, + 0.004083814565092325, + 0.004203720483928919, + 0.005230702459812164, + 0.007921663112938404, + 0.012399043887853622, + 0.012753386050462723, + 0.00788663886487484, + 0.011788231320679188, + 0.01466264110058546, + 0.01163936872035265, + 0.01775824837386608, + 0.003818843513727188, + 0.005635777488350868, + 0.005565301515161991, + 0.006332344375550747, + 0.005205221939831972, + 0.005815668497234583, + 0.005358203314244747, + 0.008323827758431435, + 0.01686020940542221, + 0.020276522263884544, + 0.023094268515706062, + 0.01580655202269554, + 0.012752717360854149, + 0.02121986635029316, + 0.011690136045217514, + 0.011239762417972088, + 0.0071953232400119305, + 0.009298508986830711, + 0.007625355385243893, + 0.011952555738389492, + 0.005576342809945345, + 0.010007515549659729, + 0.011298931203782558, + 0.007615429349243641, + 0.01373256091028452, + 0.010588113218545914, + 0.011698517017066479, + 0.00865462888032198, + 0.008562753908336163, + 0.007518457714468241, + 0.009394064545631409, + 0.005614486988633871 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 7 + ], + "selected_reward": -0.011093368753790855 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 following the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject a Siberian tiger with orange, white, and black fur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the tiger running at full speed across snow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the tiger's movement cause deep snow displacement creating a shimmering cloud of snow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the tiger's breath visible as white mist in the cold 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": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.6933536529541016 + ], + "opposite_logprob_tokens": [ + -0.6933536529541016 + ], + "logprob_reward": -0.6933536529541016, + "yes_no_margin_reward": 0.0, + "token_credit_reward": -0.6933536529541016, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.046606361865997314, + 0.5896943211555481, + 0.03361339494585991, + 0.08237800747156143, + 0.022589731961488724, + 0.024795599281787872, + 0.03859575092792511, + 0.015924710780382156, + 0.07306624948978424, + 0.07273588329553604 + ], + "credit_map": [ + 0.008341720327734947, + 0.003766447538509965, + 0.006712533067911863, + 0.005063268356025219, + 0.0035679182037711143, + 0.003541350131854415, + 0.0052961017936468124, + 0.010317020118236542, + 0.15434294939041138, + 0.03821761533617973, + 0.09730691462755203, + 0.05520884320139885, + 0.028197849169373512, + 0.05523497238755226, + 0.03400155156850815, + 0.12718361616134644, + 0.005995059851557016, + 0.0038026352412998676, + 0.00471497280523181, + 0.005068651400506496, + 0.003978533670306206, + 0.002717954572290182, + 0.004178420640528202, + 0.003157165599986911, + 0.012821348384022713, + 0.007719381246715784, + 0.010941998101770878, + 0.01473377924412489, + 0.006093790289014578, + 0.00767924590036273, + 0.013050567358732224, + 0.009337887167930603, + 0.002537323394790292, + 0.0024036653339862823, + 0.0037966242525726557, + 0.003168234834447503, + 0.003183143911883235, + 0.00228038989007473, + 0.002492496045306325, + 0.002727852901443839, + 0.0027424178551882505, + 0.0025056221056729555, + 0.0025674807839095592, + 0.0026574074290692806, + 0.0026810714043676853, + 0.0038979172240942717, + 0.003223962150514126, + 0.004519718699157238, + 0.004713186528533697, + 0.005042121745646, + 0.004758045077323914, + 0.005829319357872009, + 0.0038580026011914015, + 0.003995215520262718, + 0.005775162950158119, + 0.0046246969141066074, + 0.0016136056510731578, + 0.001054585911333561, + 0.0032235472463071346, + 0.0013112945016473532, + 0.0009477382409386337, + 0.0030794632621109486, + 0.0021103518083691597, + 0.002584123518317938, + 0.007082436699420214, + 0.005791376810520887, + 0.014748325571417809, + 0.010140985250473022, + 0.006087639834731817, + 0.012921548448503017, + 0.007493218407034874, + 0.00880071334540844, + 0.007830905728042126, + 0.0131745720282197, + 0.009360413067042828, + 0.00986957736313343, + 0.00827283039689064, + 0.007962527684867382, + 0.008319499902427197, + 0.007945551536977291 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 3, + 8 + ], + "selected_reward": -0.6933536529541016 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 following the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject a Siberian tiger with orange, white, and black fur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the tiger running at full speed across snow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the tiger's movement cause deep snow displacement creating a shimmering cloud of snow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the tiger's breath visible as white mist in the cold 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": [ + -0.011079221963882446 + ], + "opposite_logprob_tokens": [ + -4.51107931137085 + ], + "logprob_reward": -0.011079221963882446, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011079221963882446, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07790526002645493, + 0.43934738636016846, + 0.04517590254545212, + 0.06183050945401192, + 0.04829053953289986, + 0.09370766580104828, + 0.09705810248851776, + 0.025794442743062973, + 0.044859807938337326, + 0.06603038311004639 + ], + "credit_map": [ + 0.012581649236381054, + 0.0065949950367212296, + 0.010916749015450478, + 0.011239447630941868, + 0.003982424736022949, + 0.006005511619150639, + 0.008282801136374474, + 0.018301676958799362, + 0.08421352505683899, + 0.040957313030958176, + 0.0653616264462471, + 0.06509991735219955, + 0.02849995717406273, + 0.04145929962396622, + 0.040916960686445236, + 0.07283879816532135, + 0.007258248049765825, + 0.004571130033582449, + 0.006496003828942776, + 0.008195689879357815, + 0.004673787392675877, + 0.0033717036712914705, + 0.005521951708942652, + 0.005087386816740036, + 0.009415376000106335, + 0.005936009343713522, + 0.007922952063381672, + 0.008614959195256233, + 0.005417743697762489, + 0.006172713823616505, + 0.010416083969175816, + 0.007934670895338058, + 0.005011765751987696, + 0.006214294116944075, + 0.007868429645895958, + 0.0065821451134979725, + 0.006909176241606474, + 0.004926502704620361, + 0.0052328454330563545, + 0.005545379593968391, + 0.008176887407898903, + 0.011525214649736881, + 0.012421413324773312, + 0.008397637866437435, + 0.010707497596740723, + 0.014521759934723377, + 0.011704320088028908, + 0.016252929344773293, + 0.00980786420404911, + 0.007350241299718618, + 0.01716313511133194, + 0.016359325498342514, + 0.007138589397072792, + 0.011990318074822426, + 0.016342123970389366, + 0.010906504467129707, + 0.0021704761311411858, + 0.002044531749561429, + 0.006853380240499973, + 0.002259538508951664, + 0.0015307553112506866, + 0.005182289052754641, + 0.0027124048210680485, + 0.0030410655308514833, + 0.0043828655034303665, + 0.004189261235296726, + 0.007810458540916443, + 0.006261398550122976, + 0.005170484539121389, + 0.007382520940154791, + 0.005859137512743473, + 0.003803683677688241, + 0.005493086762726307, + 0.006260709371417761, + 0.009585471823811531, + 0.00724074337631464, + 0.00490350229665637, + 0.009923413395881653, + 0.007813445292413235, + 0.01481001079082489 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 6, + 5 + ], + "selected_reward": -0.011079221963882446 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 following the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject a Siberian tiger with orange, white, and black fur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the tiger running at full speed across snow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the tiger's movement cause deep snow displacement creating a shimmering cloud of snow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the tiger's breath visible as white mist in the cold 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": [ + -0.14278970658779144 + ], + "opposite_logprob_tokens": [ + -2.017789602279663 + ], + "logprob_reward": -0.14278970658779144, + "yes_no_margin_reward": 1.8749998807907104, + "token_credit_reward": -0.14278970658779144, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1321273297071457, + 0.1084674671292305, + 0.05191018432378769, + 0.07498714327812195, + 0.07349063456058502, + 0.16352243721485138, + 0.06695161014795303, + 0.030109846964478493, + 0.07614351063966751, + 0.22228984534740448 + ], + "credit_map": [ + 0.02158464677631855, + 0.013887882232666016, + 0.018107501789927483, + 0.01716967113316059, + 0.010016309097409248, + 0.008950608782470226, + 0.02042045071721077, + 0.02199026197195053, + 0.023111050948500633, + 0.008376475423574448, + 0.01696113310754299, + 0.014029419049620628, + 0.007520291954278946, + 0.009401218965649605, + 0.008155735209584236, + 0.02091214247047901, + 0.005700751673430204, + 0.007156779523938894, + 0.006274963263422251, + 0.008466177619993687, + 0.006578127853572369, + 0.0052795568481087685, + 0.007814931683242321, + 0.004638894461095333, + 0.014144452288746834, + 0.006727767642587423, + 0.012087304145097733, + 0.008462216705083847, + 0.005440377630293369, + 0.007189503405243158, + 0.007827994413673878, + 0.013107520528137684, + 0.00815312098711729, + 0.009681005030870438, + 0.011533689685165882, + 0.009430681355297565, + 0.011392109096050262, + 0.007432818878442049, + 0.0069854119792580605, + 0.008881795220077038, + 0.014371993951499462, + 0.018092315644025803, + 0.023457840085029602, + 0.01495441421866417, + 0.016941694542765617, + 0.025422144681215286, + 0.02004205249249935, + 0.03023998998105526, + 0.007196048274636269, + 0.0074354978278279305, + 0.009948693215847015, + 0.011106555350124836, + 0.006043675821274519, + 0.00871637836098671, + 0.009266650304198265, + 0.0072381081990897655, + 0.004770493134856224, + 0.0029823938384652138, + 0.007106855511665344, + 0.002388554159551859, + 0.002046434674412012, + 0.006100799422711134, + 0.002393125556409359, + 0.0023211913648992777, + 0.0077421157620847225, + 0.007300600875169039, + 0.008806312456727028, + 0.010155679658055305, + 0.008351009339094162, + 0.011157297529280186, + 0.010242568328976631, + 0.012387929484248161, + 0.021946322172880173, + 0.028255688026547432, + 0.027782827615737915, + 0.024732030928134918, + 0.019328545778989792, + 0.03087209165096283, + 0.019870709627866745, + 0.049501631408929825 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 5, + 0 + ], + "selected_reward": -0.14278970658779144 + } + ], + "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/0000026/group_13/index_tensor([3139, 3139])_group_13_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000026/group_13/index_tensor([3139, 3139])_group_13_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..aa2f4a7e2e255ae4f99291cdc2e89df5e614e9d9 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000026/group_13/index_tensor([3139, 3139])_group_13_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "static, eye-level medium shot captures the scene in photoreal detail, focusing on a stretch of worn, cracked asphalt in a narrow city alley. A small, vibrant yellow rubber ball, its surface displaying a subtle matte sheen and faint signs of wear, plummets into the frame, pulled by gravity. As it quickly makes contact with the ground, its elastic surface visibly and dramatically deforms, instantly compressing and flattening against the hard pavement in a surreal display of physics. This stored energy is instantly released, propelling the ball back into the air with a powerful primary rebound, its spherical shape perfectly restored during flight. Pulled down again, it undergoes a much smaller, faster secondary rebound, losing its remaining momentum before slowly rolling a few inches before coming to a perfectly stable rest in a slight crack in the asphalt. The warm, golden sunlight of the late afternoon streams in from the side, casting long, dynamic shadows that mimic the ball's every movement and contrast against the cool gray tones of the ground, creating an atmosphere of quiet, focused observation and highlighting the pure mechanics of movement and elasticity.", + "seed": 4506, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000026/group_13/index_3139_seed_4506_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": -0.08983949495013803, + "vlm_logprob": -0.08983949495013803, + "vlm_yes_no_margin": 4.075, + "vlm_token_credit": -0.08983949495013803, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot static and at eye-level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set on a stretch of worn, cracked asphalt in a narrow city alley?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a small, vibrant yellow rubber ball with a matte sheen visible in the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the ball undergo a primary rebound where it compresses upon impact and then restores its spherical shape in flight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting characterized by warm, golden late afternoon sunlight casting long, dynamic 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": [ + -0.39033612608909607 + ], + "opposite_logprob_tokens": [ + -1.3903361558914185 + ], + "logprob_reward": -0.39033612608909607, + "yes_no_margin_reward": 1.0, + "token_credit_reward": -0.39033612608909607, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3306710422039032, + 0.1162622794508934, + 0.18866343796253204, + 0.08567459136247635, + 0.04954442009329796, + 0.07129132747650146, + 0.06771060824394226, + 0.035470765084028244, + 0.023875020444393158, + 0.030836481600999832 + ], + "credit_map": [ + 0.019205085933208466, + 0.02297404035925865, + 0.024641506373882294, + 0.04233590140938759, + 0.019048616290092468, + 0.019141461700201035, + 0.09797486662864685, + 0.08534958958625793, + 0.008203285746276379, + 0.009195360355079174, + 0.025897646322846413, + 0.015306524932384491, + 0.010148126631975174, + 0.013090336695313454, + 0.011666661128401756, + 0.022754337638616562, + 0.011387494392693043, + 0.010672145523130894, + 0.05941123515367508, + 0.014953922480344772, + 0.0119862612336874, + 0.014104594476521015, + 0.013474512845277786, + 0.05267327278852463, + 0.01028781570494175, + 0.009051122702658176, + 0.017195172607898712, + 0.011502531357109547, + 0.011137127876281738, + 0.008516523987054825, + 0.009086932055652142, + 0.008897361345589161, + 0.0052574630826711655, + 0.005295029375702143, + 0.009412615559995174, + 0.00972739513963461, + 0.004249403718858957, + 0.004186328500509262, + 0.00604590866714716, + 0.005370276048779488, + 0.0061136833392083645, + 0.005831246264278889, + 0.023376716300845146, + 0.017210766673088074, + 0.00445444043725729, + 0.0070413523353636265, + 0.004112785216420889, + 0.003150335978716612, + 0.0046332841739058495, + 0.0032185260206460953, + 0.0251809973269701, + 0.022256480529904366, + 0.0029498275835067034, + 0.00308754900470376, + 0.0036404915153980255, + 0.002743450691923499, + 0.003042126540094614, + 0.0025934011209756136, + 0.009416808374226093, + 0.008278689347207546, + 0.002244418952614069, + 0.0029578176327049732, + 0.0034316701348870993, + 0.0035058322828263044, + 0.0024447368923574686, + 0.0028047605883330107, + 0.003575799288228154, + 0.002760214265435934, + 0.0023355772718787193, + 0.0024837672244757414, + 0.004898770246654749, + 0.00257139652967453, + 0.0032341368496418, + 0.0031371901277452707, + 0.009578394703567028, + 0.004683947656303644, + 0.002098122611641884, + 0.0028715210501104593, + 0.0018465350149199367, + 0.0033866327721625566 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 1 + ], + "selected_reward": -0.39033612608909607 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot static and at eye-level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set on a stretch of worn, cracked asphalt in a narrow city alley?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a small, vibrant yellow rubber ball with a matte sheen visible in the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the ball undergo a primary rebound where it compresses upon impact and then restores its spherical shape in flight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting characterized by warm, golden late afternoon sunlight casting long, dynamic 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": [ + -0.02982828952372074 + ], + "opposite_logprob_tokens": [ + -3.5298283100128174 + ], + "logprob_reward": -0.02982828952372074, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.02982828952372074, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20351259410381317, + 0.05915975943207741, + 0.1015322282910347, + 0.16988295316696167, + 0.1095215305685997, + 0.19206510484218597, + 0.09080637246370316, + 0.032583508640527725, + 0.026436442509293556, + 0.014499536715447903 + ], + "credit_map": [ + 0.014993855729699135, + 0.035016972571611404, + 0.027492856606841087, + 0.04335326701402664, + 0.01938241347670555, + 0.014260413125157356, + 0.02291199006140232, + 0.02610081247985363, + 0.004987919237464666, + 0.004149608314037323, + 0.01032879389822483, + 0.01241112221032381, + 0.0045189810916781425, + 0.0066956342197954655, + 0.007744153495877981, + 0.008323552086949348, + 0.00808748323470354, + 0.010232553817331791, + 0.01845814287662506, + 0.018008556216955185, + 0.010770617052912712, + 0.011541733518242836, + 0.012398279272019863, + 0.01203486043959856, + 0.015614984557032585, + 0.013835834339261055, + 0.04957910254597664, + 0.03226843848824501, + 0.010962352156639099, + 0.015781408175826073, + 0.015580426901578903, + 0.01626039855182171, + 0.015117606148123741, + 0.013584037311375141, + 0.019308116286993027, + 0.019964883103966713, + 0.010468393564224243, + 0.01191607303917408, + 0.009861364029347897, + 0.00930105522274971, + 0.018100537359714508, + 0.017943566665053368, + 0.04878334328532219, + 0.035624705255031586, + 0.015309895388782024, + 0.024304673075675964, + 0.018315481022000313, + 0.013682902790606022, + 0.008943813852965832, + 0.00973834004253149, + 0.016559328883886337, + 0.017644276842474937, + 0.007203832268714905, + 0.01092943362891674, + 0.01181348878890276, + 0.007973853498697281, + 0.00303073157556355, + 0.003138680709525943, + 0.006015995051711798, + 0.005918726325035095, + 0.003098118584603071, + 0.003928929567337036, + 0.004373712930828333, + 0.0030786122661083937, + 0.001861105440184474, + 0.0027131056413054466, + 0.006083483342081308, + 0.003734520636498928, + 0.002497142180800438, + 0.0032594839576631784, + 0.003475270001217723, + 0.0028123310767114162, + 0.0012599860783666372, + 0.0014085517032071948, + 0.002615987556055188, + 0.002359963720664382, + 0.0014992940705269575, + 0.00209924578666687, + 0.001613053260371089, + 0.0016434548888355494 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 3 + ], + "selected_reward": -0.02982828952372074 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot static and at eye-level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set on a stretch of worn, cracked asphalt in a narrow city alley?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a small, vibrant yellow rubber ball with a matte sheen visible in the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the ball undergo a primary rebound where it compresses upon impact and then restores its spherical shape in flight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting characterized by warm, golden late afternoon sunlight casting long, dynamic 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": [ + -0.0008279233006760478 + ], + "opposite_logprob_tokens": [ + -7.125827789306641 + ], + "logprob_reward": -0.0008279233006760478, + "yes_no_margin_reward": 7.125, + "token_credit_reward": -0.0008279233006760478, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5039351582527161, + 0.11723046749830246, + 0.0950031578540802, + 0.07406416535377502, + 0.02691054157912731, + 0.024085313081741333, + 0.05045574903488159, + 0.049437861889600754, + 0.03473629802465439, + 0.02414127066731453 + ], + "credit_map": [ + 0.02862536534667015, + 0.04600914567708969, + 0.04615669697523117, + 0.05894852057099342, + 0.023651251569390297, + 0.021415797993540764, + 0.17901511490345, + 0.10011328756809235, + 0.0081383241340518, + 0.0074547347612679005, + 0.019490791484713554, + 0.01760694943368435, + 0.006880358327180147, + 0.011619241908192635, + 0.0244150273501873, + 0.02162504568696022, + 0.008590798825025558, + 0.005836795549839735, + 0.015642590820789337, + 0.010293158702552319, + 0.009496795013546944, + 0.008367282338440418, + 0.024099329486489296, + 0.01267640944570303, + 0.0063351821154356, + 0.008015302009880543, + 0.01764252409338951, + 0.010555591434240341, + 0.007644715718924999, + 0.006428765133023262, + 0.008685383945703506, + 0.008756696246564388, + 0.0027211466804146767, + 0.0027865914162248373, + 0.003958378452807665, + 0.004348850809037685, + 0.0026855713222175837, + 0.002539912471547723, + 0.003769446862861514, + 0.004100644960999489, + 0.0024432726204395294, + 0.0032737755682319403, + 0.0030651201959699392, + 0.003893618006259203, + 0.0026622777804732323, + 0.002609824063256383, + 0.0033930912613868713, + 0.002744333352893591, + 0.003079512855038047, + 0.0024975030682981014, + 0.01807989552617073, + 0.01542629487812519, + 0.0027313686441630125, + 0.002909792587161064, + 0.003226746106520295, + 0.002504633739590645, + 0.003717218991369009, + 0.002939303871244192, + 0.015527825802564621, + 0.013767486438155174, + 0.002551781013607979, + 0.0037359935231506824, + 0.0033133712131530046, + 0.0038848817348480225, + 0.003250849200412631, + 0.0037872681859880686, + 0.008148678578436375, + 0.004254419822245836, + 0.003510990645736456, + 0.0036740831565111876, + 0.0047915019094944, + 0.0033185076899826527, + 0.002197712892666459, + 0.0020022443495690823, + 0.006252932362258434, + 0.0036618486046791077, + 0.0018048804486170411, + 0.0022840118035674095, + 0.003045802004635334, + 0.0028918383177369833 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.0008279233006760478 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot static and at eye-level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set on a stretch of worn, cracked asphalt in a narrow city alley?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a small, vibrant yellow rubber ball with a matte sheen visible in the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the ball undergo a primary rebound where it compresses upon impact and then restores its spherical shape in flight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting characterized by warm, golden late afternoon sunlight casting long, dynamic 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": [ + -0.020579854026436806 + ], + "opposite_logprob_tokens": [ + -3.8955798149108887 + ], + "logprob_reward": -0.020579854026436806, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020579854026436806, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4348233640193939, + 0.08301287889480591, + 0.08682014048099518, + 0.14753389358520508, + 0.02240336872637272, + 0.041807401925325394, + 0.040235403925180435, + 0.03378476947546005, + 0.08511926233768463, + 0.02445957064628601 + ], + "credit_map": [ + 0.031261373311281204, + 0.060892123728990555, + 0.03380339592695236, + 0.0760132446885109, + 0.029917018488049507, + 0.02126108482480049, + 0.09359271079301834, + 0.08808238804340363, + 0.005312898196280003, + 0.0051127271726727486, + 0.01392565481364727, + 0.012683301232755184, + 0.006249678786844015, + 0.008373275399208069, + 0.016828007996082306, + 0.014527329243719578, + 0.006759094540029764, + 0.005380806513130665, + 0.022000407800078392, + 0.010310320183634758, + 0.006694088689982891, + 0.005745467729866505, + 0.014367258176207542, + 0.015562687069177628, + 0.012229843996465206, + 0.010568286292254925, + 0.049329373985528946, + 0.02760390006005764, + 0.009003328159451485, + 0.010501139797270298, + 0.013521919026970863, + 0.014776080846786499, + 0.0023548887111246586, + 0.0016759925056248903, + 0.0029262977186590433, + 0.0033452522475272417, + 0.0020016671624034643, + 0.0022794720716774464, + 0.004335516132414341, + 0.003484281012788415, + 0.0030679721385240555, + 0.00377282383851707, + 0.011737354099750519, + 0.010165235958993435, + 0.0030138089787214994, + 0.003401760244742036, + 0.0038216253742575645, + 0.0028268215246498585, + 0.002844056114554405, + 0.002619280945509672, + 0.013793681748211384, + 0.010624094866216183, + 0.0020365866366773844, + 0.0030811966862529516, + 0.0031192435417324305, + 0.00211725989356637, + 0.0022084424272179604, + 0.002309159142896533, + 0.008615479804575443, + 0.007587315049022436, + 0.002524281619116664, + 0.002545709256082773, + 0.003948022145777941, + 0.0040463549084961414, + 0.007626232225447893, + 0.00779271824285388, + 0.018462486565113068, + 0.011563369072973728, + 0.008204937912523746, + 0.008905247785151005, + 0.014452265575528145, + 0.008112000301480293, + 0.0026487596333026886, + 0.0025383862666785717, + 0.006792310159653425, + 0.003425553673878312, + 0.002102355705574155, + 0.002615425968542695, + 0.0012539061717689037, + 0.0030828716699033976 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 2 + ], + "selected_reward": -0.020579854026436806 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot static and at eye-level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set on a stretch of worn, cracked asphalt in a narrow city alley?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a small, vibrant yellow rubber ball with a matte sheen visible in the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the ball undergo a primary rebound where it compresses upon impact and then restores its spherical shape in flight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting characterized by warm, golden late afternoon sunlight casting long, dynamic 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": [ + -0.007625281810760498 + ], + "opposite_logprob_tokens": [ + -4.882625102996826 + ], + "logprob_reward": -0.007625281810760498, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007625281810760498, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.7037805318832397, + 0.040782857686281204, + 0.08073489367961884, + 0.017764855176210403, + 0.01879531890153885, + 0.05871196091175079, + 0.02062593400478363, + 0.019408773630857468, + 0.010177007876336575, + 0.029217848554253578 + ], + "credit_map": [ + 0.04103514179587364, + 0.03126638010144234, + 0.05385022982954979, + 0.034340232610702515, + 0.029114102944731712, + 0.025472285225987434, + 0.28753772377967834, + 0.2011643946170807, + 0.004019171465188265, + 0.004540977068245411, + 0.007255419623106718, + 0.006737387273460627, + 0.003050998318940401, + 0.0032312823459506035, + 0.006148673593997955, + 0.005798947531729937, + 0.006603988818824291, + 0.007601771038025618, + 0.022759225219488144, + 0.009953168220818043, + 0.006439839489758015, + 0.007283350918442011, + 0.00727276923134923, + 0.01282077468931675, + 0.0016295929672196507, + 0.0018473082454875112, + 0.0027691598515957594, + 0.0025433755945414305, + 0.0020520875696092844, + 0.0018633059225976467, + 0.002832087455317378, + 0.0022279368713498116, + 0.0017071082256734371, + 0.0015640335623174906, + 0.0037304058205336332, + 0.003543283324688673, + 0.0014807329280301929, + 0.0019113338785246015, + 0.001646732329390943, + 0.003211689880117774, + 0.007304163184016943, + 0.008507976308465004, + 0.00892853643745184, + 0.010610874742269516, + 0.007123372983187437, + 0.005365782883018255, + 0.006341255735605955, + 0.00452999584376812, + 0.0022689627949148417, + 0.0030909166671335697, + 0.003920004703104496, + 0.002094660187140107, + 0.0021866022143512964, + 0.0030616684816777706, + 0.002228440484032035, + 0.001774679170921445, + 0.0021395741496235132, + 0.0017312834970653057, + 0.004576609004288912, + 0.0044783917255699635, + 0.0013372135581448674, + 0.0016175953205674887, + 0.0019317104015499353, + 0.0015963953919708729, + 0.001026476384140551, + 0.0008372209267690778, + 0.002089237794280052, + 0.0015528332442045212, + 0.0008705805521458387, + 0.0014774724841117859, + 0.0011679923627525568, + 0.0011551943607628345, + 0.0032237365376204252, + 0.0032188682816922665, + 0.007528637070208788, + 0.004809523932635784, + 0.002330104121938348, + 0.0030618919990956783, + 0.0015274639008566737, + 0.0035176221281290054 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 5 + ], + "selected_reward": -0.007625281810760498 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, eye-level medium shot captures the scene in photoreal detail, focusing on a stretch of worn, cracked asphalt in a narrow city alley. A small, vibrant yellow rubber ball, its surface displaying a subtle matte sheen and faint signs of wear, plummets into the frame, pulled by gravity. As it quickly makes contact with the ground, its elastic surface visibly and dramatically deforms, instantly compressing and flattening against the hard pavement in a surreal display of physics. This stored energy is instantly released, propelling the ball back into the air with a powerful primary rebound, its spherical shape perfectly restored during flight. Pulled down again, it undergoes a much smaller, faster secondary rebound, losing its remaining momentum before slowly rolling a few inches before coming to a perfectly stable rest in a slight crack in the asphalt. The warm, golden sunlight of the late afternoon streams in from the side, casting long, dynamic shadows that mimic the ball's every movement and contrast against the cool gray tones of the ground, creating an atmosphere of quiet, focused observation and highlighting the pure mechanics of movement and elasticity.", + "seed": 4507, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000026/group_13/index_3139_seed_4507_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": -0.07929263971745967, + "vlm_logprob": -0.07929263971745967, + "vlm_yes_no_margin": 4.025, + "vlm_token_credit": -0.07929263971745967, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot static and at eye-level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set on a stretch of worn, cracked asphalt in a narrow city alley?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a small, vibrant yellow rubber ball with a matte sheen visible in the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the ball undergo a primary rebound where it compresses upon impact and then restores its spherical shape in flight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting characterized by warm, golden late afternoon sunlight casting long, dynamic 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": [ + -0.3203485310077667 + ], + "opposite_logprob_tokens": [ + -1.5703485012054443 + ], + "logprob_reward": -0.3203485310077667, + "yes_no_margin_reward": 1.25, + "token_credit_reward": -0.3203485310077667, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18415142595767975, + 0.20995596051216125, + 0.12528710067272186, + 0.10717067867517471, + 0.08452864736318588, + 0.1547706127166748, + 0.05857526510953903, + 0.0275576189160347, + 0.028302066028118134, + 0.01970062032341957 + ], + "credit_map": [ + 0.015945512801408768, + 0.020374715328216553, + 0.034038566052913666, + 0.028692372143268585, + 0.02064305916428566, + 0.0156552791595459, + 0.023644836619496346, + 0.025157075375318527, + 0.007960453629493713, + 0.022814663127064705, + 0.03595219552516937, + 0.04746483266353607, + 0.015625212341547012, + 0.009156791493296623, + 0.03075088933110237, + 0.04023092985153198, + 0.011764797382056713, + 0.01039643120020628, + 0.022310571745038033, + 0.020659593865275383, + 0.01043017115443945, + 0.010748876258730888, + 0.016643112525343895, + 0.022333545610308647, + 0.010465854778885841, + 0.01756858453154564, + 0.016202129423618317, + 0.011698656715452671, + 0.017386266961693764, + 0.008642740547657013, + 0.014218092896044254, + 0.010988353751599789, + 0.011585313826799393, + 0.00862500537186861, + 0.011954059824347496, + 0.01024385541677475, + 0.009927980601787567, + 0.006680791266262531, + 0.015112760476768017, + 0.010398882441222668, + 0.013983821496367455, + 0.018141984939575195, + 0.02523156814277172, + 0.03192249312996864, + 0.013437696732580662, + 0.011807763017714024, + 0.02435694821178913, + 0.01588832587003708, + 0.006411276757717133, + 0.007457912899553776, + 0.009593119844794273, + 0.009620183147490025, + 0.006048993207514286, + 0.005627988837659359, + 0.007418635301291943, + 0.006397154647856951, + 0.002670183777809143, + 0.0023443938698619604, + 0.003937056288123131, + 0.003377652494236827, + 0.003053313121199608, + 0.001972063910216093, + 0.006524078547954559, + 0.003678877605125308, + 0.00270170159637928, + 0.0027353668119758368, + 0.004042913671582937, + 0.003272940870374441, + 0.0032983587589114904, + 0.0018583773635327816, + 0.0070291138254106045, + 0.003363292431458831, + 0.0018341324757784605, + 0.0020232684910297394, + 0.002904455177485943, + 0.002462890697643161, + 0.0027766721323132515, + 0.001644941046833992, + 0.0034549706615507603, + 0.0025992891751229763 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 5 + ], + "selected_reward": -0.3203485310077667 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot static and at eye-level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set on a stretch of worn, cracked asphalt in a narrow city alley?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a small, vibrant yellow rubber ball with a matte sheen visible in the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the ball undergo a primary rebound where it compresses upon impact and then restores its spherical shape in flight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting characterized by warm, golden late afternoon sunlight casting long, dynamic 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": [ + -0.038126904517412186 + ], + "opposite_logprob_tokens": [ + -3.2881269454956055 + ], + "logprob_reward": -0.038126904517412186, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.038126904517412186, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.46811962127685547, + 0.06979764997959137, + 0.14932836592197418, + 0.04635344445705414, + 0.05865087732672691, + 0.06879685819149017, + 0.041246261447668076, + 0.02439110539853573, + 0.04111015424132347, + 0.03220566362142563 + ], + "credit_map": [ + 0.033163029700517654, + 0.04190603271126747, + 0.08500777184963226, + 0.08310693502426147, + 0.039781879633665085, + 0.05305003747344017, + 0.08006885647773743, + 0.05203508958220482, + 0.0041172197088599205, + 0.007664690725505352, + 0.011374363675713539, + 0.01306252833455801, + 0.006389551796019077, + 0.004210745450109243, + 0.010453163646161556, + 0.012525387108325958, + 0.03200451657176018, + 0.018933774903416634, + 0.021020330488681793, + 0.015665311366319656, + 0.011452628299593925, + 0.015731127932667732, + 0.01671396568417549, + 0.017806721851229668, + 0.004319505300372839, + 0.006146441679447889, + 0.0064673153683543205, + 0.006008058786392212, + 0.004936614073812962, + 0.0030216951854526997, + 0.00750168738886714, + 0.007952125743031502, + 0.005844444036483765, + 0.006590287666767836, + 0.008958377875387669, + 0.008764891885221004, + 0.006688248831778765, + 0.005488031543791294, + 0.0071413107216358185, + 0.009175281971693039, + 0.006978044752031565, + 0.007149485871195793, + 0.011387111619114876, + 0.010013719089329243, + 0.00608133664354682, + 0.007884225808084011, + 0.012390337884426117, + 0.006912599317729473, + 0.006726536434143782, + 0.005990141071379185, + 0.006827134173363447, + 0.005146119277924299, + 0.004654689226299524, + 0.0033235985320061445, + 0.004420887213200331, + 0.004157152026891708, + 0.004720147233456373, + 0.0029689846560359, + 0.0033857659436762333, + 0.002080008154734969, + 0.0020268275402486324, + 0.0027158476877957582, + 0.003312091575935483, + 0.0031814337708055973, + 0.008493316359817982, + 0.005081367678940296, + 0.005934861954301596, + 0.0031312766950577497, + 0.0033925871830433607, + 0.005369473714381456, + 0.00447123683989048, + 0.005236030090600252, + 0.0028522086795419455, + 0.003500108141452074, + 0.005218188744038343, + 0.003088046330958605, + 0.005396708380430937, + 0.002969921100884676, + 0.004353165626525879, + 0.004827320110052824 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 1 + ], + "selected_reward": -0.038126904517412186 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot static and at eye-level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set on a stretch of worn, cracked asphalt in a narrow city alley?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a small, vibrant yellow rubber ball with a matte sheen visible in the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the ball undergo a primary rebound where it compresses upon impact and then restores its spherical shape in flight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting characterized by warm, golden late afternoon sunlight casting long, dynamic 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": [ + -0.0005757343024015427 + ], + "opposite_logprob_tokens": [ + -7.500575542449951 + ], + "logprob_reward": -0.0005757343024015427, + "yes_no_margin_reward": 7.5, + "token_credit_reward": -0.0005757343024015427, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2710053324699402, + 0.10664483159780502, + 0.14002390205860138, + 0.054119061678647995, + 0.129575714468956, + 0.07589239627122879, + 0.08627311140298843, + 0.0476376935839653, + 0.056875698268413544, + 0.03195231407880783 + ], + "credit_map": [ + 0.02267342247068882, + 0.0238787904381752, + 0.051930587738752365, + 0.049260884523391724, + 0.027436312288045883, + 0.03052511252462864, + 0.04248443618416786, + 0.02281578816473484, + 0.005740573164075613, + 0.009309052489697933, + 0.01766548678278923, + 0.026787007227540016, + 0.009730968624353409, + 0.004207308404147625, + 0.01942742057144642, + 0.013777009211480618, + 0.010445142164826393, + 0.010460428893566132, + 0.023305581882596016, + 0.024675998836755753, + 0.012850039638578892, + 0.008355668745934963, + 0.025998463854193687, + 0.023932570591568947, + 0.0036498436238616705, + 0.003851613961160183, + 0.00822390802204609, + 0.00724569009616971, + 0.006415741518139839, + 0.003803979605436325, + 0.012987461872398853, + 0.007940821349620819, + 0.008076717145740986, + 0.009022933430969715, + 0.02261205203831196, + 0.027888894081115723, + 0.011950029991567135, + 0.009644926525652409, + 0.013499130494892597, + 0.026881029829382896, + 0.006897271610796452, + 0.007888547144830227, + 0.014117305167019367, + 0.01036960817873478, + 0.0075195832177996635, + 0.005408328026533127, + 0.016093192622065544, + 0.007598565425723791, + 0.010134205222129822, + 0.007700151298195124, + 0.011275364086031914, + 0.013176839798688889, + 0.013424389064311981, + 0.005825861822813749, + 0.013648584485054016, + 0.011087712831795216, + 0.004210271406918764, + 0.003816040465608239, + 0.006988207809627056, + 0.00493326410651207, + 0.007118151523172855, + 0.0038250968791544437, + 0.01112030353397131, + 0.00562635762616992, + 0.0035935568157583475, + 0.004966569133102894, + 0.009604428894817829, + 0.007317231502383947, + 0.009475678205490112, + 0.003985458519309759, + 0.012257938273251057, + 0.005674836225807667, + 0.0057846843264997005, + 0.0031180507503449917, + 0.004557940177619457, + 0.0031434218399226665, + 0.0038598813116550446, + 0.00320571125485003, + 0.004963735584169626, + 0.0033188883680850267 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 4 + ], + "selected_reward": -0.0005757343024015427 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot static and at eye-level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set on a stretch of worn, cracked asphalt in a narrow city alley?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a small, vibrant yellow rubber ball with a matte sheen visible in the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the ball undergo a primary rebound where it compresses upon impact and then restores its spherical shape in flight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting characterized by warm, golden late afternoon sunlight casting long, dynamic 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": [ + -0.01107686385512352 + ], + "opposite_logprob_tokens": [ + -4.511076927185059 + ], + "logprob_reward": -0.01107686385512352, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.01107686385512352, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1373782604932785, + 0.44223055243492126, + 0.12697646021842957, + 0.06848472356796265, + 0.06261268258094788, + 0.033783815801143646, + 0.02387857250869274, + 0.05830716714262962, + 0.02057386003434658, + 0.02577391266822815 + ], + "credit_map": [ + 0.011934618465602398, + 0.013317822478711605, + 0.030913209542632103, + 0.026003718376159668, + 0.011408839374780655, + 0.00953409168869257, + 0.018379703164100647, + 0.015886249020695686, + 0.015707004815340042, + 0.04216422885656357, + 0.0819990336894989, + 0.09203194081783295, + 0.026911981403827667, + 0.01922612637281418, + 0.07211719453334808, + 0.09207300841808319, + 0.010791962035000324, + 0.011561769060790539, + 0.022392578423023224, + 0.021264677867293358, + 0.01223030500113964, + 0.00663259020075202, + 0.019834119826555252, + 0.022268448024988174, + 0.005291915964335203, + 0.006506061181426048, + 0.009438934735953808, + 0.005852786358445883, + 0.010545208118855953, + 0.00810479000210762, + 0.01619730144739151, + 0.006547723431140184, + 0.004422253929078579, + 0.005318487528711557, + 0.010941301472485065, + 0.009379281662404537, + 0.006548506673425436, + 0.0037048151716589928, + 0.01066561695188284, + 0.011632420122623444, + 0.0028598769567906857, + 0.004107942339032888, + 0.008012025617063046, + 0.0065649645403027534, + 0.0037058705929666758, + 0.0024487634655088186, + 0.003103654831647873, + 0.002980718854814768, + 0.001991793280467391, + 0.00224918476305902, + 0.003353768726810813, + 0.0033890472259372473, + 0.0035432735458016396, + 0.0017837767954915762, + 0.004752534441649914, + 0.0028151923324912786, + 0.004084827844053507, + 0.004637933801859617, + 0.008358743973076344, + 0.006115256808698177, + 0.009458613581955433, + 0.004558417946100235, + 0.01493387296795845, + 0.006159497890621424, + 0.002714757574722171, + 0.00175530556589365, + 0.0028903144411742687, + 0.0026010172441601753, + 0.00222716573625803, + 0.0023487922735512257, + 0.0033603638876229525, + 0.002676143543794751, + 0.0028122246731072664, + 0.002413863316178322, + 0.003923922311514616, + 0.002830501412972808, + 0.0038680313155055046, + 0.0024112872779369354, + 0.004488280043005943, + 0.003025803482159972 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 2 + ], + "selected_reward": -0.01107686385512352 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot static and at eye-level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set on a stretch of worn, cracked asphalt in a narrow city alley?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a small, vibrant yellow rubber ball with a matte sheen visible in the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the ball undergo a primary rebound where it compresses upon impact and then restores its spherical shape in flight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting characterized by warm, golden late afternoon sunlight casting long, dynamic 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": [ + -0.02633516490459442 + ], + "opposite_logprob_tokens": [ + -3.6513352394104004 + ], + "logprob_reward": -0.02633516490459442, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.02633516490459442, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.38907623291015625, + 0.1361774206161499, + 0.10597432404756546, + 0.08182074874639511, + 0.04605817794799805, + 0.12332215905189514, + 0.015576640143990517, + 0.07037138193845749, + 0.018148187547922134, + 0.013474729843437672 + ], + "credit_map": [ + 0.031548742204904556, + 0.046525802463293076, + 0.07300686836242676, + 0.06287287920713425, + 0.05551985278725624, + 0.02854197286069393, + 0.05435189977288246, + 0.036708202213048935, + 0.005076956935226917, + 0.01455668918788433, + 0.024086235091090202, + 0.025147533044219017, + 0.009226162917912006, + 0.006327284500002861, + 0.01994992420077324, + 0.031806644052267075, + 0.017060771584510803, + 0.01062304899096489, + 0.015677640214562416, + 0.01190731581300497, + 0.012006472796201706, + 0.009300791658461094, + 0.014842427335679531, + 0.014555854722857475, + 0.007260539568960667, + 0.0078081353567540646, + 0.011655446141958237, + 0.008736085146665573, + 0.012724684551358223, + 0.01010118331760168, + 0.014872154220938683, + 0.00866252277046442, + 0.007567902561277151, + 0.0060190181247889996, + 0.00647982582449913, + 0.005184714216738939, + 0.0053454735316336155, + 0.005420893896371126, + 0.0050382427871227264, + 0.005002105608582497, + 0.01302374992519617, + 0.019093893468379974, + 0.02093878760933876, + 0.020143885165452957, + 0.012885689735412598, + 0.012981191277503967, + 0.01286978181451559, + 0.011385179124772549, + 0.001143651781603694, + 0.002364266663789749, + 0.002367942361161113, + 0.0018992375116795301, + 0.0016758241690695286, + 0.001531838672235608, + 0.002648816676810384, + 0.001945062424056232, + 0.011098855175077915, + 0.0077807060442864895, + 0.00962319690734148, + 0.0064688087441027164, + 0.010042499750852585, + 0.005707590375095606, + 0.01020131353288889, + 0.00944841094315052, + 0.0023593390360474586, + 0.0021815309301018715, + 0.0028690826147794724, + 0.0019368174253031611, + 0.002229386940598488, + 0.001873972243629396, + 0.0026188010815531015, + 0.0020792586728930473, + 0.002121672034263611, + 0.0017297561280429363, + 0.0018460254650563002, + 0.0013670725747942924, + 0.0019464108627289534, + 0.0012919261353090405, + 0.001806094078347087, + 0.0013657723320648074 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 5 + ], + "selected_reward": -0.02633516490459442 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, eye-level medium shot captures the scene in photoreal detail, focusing on a stretch of worn, cracked asphalt in a narrow city alley. A small, vibrant yellow rubber ball, its surface displaying a subtle matte sheen and faint signs of wear, plummets into the frame, pulled by gravity. As it quickly makes contact with the ground, its elastic surface visibly and dramatically deforms, instantly compressing and flattening against the hard pavement in a surreal display of physics. This stored energy is instantly released, propelling the ball back into the air with a powerful primary rebound, its spherical shape perfectly restored during flight. Pulled down again, it undergoes a much smaller, faster secondary rebound, losing its remaining momentum before slowly rolling a few inches before coming to a perfectly stable rest in a slight crack in the asphalt. The warm, golden sunlight of the late afternoon streams in from the side, casting long, dynamic shadows that mimic the ball's every movement and contrast against the cool gray tones of the ground, creating an atmosphere of quiet, focused observation and highlighting the pure mechanics of movement and elasticity.", + "seed": 4508, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000026/group_13/index_3139_seed_4508_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": -0.09430917947320268, + "vlm_logprob": -0.09430917947320268, + "vlm_yes_no_margin": 4.25, + "vlm_token_credit": -0.09430917947320268, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot static and at eye-level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set on a stretch of worn, cracked asphalt in a narrow city alley?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a small, vibrant yellow rubber ball with a matte sheen visible in the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the ball undergo a primary rebound where it compresses upon impact and then restores its spherical shape in flight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting characterized by warm, golden late afternoon sunlight casting long, dynamic 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": [ + -0.4323993921279907 + ], + "opposite_logprob_tokens": [ + -1.3073993921279907 + ], + "logprob_reward": -0.4323993921279907, + "yes_no_margin_reward": 0.875, + "token_credit_reward": -0.4323993921279907, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1699720174074173, + 0.10294973850250244, + 0.03678661957383156, + 0.07862240076065063, + 0.07606006413698196, + 0.1081676110625267, + 0.13610629737377167, + 0.12696413695812225, + 0.12110809981822968, + 0.043263014405965805 + ], + "credit_map": [ + 0.01562603749334812, + 0.019412243738770485, + 0.019168445840477943, + 0.027054723352193832, + 0.02054869756102562, + 0.01962948963046074, + 0.031493786722421646, + 0.017038583755493164, + 0.009611990302801132, + 0.010135989636182785, + 0.012120364233851433, + 0.016627756878733635, + 0.008473822847008705, + 0.010458612814545631, + 0.013088501989841461, + 0.02243269421160221, + 0.0032160531263798475, + 0.0027124700136482716, + 0.0050871288403868675, + 0.006482318509370089, + 0.0031564845703542233, + 0.0031048760283738375, + 0.0064741033129394054, + 0.00655318470671773, + 0.007574424147605896, + 0.009881447069346905, + 0.008992962539196014, + 0.01650441624224186, + 0.007930702529847622, + 0.007696258369833231, + 0.012142070569097996, + 0.00790011789649725, + 0.007321570534259081, + 0.010062034241855145, + 0.009647220373153687, + 0.010189453139901161, + 0.006470872554928064, + 0.007860114797949791, + 0.012524784542620182, + 0.011984015814960003, + 0.008282055146992207, + 0.010833282954990864, + 0.0247836634516716, + 0.019790668040513992, + 0.012381955981254578, + 0.0079370541498065, + 0.011943547055125237, + 0.012215379625558853, + 0.0087087107822299, + 0.011462231166660786, + 0.013909934088587761, + 0.022090310230851173, + 0.025386132299900055, + 0.013096657581627369, + 0.02860269509255886, + 0.01284962147474289, + 0.00972751621156931, + 0.017619740217924118, + 0.013720891438424587, + 0.029812835156917572, + 0.011710921302437782, + 0.010990976355969906, + 0.010889659635722637, + 0.022491594776511192, + 0.01819566637277603, + 0.01157296635210514, + 0.013234050944447517, + 0.015480105765163898, + 0.018609140068292618, + 0.01389245968312025, + 0.013142331503331661, + 0.016981380060315132, + 0.00391111895442009, + 0.003607172518968582, + 0.0037344652228057384, + 0.004623020067811012, + 0.010235454887151718, + 0.004192633554339409, + 0.0070460024289786816, + 0.005913146305829287 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 7 + ], + "selected_reward": -0.4323993921279907 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot static and at eye-level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set on a stretch of worn, cracked asphalt in a narrow city alley?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a small, vibrant yellow rubber ball with a matte sheen visible in the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the ball undergo a primary rebound where it compresses upon impact and then restores its spherical shape in flight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting characterized by warm, golden late afternoon sunlight casting long, dynamic 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": [ + -0.016096169129014015 + ], + "opposite_logprob_tokens": [ + -4.141096115112305 + ], + "logprob_reward": -0.016096169129014015, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016096169129014015, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5499631762504578, + 0.043730635195970535, + 0.020615139976143837, + 0.01547910738736391, + 0.016459057107567787, + 0.030210040509700775, + 0.05981336161494255, + 0.1537470519542694, + 0.044352348893880844, + 0.06563010066747665 + ], + "credit_map": [ + 0.04959029704332352, + 0.05370643734931946, + 0.0809713676571846, + 0.0958748608827591, + 0.06065618619322777, + 0.07193102687597275, + 0.06674445420503616, + 0.07048854976892471, + 0.003732123179361224, + 0.0036998754367232323, + 0.006625186186283827, + 0.007253063376992941, + 0.003816327778622508, + 0.0030867327004671097, + 0.007425799034535885, + 0.008091527037322521, + 0.001621831557713449, + 0.001515422947704792, + 0.0031111727003008127, + 0.0038710369262844324, + 0.0016085697570815682, + 0.0015394778456538916, + 0.003995950799435377, + 0.003351678606122732, + 0.0012867233017459512, + 0.0019138841889798641, + 0.0025420705787837505, + 0.0022603722754865885, + 0.001880575786344707, + 0.0017246035858988762, + 0.0017907782457768917, + 0.0020800998900085688, + 0.0017572431825101376, + 0.0018203096697106957, + 0.002360146725550294, + 0.002715547801926732, + 0.0013718263944610953, + 0.002128514926880598, + 0.002285978989675641, + 0.002019488951191306, + 0.0027700436767190695, + 0.0032917906064540148, + 0.005775961093604565, + 0.006012170109897852, + 0.0022570679429918528, + 0.00325007108040154, + 0.0034212807659059763, + 0.0034316550008952618, + 0.006861844565719366, + 0.007776935119181871, + 0.010807671584188938, + 0.009831356815993786, + 0.005047387909144163, + 0.007385419216006994, + 0.006027924828231335, + 0.006074823904782534, + 0.010730623267591, + 0.009044752456247807, + 0.013979949057102203, + 0.05588832497596741, + 0.004521236754953861, + 0.007742011919617653, + 0.006346283480525017, + 0.04549385607242584, + 0.005437997169792652, + 0.004073385614901781, + 0.0036829249002039433, + 0.0066078160889446735, + 0.00760938273742795, + 0.0046489425003528595, + 0.004463230259716511, + 0.007828669622540474, + 0.011233171448111534, + 0.0066528902389109135, + 0.009716801345348358, + 0.010010194964706898, + 0.004311791155487299, + 0.009693784639239311, + 0.004506717901676893, + 0.009504743851721287 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 9 + ], + "selected_reward": -0.016096169129014015 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot static and at eye-level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set on a stretch of worn, cracked asphalt in a narrow city alley?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a small, vibrant yellow rubber ball with a matte sheen visible in the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the ball undergo a primary rebound where it compresses upon impact and then restores its spherical shape in flight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting characterized by warm, golden late afternoon sunlight casting long, dynamic 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": [ + -0.0007323920144699514 + ], + "opposite_logprob_tokens": [ + -7.250732421875 + ], + "logprob_reward": -0.0007323920144699514, + "yes_no_margin_reward": 7.25, + "token_credit_reward": -0.0007323920144699514, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4397987127304077, + 0.12550869584083557, + 0.10639563202857971, + 0.020971789956092834, + 0.04551943019032478, + 0.0566924586892128, + 0.05118358135223389, + 0.08320679515600204, + 0.028852257877588272, + 0.04187065362930298 + ], + "credit_map": [ + 0.020239662379026413, + 0.027248704805970192, + 0.03892926126718521, + 0.06930096447467804, + 0.08751323819160461, + 0.04455062001943588, + 0.07501375675201416, + 0.07700251787900925, + 0.0077547840774059296, + 0.006678035948425531, + 0.01258005015552044, + 0.022196168079972267, + 0.007468498777598143, + 0.012411107309162617, + 0.017983855679631233, + 0.038436196744441986, + 0.0077768899500370026, + 0.006507720798254013, + 0.01077231578528881, + 0.02437286265194416, + 0.006788940634578466, + 0.010301648639142513, + 0.020834581926465034, + 0.019040677696466446, + 0.0012065756600350142, + 0.001344577525742352, + 0.002847525291144848, + 0.005659894552081823, + 0.0015407315222546458, + 0.0018681908259168267, + 0.0041190823540091515, + 0.002385210944339633, + 0.00380039238370955, + 0.0048505463637411594, + 0.003835879033431411, + 0.00833897851407528, + 0.004849608056247234, + 0.0032000369392335415, + 0.009648330509662628, + 0.006995656527578831, + 0.0060732122510671616, + 0.00891776755452156, + 0.006366906687617302, + 0.009416707791388035, + 0.0054516298696398735, + 0.00816227588802576, + 0.006963164079934359, + 0.005340798757970333, + 0.003369631478562951, + 0.005333296954631805, + 0.004898439161479473, + 0.006019610911607742, + 0.008647524751722813, + 0.0034350405912846327, + 0.013453199528157711, + 0.006026839837431908, + 0.005498540122061968, + 0.006120278034359217, + 0.008910700678825378, + 0.018640214577317238, + 0.016459915786981583, + 0.006079722661525011, + 0.012078240513801575, + 0.009419184178113937, + 0.0019655434880405664, + 0.002285612979903817, + 0.0030828474555164576, + 0.0041825720109045506, + 0.006105111446231604, + 0.002711392240598798, + 0.00445356173440814, + 0.004065617453306913, + 0.004022517241537571, + 0.003735925070941448, + 0.0025520576164126396, + 0.0039433990605175495, + 0.010618245229125023, + 0.003856769297271967, + 0.007375083863735199, + 0.005766655784100294 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.0007323920144699514 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot static and at eye-level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set on a stretch of worn, cracked asphalt in a narrow city alley?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a small, vibrant yellow rubber ball with a matte sheen visible in the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the ball undergo a primary rebound where it compresses upon impact and then restores its spherical shape in flight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting characterized by warm, golden late afternoon sunlight casting long, dynamic 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": [ + -0.009783296845853329 + ], + "opposite_logprob_tokens": [ + -4.6347832679748535 + ], + "logprob_reward": -0.009783296845853329, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009783296845853329, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14890559017658234, + 0.12064246088266373, + 0.046007782220840454, + 0.018277663737535477, + 0.04514070227742195, + 0.08278762549161911, + 0.05946364626288414, + 0.1644126921892166, + 0.21710669994354248, + 0.09725511819124222 + ], + "credit_map": [ + 0.00750349648296833, + 0.008733803406357765, + 0.014206331223249435, + 0.02880934253334999, + 0.027929961681365967, + 0.014534849673509598, + 0.019619407132267952, + 0.027568409219384193, + 0.0073442053981125355, + 0.006512628402560949, + 0.017751777544617653, + 0.018700188025832176, + 0.0066042435355484486, + 0.013459251262247562, + 0.012266508303582668, + 0.038003671914339066, + 0.0035088846925646067, + 0.004802450072020292, + 0.0043669347651302814, + 0.007552038878202438, + 0.0035780223552137613, + 0.00508650578558445, + 0.0066406563855707645, + 0.010472290217876434, + 0.0010372886899858713, + 0.0015222247457131743, + 0.003765439847484231, + 0.003015226684510708, + 0.001514393836259842, + 0.0014898646622896194, + 0.0028883428312838078, + 0.0030448827892541885, + 0.003986031282693148, + 0.003436685772612691, + 0.005343264434486628, + 0.007813137955963612, + 0.003960859030485153, + 0.004638755228370428, + 0.009498286992311478, + 0.006463685538619757, + 0.005396197084337473, + 0.007589510641992092, + 0.010933998972177505, + 0.010686318390071392, + 0.010347618721425533, + 0.007076428271830082, + 0.020287536084651947, + 0.010470017790794373, + 0.004651741590350866, + 0.0061520966701209545, + 0.003933617379516363, + 0.01063266210258007, + 0.008977769874036312, + 0.004120085388422012, + 0.014408177696168423, + 0.0065875002183020115, + 0.009711788035929203, + 0.01184051763266325, + 0.016406090930104256, + 0.026966512203216553, + 0.0357048436999321, + 0.014495222829282284, + 0.02920985408127308, + 0.020077873021364212, + 0.011258312501013279, + 0.018540650606155396, + 0.011949469335377216, + 0.02502775564789772, + 0.06529686599969864, + 0.01825893297791481, + 0.03389560803771019, + 0.03287910297513008, + 0.00574714969843626, + 0.008201077580451965, + 0.004753579385578632, + 0.009824199602007866, + 0.02751048095524311, + 0.008175736293196678, + 0.018341923132538795, + 0.014700975269079208 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 0 + ], + "selected_reward": -0.009783296845853329 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot static and at eye-level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set on a stretch of worn, cracked asphalt in a narrow city alley?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a small, vibrant yellow rubber ball with a matte sheen visible in the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the ball undergo a primary rebound where it compresses upon impact and then restores its spherical shape in flight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting characterized by warm, golden late afternoon sunlight casting long, dynamic 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": [ + -0.01253464724868536 + ], + "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": [ + 0.4680914878845215, + 0.06264614313840866, + 0.03301894664764404, + 0.03206603229045868, + 0.028218446299433708, + 0.028488764539361, + 0.05443380028009415, + 0.21574567258358002, + 0.06252296268939972, + 0.014767707325518131 + ], + "credit_map": [ + 0.036471642553806305, + 0.04347335919737816, + 0.09271042793989182, + 0.058847106993198395, + 0.06330374628305435, + 0.06383199244737625, + 0.05210275575518608, + 0.05735045671463013, + 0.0059440601617097855, + 0.006216582376509905, + 0.006590330973267555, + 0.008391926996409893, + 0.005605949554592371, + 0.005710117984563112, + 0.0103771286085248, + 0.013810047879815102, + 0.0024739340879023075, + 0.0024586657527834177, + 0.003801907179877162, + 0.006643643602728844, + 0.0021127776708453894, + 0.0024714742321521044, + 0.006778140552341938, + 0.006278404500335455, + 0.0020197725389152765, + 0.0027967020869255066, + 0.005041812546551228, + 0.006498724687844515, + 0.0026982093695551157, + 0.0027739531360566616, + 0.00632742652669549, + 0.0039094313979148865, + 0.002859989646822214, + 0.004006606061011553, + 0.004882410168647766, + 0.0036557093262672424, + 0.002175797475501895, + 0.0037380103021860123, + 0.0039059326518326998, + 0.0029939927626401186, + 0.003234999720007181, + 0.003344595432281494, + 0.005284217186272144, + 0.005018559284508228, + 0.0021665364038199186, + 0.003093835199251771, + 0.0033176392316818237, + 0.003028381150215864, + 0.003879640018567443, + 0.004330138675868511, + 0.006614765617996454, + 0.009294661693274975, + 0.009048353880643845, + 0.005518690217286348, + 0.007959766313433647, + 0.007787782698869705, + 0.012437481433153152, + 0.015895016491413116, + 0.024215521290898323, + 0.07225874066352844, + 0.013286210596561432, + 0.013343018479645252, + 0.016300439834594727, + 0.048009242862463, + 0.009244478307664394, + 0.0065876212902367115, + 0.007364933844655752, + 0.008043570443987846, + 0.008674628101289272, + 0.008812437765300274, + 0.005633496679365635, + 0.008161798119544983, + 0.0011948664905503392, + 0.0017567883478477597, + 0.001867078011855483, + 0.001885704812593758, + 0.0026089176535606384, + 0.0011093829525634646, + 0.0023909923620522022, + 0.0019539755303412676 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 1 + ], + "selected_reward": -0.01253464724868536 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, eye-level medium shot captures the scene in photoreal detail, focusing on a stretch of worn, cracked asphalt in a narrow city alley. A small, vibrant yellow rubber ball, its surface displaying a subtle matte sheen and faint signs of wear, plummets into the frame, pulled by gravity. As it quickly makes contact with the ground, its elastic surface visibly and dramatically deforms, instantly compressing and flattening against the hard pavement in a surreal display of physics. This stored energy is instantly released, propelling the ball back into the air with a powerful primary rebound, its spherical shape perfectly restored during flight. Pulled down again, it undergoes a much smaller, faster secondary rebound, losing its remaining momentum before slowly rolling a few inches before coming to a perfectly stable rest in a slight crack in the asphalt. The warm, golden sunlight of the late afternoon streams in from the side, casting long, dynamic shadows that mimic the ball's every movement and contrast against the cool gray tones of the ground, creating an atmosphere of quiet, focused observation and highlighting the pure mechanics of movement and elasticity.", + "seed": 4509, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000026/group_13/index_3139_seed_4509_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": -0.08126944426912815, + "vlm_logprob": -0.08126944426912815, + "vlm_yes_no_margin": 4.375, + "vlm_token_credit": -0.08126944426912815, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot static and at eye-level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set on a stretch of worn, cracked asphalt in a narrow city alley?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a small, vibrant yellow rubber ball with a matte sheen visible in the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the ball undergo a primary rebound where it compresses upon impact and then restores its spherical shape in flight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting characterized by warm, golden late afternoon sunlight casting long, dynamic 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": [ + -0.36676234006881714 + ], + "opposite_logprob_tokens": [ + -1.491762399673462 + ], + "logprob_reward": -0.36676234006881714, + "yes_no_margin_reward": 1.125, + "token_credit_reward": -0.36676234006881714, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.33331185579299927, + 0.0791555717587471, + 0.054994143545627594, + 0.06720291823148727, + 0.12798137962818146, + 0.16218964755535126, + 0.09077179431915283, + 0.028359776362776756, + 0.03978538140654564, + 0.01624753139913082 + ], + "credit_map": [ + 0.02654544822871685, + 0.05969956889748573, + 0.029798680916428566, + 0.030081044882535934, + 0.09352094680070877, + 0.03308780863881111, + 0.022415293380618095, + 0.03816306218504906, + 0.008215281181037426, + 0.00945170409977436, + 0.010280185379087925, + 0.009187706746160984, + 0.007898280397057533, + 0.0101410411298275, + 0.007851040922105312, + 0.01613032817840576, + 0.008055426180362701, + 0.0052954270504415035, + 0.009104573167860508, + 0.00409597996622324, + 0.004778225440531969, + 0.006606303621083498, + 0.007666459772735834, + 0.009391753003001213, + 0.00765221519395709, + 0.011583421379327774, + 0.009927467443048954, + 0.007167784031480551, + 0.00713836494833231, + 0.007619691547006369, + 0.005449820775538683, + 0.010664156638085842, + 0.018511995673179626, + 0.016976937651634216, + 0.021593855693936348, + 0.012079001404345036, + 0.012998728081583977, + 0.017869945615530014, + 0.012102467007935047, + 0.015848446637392044, + 0.017636628821492195, + 0.018802056089043617, + 0.017489157617092133, + 0.018853994086384773, + 0.015981175005435944, + 0.03864825516939163, + 0.013574348762631416, + 0.021204033866524696, + 0.011019484139978886, + 0.008493621833622456, + 0.015436193905770779, + 0.015237255953252316, + 0.00980096124112606, + 0.009543528780341148, + 0.008618202991783619, + 0.012622549198567867, + 0.0025523186195641756, + 0.002588811097666621, + 0.004846499301493168, + 0.0053413319401443005, + 0.0031659684609621763, + 0.0027565942145884037, + 0.003311384469270706, + 0.0037968684919178486, + 0.0029822238720953465, + 0.005886388011276722, + 0.007144811097532511, + 0.004427857231348753, + 0.004848111420869827, + 0.004147322848439217, + 0.005484979134052992, + 0.004863684065639973, + 0.0012576397275552154, + 0.0017919708043336868, + 0.001994543941691518, + 0.002445049351081252, + 0.002110664499923587, + 0.0029136461671441793, + 0.0016105903778225183, + 0.002123425714671612 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 4 + ], + "selected_reward": -0.36676234006881714 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot static and at eye-level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set on a stretch of worn, cracked asphalt in a narrow city alley?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a small, vibrant yellow rubber ball with a matte sheen visible in the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the ball undergo a primary rebound where it compresses upon impact and then restores its spherical shape in flight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting characterized by warm, golden late afternoon sunlight casting long, dynamic 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": [ + -0.009815759025514126 + ], + "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": [ + 0.2666687071323395, + 0.08018123358488083, + 0.20694442093372345, + 0.11614322662353516, + 0.17805173993110657, + 0.04370971396565437, + 0.029272250831127167, + 0.03430095314979553, + 0.030578548088669777, + 0.01414920948445797 + ], + "credit_map": [ + 0.02386380359530449, + 0.041019078344106674, + 0.03507035970687866, + 0.027751106768846512, + 0.05140262097120285, + 0.025702951475977898, + 0.027528245002031326, + 0.0343305729329586, + 0.007563047111034393, + 0.009068446233868599, + 0.018462736159563065, + 0.009774352423846722, + 0.009798100218176842, + 0.006507705897092819, + 0.006297609768807888, + 0.012709234841167927, + 0.04357179254293442, + 0.03270166367292404, + 0.027827907353639603, + 0.010303794406354427, + 0.02389867790043354, + 0.029266435652971268, + 0.012487408705055714, + 0.026886755600571632, + 0.015653712674975395, + 0.02057013101875782, + 0.014701909385621548, + 0.01034533977508545, + 0.01623118296265602, + 0.012932445853948593, + 0.009638819843530655, + 0.0160696841776371, + 0.026787586510181427, + 0.022518005222082138, + 0.02887471206486225, + 0.013448530808091164, + 0.024722320958971977, + 0.029142549261450768, + 0.008863728493452072, + 0.023694295436143875, + 0.005299519747495651, + 0.004014813806861639, + 0.005394951440393925, + 0.006018218584358692, + 0.003969368990510702, + 0.007505649235099554, + 0.0052560437470674515, + 0.006251146551221609, + 0.0025256136432290077, + 0.00298731098882854, + 0.0039505125023424625, + 0.0027828658930957317, + 0.0050855097360908985, + 0.005076237954199314, + 0.002091279486194253, + 0.004772923421114683, + 0.0034201310481876135, + 0.0031540568452328444, + 0.004180424381047487, + 0.00321220513433218, + 0.004748844075948, + 0.005765196401625872, + 0.0036728065460920334, + 0.006147289648652077, + 0.0032701250165700912, + 0.004148665815591812, + 0.0047052414156496525, + 0.004536531399935484, + 0.0027873115614056587, + 0.003970141522586346, + 0.0034499005414545536, + 0.0037106340751051903, + 0.0015387698076665401, + 0.0023341027554124594, + 0.0018481225706636906, + 0.0012187690008431673, + 0.001755504054017365, + 0.0024334832560271025, + 0.0012569853570312262, + 0.0017634727992117405 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 4 + ], + "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 shot static and at eye-level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set on a stretch of worn, cracked asphalt in a narrow city alley?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a small, vibrant yellow rubber ball with a matte sheen visible in the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the ball undergo a primary rebound where it compresses upon impact and then restores its spherical shape in flight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting characterized by warm, golden late afternoon sunlight casting long, dynamic 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": [ + -0.0005112771177664399 + ], + "opposite_logprob_tokens": [ + -7.625511169433594 + ], + "logprob_reward": -0.0005112771177664399, + "yes_no_margin_reward": 7.625, + "token_credit_reward": -0.0005112771177664399, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.44896259903907776, + 0.07168806344270706, + 0.09766367077827454, + 0.11655333638191223, + 0.0666395053267479, + 0.1189299002289772, + 0.019427085295319557, + 0.0247543565928936, + 0.020893914625048637, + 0.014487583190202713 + ], + "credit_map": [ + 0.04050355404615402, + 0.08494560420513153, + 0.040354080498218536, + 0.037139877676963806, + 0.12698209285736084, + 0.04697982221841812, + 0.028969405218958855, + 0.0430881641805172, + 0.006421541795134544, + 0.008037413470447063, + 0.011950382962822914, + 0.008024850860238075, + 0.01176257710903883, + 0.006422638427466154, + 0.009624186903238297, + 0.00944447424262762, + 0.015921656042337418, + 0.014893552288413048, + 0.012620765715837479, + 0.005977185908704996, + 0.01064439956098795, + 0.01152288168668747, + 0.010175053030252457, + 0.01590818352997303, + 0.02083228901028633, + 0.011650800704956055, + 0.018125059083104134, + 0.013534065335988998, + 0.010399013757705688, + 0.013584217987954617, + 0.01125167217105627, + 0.01717621646821499, + 0.007154474966228008, + 0.00977629516273737, + 0.012494219467043877, + 0.0064216176979243755, + 0.007554818410426378, + 0.008794764056801796, + 0.005361077841371298, + 0.00908223818987608, + 0.011754055507481098, + 0.013626949861645699, + 0.013252166099846363, + 0.012591701932251453, + 0.015982231125235558, + 0.026217496022582054, + 0.008583253249526024, + 0.016922060400247574, + 0.0023779296316206455, + 0.0022852462716400623, + 0.0027014631778001785, + 0.0034682496916502714, + 0.002189013408496976, + 0.00238241208717227, + 0.0015426533063873649, + 0.0024801180697977543, + 0.002801246475428343, + 0.0025700340047478676, + 0.0036393713671714067, + 0.0037856325507164, + 0.0032282108440995216, + 0.003032015636563301, + 0.002196318469941616, + 0.0035015270113945007, + 0.001803237828426063, + 0.0024680383503437042, + 0.0024469331838190556, + 0.002311559859663248, + 0.002884314628317952, + 0.003746037371456623, + 0.002499248133972287, + 0.0027345451526343822, + 0.0013181963004171848, + 0.0018930446822196245, + 0.0021379832178354263, + 0.001780215883627534, + 0.0014685848727822304, + 0.0025030924007296562, + 0.001541410805657506, + 0.0018450553761795163 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 3 + ], + "selected_reward": -0.0005112771177664399 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot static and at eye-level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set on a stretch of worn, cracked asphalt in a narrow city alley?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a small, vibrant yellow rubber ball with a matte sheen visible in the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the ball undergo a primary rebound where it compresses upon impact and then restores its spherical shape in flight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting characterized by warm, golden late afternoon sunlight casting long, dynamic 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": [ + -0.011076509952545166 + ], + "opposite_logprob_tokens": [ + -4.5110764503479 + ], + "logprob_reward": -0.011076509952545166, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011076509952545166, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.49534568190574646, + 0.08927512168884277, + 0.09071654081344604, + 0.0898670181632042, + 0.07665576785802841, + 0.08951461315155029, + 0.02532185986638069, + 0.017528649419546127, + 0.014010711573064327, + 0.01176395732909441 + ], + "credit_map": [ + 0.04164521396160126, + 0.09224572777748108, + 0.04566628485918045, + 0.03759308159351349, + 0.15034742653369904, + 0.04906970635056496, + 0.03187279775738716, + 0.046905480325222015, + 0.005215037148445845, + 0.010864960961043835, + 0.010615093633532524, + 0.009683179669082165, + 0.020958542823791504, + 0.007193822879344225, + 0.01332231517881155, + 0.011422178708016872, + 0.00898603443056345, + 0.01187737938016653, + 0.01753106713294983, + 0.00879406277090311, + 0.008581247180700302, + 0.009480993263423443, + 0.011348373256623745, + 0.014117393642663956, + 0.009500650689005852, + 0.016063226386904716, + 0.012186930514872074, + 0.01195921003818512, + 0.009892342612147331, + 0.008236167952418327, + 0.008157515898346901, + 0.013870985247194767, + 0.013630738481879234, + 0.007922669872641563, + 0.009834056720137596, + 0.008624670095741749, + 0.007952583022415638, + 0.008552303537726402, + 0.009437314234673977, + 0.010701430961489677, + 0.010913942009210587, + 0.010554196313023567, + 0.009545057080686092, + 0.014519593678414822, + 0.009525037370622158, + 0.012466677464544773, + 0.009850086644291878, + 0.012140029110014439, + 0.003554380964487791, + 0.0021965934429317713, + 0.004288488533347845, + 0.003960358444601297, + 0.002400818280875683, + 0.0033051413483917713, + 0.0023736669681966305, + 0.003242415376007557, + 0.0022734426893293858, + 0.001447441871277988, + 0.0028357647825032473, + 0.00220598466694355, + 0.0019512416329234838, + 0.0025818059220910072, + 0.0018016189569607377, + 0.002431352622807026, + 0.0025082866195589304, + 0.0015696394257247448, + 0.0019206753931939602, + 0.002183374483138323, + 0.0014173464151099324, + 0.0014719614991918206, + 0.001275540329515934, + 0.001663888804614544, + 0.0013353204121813178, + 0.0013302518054842949, + 0.001738826627843082, + 0.0016549876891076565, + 0.0017938913078978658, + 0.0015438285190612078, + 0.0008338289917446673, + 0.0015330229653045535 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 3 + ], + "selected_reward": -0.011076509952545166 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot static and at eye-level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set on a stretch of worn, cracked asphalt in a narrow city alley?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a small, vibrant yellow rubber ball with a matte sheen visible in the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the ball undergo a primary rebound where it compresses upon impact and then restores its spherical shape in flight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting characterized by warm, golden late afternoon sunlight casting long, dynamic 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": [ + -0.01818133518099785 + ], + "opposite_logprob_tokens": [ + -4.018181324005127 + ], + "logprob_reward": -0.01818133518099785, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.01818133518099785, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.30896493792533875, + 0.1466258019208908, + 0.07737206667661667, + 0.0756516307592392, + 0.12150705605745316, + 0.08709944039583206, + 0.08842623233795166, + 0.053039878606796265, + 0.01824313774704933, + 0.023069817572832108 + ], + "credit_map": [ + 0.03204956650733948, + 0.022587601095438004, + 0.056224673986434937, + 0.036189351230859756, + 0.035408422350883484, + 0.03838779032230377, + 0.039355725049972534, + 0.04876181110739708, + 0.020920736715197563, + 0.017816290259361267, + 0.02656608261168003, + 0.020413080230355263, + 0.012489763088524342, + 0.016139568760991096, + 0.011322901584208012, + 0.020957378670573235, + 0.012027079239487648, + 0.011301430873572826, + 0.01198456808924675, + 0.005526110529899597, + 0.010484729893505573, + 0.00997871719300747, + 0.005534036085009575, + 0.01053539291024208, + 0.008959362283349037, + 0.0143095338717103, + 0.007757030427455902, + 0.008159996010363102, + 0.009307174943387508, + 0.010024026036262512, + 0.007395039778202772, + 0.009739469736814499, + 0.01729460246860981, + 0.01320603396743536, + 0.01601249910891056, + 0.010152526199817657, + 0.015136774629354477, + 0.019905857741832733, + 0.01238131895661354, + 0.017417438328266144, + 0.007580060511827469, + 0.00791320763528347, + 0.011030658148229122, + 0.01775963045656681, + 0.007913538254797459, + 0.012281741015613079, + 0.009997708722949028, + 0.012622902169823647, + 0.01145933847874403, + 0.010769731365144253, + 0.015602865256369114, + 0.0070065767504274845, + 0.0070427535101771355, + 0.012916423380374908, + 0.008862288668751717, + 0.01476624608039856, + 0.0059093390591442585, + 0.004899094812572002, + 0.009562140330672264, + 0.00657048961147666, + 0.0038647453766316175, + 0.006830020807683468, + 0.007133109495043755, + 0.008270940743386745, + 0.003577187890186906, + 0.0019574847538024187, + 0.002518927212804556, + 0.0024286191910505295, + 0.001886703772470355, + 0.0021203511860221624, + 0.001330872531980276, + 0.00242299004457891, + 0.0034906279761344194, + 0.0022679343819618225, + 0.0032311573158949614, + 0.002383905928581953, + 0.0026778150349855423, + 0.0036418247036635876, + 0.0020541732665151358, + 0.003322379430755973 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 4 + ], + "selected_reward": -0.01818133518099785 + } + ], + "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/0000026/group_14/index_tensor([2166, 2166])_group_14_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000026/group_14/index_tensor([2166, 2166])_group_14_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..d9d28866e77914456afb1cf3b42df11941850d94 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000026/group_14/index_tensor([2166, 2166])_group_14_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "Visceral first-person perspective footage shot from the confines of a heavily used off-road helmet thrusts the viewer directly into the high-speed sequence. The helmet's visor frames the world in a circle, deeply scratched and splattered with dry, cracked dirt, partially obscuring the field of vision. Through the clearer sections of plexiglass, the rust and concrete surroundings transform into a disorienting motion blur, hinting at the immense speeds being driven through the abandoned industrial area. Suddenly, a jagged shard of concrete, with a twisted piece of rebar protruding from it, hurtles into the frame and violently deflects from the right side of the helmet with a deafening metallic clang. In immediate reaction to the impact, the entire field of vision shudders violently, a violent and disorienting jolt that momentarily plunges the scene into a chaotic smear of textures and colors. The entire scene is rendered in a raw, gritty industrial style, characterized by a desaturated palette of cold concrete grays and oxidized rust browns, all set under harsh, high-contrast lighting, creating an atmosphere of pure, physical tension and danger.", + "seed": 11569, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000026/group_14/index_2166_seed_11569_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": -0.0534726832062006, + "vlm_logprob": -0.0534726832062006, + "vlm_yes_no_margin": 3.3, + "vlm_token_credit": -0.0534726832062006, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot from a first-person perspective inside a heavily used off-road helmet with a scratched and dirt-splattered visor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment an abandoned industrial area featuring rust and concrete surroundings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a jagged shard of concrete with a twisted piece of rebar hurtle into the frame from the right side and deflect with a metallic clang?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene include a violent shudder and chaotic smear of textures immediately following the impact?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style raw and gritty with a desaturated palette of cold concrete grays and oxidized rust browns under harsh, high-contrast 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": [ + -0.024743063375353813 + ], + "opposite_logprob_tokens": [ + -4.77474308013916 + ], + "logprob_reward": -0.024743063375353813, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.024743063375353813, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1579909771680832, + 0.11671377718448639, + 0.051209188997745514, + 0.09264969080686569, + 0.043531451374292374, + 0.06309380382299423, + 0.08992695063352585, + 0.07128036767244339, + 0.19266827404499054, + 0.12093555182218552 + ], + "credit_map": [ + 0.014207566156983376, + 0.02001873031258583, + 0.020506057888269424, + 0.03185199946165085, + 0.016223641112446785, + 0.01879393868148327, + 0.02053455263376236, + 0.015854492783546448, + 0.014299488626420498, + 0.015695534646511078, + 0.014362559653818607, + 0.017640259116888046, + 0.011047649197280407, + 0.014295803382992744, + 0.02073679119348526, + 0.0086356857791543, + 0.004583421163260937, + 0.005004948005080223, + 0.005806818604469299, + 0.005806253291666508, + 0.003982766531407833, + 0.007119069807231426, + 0.0069896019995212555, + 0.011916306801140308, + 0.008816882036626339, + 0.00802771095186472, + 0.012208528816699982, + 0.013340969569981098, + 0.017794180661439896, + 0.009743966162204742, + 0.014091088436543941, + 0.008626356720924377, + 0.005194340832531452, + 0.005341741722077131, + 0.004981060512363911, + 0.008776010014116764, + 0.004565377254039049, + 0.004633003380149603, + 0.005566122476011515, + 0.0044737933203577995, + 0.005449706222862005, + 0.014927796088159084, + 0.005500831175595522, + 0.00678764283657074, + 0.004448228515684605, + 0.010923657566308975, + 0.0053517185151577, + 0.009704225696623325, + 0.007783266715705395, + 0.016449308022856712, + 0.011200745590031147, + 0.010925199836492538, + 0.01136210560798645, + 0.008528618142008781, + 0.016453539952635765, + 0.007224164437502623, + 0.008701357990503311, + 0.0145315146073699, + 0.011379843577742577, + 0.00989251397550106, + 0.005707453936338425, + 0.007185799069702625, + 0.008243254385888577, + 0.005638623144477606, + 0.017761897295713425, + 0.07181395590305328, + 0.01541190966963768, + 0.014324704185128212, + 0.04600084573030472, + 0.0069211916998028755, + 0.012973662465810776, + 0.007460111752152443, + 0.026955798268318176, + 0.013515411876142025, + 0.012232373468577862, + 0.010636474005877972, + 0.018081266433000565, + 0.013990269973874092, + 0.011812007986009121, + 0.013711952604353428 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 9 + ], + "selected_reward": -0.024743063375353813 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot from a first-person perspective inside a heavily used off-road helmet with a scratched and dirt-splattered visor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment an abandoned industrial area featuring rust and concrete surroundings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a jagged shard of concrete with a twisted piece of rebar hurtle into the frame from the right side and deflect with a metallic clang?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene include a violent shudder and chaotic smear of textures immediately following the impact?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style raw and gritty with a desaturated palette of cold concrete grays and oxidized rust browns under harsh, high-contrast 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": [ + -0.11296716332435608 + ], + "opposite_logprob_tokens": [ + -2.2379672527313232 + ], + "logprob_reward": -0.11296716332435608, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11296716332435608, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16887502372264862, + 0.15365493297576904, + 0.2504388988018036, + 0.05507608503103256, + 0.024927115067839622, + 0.02618017978966236, + 0.05195347219705582, + 0.029221221804618835, + 0.05960206314921379, + 0.18007096648216248 + ], + "credit_map": [ + 0.015725549310445786, + 0.01825949363410473, + 0.01834234967827797, + 0.027999073266983032, + 0.017962941899895668, + 0.02098623290657997, + 0.020330820232629776, + 0.02926856093108654, + 0.014697851613163948, + 0.022940607741475105, + 0.019904201850295067, + 0.028460802510380745, + 0.011008953675627708, + 0.019040217623114586, + 0.022785689681768417, + 0.01481659710407257, + 0.014260141178965569, + 0.017676223069429398, + 0.027317872270941734, + 0.03430579602718353, + 0.020746706053614616, + 0.04069637879729271, + 0.039721108973026276, + 0.05571466684341431, + 0.005380822345614433, + 0.005960302427411079, + 0.006146693602204323, + 0.006949388422071934, + 0.01156316976994276, + 0.005913319531828165, + 0.009302541613578796, + 0.0038598463870584965, + 0.0019245101138949394, + 0.003099011490121484, + 0.003515751101076603, + 0.003778927493840456, + 0.003200471168383956, + 0.0025286711752414703, + 0.004168652929365635, + 0.002711119595915079, + 0.004628884606063366, + 0.003156938124448061, + 0.0026287713553756475, + 0.00372289284132421, + 0.0021349797025322914, + 0.002551476936787367, + 0.0030009064357727766, + 0.004355329088866711, + 0.005072935484349728, + 0.007654630113393068, + 0.00689389742910862, + 0.0068848212249577045, + 0.005826902575790882, + 0.005744547117501497, + 0.00873336661607027, + 0.005142370704561472, + 0.003535497235134244, + 0.005799393635243177, + 0.004626816138625145, + 0.004463877994567156, + 0.0021661773789674044, + 0.0022783188614994287, + 0.004137599840760231, + 0.002213542116805911, + 0.005127900280058384, + 0.015964461490511894, + 0.009647565893828869, + 0.005659437272697687, + 0.01294004637748003, + 0.0033200657926499844, + 0.004680179990828037, + 0.00226240954361856, + 0.043144334107637405, + 0.01893051527440548, + 0.020995786413550377, + 0.0175690446048975, + 0.028737017884850502, + 0.019052892923355103, + 0.015259127132594585, + 0.0163822453469038 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 9, + 0 + ], + "selected_reward": -0.11296716332435608 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot from a first-person perspective inside a heavily used off-road helmet with a scratched and dirt-splattered visor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment an abandoned industrial area featuring rust and concrete surroundings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a jagged shard of concrete with a twisted piece of rebar hurtle into the frame from the right side and deflect with a metallic clang?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene include a violent shudder and chaotic smear of textures immediately following the impact?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style raw and gritty with a desaturated palette of cold concrete grays and oxidized rust browns under harsh, high-contrast 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": [ + -0.07004115730524063 + ], + "opposite_logprob_tokens": [ + -2.6950411796569824 + ], + "logprob_reward": -0.07004115730524063, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07004115730524063, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14438496530056, + 0.06651050597429276, + 0.09419042617082596, + 0.04162810742855072, + 0.04522371664643288, + 0.2209974080324173, + 0.10744091868400574, + 0.03354722261428833, + 0.053512781858444214, + 0.1925639808177948 + ], + "credit_map": [ + 0.011673952452838421, + 0.01640746369957924, + 0.016559455543756485, + 0.023836230859160423, + 0.015391494147479534, + 0.01933784782886505, + 0.015423735603690147, + 0.025754788890480995, + 0.008181342855095863, + 0.009051994420588017, + 0.00833338312804699, + 0.010723263956606388, + 0.0068511259742081165, + 0.007890792563557625, + 0.010547718964517117, + 0.0049308836460113525, + 0.0051718601025640965, + 0.00935803260654211, + 0.010373328812420368, + 0.016103005036711693, + 0.008307026699185371, + 0.014584322459995747, + 0.012570449151098728, + 0.017722398042678833, + 0.0043910592794418335, + 0.003996090032160282, + 0.004948780871927738, + 0.005712682381272316, + 0.005197088699787855, + 0.0038561259862035513, + 0.00724396388977766, + 0.006282313261181116, + 0.005793160758912563, + 0.00485287606716156, + 0.004889866337180138, + 0.00859243143349886, + 0.005547450855374336, + 0.004838855005800724, + 0.006682385690510273, + 0.004026693757623434, + 0.021672315895557404, + 0.03387761861085892, + 0.023445049300789833, + 0.0248324666172266, + 0.017167940735816956, + 0.020931418985128403, + 0.02568957209587097, + 0.053381018340587616, + 0.010218714363873005, + 0.014859567396342754, + 0.014785212464630604, + 0.014007988385856152, + 0.013531859964132309, + 0.01103433407843113, + 0.019940277561545372, + 0.00906296819448471, + 0.003377191722393036, + 0.0070015923120081425, + 0.005897401832044125, + 0.004658914636820555, + 0.0027552731335163116, + 0.0038672802038490772, + 0.003376844571903348, + 0.0026127223391085863, + 0.006295615807175636, + 0.015225924551486969, + 0.008283501490950584, + 0.004945801105350256, + 0.008586778305470943, + 0.0030942175071686506, + 0.004168415442109108, + 0.0029125274159014225, + 0.04482308030128479, + 0.019271625205874443, + 0.023003211244940758, + 0.02148592472076416, + 0.031473252922296524, + 0.018656468018889427, + 0.01769099198281765, + 0.0161594208329916 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 9, + 0 + ], + "selected_reward": -0.07004115730524063 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot from a first-person perspective inside a heavily used off-road helmet with a scratched and dirt-splattered visor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment an abandoned industrial area featuring rust and concrete surroundings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a jagged shard of concrete with a twisted piece of rebar hurtle into the frame from the right side and deflect with a metallic clang?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene include a violent shudder and chaotic smear of textures immediately following the impact?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style raw and gritty with a desaturated palette of cold concrete grays and oxidized rust browns under harsh, high-contrast 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": [ + -0.029808735474944115 + ], + "opposite_logprob_tokens": [ + -3.529808759689331 + ], + "logprob_reward": -0.029808735474944115, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029808735474944115, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18025393784046173, + 0.08826187998056412, + 0.07049667835235596, + 0.039829038083553314, + 0.05367675796151161, + 0.08401835709810257, + 0.05904151126742363, + 0.07742682844400406, + 0.044166527688503265, + 0.30282843112945557 + ], + "credit_map": [ + 0.013584683649241924, + 0.024955714121460915, + 0.026533929631114006, + 0.035051148384809494, + 0.018960479646921158, + 0.021323220804333687, + 0.01871214434504509, + 0.021132629364728928, + 0.009267798624932766, + 0.012689449824392796, + 0.013376712799072266, + 0.013901249505579472, + 0.007363158743828535, + 0.009578617289662361, + 0.014103084802627563, + 0.007981803268194199, + 0.004822211805731058, + 0.007104775868356228, + 0.007946907542645931, + 0.008528114296495914, + 0.008020337671041489, + 0.011676456779241562, + 0.011342772282660007, + 0.011055100709199905, + 0.004208355210721493, + 0.004518317058682442, + 0.004492509178817272, + 0.006239726208150387, + 0.0041375053115189075, + 0.003978692926466465, + 0.005717669613659382, + 0.0065362597815692425, + 0.004517238587141037, + 0.005991007667034864, + 0.007023207377642393, + 0.012859981507062912, + 0.004788958001881838, + 0.00490619195625186, + 0.008773425593972206, + 0.0048167468048632145, + 0.012531263753771782, + 0.01118447631597519, + 0.01222846657037735, + 0.013003646396100521, + 0.007091874722391367, + 0.00842166319489479, + 0.007383523974567652, + 0.012173444963991642, + 0.005775758996605873, + 0.0066321901977062225, + 0.007606974802911282, + 0.009320005774497986, + 0.006095097400248051, + 0.0074432482942938805, + 0.009146361611783504, + 0.007021877449005842, + 0.010506115853786469, + 0.011668526567518711, + 0.00999409705400467, + 0.010915924794971943, + 0.006668193731456995, + 0.012224285863339901, + 0.008562997914850712, + 0.00688667967915535, + 0.004666831810027361, + 0.005761389154940844, + 0.00951567105948925, + 0.004315515980124474, + 0.008159644901752472, + 0.004277127329260111, + 0.0042113675735890865, + 0.003258981043472886, + 0.06452669948339462, + 0.02939750999212265, + 0.036530666053295135, + 0.04114252328872681, + 0.04575291648507118, + 0.029234329238533974, + 0.031149553135037422, + 0.025094233453273773 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 1 + ], + "selected_reward": -0.029808735474944115 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot from a first-person perspective inside a heavily used off-road helmet with a scratched and dirt-splattered visor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment an abandoned industrial area featuring rust and concrete surroundings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a jagged shard of concrete with a twisted piece of rebar hurtle into the frame from the right side and deflect with a metallic clang?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene include a violent shudder and chaotic smear of textures immediately following the impact?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style raw and gritty with a desaturated palette of cold concrete grays and oxidized rust browns under harsh, high-contrast 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": [ + -0.02980329655110836 + ], + "opposite_logprob_tokens": [ + -3.5298032760620117 + ], + "logprob_reward": -0.02980329655110836, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.02980329655110836, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2001059502363205, + 0.10280947387218475, + 0.1821880340576172, + 0.04727698117494583, + 0.022274373099207878, + 0.07736624032258987, + 0.03230627253651619, + 0.03942074999213219, + 0.08894622325897217, + 0.20730571448802948 + ], + "credit_map": [ + 0.023822415620088577, + 0.019250983372330666, + 0.02562028169631958, + 0.03307490423321724, + 0.01963929459452629, + 0.021304896101355553, + 0.026425663381814957, + 0.030967505648732185, + 0.010306109674274921, + 0.015794066712260246, + 0.010689156129956245, + 0.020142974331974983, + 0.006310104392468929, + 0.01378690917044878, + 0.015336093492805958, + 0.010444057174026966, + 0.012793215923011303, + 0.013827505521476269, + 0.021324889734387398, + 0.02291121706366539, + 0.014763926155865192, + 0.028926437720656395, + 0.025613831356167793, + 0.04202702268958092, + 0.0037349271588027477, + 0.003885669633746147, + 0.006633426062762737, + 0.007998528890311718, + 0.006042292341589928, + 0.004489490296691656, + 0.008919143117964268, + 0.0055735004134476185, + 0.0016299893613904715, + 0.0024186770897358656, + 0.0030228523537516594, + 0.0050561572425067425, + 0.0021485150791704655, + 0.002086474560201168, + 0.003927835263311863, + 0.0019838721491396427, + 0.007645407225936651, + 0.011945818550884724, + 0.01015265192836523, + 0.009283539839088917, + 0.006655323319137096, + 0.007294093258678913, + 0.009447349235415459, + 0.014942053705453873, + 0.0025218974333256483, + 0.005338183604180813, + 0.002955653704702854, + 0.005680416710674763, + 0.0035910294391214848, + 0.004011122509837151, + 0.003372511826455593, + 0.004835458938032389, + 0.004883539862930775, + 0.007802111096680164, + 0.005626676604151726, + 0.004894097801297903, + 0.0029418235644698143, + 0.005749501753598452, + 0.004366704262793064, + 0.003156292950734496, + 0.009075057692825794, + 0.01696734130382538, + 0.01944093406200409, + 0.009090687148272991, + 0.015189651399850845, + 0.005723691079765558, + 0.009220357984304428, + 0.004238503985106945, + 0.04193330556154251, + 0.02126786857843399, + 0.018227029591798782, + 0.022676028311252594, + 0.02771814912557602, + 0.026655878871679306, + 0.02642623521387577, + 0.022401219233870506 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 2 + ], + "selected_reward": -0.02980329655110836 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "Visceral first-person perspective footage shot from the confines of a heavily used off-road helmet thrusts the viewer directly into the high-speed sequence. The helmet's visor frames the world in a circle, deeply scratched and splattered with dry, cracked dirt, partially obscuring the field of vision. Through the clearer sections of plexiglass, the rust and concrete surroundings transform into a disorienting motion blur, hinting at the immense speeds being driven through the abandoned industrial area. Suddenly, a jagged shard of concrete, with a twisted piece of rebar protruding from it, hurtles into the frame and violently deflects from the right side of the helmet with a deafening metallic clang. In immediate reaction to the impact, the entire field of vision shudders violently, a violent and disorienting jolt that momentarily plunges the scene into a chaotic smear of textures and colors. The entire scene is rendered in a raw, gritty industrial style, characterized by a desaturated palette of cold concrete grays and oxidized rust browns, all set under harsh, high-contrast lighting, creating an atmosphere of pure, physical tension and danger.", + "seed": 11570, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000026/group_14/index_2166_seed_11570_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": -0.032784252986311914, + "vlm_logprob": -0.032784252986311914, + "vlm_yes_no_margin": 3.75, + "vlm_token_credit": -0.032784252986311914, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot from a first-person perspective inside a heavily used off-road helmet with a scratched and dirt-splattered visor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment an abandoned industrial area featuring rust and concrete surroundings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a jagged shard of concrete with a twisted piece of rebar hurtle into the frame from the right side and deflect with a metallic clang?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene include a violent shudder and chaotic smear of textures immediately following the impact?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style raw and gritty with a desaturated palette of cold concrete grays and oxidized rust browns under harsh, high-contrast 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": [ + -0.02590116485953331 + ], + "opposite_logprob_tokens": [ + -4.0259013175964355 + ], + "logprob_reward": -0.02590116485953331, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.02590116485953331, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11018334329128265, + 0.12123088538646698, + 0.061504632234573364, + 0.3187355697154999, + 0.019939398393034935, + 0.0770910382270813, + 0.11745411157608032, + 0.06294476985931396, + 0.044259410351514816, + 0.06665682047605515 + ], + "credit_map": [ + 0.014885688200592995, + 0.02582419477403164, + 0.02235938236117363, + 0.017584772780537605, + 0.006905328016728163, + 0.00817701406776905, + 0.010330184362828732, + 0.004116784781217575, + 0.00872771441936493, + 0.0063421656377613544, + 0.04412400722503662, + 0.033177196979522705, + 0.0027737789787352085, + 0.008112655021250248, + 0.005773734301328659, + 0.012199637480080128, + 0.006338043604046106, + 0.008663431741297245, + 0.01124644372612238, + 0.012292729690670967, + 0.004457366652786732, + 0.004513402935117483, + 0.00836117286235094, + 0.005632044747471809, + 0.050254251807928085, + 0.06331691145896912, + 0.04049643874168396, + 0.038086432963609695, + 0.03287866711616516, + 0.03322219476103783, + 0.038711585104465485, + 0.021769104525446892, + 0.0030016324017196894, + 0.0028475080616772175, + 0.0026153174694627523, + 0.002826621988788247, + 0.001292780856601894, + 0.002631677547469735, + 0.0024553395342081785, + 0.0022685208823531866, + 0.013661335222423077, + 0.010317609645426273, + 0.01575075462460518, + 0.014980918727815151, + 0.004834019113332033, + 0.007251441944390535, + 0.0057421536184847355, + 0.00455280439928174, + 0.017084376886487007, + 0.024798447266221046, + 0.014333946630358696, + 0.015862829983234406, + 0.009225317277014256, + 0.011938514187932014, + 0.010686223395168781, + 0.013524462468922138, + 0.014258991926908493, + 0.011581314727663994, + 0.007415407337248325, + 0.007626643870025873, + 0.005427096039056778, + 0.0032331242691725492, + 0.004474738147109747, + 0.008927450515329838, + 0.003706939285621047, + 0.004191642627120018, + 0.006594798061996698, + 0.007199855986982584, + 0.006147308275103569, + 0.0040810503996908665, + 0.00808874610811472, + 0.004249068908393383, + 0.009626494720578194, + 0.01385976281017065, + 0.010370484553277493, + 0.01111972238868475, + 0.006545075215399265, + 0.004639119375497103, + 0.006457203533500433, + 0.004038956481963396 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 1, + 6 + ], + "selected_reward": -0.02590116485953331 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot from a first-person perspective inside a heavily used off-road helmet with a scratched and dirt-splattered visor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment an abandoned industrial area featuring rust and concrete surroundings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a jagged shard of concrete with a twisted piece of rebar hurtle into the frame from the right side and deflect with a metallic clang?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene include a violent shudder and chaotic smear of textures immediately following the impact?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style raw and gritty with a desaturated palette of cold concrete grays and oxidized rust browns under harsh, high-contrast 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": [ + -0.08909747749567032 + ], + "opposite_logprob_tokens": [ + -2.464097499847412 + ], + "logprob_reward": -0.08909747749567032, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08909747749567032, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3226703107357025, + 0.02495049498975277, + 0.056208256632089615, + 0.1704835593700409, + 0.022413788363337517, + 0.03807757794857025, + 0.19807298481464386, + 0.05556691437959671, + 0.04650914669036865, + 0.06504694372415543 + ], + "credit_map": [ + 0.05000504106283188, + 0.08518246561288834, + 0.044614408165216446, + 0.04757889732718468, + 0.025244319811463356, + 0.027027936652302742, + 0.029796333983540535, + 0.013220894150435925, + 0.00258479337207973, + 0.0028618883807212114, + 0.004721068777143955, + 0.0038929444272071123, + 0.002220229944214225, + 0.0034964580554515123, + 0.00286219734698534, + 0.0023109137546271086, + 0.005989348515868187, + 0.007959894835948944, + 0.005805484019219875, + 0.0068869516253471375, + 0.008938784711062908, + 0.011071214452385902, + 0.0059649343602359295, + 0.003591642016544938, + 0.025689058005809784, + 0.031631190329790115, + 0.021356284618377686, + 0.021116800606250763, + 0.017187757417559624, + 0.016809627413749695, + 0.02290608361363411, + 0.013786762952804565, + 0.003201593877747655, + 0.0023849736899137497, + 0.002186958445236087, + 0.00258141802623868, + 0.0018918407149612904, + 0.0033935303799808025, + 0.0031563718803226948, + 0.0036170990206301212, + 0.006784656550735235, + 0.005322187673300505, + 0.0036881405394524336, + 0.007029123138636351, + 0.0029219593852758408, + 0.004398157354444265, + 0.004653633572161198, + 0.003279720665886998, + 0.03545225039124489, + 0.02183889038860798, + 0.04080896079540253, + 0.021113893017172813, + 0.015242369845509529, + 0.02390093356370926, + 0.014973736368119717, + 0.024741945788264275, + 0.008638340048491955, + 0.006821077316999435, + 0.007043246645480394, + 0.008880877867341042, + 0.004444705322384834, + 0.0047955443151295185, + 0.004463727120310068, + 0.010479395277798176, + 0.004088024143129587, + 0.006288055330514908, + 0.007035649847239256, + 0.007774274330586195, + 0.0051411050371825695, + 0.004019100219011307, + 0.0078102289699018, + 0.004352708347141743, + 0.007116342429071665, + 0.005635393783450127, + 0.005311275366693735, + 0.009658614173531532, + 0.005808679386973381, + 0.005642375443130732, + 0.009246423840522766, + 0.016627833247184753 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 3 + ], + "selected_reward": -0.08909747749567032 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot from a first-person perspective inside a heavily used off-road helmet with a scratched and dirt-splattered visor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment an abandoned industrial area featuring rust and concrete surroundings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a jagged shard of concrete with a twisted piece of rebar hurtle into the frame from the right side and deflect with a metallic clang?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene include a violent shudder and chaotic smear of textures immediately following the impact?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style raw and gritty with a desaturated palette of cold concrete grays and oxidized rust browns under harsh, high-contrast 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": [ + -0.018194211646914482 + ], + "opposite_logprob_tokens": [ + -4.018194198608398 + ], + "logprob_reward": -0.018194211646914482, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018194211646914482, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03789634257555008, + 0.09725817292928696, + 0.04640960320830345, + 0.5362136960029602, + 0.02972000278532505, + 0.06737460196018219, + 0.030980518087744713, + 0.10702208429574966, + 0.028067372739315033, + 0.019057629629969597 + ], + "credit_map": [ + 0.0032565491273999214, + 0.004014382138848305, + 0.005392302293330431, + 0.0056542991660535336, + 0.0020316357258707285, + 0.006628141272813082, + 0.007999357767403126, + 0.0029196767136454582, + 0.00700019858777523, + 0.004692899528890848, + 0.03433319181203842, + 0.026217760518193245, + 0.003121728543192148, + 0.007360171992331743, + 0.0045848130248487, + 0.009947399608790874, + 0.004609250463545322, + 0.0056491126306355, + 0.009330743923783302, + 0.009307222440838814, + 0.004526220727711916, + 0.00449887802824378, + 0.004057047888636589, + 0.004431128967553377, + 0.08000239729881287, + 0.09995243698358536, + 0.06687731295824051, + 0.06761649996042252, + 0.05421103164553642, + 0.04978890344500542, + 0.07408426702022552, + 0.0436808280646801, + 0.0029657259583473206, + 0.0034582011867314577, + 0.004025445319712162, + 0.0057217711582779884, + 0.0020768495742231607, + 0.002979924203827977, + 0.005633633583784103, + 0.0028584501706063747, + 0.009218381717801094, + 0.0070860860869288445, + 0.015540480613708496, + 0.014653108082711697, + 0.004120637197047472, + 0.005583634600043297, + 0.0067744567058980465, + 0.0043978155590593815, + 0.004428549204021692, + 0.005357890389859676, + 0.0037617485504597425, + 0.0042010280303657055, + 0.002816511318087578, + 0.002893269993364811, + 0.0025471914559602737, + 0.00497432891279459, + 0.030070679262280464, + 0.015299588441848755, + 0.006913298275321722, + 0.007410894148051739, + 0.005030507687479258, + 0.00812623929232359, + 0.008074396289885044, + 0.02609647810459137, + 0.0028957666363567114, + 0.003347977064549923, + 0.004866365343332291, + 0.004064278211444616, + 0.002977729542180896, + 0.003721865126863122, + 0.0035673831589519978, + 0.002626009052619338, + 0.00280637014657259, + 0.0021225465461611748, + 0.0020068101584911346, + 0.0031321034766733646, + 0.0018594057764858007, + 0.001471782336011529, + 0.003088513622060418, + 0.0025700987316668034 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 7, + 1 + ], + "selected_reward": -0.018194211646914482 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot from a first-person perspective inside a heavily used off-road helmet with a scratched and dirt-splattered visor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment an abandoned industrial area featuring rust and concrete surroundings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a jagged shard of concrete with a twisted piece of rebar hurtle into the frame from the right side and deflect with a metallic clang?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene include a violent shudder and chaotic smear of textures immediately following the impact?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style raw and gritty with a desaturated palette of cold concrete grays and oxidized rust browns under harsh, high-contrast 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": [ + -0.012538531795144081 + ], + "opposite_logprob_tokens": [ + -4.387538433074951 + ], + "logprob_reward": -0.012538531795144081, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012538531795144081, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07586675137281418, + 0.08855520933866501, + 0.02798403613269329, + 0.3088337182998657, + 0.020281808450818062, + 0.2487553060054779, + 0.16123969852924347, + 0.024644875898957253, + 0.02802259474992752, + 0.01581600308418274 + ], + "credit_map": [ + 0.011390709318220615, + 0.017722679302096367, + 0.011450155638158321, + 0.012251149863004684, + 0.005763457156717777, + 0.006615870166569948, + 0.006818510591983795, + 0.0038542174734175205, + 0.006536324508488178, + 0.004319826606661081, + 0.02885443903505802, + 0.022488879039883614, + 0.002934100339189172, + 0.0070350281894207, + 0.0059877485036849976, + 0.010398861952126026, + 0.0029826494865119457, + 0.003935531247407198, + 0.004390171729028225, + 0.004874029662460089, + 0.0025140130892395973, + 0.0034357565455138683, + 0.003525937907397747, + 0.0023259457666426897, + 0.047430846840143204, + 0.0591004453599453, + 0.0385177843272686, + 0.03816837817430496, + 0.031015001237392426, + 0.0297857653349638, + 0.043149929493665695, + 0.02166556566953659, + 0.002394425682723522, + 0.0025064065121114254, + 0.0025751465000212193, + 0.004423034843057394, + 0.001333009684458375, + 0.0016532219015061855, + 0.0035827120300382376, + 0.001813850598409772, + 0.0314486026763916, + 0.026104766875505447, + 0.059964362531900406, + 0.057285647839307785, + 0.01256569940596819, + 0.018110455945134163, + 0.026050200685858727, + 0.017225580289959908, + 0.021524131298065186, + 0.032177556306123734, + 0.03563303127884865, + 0.019603891298174858, + 0.012485763058066368, + 0.013860986568033695, + 0.01664901338517666, + 0.009305313229560852, + 0.003501239465549588, + 0.0028010259848088026, + 0.003621595911681652, + 0.0027060681022703648, + 0.001849826890975237, + 0.0021148542873561382, + 0.002519962377846241, + 0.005530303809791803, + 0.0030415200162678957, + 0.003098036628216505, + 0.004701307974755764, + 0.003523933468386531, + 0.00289848237298429, + 0.003969130571931601, + 0.004232365172356367, + 0.002557817380875349, + 0.0022026372607797384, + 0.00196076394058764, + 0.001849762862548232, + 0.0028290406335145235, + 0.0012153716525062919, + 0.0011199602158740163, + 0.0027745049446821213, + 0.0018639607587829232 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 5, + 6 + ], + "selected_reward": -0.012538531795144081 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot from a first-person perspective inside a heavily used off-road helmet with a scratched and dirt-splattered visor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment an abandoned industrial area featuring rust and concrete surroundings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a jagged shard of concrete with a twisted piece of rebar hurtle into the frame from the right side and deflect with a metallic clang?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene include a violent shudder and chaotic smear of textures immediately following the impact?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style raw and gritty with a desaturated palette of cold concrete grays and oxidized rust browns under harsh, high-contrast 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": [ + -0.01818987913429737 + ], + "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": [ + 0.3660396337509155, + 0.06459065526723862, + 0.05212203785777092, + 0.03534718230366707, + 0.02659391239285469, + 0.08145136386156082, + 0.16053321957588196, + 0.061669815331697464, + 0.04711136966943741, + 0.10454083979129791 + ], + "credit_map": [ + 0.035047221928834915, + 0.05765543133020401, + 0.10323918610811234, + 0.08000292629003525, + 0.02063535340130329, + 0.029557403177022934, + 0.025996744632720947, + 0.013905376195907593, + 0.005001524463295937, + 0.005140019580721855, + 0.015259933657944202, + 0.010757198557257652, + 0.0046974848955869675, + 0.00716904504224658, + 0.010228303261101246, + 0.006337148603051901, + 0.006468704901635647, + 0.012378212995827198, + 0.0065015763975679874, + 0.007385535631328821, + 0.005459730047732592, + 0.0065428391098976135, + 0.004549460019916296, + 0.002835974795743823, + 0.004380351863801479, + 0.0068589444272220135, + 0.004155375994741917, + 0.004749995656311512, + 0.004353655967861414, + 0.004235113505274057, + 0.004620757419615984, + 0.001992987236008048, + 0.005113775841891766, + 0.0037779989652335644, + 0.0025605522096157074, + 0.003280024975538254, + 0.0021191814448684454, + 0.002431093016639352, + 0.00419069267809391, + 0.0031205934938043356, + 0.012754879891872406, + 0.006768117193132639, + 0.015867112204432487, + 0.020001526921987534, + 0.0039721340872347355, + 0.007910137996077538, + 0.009463785216212273, + 0.004713670816272497, + 0.01586836576461792, + 0.02091078646481037, + 0.03383542224764824, + 0.019451314583420753, + 0.013545994646847248, + 0.024366002529859543, + 0.013623555190861225, + 0.018931761384010315, + 0.012784798629581928, + 0.00839021522551775, + 0.007455149199813604, + 0.009662154130637646, + 0.0039282930083572865, + 0.004416897892951965, + 0.004298797808587551, + 0.010733511298894882, + 0.005536965560168028, + 0.008998026140034199, + 0.0077175297774374485, + 0.007573316805064678, + 0.004003702662885189, + 0.0034577948972582817, + 0.005745223257690668, + 0.004078808706253767, + 0.014804734848439693, + 0.010331524536013603, + 0.011750451289117336, + 0.015743490308523178, + 0.00915464200079441, + 0.007963270880281925, + 0.018000977113842964, + 0.016791747882962227 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 9 + ], + "selected_reward": -0.01818987913429737 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "Visceral first-person perspective footage shot from the confines of a heavily used off-road helmet thrusts the viewer directly into the high-speed sequence. The helmet's visor frames the world in a circle, deeply scratched and splattered with dry, cracked dirt, partially obscuring the field of vision. Through the clearer sections of plexiglass, the rust and concrete surroundings transform into a disorienting motion blur, hinting at the immense speeds being driven through the abandoned industrial area. Suddenly, a jagged shard of concrete, with a twisted piece of rebar protruding from it, hurtles into the frame and violently deflects from the right side of the helmet with a deafening metallic clang. In immediate reaction to the impact, the entire field of vision shudders violently, a violent and disorienting jolt that momentarily plunges the scene into a chaotic smear of textures and colors. The entire scene is rendered in a raw, gritty industrial style, characterized by a desaturated palette of cold concrete grays and oxidized rust browns, all set under harsh, high-contrast lighting, creating an atmosphere of pure, physical tension and danger.", + "seed": 11571, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000026/group_14/index_2166_seed_11571_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": -0.028163571283221243, + "vlm_logprob": -0.028163571283221243, + "vlm_yes_no_margin": 3.975, + "vlm_token_credit": -0.028163571283221243, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot from a first-person perspective inside a heavily used off-road helmet with a scratched and dirt-splattered visor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment an abandoned industrial area featuring rust and concrete surroundings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a jagged shard of concrete with a twisted piece of rebar hurtle into the frame from the right side and deflect with a metallic clang?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene include a violent shudder and chaotic smear of textures immediately following the impact?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style raw and gritty with a desaturated palette of cold concrete grays and oxidized rust browns under harsh, high-contrast 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": [ + -0.015397251583635807 + ], + "opposite_logprob_tokens": [ + -4.765397071838379 + ], + "logprob_reward": -0.015397251583635807, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.015397251583635807, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14145389199256897, + 0.10978996753692627, + 0.03400884568691254, + 0.03116362914443016, + 0.05352557823061943, + 0.16731519997119904, + 0.3014585077762604, + 0.06779155135154724, + 0.04240122064948082, + 0.05109163373708725 + ], + "credit_map": [ + 0.019599173218011856, + 0.018465090543031693, + 0.015880079939961433, + 0.01787516102194786, + 0.01589520089328289, + 0.016744859516620636, + 0.02078184485435486, + 0.01621248573064804, + 0.024195190519094467, + 0.016950540244579315, + 0.01788550242781639, + 0.015844203531742096, + 0.010128705762326717, + 0.00688918586820364, + 0.009426949545741081, + 0.008469698019325733, + 0.004084005020558834, + 0.004241638816893101, + 0.003829909721389413, + 0.004374659154564142, + 0.0032372043933719397, + 0.0037118899635970592, + 0.007163978181779385, + 0.0033655615989118814, + 0.004630801267921925, + 0.00555129861459136, + 0.004269035067409277, + 0.004667374305427074, + 0.0031956336461007595, + 0.0018302578246220946, + 0.0036698689218610525, + 0.0033493598457425833, + 0.006891197990626097, + 0.007304435595870018, + 0.009194944053888321, + 0.010021780617535114, + 0.004848828539252281, + 0.004848197102546692, + 0.005353796761482954, + 0.005062401294708252, + 0.015966197475790977, + 0.010490686632692814, + 0.04278258979320526, + 0.01772262342274189, + 0.04648277536034584, + 0.022514905780553818, + 0.00643549719825387, + 0.0049199303612113, + 0.06542740762233734, + 0.043639473617076874, + 0.034151479601860046, + 0.0353204682469368, + 0.03822069987654686, + 0.03859582170844078, + 0.021421099081635475, + 0.024682048708200455, + 0.009890800341963768, + 0.012668794952332973, + 0.004979991354048252, + 0.00610469002276659, + 0.01180174108594656, + 0.00775996083393693, + 0.008854596875607967, + 0.005730972625315189, + 0.0050025153905153275, + 0.004266463685780764, + 0.004491006024181843, + 0.00905902311205864, + 0.004899346269667149, + 0.003429270815104246, + 0.005175724159926176, + 0.006077873986214399, + 0.0065112607553601265, + 0.007867151871323586, + 0.004057054407894611, + 0.006827619392424822, + 0.00827715639024973, + 0.00483670411631465, + 0.008123033680021763, + 0.004591648932546377 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 5, + 0 + ], + "selected_reward": -0.015397251583635807 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot from a first-person perspective inside a heavily used off-road helmet with a scratched and dirt-splattered visor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment an abandoned industrial area featuring rust and concrete surroundings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a jagged shard of concrete with a twisted piece of rebar hurtle into the frame from the right side and deflect with a metallic clang?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene include a violent shudder and chaotic smear of textures immediately following the impact?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style raw and gritty with a desaturated palette of cold concrete grays and oxidized rust browns under harsh, high-contrast 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": [ + -0.0789872407913208 + ], + "opposite_logprob_tokens": [ + -2.5789871215820312 + ], + "logprob_reward": -0.0789872407913208, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.0789872407913208, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2549649178981781, + 0.12962985038757324, + 0.053595852106809616, + 0.11375240236520767, + 0.04249186813831329, + 0.031911421567201614, + 0.16100987792015076, + 0.06808310002088547, + 0.10921624302864075, + 0.035344477742910385 + ], + "credit_map": [ + 0.0446162074804306, + 0.029965480789542198, + 0.03104628436267376, + 0.0423370897769928, + 0.021116476505994797, + 0.023874016478657722, + 0.032618146389722824, + 0.02939119189977646, + 0.021436307579278946, + 0.020727235823869705, + 0.018611043691635132, + 0.020371368154883385, + 0.015368407592177391, + 0.00858156755566597, + 0.014101334847509861, + 0.010432591661810875, + 0.008320474065840244, + 0.00783457886427641, + 0.00652015395462513, + 0.009708283469080925, + 0.006670363247394562, + 0.004361252766102552, + 0.006432565860450268, + 0.0037481794133782387, + 0.019713057205080986, + 0.01740231364965439, + 0.016176221892237663, + 0.018113652244210243, + 0.010141542181372643, + 0.006854535546153784, + 0.012928945012390614, + 0.012422134168446064, + 0.005719986744225025, + 0.006627883296459913, + 0.007661832496523857, + 0.00728246383368969, + 0.0040246122516691685, + 0.003946115728467703, + 0.003482647007331252, + 0.003746327944099903, + 0.003444660920649767, + 0.0028937540482729673, + 0.0026528213638812304, + 0.0033327399287372828, + 0.007638598792254925, + 0.0025458447635173798, + 0.0025897540617734194, + 0.00681324815377593, + 0.037759099155664444, + 0.02099151536822319, + 0.021322421729564667, + 0.022122742608189583, + 0.01647617667913437, + 0.012451842427253723, + 0.016512179747223854, + 0.013373895548284054, + 0.0058131436817348, + 0.012098529376089573, + 0.005994747392833233, + 0.009082350879907608, + 0.008355802856385708, + 0.005882319062948227, + 0.007960648275911808, + 0.012895557098090649, + 0.013261551037430763, + 0.012643822468817234, + 0.010795487090945244, + 0.02059711515903473, + 0.01755615696310997, + 0.008419008925557137, + 0.014031734317541122, + 0.011911369860172272, + 0.0027834628708660603, + 0.004606146831065416, + 0.002256202744320035, + 0.0026611562352627516, + 0.00988744106143713, + 0.0031309183686971664, + 0.0022187535651028156, + 0.0078003923408687115 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 1 + ], + "selected_reward": -0.0789872407913208 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot from a first-person perspective inside a heavily used off-road helmet with a scratched and dirt-splattered visor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment an abandoned industrial area featuring rust and concrete surroundings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a jagged shard of concrete with a twisted piece of rebar hurtle into the frame from the right side and deflect with a metallic clang?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene include a violent shudder and chaotic smear of textures immediately following the impact?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style raw and gritty with a desaturated palette of cold concrete grays and oxidized rust browns under harsh, high-contrast 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": [ + -0.020585693418979645 + ], + "opposite_logprob_tokens": [ + -3.895585775375366 + ], + "logprob_reward": -0.020585693418979645, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020585693418979645, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2829705476760864, + 0.0905366837978363, + 0.1671367585659027, + 0.04545418918132782, + 0.044123537838459015, + 0.11706162244081497, + 0.1134062111377716, + 0.05597289651632309, + 0.028753137215971947, + 0.05458441376686096 + ], + "credit_map": [ + 0.029515571892261505, + 0.03412206843495369, + 0.02489349991083145, + 0.0356774777173996, + 0.033239711076021194, + 0.036244627088308334, + 0.060457512736320496, + 0.02882007695734501, + 0.01339871995151043, + 0.01604720763862133, + 0.011137220077216625, + 0.014182781800627708, + 0.012452066875994205, + 0.006860456895083189, + 0.009639122523367405, + 0.00681910989806056, + 0.022533146664500237, + 0.020767921581864357, + 0.01838097535073757, + 0.021828893572092056, + 0.020162824541330338, + 0.01551494188606739, + 0.029836468398571014, + 0.018111566081643105, + 0.007866737432777882, + 0.008176597766578197, + 0.005034166853874922, + 0.006157076917588711, + 0.005414504092186689, + 0.0034030298702418804, + 0.004587268456816673, + 0.004814804997295141, + 0.004885922186076641, + 0.0073702409863471985, + 0.006070542149245739, + 0.007733923848718405, + 0.003521422389894724, + 0.005450316239148378, + 0.003465068992227316, + 0.005626101046800613, + 0.014150919392704964, + 0.01026561763137579, + 0.012289167381823063, + 0.010815608315169811, + 0.027415020391345024, + 0.009951797313988209, + 0.00755680026486516, + 0.024616684764623642, + 0.024909427389502525, + 0.016184987500309944, + 0.013354143127799034, + 0.012614472769200802, + 0.014469155110418797, + 0.014598728157579899, + 0.008156437426805496, + 0.009118854999542236, + 0.006063773296773434, + 0.010663107968866825, + 0.005349029321223497, + 0.006573123391717672, + 0.006235054694116116, + 0.005432541016489267, + 0.006358090788125992, + 0.009298169985413551, + 0.0028510375414043665, + 0.003311422187834978, + 0.002857912564650178, + 0.005614152178168297, + 0.0036106170155107975, + 0.0016472275601699948, + 0.005795632489025593, + 0.0030651320703327656, + 0.005548339337110519, + 0.0067192609421908855, + 0.0047225733287632465, + 0.008026348426938057, + 0.009766307659447193, + 0.0054376800544559956, + 0.005922514013946056, + 0.008441385813057423 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 5 + ], + "selected_reward": -0.020585693418979645 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot from a first-person perspective inside a heavily used off-road helmet with a scratched and dirt-splattered visor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment an abandoned industrial area featuring rust and concrete surroundings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a jagged shard of concrete with a twisted piece of rebar hurtle into the frame from the right side and deflect with a metallic clang?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene include a violent shudder and chaotic smear of textures immediately following the impact?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style raw and gritty with a desaturated palette of cold concrete grays and oxidized rust browns under harsh, high-contrast 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": [ + -0.00978246983140707 + ], + "opposite_logprob_tokens": [ + -4.634782314300537 + ], + "logprob_reward": -0.00978246983140707, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.00978246983140707, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2438928335905075, + 0.09791098535060883, + 0.1453358680009842, + 0.073702372610569, + 0.05421886593103409, + 0.051511988043785095, + 0.20727886259555817, + 0.035654060542583466, + 0.046511922031641006, + 0.04398218169808388 + ], + "credit_map": [ + 0.030680108815431595, + 0.029100526124238968, + 0.02190869301557541, + 0.028890300542116165, + 0.029859056696295738, + 0.03144722431898117, + 0.04672343283891678, + 0.02528349496424198, + 0.01400406938046217, + 0.02060803398489952, + 0.011441653594374657, + 0.013466710224747658, + 0.012720989994704723, + 0.007830319926142693, + 0.009967404417693615, + 0.007871795445680618, + 0.018894167616963387, + 0.015941999852657318, + 0.016512984409928322, + 0.015358335338532925, + 0.01347177941352129, + 0.016399364918470383, + 0.03013279289007187, + 0.018624451011419296, + 0.01382377091795206, + 0.010091261938214302, + 0.010687494650483131, + 0.009912729263305664, + 0.008138326928019524, + 0.0071659698151052, + 0.005779177416115999, + 0.008103643544018269, + 0.006156213581562042, + 0.010833201929926872, + 0.0077186692506074905, + 0.009236415848135948, + 0.005632751155644655, + 0.00407111132517457, + 0.005888697225600481, + 0.004681806080043316, + 0.0061745657585561275, + 0.006304895970970392, + 0.007696440443396568, + 0.007645151112228632, + 0.004730324260890484, + 0.005811227019876242, + 0.0061545842327177525, + 0.0069947997108101845, + 0.04638350009918213, + 0.03281218186020851, + 0.015912789851427078, + 0.018552938476204872, + 0.02780897170305252, + 0.03401368111371994, + 0.013351297937333584, + 0.01844351552426815, + 0.0031380197033286095, + 0.005852613132447004, + 0.0033674854785203934, + 0.005339309573173523, + 0.003867913968861103, + 0.003103897674009204, + 0.007037903182208538, + 0.003946920391172171, + 0.005099070258438587, + 0.005493749864399433, + 0.005078576505184174, + 0.009045389480888844, + 0.006363052874803543, + 0.00329407537356019, + 0.00684339040890336, + 0.005294616799801588, + 0.004741390235722065, + 0.005308253224939108, + 0.0036398707889020443, + 0.004965620115399361, + 0.007538128178566694, + 0.007031964138150215, + 0.004634607117623091, + 0.006122348364442587 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 2 + ], + "selected_reward": -0.00978246983140707 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot from a first-person perspective inside a heavily used off-road helmet with a scratched and dirt-splattered visor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment an abandoned industrial area featuring rust and concrete surroundings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a jagged shard of concrete with a twisted piece of rebar hurtle into the frame from the right side and deflect with a metallic clang?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene include a violent shudder and chaotic smear of textures immediately following the impact?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style raw and gritty with a desaturated palette of cold concrete grays and oxidized rust browns under harsh, high-contrast 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": [ + -0.0160652007907629 + ], + "opposite_logprob_tokens": [ + -4.1410651206970215 + ], + "logprob_reward": -0.0160652007907629, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.0160652007907629, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24608027935028076, + 0.09865827113389969, + 0.07337632030248642, + 0.03548262268304825, + 0.0737079605460167, + 0.06148703768849373, + 0.2396099716424942, + 0.0455034077167511, + 0.05585797503590584, + 0.0702362060546875 + ], + "credit_map": [ + 0.03368096053600311, + 0.02914584055542946, + 0.022460799664258957, + 0.02742580696940422, + 0.02733502723276615, + 0.028172360733151436, + 0.04705623909831047, + 0.030803220346570015, + 0.016854137182235718, + 0.018773667514324188, + 0.014005868695676327, + 0.014975094236433506, + 0.010341363027691841, + 0.0063530560582876205, + 0.009339256212115288, + 0.008015838451683521, + 0.012706730514764786, + 0.011082380078732967, + 0.010249646380543709, + 0.012253749184310436, + 0.007396379951387644, + 0.006055861711502075, + 0.00889329332858324, + 0.004738271236419678, + 0.0068124630488455296, + 0.0052836379036307335, + 0.005131993442773819, + 0.006372502073645592, + 0.0033076833933591843, + 0.0022034686990082264, + 0.003397057531401515, + 0.002973814494907856, + 0.010498021729290485, + 0.00833031814545393, + 0.018368389457464218, + 0.013163306750357151, + 0.004681643098592758, + 0.006293342448771, + 0.004968096502125263, + 0.0074048419483006, + 0.005894423928111792, + 0.004223666153848171, + 0.01226991880685091, + 0.005487590096890926, + 0.01759534515440464, + 0.008332808502018452, + 0.003517907578498125, + 0.004165374208241701, + 0.051372230052948, + 0.03565436229109764, + 0.024536635726690292, + 0.02606934681534767, + 0.031134625896811485, + 0.035505156964063644, + 0.015877867117524147, + 0.019459735602140427, + 0.004830620717257261, + 0.007742604240775108, + 0.00392678938806057, + 0.0038716127164661884, + 0.008839723654091358, + 0.005772958975285292, + 0.0047401501797139645, + 0.005778944585472345, + 0.0069693694822490215, + 0.006095564924180508, + 0.003839725162833929, + 0.007400956936180592, + 0.007438841741532087, + 0.005301482044160366, + 0.011211122386157513, + 0.007600913755595684, + 0.006395892705768347, + 0.007275737822055817, + 0.005009759683161974, + 0.008672015741467476, + 0.015703996643424034, + 0.005123755428940058, + 0.007349930703639984, + 0.014705117791891098 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 1 + ], + "selected_reward": -0.0160652007907629 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "Visceral first-person perspective footage shot from the confines of a heavily used off-road helmet thrusts the viewer directly into the high-speed sequence. The helmet's visor frames the world in a circle, deeply scratched and splattered with dry, cracked dirt, partially obscuring the field of vision. Through the clearer sections of plexiglass, the rust and concrete surroundings transform into a disorienting motion blur, hinting at the immense speeds being driven through the abandoned industrial area. Suddenly, a jagged shard of concrete, with a twisted piece of rebar protruding from it, hurtles into the frame and violently deflects from the right side of the helmet with a deafening metallic clang. In immediate reaction to the impact, the entire field of vision shudders violently, a violent and disorienting jolt that momentarily plunges the scene into a chaotic smear of textures and colors. The entire scene is rendered in a raw, gritty industrial style, characterized by a desaturated palette of cold concrete grays and oxidized rust browns, all set under harsh, high-contrast lighting, creating an atmosphere of pure, physical tension and danger.", + "seed": 11572, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000026/group_14/index_2166_seed_11572_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": -0.018483055103570224, + "vlm_logprob": -0.018483055103570224, + "vlm_yes_no_margin": 4.5, + "vlm_token_credit": -0.018483055103570224, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot from a first-person perspective inside a heavily used off-road helmet with a scratched and dirt-splattered visor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment an abandoned industrial area featuring rust and concrete surroundings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a jagged shard of concrete with a twisted piece of rebar hurtle into the frame from the right side and deflect with a metallic clang?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene include a violent shudder and chaotic smear of textures immediately following the impact?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style raw and gritty with a desaturated palette of cold concrete grays and oxidized rust browns under harsh, high-contrast 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": [ + -0.009363191202282906 + ], + "opposite_logprob_tokens": [ + -5.509363174438477 + ], + "logprob_reward": -0.009363191202282906, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.009363191202282906, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12205079197883606, + 0.08574361354112625, + 0.04793347045779228, + 0.0455528162419796, + 0.06709195673465729, + 0.060666363686323166, + 0.03929351270198822, + 0.3645583987236023, + 0.08476553112268448, + 0.08234357833862305 + ], + "credit_map": [ + 0.015929045155644417, + 0.016634885221719742, + 0.020581398159265518, + 0.013411272317171097, + 0.011900827288627625, + 0.01743999868631363, + 0.010983369313180447, + 0.015169995836913586, + 0.006842317525297403, + 0.012904325500130653, + 0.016276007518172264, + 0.007749381009489298, + 0.013852132484316826, + 0.012822345830500126, + 0.00713203614577651, + 0.008165067061781883, + 0.005352819804102182, + 0.005343851167708635, + 0.005804934538900852, + 0.007043888326734304, + 0.0054761264473199844, + 0.005967320874333382, + 0.006687752902507782, + 0.006256780121475458, + 0.005255941767245531, + 0.004122012294828892, + 0.004444427322596312, + 0.006728955078870058, + 0.005520340520888567, + 0.00673557398840785, + 0.0074944100342690945, + 0.005251155234873295, + 0.005990046542137861, + 0.0076095773838460445, + 0.007605803664773703, + 0.006610002834349871, + 0.009802297689020634, + 0.01049430575221777, + 0.007997383363544941, + 0.010982539504766464, + 0.006126464810222387, + 0.01772543601691723, + 0.0031780872959643602, + 0.0035795136354863644, + 0.007171096745878458, + 0.009555796161293983, + 0.007973706349730492, + 0.005356263369321823, + 0.003146115457639098, + 0.006981871090829372, + 0.004247128497809172, + 0.004741986282169819, + 0.004174640867859125, + 0.006546138785779476, + 0.006870304234325886, + 0.0025853256229311228, + 0.015592064708471298, + 0.024009138345718384, + 0.13239531219005585, + 0.03407498449087143, + 0.021816259250044823, + 0.0924956202507019, + 0.0303829163312912, + 0.013792099431157112, + 0.009717832319438457, + 0.009689616970717907, + 0.008513402193784714, + 0.0108509985730052, + 0.010593214072287083, + 0.008679638616740704, + 0.011680188588798046, + 0.015040633268654346, + 0.008164798840880394, + 0.00846589170396328, + 0.008451906964182854, + 0.01238161139190197, + 0.007169533520936966, + 0.008706502616405487, + 0.01258588582277298, + 0.016417445614933968 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 1 + ], + "selected_reward": -0.009363191202282906 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot from a first-person perspective inside a heavily used off-road helmet with a scratched and dirt-splattered visor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment an abandoned industrial area featuring rust and concrete surroundings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a jagged shard of concrete with a twisted piece of rebar hurtle into the frame from the right side and deflect with a metallic clang?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene include a violent shudder and chaotic smear of textures immediately following the impact?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style raw and gritty with a desaturated palette of cold concrete grays and oxidized rust browns under harsh, high-contrast 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": [ + -0.011077099479734898 + ], + "opposite_logprob_tokens": [ + -4.511076927185059 + ], + "logprob_reward": -0.011077099479734898, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011077099479734898, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.062067482620477676, + 0.041214365512132645, + 0.03466019034385681, + 0.1597340703010559, + 0.08958575874567032, + 0.21684467792510986, + 0.0802401453256607, + 0.21000345051288605, + 0.08672798424959183, + 0.018921835348010063 + ], + "credit_map": [ + 0.006879803724586964, + 0.009376387111842632, + 0.009971851482987404, + 0.00833884533494711, + 0.005690805148333311, + 0.009287054650485516, + 0.00556426914408803, + 0.006958462297916412, + 0.0033451570197939873, + 0.003764896420761943, + 0.010860580019652843, + 0.006407327484339476, + 0.004259839188307524, + 0.003913146443665028, + 0.004542302340269089, + 0.004121117759495974, + 0.003608672646805644, + 0.0050111678428947926, + 0.004192633554339409, + 0.004986901301890612, + 0.004095613956451416, + 0.004076201934367418, + 0.004494954831898212, + 0.004194042645394802, + 0.020773036405444145, + 0.016547812148928642, + 0.041629187762737274, + 0.024716326966881752, + 0.010731485672295094, + 0.012892985716462135, + 0.01757816970348358, + 0.014865049161016941, + 0.008977709338068962, + 0.010510069318115711, + 0.009643949568271637, + 0.010535094887018204, + 0.013533870689570904, + 0.01283713337033987, + 0.009058662690222263, + 0.014489264227449894, + 0.02501213178038597, + 0.080101378262043, + 0.009538663551211357, + 0.010990620590746403, + 0.018214304000139236, + 0.049543317407369614, + 0.005694581661373377, + 0.01774967834353447, + 0.0032372360583394766, + 0.013313552364706993, + 0.0061586652882397175, + 0.007656108122318983, + 0.006162789650261402, + 0.01661607436835766, + 0.02269497513771057, + 0.004400738049298525, + 0.008832267485558987, + 0.01203736662864685, + 0.07980196177959442, + 0.019791943952441216, + 0.00912278238683939, + 0.05466000363230705, + 0.01796748675405979, + 0.007789621129631996, + 0.015005647204816341, + 0.012879402376711369, + 0.009590482339262962, + 0.013181095942854881, + 0.0073529514484107494, + 0.00687048677355051, + 0.010589538142085075, + 0.011258378624916077, + 0.0027454362716525793, + 0.002248553093522787, + 0.002001037122681737, + 0.0029578725807368755, + 0.0017170346109196544, + 0.0016137862112373114, + 0.002997002098709345, + 0.0026411141734570265 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 7, + 3 + ], + "selected_reward": -0.011077099479734898 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot from a first-person perspective inside a heavily used off-road helmet with a scratched and dirt-splattered visor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment an abandoned industrial area featuring rust and concrete surroundings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a jagged shard of concrete with a twisted piece of rebar hurtle into the frame from the right side and deflect with a metallic clang?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene include a violent shudder and chaotic smear of textures immediately following the impact?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style raw and gritty with a desaturated palette of cold concrete grays and oxidized rust browns under harsh, high-contrast 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": [ + -0.054957568645477295 + ], + "opposite_logprob_tokens": [ + -2.929957628250122 + ], + "logprob_reward": -0.054957568645477295, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.054957568645477295, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2059796005487442, + 0.09219378978013992, + 0.07129085063934326, + 0.04963407665491104, + 0.07222652435302734, + 0.06626918166875839, + 0.14658798277378082, + 0.16328829526901245, + 0.07391608506441116, + 0.058613646775484085 + ], + "credit_map": [ + 0.024120325222611427, + 0.032128285616636276, + 0.023229124024510384, + 0.020195046439766884, + 0.025556767359375954, + 0.022265078499913216, + 0.024272896349430084, + 0.034212078899145126, + 0.007494708988815546, + 0.010223926976323128, + 0.020730698481202126, + 0.00956487376242876, + 0.01437478605657816, + 0.013090305030345917, + 0.007980719208717346, + 0.008733765222132206, + 0.006645281799137592, + 0.009158587083220482, + 0.008751649409532547, + 0.009848206304013729, + 0.007768182083964348, + 0.008816677145659924, + 0.010343058966100216, + 0.0099592050537467, + 0.0042229085229337215, + 0.00383922946639359, + 0.006461634300649166, + 0.007187452632933855, + 0.004895919468253851, + 0.006458573043346405, + 0.007815518416464329, + 0.008752843365073204, + 0.00858002994209528, + 0.011759686283767223, + 0.006138780154287815, + 0.005505775101482868, + 0.013010210357606411, + 0.009754027239978313, + 0.008500739932060242, + 0.00897727720439434, + 0.006793245207518339, + 0.022141356021165848, + 0.0045986175537109375, + 0.0038010021671652794, + 0.006210187915712595, + 0.01365671493113041, + 0.003527470864355564, + 0.005540585611015558, + 0.006643100641667843, + 0.026198038831353188, + 0.010554338805377483, + 0.011644301936030388, + 0.011159088462591171, + 0.030145199969410896, + 0.04117734730243683, + 0.009066569618880749, + 0.008539820089936256, + 0.01336618885397911, + 0.05874365195631981, + 0.01634989306330681, + 0.006412987131625414, + 0.035846877843141556, + 0.018798217177391052, + 0.0052306619472801685, + 0.007977308705449104, + 0.007760556880384684, + 0.0074096606113016605, + 0.010126295499503613, + 0.009769883006811142, + 0.00814720056951046, + 0.01083611510694027, + 0.011889061890542507, + 0.005458596162497997, + 0.00446314224973321, + 0.005522188730537891, + 0.0066167376935482025, + 0.00541991600766778, + 0.007173295598477125, + 0.01046693604439497, + 0.013492831028997898 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 6 + ], + "selected_reward": -0.054957568645477295 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot from a first-person perspective inside a heavily used off-road helmet with a scratched and dirt-splattered visor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment an abandoned industrial area featuring rust and concrete surroundings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a jagged shard of concrete with a twisted piece of rebar hurtle into the frame from the right side and deflect with a metallic clang?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene include a violent shudder and chaotic smear of textures immediately following the impact?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style raw and gritty with a desaturated palette of cold concrete grays and oxidized rust browns under harsh, high-contrast 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": [ + -0.005946918856352568 + ], + "opposite_logprob_tokens": [ + -5.130947113037109 + ], + "logprob_reward": -0.005946918856352568, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005946918856352568, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08158554136753082, + 0.02498197741806507, + 0.0290488563477993, + 0.024968033656477928, + 0.012491550296545029, + 0.09563817828893661, + 0.032266389578580856, + 0.6661909818649292, + 0.014992267824709415, + 0.017836248502135277 + ], + "credit_map": [ + 0.010113371536135674, + 0.01534632220864296, + 0.008812651969492435, + 0.010398905724287033, + 0.010981633327901363, + 0.008945564739406109, + 0.008111692033708096, + 0.008875403553247452, + 0.0019403152400627732, + 0.002958051860332489, + 0.0032629447523504496, + 0.0025170231238007545, + 0.004626414272934198, + 0.004076972138136625, + 0.002247365890070796, + 0.0033528893254697323, + 0.0028563786763697863, + 0.003459325060248375, + 0.0036762976087629795, + 0.004070660565048456, + 0.0028331493958830833, + 0.0035871868021786213, + 0.0038235762622207403, + 0.004742281045764685, + 0.0024264035746455193, + 0.0021766689606010914, + 0.0051038069650530815, + 0.00263697630725801, + 0.0022136897314339876, + 0.0024870040360838175, + 0.0037280265241861343, + 0.00419545779004693, + 0.0013483630027621984, + 0.0016030387487262487, + 0.0011384745594114065, + 0.0011222853790968657, + 0.0020196980331093073, + 0.0019403407350182533, + 0.0018962942995131016, + 0.001423054258339107, + 0.011416279710829258, + 0.03661516681313515, + 0.0023297504521906376, + 0.005004481412470341, + 0.007941681891679764, + 0.022597581148147583, + 0.0020707694347947836, + 0.007662462070584297, + 0.001753964927047491, + 0.006234719417989254, + 0.002218574983999133, + 0.003353698644787073, + 0.0031701254192739725, + 0.005638335831463337, + 0.0072811986319720745, + 0.002615772420540452, + 0.027441777288913727, + 0.04124564677476883, + 0.2488103061914444, + 0.06150412932038307, + 0.03418080136179924, + 0.17003929615020752, + 0.057377129793167114, + 0.025591867044568062, + 0.0020759599283337593, + 0.002115898299962282, + 0.001648886944167316, + 0.003004518337547779, + 0.0016964843962341547, + 0.001471724477596581, + 0.0015247680712491274, + 0.0014540278352797031, + 0.0033789388835430145, + 0.002947645727545023, + 0.002020368119701743, + 0.0037575282622128725, + 0.0010999712394550443, + 0.0013394528068602085, + 0.001452603843063116, + 0.0018397391540929675 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 5, + 0 + ], + "selected_reward": -0.005946918856352568 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot from a first-person perspective inside a heavily used off-road helmet with a scratched and dirt-splattered visor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment an abandoned industrial area featuring rust and concrete surroundings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a jagged shard of concrete with a twisted piece of rebar hurtle into the frame from the right side and deflect with a metallic clang?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene include a violent shudder and chaotic smear of textures immediately following the impact?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style raw and gritty with a desaturated palette of cold concrete grays and oxidized rust browns under harsh, high-contrast 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": [ + -0.011070497334003448 + ], + "opposite_logprob_tokens": [ + -4.511070728302002 + ], + "logprob_reward": -0.011070497334003448, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011070497334003448, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11910982429981232, + 0.059207484126091, + 0.058009568601846695, + 0.1062433123588562, + 0.05917229503393173, + 0.06511449813842773, + 0.13925683498382568, + 0.25417283177375793, + 0.08900442719459534, + 0.05070892721414566 + ], + "credit_map": [ + 0.013199050910770893, + 0.0190887488424778, + 0.02131367102265358, + 0.013984539546072483, + 0.011057133786380291, + 0.016447940841317177, + 0.008972585201263428, + 0.015046156011521816, + 0.00696837343275547, + 0.013180295936763287, + 0.010233549401164055, + 0.006791780237108469, + 0.005138194654136896, + 0.0042259423062205315, + 0.005508475936949253, + 0.007160876411944628, + 0.006599684711545706, + 0.00875173881649971, + 0.00703946128487587, + 0.00761756394058466, + 0.005840965546667576, + 0.004830211866647005, + 0.009158915840089321, + 0.00817103311419487, + 0.0077202762477099895, + 0.008981339633464813, + 0.01289290003478527, + 0.014212718233466148, + 0.013054016046226025, + 0.017365606501698494, + 0.018053539097309113, + 0.013962921686470509, + 0.007477139588445425, + 0.005751118529587984, + 0.007220379542559385, + 0.007804215420037508, + 0.008770276792347431, + 0.007076505105942488, + 0.005991961341351271, + 0.009080703370273113, + 0.007585024926811457, + 0.022794276475906372, + 0.0031269777100533247, + 0.0034264593850821257, + 0.005807670298963785, + 0.013594383373856544, + 0.0035736358258873224, + 0.0052060759626328945, + 0.006774463225156069, + 0.023552099242806435, + 0.010392505675554276, + 0.014613931998610497, + 0.013558369129896164, + 0.02634347602725029, + 0.036889366805553436, + 0.00713263638317585, + 0.011171435937285423, + 0.01918482594192028, + 0.08601196855306625, + 0.02607390470802784, + 0.01705743931233883, + 0.06188422068953514, + 0.020746326074004173, + 0.01204272173345089, + 0.013674277812242508, + 0.01368176843971014, + 0.009446049109101295, + 0.01223209872841835, + 0.008595200255513191, + 0.006504422053694725, + 0.01111123338341713, + 0.013759384863078594, + 0.0071788569912314415, + 0.00825829990208149, + 0.005308421328663826, + 0.006567026488482952, + 0.00457646744325757, + 0.005196269601583481, + 0.007132911588996649, + 0.006490679457783699 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 6, + 0 + ], + "selected_reward": -0.011070497334003448 + } + ], + "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/0000026/group_15/index_tensor([1868, 1868])_group_15_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000026/group_15/index_tensor([1868, 1868])_group_15_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..5dffa946b274523f5f4dd0cdd186d29ac805fbbc --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000026/group_15/index_tensor([1868, 1868])_group_15_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "stunning, close-up, static shot, focused at near-eye level, captures the face of a sleeping young woman, the entire scene steeped in the opulence and softness of Rococo style. Soft candlelight intertwines with the warm afternoon sun filtering through silk curtains, casting a warm, three-dimensional light on the girl's delicate, fair skin. Her pale pink eyeshadow adds a touch of sweetness. Initially, the subtle flutter of her thick, butterfly-like, dark brown eyelashes shatters the stillness of the scene. Then, her eyes snap open, her clear, light brown pupils dilating with sudden surprise. The camera captures, with pinpoint accuracy, the perfect reflection of a delicate, multi-layered cake in the foreground, adorned with sugar roses, fresh berries, and gold foil. Even the sheen of the buttercream is clearly visible. The scene, dominated by a palette of ivory, pale pink, and gold, creates a dramatic atmosphere of innocence, dreaminess, and sudden awakening, like a magnificent painting in motion.", + "seed": 8461, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000026/group_15/index_1868_seed_8461_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": -0.06628579776734114, + "vlm_logprob": -0.06628579776734114, + "vlm_yes_no_margin": 3.8, + "vlm_token_credit": -0.06628579776734114, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 focused at near-eye level showing the face of a sleeping young woman?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene illuminated by soft candlelight and warm afternoon sun filtering through silk curtains?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the woman have pale pink eyeshadow and thick dark brown eyelashes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the woman's eyes snap open with dilated light brown pupils showing sudden surprise?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a multi-layered cake with sugar roses, fresh berries, and gold foil visible in the foreground?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012005571275949478 + ], + "opposite_logprob_tokens": [ + -4.637005805969238 + ], + "logprob_reward": -0.012005571275949478, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.012005571275949478, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0583939328789711, + 0.022231299430131912, + 0.02404261939227581, + 0.06259649991989136, + 0.013120311312377453, + 0.08076924085617065, + 0.04020427539944649, + 0.6127977967262268, + 0.049026187509298325, + 0.03681781515479088 + ], + "credit_map": [ + 0.006561536807566881, + 0.0056313639506697655, + 0.003809757297858596, + 0.004420209676027298, + 0.006351571064442396, + 0.007971669547259808, + 0.01244040485471487, + 0.011207425966858864, + 0.002512585138902068, + 0.0031277844682335854, + 0.0024922778829932213, + 0.0028295007068663836, + 0.002681946149095893, + 0.003392283571884036, + 0.0024648047983646393, + 0.0027301213704049587, + 0.002229322912171483, + 0.0026174781378358603, + 0.0023282086476683617, + 0.0029593699146062136, + 0.002919440856203437, + 0.003951871767640114, + 0.00349962105974555, + 0.003537307260558009, + 0.007272885646671057, + 0.010617642663419247, + 0.007611383683979511, + 0.00817770604044199, + 0.005618801340460777, + 0.009375284425914288, + 0.00599160511046648, + 0.007931197993457317, + 0.0012528715888038278, + 0.001273186644539237, + 0.0014517911477014422, + 0.0024783031549304724, + 0.0014624459436163306, + 0.0020497699733823538, + 0.0014593638479709625, + 0.0016925801755860448, + 0.007889027707278728, + 0.012785975821316242, + 0.007770510856062174, + 0.009419827722012997, + 0.009512820281088352, + 0.015069345012307167, + 0.008120395243167877, + 0.01020134799182415, + 0.004643967840820551, + 0.007086764555424452, + 0.004632148891687393, + 0.004923044703900814, + 0.004264796152710915, + 0.005710161291062832, + 0.004236271604895592, + 0.0047071222215890884, + 0.01502437423914671, + 0.009555640630424023, + 0.01198752224445343, + 0.02066039852797985, + 0.1203809529542923, + 0.0614301897585392, + 0.2231527864933014, + 0.1506059616804123, + 0.003573380410671234, + 0.0037619774229824543, + 0.00504289660602808, + 0.004224656615406275, + 0.007174069993197918, + 0.008200154639780521, + 0.01037247572094202, + 0.00667657982558012, + 0.0034193648025393486, + 0.0016291133360937238, + 0.0036264993250370026, + 0.0024958483409136534, + 0.005708626005798578, + 0.004234109073877335, + 0.009340397082269192, + 0.006363862194120884 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 5, + 3 + ], + "selected_reward": -0.012005571275949478 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 focused at near-eye level showing the face of a sleeping young woman?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene illuminated by soft candlelight and warm afternoon sun filtering through silk curtains?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the woman have pale pink eyeshadow and thick dark brown eyelashes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the woman's eyes snap open with dilated light brown pupils showing sudden surprise?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a multi-layered cake with sugar roses, fresh berries, and gold foil visible in the foreground?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2815154194831848 + ], + "opposite_logprob_tokens": [ + -1.40651535987854 + ], + "logprob_reward": -0.2815154194831848, + "yes_no_margin_reward": 1.125, + "token_credit_reward": -0.2815154194831848, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17947368323802948, + 0.026471789926290512, + 0.02800908125936985, + 0.03307776898145676, + 0.0435599684715271, + 0.16000302135944366, + 0.029659757390618324, + 0.44707220792770386, + 0.014490680769085884, + 0.038182031363248825 + ], + "credit_map": [ + 0.010725194588303566, + 0.00508115766569972, + 0.008862370625138283, + 0.010913696140050888, + 0.022387776523828506, + 0.023503486067056656, + 0.052055176347494125, + 0.04594481736421585, + 0.0026818551123142242, + 0.0045409114100039005, + 0.004172545857727528, + 0.004440034739673138, + 0.0024602068588137627, + 0.0031161096412688494, + 0.0023409458808600903, + 0.002719178795814514, + 0.003035909729078412, + 0.00423654867336154, + 0.003726966679096222, + 0.003824594896286726, + 0.0028310136403888464, + 0.004328134469687939, + 0.0027011935599148273, + 0.003324720077216625, + 0.0038206048775464296, + 0.005046093836426735, + 0.004132958594709635, + 0.0051685902290046215, + 0.002850067103281617, + 0.004478461109101772, + 0.0034520842600613832, + 0.004128910601139069, + 0.004631466697901487, + 0.004821879789233208, + 0.0072123766876757145, + 0.006584860850125551, + 0.005505524575710297, + 0.005344346631318331, + 0.005507979542016983, + 0.003951534163206816, + 0.018940502777695656, + 0.028832223266363144, + 0.01693163439631462, + 0.017185255885124207, + 0.014695868827402592, + 0.02582314983010292, + 0.01745731197297573, + 0.02013707347214222, + 0.003935113549232483, + 0.005145041272044182, + 0.0031186824198812246, + 0.003244120394811034, + 0.0031799126882106066, + 0.0033874628134071827, + 0.004364275839179754, + 0.0032851477153599262, + 0.0108456676825881, + 0.00620305398479104, + 0.008907962590456009, + 0.014754514209926128, + 0.08782826364040375, + 0.04485251009464264, + 0.1632383167743683, + 0.1104419007897377, + 0.0011158438865095377, + 0.0010239823022857308, + 0.0012157553574070334, + 0.0014024949632585049, + 0.0014847165439277887, + 0.003498508594930172, + 0.002491011517122388, + 0.0022583678364753723, + 0.0021694230381399393, + 0.0016755627002567053, + 0.0014506755396723747, + 0.0025981361977756023, + 0.006491127423942089, + 0.004052427597343922, + 0.011627609841525555, + 0.00811706855893135 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 5 + ], + "selected_reward": -0.2815154194831848 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 focused at near-eye level showing the face of a sleeping young woman?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene illuminated by soft candlelight and warm afternoon sun filtering through silk curtains?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the woman have pale pink eyeshadow and thick dark brown eyelashes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the woman's eyes snap open with dilated light brown pupils showing sudden surprise?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a multi-layered cake with sugar roses, fresh berries, and gold foil visible in the foreground?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016075288876891136 + ], + "opposite_logprob_tokens": [ + -4.141075134277344 + ], + "logprob_reward": -0.016075288876891136, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016075288876891136, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11853500455617905, + 0.053387559950351715, + 0.03882965072989464, + 0.016259629279375076, + 0.07926583290100098, + 0.02193339727818966, + 0.02070552110671997, + 0.5935069918632507, + 0.018428970128297806, + 0.03914738819003105 + ], + "credit_map": [ + 0.007334467023611069, + 0.005301378201693296, + 0.006204135250300169, + 0.009769940748810768, + 0.0144228870049119, + 0.01581851951777935, + 0.031417835503816605, + 0.02826583944261074, + 0.004796176683157682, + 0.007328047417104244, + 0.003953638952225447, + 0.0044562541879713535, + 0.007090420927852392, + 0.008466481231153011, + 0.007804086431860924, + 0.009492452256381512, + 0.00500872265547514, + 0.005137088242918253, + 0.0034493771381676197, + 0.0035331586841493845, + 0.004367917310446501, + 0.008145740255713463, + 0.00380800967104733, + 0.0053796349093317986, + 0.0018009094055742025, + 0.0026318905875086784, + 0.0014329111436381936, + 0.0015521285822615027, + 0.0020500144455581903, + 0.0031131599098443985, + 0.0017242425819858909, + 0.001954371575266123, + 0.009298937395215034, + 0.01355280727148056, + 0.01030812505632639, + 0.009665891528129578, + 0.006575548090040684, + 0.012188400141894817, + 0.007426135241985321, + 0.010249986313283443, + 0.0029328118544071913, + 0.00379281141795218, + 0.0017746926750987768, + 0.0021649799309670925, + 0.0018811671761795878, + 0.0037784434389322996, + 0.002447105711326003, + 0.0031613842584192753, + 0.002585574984550476, + 0.003968958742916584, + 0.002780589973554015, + 0.0021859449334442616, + 0.002156925620511174, + 0.002152812434360385, + 0.0028268322348594666, + 0.0020478833466768265, + 0.014945320785045624, + 0.009839345701038837, + 0.01201533805578947, + 0.02055862918496132, + 0.11564759910106659, + 0.06008526682853699, + 0.21454045176506042, + 0.14587508141994476, + 0.0013153282925486565, + 0.0015266056871041656, + 0.0013539382489398122, + 0.0014400199288502336, + 0.002964707324281335, + 0.0021798836532980204, + 0.004689858760684729, + 0.002958628349006176, + 0.0034134939778596163, + 0.002787655685096979, + 0.004239345900714397, + 0.0028946050442755222, + 0.006050885654985905, + 0.004401165060698986, + 0.009355835616588593, + 0.006004400085657835 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 4 + ], + "selected_reward": -0.016075288876891136 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 focused at near-eye level showing the face of a sleeping young woman?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene illuminated by soft candlelight and warm afternoon sun filtering through silk curtains?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the woman have pale pink eyeshadow and thick dark brown eyelashes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the woman's eyes snap open with dilated light brown pupils showing sudden surprise?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a multi-layered cake with sugar roses, fresh berries, and gold foil visible in the foreground?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01420411467552185 + ], + "opposite_logprob_tokens": [ + -4.264204025268555 + ], + "logprob_reward": -0.01420411467552185, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.01420411467552185, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08766254037618637, + 0.015172159299254417, + 0.01757788471877575, + 0.023622464388608932, + 0.017282191663980484, + 0.07367076724767685, + 0.04282021149992943, + 0.6328632831573486, + 0.0644979402422905, + 0.024830494076013565 + ], + "credit_map": [ + 0.003941485658288002, + 0.0030377614311873913, + 0.0028615747578442097, + 0.003671898739412427, + 0.011677918955683708, + 0.010719545185565948, + 0.02712216041982174, + 0.024630190804600716, + 0.0014049977762624621, + 0.0018934572581201792, + 0.0010874997824430466, + 0.0015526644419878721, + 0.0015179964248090982, + 0.002513241721317172, + 0.0023583367001265287, + 0.002843965543434024, + 0.0017030943417921662, + 0.0025036160368472338, + 0.0025250304024666548, + 0.0030329623259603977, + 0.001712515950202942, + 0.002405876526609063, + 0.0016565172700211406, + 0.002038272563368082, + 0.002499106340110302, + 0.003036248032003641, + 0.002813509199768305, + 0.0033768152352422476, + 0.0024510896764695644, + 0.0036550017539411783, + 0.0028865837957710028, + 0.002904109889641404, + 0.0015981286996975541, + 0.0019454492721706629, + 0.0032152836211025715, + 0.003206669120118022, + 0.001554963761009276, + 0.002232783008366823, + 0.001937762019224465, + 0.0015911527443677187, + 0.00818116869777441, + 0.013869457878172398, + 0.010444124229252338, + 0.00961984507739544, + 0.006185285747051239, + 0.009922435507178307, + 0.007417311426252127, + 0.008031134493649006, + 0.003662001807242632, + 0.0055976491421461105, + 0.0051222676411271095, + 0.004671781323850155, + 0.004947868641465902, + 0.0047766040079295635, + 0.008183767087757587, + 0.005858271848410368, + 0.01630101539194584, + 0.010495021007955074, + 0.014934001490473747, + 0.021526021882891655, + 0.123872309923172, + 0.0639696717262268, + 0.22797924280166626, + 0.1537860482931137, + 0.0043385811150074005, + 0.005870314780622721, + 0.006425630301237106, + 0.0070757512003183365, + 0.008103491738438606, + 0.008752724155783653, + 0.013627332635223866, + 0.010304118506610394, + 0.002774710301309824, + 0.0028049990069121122, + 0.0046553416177630424, + 0.0029143174178898335, + 0.0027899884153157473, + 0.0026406587567180395, + 0.003661497961729765, + 0.0025889805983752012 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 5 + ], + "selected_reward": -0.01420411467552185 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up focused at near-eye level showing the face of a sleeping young woman?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene illuminated by soft candlelight and warm afternoon sun filtering through silk curtains?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the woman have pale pink eyeshadow and thick dark brown eyelashes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the woman's eyes snap open with dilated light brown pupils showing sudden surprise?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a multi-layered cake with sugar roses, fresh berries, and gold foil visible in the foreground?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007628594525158405 + ], + "opposite_logprob_tokens": [ + -4.882628440856934 + ], + "logprob_reward": -0.007628594525158405, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007628594525158405, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08017206937074661, + 0.05576910451054573, + 0.023711366578936577, + 0.022117458283901215, + 0.08075171709060669, + 0.13088928163051605, + 0.02268671616911888, + 0.5092238187789917, + 0.03007938154041767, + 0.04459904879331589 + ], + "credit_map": [ + 0.00788860209286213, + 0.005242963321506977, + 0.010081404820084572, + 0.01129953097552061, + 0.00858190469443798, + 0.012530962005257607, + 0.01254774909466505, + 0.011998949572443962, + 0.005748963449150324, + 0.009145920164883137, + 0.005131957121193409, + 0.0089205177500844, + 0.006543333176523447, + 0.00805389229208231, + 0.005295028444379568, + 0.006929493974894285, + 0.0028026998043060303, + 0.0036384365521371365, + 0.0029941571410745382, + 0.0030158120207488537, + 0.0021927175112068653, + 0.0038074476178735495, + 0.002255417639389634, + 0.0030046801548451185, + 0.003044783603399992, + 0.003673225175589323, + 0.0018477421253919601, + 0.0030803591944277287, + 0.0019878868479281664, + 0.003253314644098282, + 0.0025031394325196743, + 0.002727008890360594, + 0.009589950554072857, + 0.0123534444719553, + 0.013007066212594509, + 0.011922860518097878, + 0.006852213758975267, + 0.010344772599637508, + 0.007616318296641111, + 0.009065089747309685, + 0.014175739139318466, + 0.02384989522397518, + 0.009965005330741405, + 0.015033402480185032, + 0.014432430267333984, + 0.021654950454831123, + 0.01450091041624546, + 0.017276940867304802, + 0.002161437412723899, + 0.0035486717242747545, + 0.0023180253338068724, + 0.0030664410442113876, + 0.0025533949956297874, + 0.002981428522616625, + 0.0034960294142365456, + 0.002561286324635148, + 0.012426255270838737, + 0.007845671847462654, + 0.009671342559158802, + 0.016847627237439156, + 0.1004844531416893, + 0.05254428833723068, + 0.18473735451698303, + 0.12466683238744736, + 0.0037278702948242426, + 0.0025809097569435835, + 0.00486782006919384, + 0.004212273750454187, + 0.0026541363913565874, + 0.005239931866526604, + 0.0033321448136121035, + 0.003464294597506523, + 0.003138981992378831, + 0.001701122266240418, + 0.0037624724209308624, + 0.003576688002794981, + 0.0067189582623541355, + 0.0050155301578342915, + 0.011920853517949581, + 0.008764447644352913 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 5, + 4 + ], + "selected_reward": -0.007628594525158405 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "stunning, close-up, static shot, focused at near-eye level, captures the face of a sleeping young woman, the entire scene steeped in the opulence and softness of Rococo style. Soft candlelight intertwines with the warm afternoon sun filtering through silk curtains, casting a warm, three-dimensional light on the girl's delicate, fair skin. Her pale pink eyeshadow adds a touch of sweetness. Initially, the subtle flutter of her thick, butterfly-like, dark brown eyelashes shatters the stillness of the scene. Then, her eyes snap open, her clear, light brown pupils dilating with sudden surprise. The camera captures, with pinpoint accuracy, the perfect reflection of a delicate, multi-layered cake in the foreground, adorned with sugar roses, fresh berries, and gold foil. Even the sheen of the buttercream is clearly visible. The scene, dominated by a palette of ivory, pale pink, and gold, creates a dramatic atmosphere of innocence, dreaminess, and sudden awakening, like a magnificent painting in motion.", + "seed": 8462, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000026/group_15/index_1868_seed_8462_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": -0.42086381241679194, + "vlm_logprob": -0.42086381241679194, + "vlm_yes_no_margin": 1.75, + "vlm_token_credit": -0.42086381241679194, + "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 a close-up focused at near-eye level showing the face of a sleeping young woman?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene illuminated by soft candlelight and warm afternoon sun filtering through silk curtains?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the woman have pale pink eyeshadow and thick dark brown eyelashes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the woman's eyes snap open with dilated light brown pupils showing sudden surprise?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a multi-layered cake with sugar roses, fresh berries, and gold foil visible in the foreground?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.09054109454154968 + ], + "opposite_logprob_tokens": [ + -2.465541124343872 + ], + "logprob_reward": -0.09054109454154968, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.09054109454154968, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1242145448923111, + 0.05735548213124275, + 0.10026434063911438, + 0.06929756700992584, + 0.04444720596075058, + 0.02962765470147133, + 0.07695067673921585, + 0.03417536988854408, + 0.37564903497695923, + 0.08801811188459396 + ], + "credit_map": [ + 0.017613939940929413, + 0.016281211748719215, + 0.016570251435041428, + 0.01980341039597988, + 0.011339705437421799, + 0.0112252626568079, + 0.020165348425507545, + 0.011215397156774998, + 0.00742111261934042, + 0.00474154856055975, + 0.015775270760059357, + 0.010255690664052963, + 0.004904947709292173, + 0.003966513555496931, + 0.005543338134884834, + 0.004747055936604738, + 0.011199732311069965, + 0.00712421303614974, + 0.024065803736448288, + 0.023810353130102158, + 0.007203463930636644, + 0.00666709104552865, + 0.01064731553196907, + 0.00954636000096798, + 0.006425722502171993, + 0.006885414943099022, + 0.009899935685098171, + 0.019144758582115173, + 0.007375269662588835, + 0.005277941469103098, + 0.00922297965735197, + 0.0050655389204621315, + 0.005561082623898983, + 0.00458867009729147, + 0.0069963885471224785, + 0.008798105642199516, + 0.004122443031519651, + 0.004318710882216692, + 0.0057569947093725204, + 0.004304802510887384, + 0.003573228605091572, + 0.003121586050838232, + 0.004577349405735731, + 0.003955760505050421, + 0.0041787223890423775, + 0.0034925173968076706, + 0.0036594606935977936, + 0.003069025930017233, + 0.01085713878273964, + 0.007125722244381905, + 0.019325468689203262, + 0.01326205488294363, + 0.006942869629710913, + 0.005794124212116003, + 0.006960389669984579, + 0.0066828979179263115, + 0.003842917736619711, + 0.003092679427936673, + 0.0051329731941223145, + 0.008239619433879852, + 0.003459760919213295, + 0.0028477711603045464, + 0.004478225484490395, + 0.0030814195051789284, + 0.04601696506142616, + 0.041366178542375565, + 0.05927916616201401, + 0.05008162930607796, + 0.06611829996109009, + 0.041593037545681, + 0.0367572158575058, + 0.03443652763962746, + 0.008874671533703804, + 0.007764324080199003, + 0.018239954486489296, + 0.016484124585986137, + 0.008759657852351665, + 0.009565652348101139, + 0.008396326564252377, + 0.009933390654623508 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 2 + ], + "selected_reward": -0.09054109454154968 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 focused at near-eye level showing the face of a sleeping young woman?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene illuminated by soft candlelight and warm afternoon sun filtering through silk curtains?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the woman have pale pink eyeshadow and thick dark brown eyelashes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the woman's eyes snap open with dilated light brown pupils showing sudden surprise?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a multi-layered cake with sugar roses, fresh berries, and gold foil visible in the foreground?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"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": [ + -1.0541412830352783 + ], + "opposite_logprob_tokens": [ + -0.4291412830352783 + ], + "logprob_reward": -1.0541412830352783, + "yes_no_margin_reward": -0.625, + "token_credit_reward": -1.0541412830352783, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.1634828746318817, + 0.14795471727848053, + 0.19696614146232605, + 0.030784592032432556, + 0.06769798696041107, + 0.027200648561120033, + 0.0304096806794405, + 0.045472126454114914, + 0.14498929679393768, + 0.14504192769527435 + ], + "credit_map": [ + 0.025175044313073158, + 0.020228229463100433, + 0.027207035571336746, + 0.023813018575310707, + 0.017542608082294464, + 0.015609608963131905, + 0.017811283469200134, + 0.016096049919724464, + 0.016837749630212784, + 0.017735909670591354, + 0.041388269513845444, + 0.028820065781474113, + 0.01025841012597084, + 0.008116072975099087, + 0.01303910929709673, + 0.011759127490222454, + 0.024037515744566917, + 0.017520129680633545, + 0.04574671760201454, + 0.0441744439303875, + 0.015735695138573647, + 0.01333351619541645, + 0.01996554061770439, + 0.016452599316835403, + 0.003416783642023802, + 0.003886968595907092, + 0.0033773852046579123, + 0.003808466251939535, + 0.005187982227653265, + 0.0038323022890836, + 0.004530179314315319, + 0.0027445226442068815, + 0.008318311534821987, + 0.005256412550806999, + 0.013638417236506939, + 0.014325838536024094, + 0.00642306637018919, + 0.006103644613176584, + 0.007991792634129524, + 0.005640506278723478, + 0.0034808656200766563, + 0.002951378468424082, + 0.005548214074224234, + 0.004366185050457716, + 0.003203409956768155, + 0.0022113530430942774, + 0.003313875524327159, + 0.002125365659594536, + 0.003334704088047147, + 0.003294728696346283, + 0.0035214328672736883, + 0.005758743733167648, + 0.004150263033807278, + 0.0032337261363863945, + 0.004467228893190622, + 0.002648853464052081, + 0.004647573921829462, + 0.0034022978506982327, + 0.011180962435901165, + 0.006571456324309111, + 0.006022947374731302, + 0.004114075098186731, + 0.004823335446417332, + 0.004709480330348015, + 0.014322720468044281, + 0.017913969233632088, + 0.01645999401807785, + 0.0181700699031353, + 0.03256402537226677, + 0.019042696803808212, + 0.014799961820244789, + 0.01171585638076067, + 0.027417555451393127, + 0.012006505392491817, + 0.021142762154340744, + 0.027033133432269096, + 0.012945055030286312, + 0.012636374682188034, + 0.019305190071463585, + 0.012555359862744808 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 1 + ], + "selected_reward": -1.0541412830352783 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 focused at near-eye level showing the face of a sleeping young woman?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene illuminated by soft candlelight and warm afternoon sun filtering through silk curtains?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the woman have pale pink eyeshadow and thick dark brown eyelashes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the woman's eyes snap open with dilated light brown pupils showing sudden surprise?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a multi-layered cake with sugar roses, fresh berries, and gold foil visible in the foreground?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04868963733315468 + ], + "opposite_logprob_tokens": [ + -3.048689603805542 + ], + "logprob_reward": -0.04868963733315468, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04868963733315468, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09926740825176239, + 0.07294391095638275, + 0.05310780182480812, + 0.05171608179807663, + 0.04815993458032608, + 0.05778615549206734, + 0.028533244505524635, + 0.028499651700258255, + 0.35241198539733887, + 0.2075738161802292 + ], + "credit_map": [ + 0.01144309900701046, + 0.015543782152235508, + 0.011698885820806026, + 0.013283276930451393, + 0.015492568723857403, + 0.011286726221442223, + 0.011712930165231228, + 0.00880614947527647, + 0.007503715809434652, + 0.009160740301012993, + 0.019697144627571106, + 0.011299529112875462, + 0.0071136727929115295, + 0.005613032262772322, + 0.00750473840162158, + 0.00505133718252182, + 0.004778200760483742, + 0.004916044883430004, + 0.007204769644886255, + 0.01551235280930996, + 0.004010211676359177, + 0.004291696939617395, + 0.008910139091312885, + 0.0034843827597796917, + 0.006272752303630114, + 0.007419499568641186, + 0.006391241680830717, + 0.011015982367098331, + 0.005872971843928099, + 0.005073968321084976, + 0.005883841775357723, + 0.003785825567319989, + 0.006219083908945322, + 0.004503610543906689, + 0.006320287939161062, + 0.012007270008325577, + 0.004961383529007435, + 0.004165426827967167, + 0.006683797109872103, + 0.003299074713140726, + 0.006318511441349983, + 0.006464185193181038, + 0.005736192688345909, + 0.008318513631820679, + 0.010599244385957718, + 0.006294765044003725, + 0.007919163443148136, + 0.006135580129921436, + 0.0038559422828257084, + 0.003499515587463975, + 0.004284003749489784, + 0.003607602324336767, + 0.00453873910009861, + 0.002822399605065584, + 0.0039041764102876186, + 0.002020864514634013, + 0.003412752877920866, + 0.0028079587500542402, + 0.00590627733618021, + 0.004156690556555986, + 0.0034085342194885015, + 0.0025449765380471945, + 0.003663234179839492, + 0.0025992279406636953, + 0.042947690933942795, + 0.06860613822937012, + 0.035229288041591644, + 0.052863068878650665, + 0.05259773135185242, + 0.03342220559716225, + 0.03742092847824097, + 0.029324930161237717, + 0.030248451977968216, + 0.02139623649418354, + 0.028547199442982674, + 0.0322769358754158, + 0.023901045322418213, + 0.020263535901904106, + 0.025683922693133354, + 0.02525649219751358 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 0 + ], + "selected_reward": -0.04868963733315468 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 focused at near-eye level showing the face of a sleeping young woman?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene illuminated by soft candlelight and warm afternoon sun filtering through silk curtains?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the woman have pale pink eyeshadow and thick dark brown eyelashes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the woman's eyes snap open with dilated light brown pupils showing sudden surprise?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a multi-layered cake with sugar roses, fresh berries, and gold foil visible in the foreground?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01255878061056137 + ], + "opposite_logprob_tokens": [ + -4.387558937072754 + ], + "logprob_reward": -0.01255878061056137, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.01255878061056137, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0993860736489296, + 0.07066277414560318, + 0.04944838210940361, + 0.0639503076672554, + 0.057612624019384384, + 0.05766480043530464, + 0.05364091694355011, + 0.037447329610586166, + 0.4185267984867096, + 0.09165997058153152 + ], + "credit_map": [ + 0.013177357614040375, + 0.011793325655162334, + 0.014713413082063198, + 0.01195856835693121, + 0.013163011521100998, + 0.009729505516588688, + 0.013513290323317051, + 0.01133760716766119, + 0.009468059055507183, + 0.009561746381223202, + 0.010874491184949875, + 0.013133712112903595, + 0.005573644768446684, + 0.004675578325986862, + 0.012044059112668037, + 0.005331484135240316, + 0.0046659973450005054, + 0.006578398868441582, + 0.006603633053600788, + 0.008903386071324348, + 0.005563307087868452, + 0.0049729980528354645, + 0.007545714266598225, + 0.004614945966750383, + 0.006918487139046192, + 0.008347863331437111, + 0.008938681334257126, + 0.013352236710488796, + 0.006986487656831741, + 0.005674748215824366, + 0.008193502202630043, + 0.005538302939385176, + 0.0057981014251708984, + 0.005393308121711016, + 0.009390560910105705, + 0.013015903532505035, + 0.006354276090860367, + 0.005135655868798494, + 0.00787428393959999, + 0.004650531802326441, + 0.008140141144394875, + 0.0070043266750872135, + 0.00884088035672903, + 0.00867594126611948, + 0.006184231024235487, + 0.004672089591622353, + 0.008379153907299042, + 0.005768040660768747, + 0.007724547293037176, + 0.007274173200130463, + 0.007360334508121014, + 0.008525462821125984, + 0.007517553865909576, + 0.005200972780585289, + 0.005894188303500414, + 0.0041436851024627686, + 0.005974962841719389, + 0.005092398729175329, + 0.0033827959559857845, + 0.005004774779081345, + 0.005337586160749197, + 0.0036161127500236034, + 0.005530947353690863, + 0.0035077505744993687, + 0.058955851942300797, + 0.07005970180034637, + 0.05578747019171715, + 0.0641220435500145, + 0.060073353350162506, + 0.03776170685887337, + 0.04066918417811394, + 0.03109748475253582, + 0.015048335306346416, + 0.004847947508096695, + 0.013546577654778957, + 0.016671637073159218, + 0.008787358179688454, + 0.007912412285804749, + 0.01275181770324707, + 0.012093878351151943 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 9 + ], + "selected_reward": -0.01255878061056137 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 focused at near-eye level showing the face of a sleeping young woman?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene illuminated by soft candlelight and warm afternoon sun filtering through silk curtains?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the woman have pale pink eyeshadow and thick dark brown eyelashes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the woman's eyes snap open with dilated light brown pupils showing sudden surprise?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a multi-layered cake with sugar roses, fresh berries, and gold foil visible in the foreground?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"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": [ + -0.8983882665634155 + ], + "opposite_logprob_tokens": [ + -0.5233882665634155 + ], + "logprob_reward": -0.8983882665634155, + "yes_no_margin_reward": -0.375, + "token_credit_reward": -0.8983882665634155, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.19887582957744598, + 0.21845866739749908, + 0.18880698084831238, + 0.0430045910179615, + 0.08331269025802612, + 0.03797958791255951, + 0.05242665112018585, + 0.04474650323390961, + 0.08925206959247589, + 0.04313643276691437 + ], + "credit_map": [ + 0.026995044201612473, + 0.013973050750792027, + 0.057010311633348465, + 0.022489028051495552, + 0.020336683839559555, + 0.014105228707194328, + 0.027284840121865273, + 0.016681646928191185, + 0.025238728150725365, + 0.02741604670882225, + 0.053642839193344116, + 0.04428112134337425, + 0.014292939566075802, + 0.011953010223805904, + 0.024122020229697227, + 0.017511961981654167, + 0.021904025226831436, + 0.017602767795324326, + 0.04379892721772194, + 0.04143960401415825, + 0.016007471829652786, + 0.012610837817192078, + 0.018579591065645218, + 0.01686374470591545, + 0.0045632184483110905, + 0.0062118712812662125, + 0.006404775660485029, + 0.00921407900750637, + 0.004516178276389837, + 0.0037783838342875242, + 0.004944692365825176, + 0.003371389349922538, + 0.009337000548839569, + 0.005263575818389654, + 0.016404271125793457, + 0.019013356417417526, + 0.007754313759505749, + 0.007076028734445572, + 0.011094165034592152, + 0.00736997788771987, + 0.0048988754861056805, + 0.00455353083088994, + 0.006801983341574669, + 0.005599170923233032, + 0.004187250975519419, + 0.003879331750795245, + 0.004692539107054472, + 0.0033669027034193277, + 0.007389159873127937, + 0.006487713195383549, + 0.006001542787998915, + 0.008940381929278374, + 0.007440508343279362, + 0.005029616411775351, + 0.006365712266415358, + 0.004772016778588295, + 0.005052305292338133, + 0.004815754480659962, + 0.007474456448107958, + 0.006180981639772654, + 0.007094020955264568, + 0.004414550960063934, + 0.005202245432883501, + 0.004512187093496323, + 0.011894670315086842, + 0.013856912963092327, + 0.006822182331234217, + 0.011462079361081123, + 0.018668770790100098, + 0.010527605190873146, + 0.00952262431383133, + 0.006497220601886511, + 0.005924216005951166, + 0.002756423084065318, + 0.007433501537889242, + 0.008802702650427818, + 0.003993583377450705, + 0.004163027741014957, + 0.0055839731357991695, + 0.004479007795453072 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 2 + ], + "selected_reward": -0.8983882665634155 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "No", + "Yes", + "Yes", + "No" + ] + }, + { + "prompt": "stunning, close-up, static shot, focused at near-eye level, captures the face of a sleeping young woman, the entire scene steeped in the opulence and softness of Rococo style. Soft candlelight intertwines with the warm afternoon sun filtering through silk curtains, casting a warm, three-dimensional light on the girl's delicate, fair skin. Her pale pink eyeshadow adds a touch of sweetness. Initially, the subtle flutter of her thick, butterfly-like, dark brown eyelashes shatters the stillness of the scene. Then, her eyes snap open, her clear, light brown pupils dilating with sudden surprise. The camera captures, with pinpoint accuracy, the perfect reflection of a delicate, multi-layered cake in the foreground, adorned with sugar roses, fresh berries, and gold foil. Even the sheen of the buttercream is clearly visible. The scene, dominated by a palette of ivory, pale pink, and gold, creates a dramatic atmosphere of innocence, dreaminess, and sudden awakening, like a magnificent painting in motion.", + "seed": 8463, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000026/group_15/index_1868_seed_8463_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": -0.10203044023364782, + "vlm_logprob": -0.10203044023364782, + "vlm_yes_no_margin": 3.125, + "vlm_token_credit": -0.10203044023364782, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 focused at near-eye level showing the face of a sleeping young woman?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene illuminated by soft candlelight and warm afternoon sun filtering through silk curtains?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the woman have pale pink eyeshadow and thick dark brown eyelashes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the woman's eyes snap open with dilated light brown pupils showing sudden surprise?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a multi-layered cake with sugar roses, fresh berries, and gold foil visible in the foreground?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018343685194849968 + ], + "opposite_logprob_tokens": [ + -4.268343448638916 + ], + "logprob_reward": -0.018343685194849968, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.018343685194849968, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19827800989151, + 0.050704143941402435, + 0.06640665978193283, + 0.22108665108680725, + 0.08908528089523315, + 0.048352841287851334, + 0.06798579543828964, + 0.12558355927467346, + 0.11266322433948517, + 0.01985381543636322 + ], + "credit_map": [ + 0.030057823285460472, + 0.04100019857287407, + 0.014011955820024014, + 0.012318764813244343, + 0.019616734236478806, + 0.015876885503530502, + 0.021400680765509605, + 0.043994978070259094, + 0.009538193233311176, + 0.008809792809188366, + 0.0037516767624765635, + 0.004447546787559986, + 0.008446410298347473, + 0.00460142083466053, + 0.00434758048504591, + 0.006761522497981787, + 0.014860512688755989, + 0.00946641806513071, + 0.005373768042773008, + 0.004780501592904329, + 0.010787609033286572, + 0.007454179227352142, + 0.005644810386002064, + 0.00803886353969574, + 0.03362240269780159, + 0.02677937224507332, + 0.023048894479870796, + 0.016044270247220993, + 0.03406677022576332, + 0.02420508675277233, + 0.016457390040159225, + 0.04686248302459717, + 0.01650865748524666, + 0.011065224185585976, + 0.010331850498914719, + 0.006854596547782421, + 0.011095468886196613, + 0.007506616413593292, + 0.005758976098150015, + 0.01996389590203762, + 0.005196917336434126, + 0.005607954692095518, + 0.0032781006302684546, + 0.005177760496735573, + 0.010747113265097141, + 0.005135060753673315, + 0.003984970971941948, + 0.009224966168403625, + 0.011398457922041416, + 0.021247930824756622, + 0.0039315070025622845, + 0.004588695243000984, + 0.007992569357156754, + 0.004020975902676582, + 0.003198956837877631, + 0.011606707237660885, + 0.015124186873435974, + 0.031234463676810265, + 0.008709839545190334, + 0.010531749576330185, + 0.01435263454914093, + 0.00925254262983799, + 0.013520755805075169, + 0.02285740152001381, + 0.01324903778731823, + 0.035191185772418976, + 0.0054392386227846146, + 0.006351187359541655, + 0.014786109328269958, + 0.0076155876740813255, + 0.007407790049910545, + 0.02262309566140175, + 0.0024871407076716423, + 0.0037991926074028015, + 0.001614599721506238, + 0.0015316904755309224, + 0.0036621552426368, + 0.0020347062963992357, + 0.0020515320356935263, + 0.002672801027074456 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 7 + ], + "selected_reward": -0.018343685194849968 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 focused at near-eye level showing the face of a sleeping young woman?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene illuminated by soft candlelight and warm afternoon sun filtering through silk curtains?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the woman have pale pink eyeshadow and thick dark brown eyelashes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the woman's eyes snap open with dilated light brown pupils showing sudden surprise?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a multi-layered cake with sugar roses, fresh berries, and gold foil visible in the foreground?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.38735127449035645 + ], + "opposite_logprob_tokens": [ + -1.1373512744903564 + ], + "logprob_reward": -0.38735127449035645, + "yes_no_margin_reward": 0.75, + "token_credit_reward": -0.38735127449035645, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.31659796833992004, + 0.0660504549741745, + 0.060523174703121185, + 0.154260516166687, + 0.11770378798246384, + 0.05941978096961975, + 0.07367458939552307, + 0.04934657737612724, + 0.06749759614467621, + 0.03492552414536476 + ], + "credit_map": [ + 0.04070218279957771, + 0.060233891010284424, + 0.02540547214448452, + 0.023150747641921043, + 0.03639713302254677, + 0.026691609993577003, + 0.029617831110954285, + 0.07439912855625153, + 0.009405582211911678, + 0.007878012023866177, + 0.008731704205274582, + 0.008375421166419983, + 0.009713344275951385, + 0.006244585383683443, + 0.006164140533655882, + 0.009537664242088795, + 0.006423915736377239, + 0.00759626179933548, + 0.00680778780952096, + 0.012337732128798962, + 0.00739253917708993, + 0.0041366079822182655, + 0.007794801611453295, + 0.008033530786633492, + 0.02481943927705288, + 0.018503496423363686, + 0.019382989034056664, + 0.012605180963873863, + 0.019215194508433342, + 0.015085505321621895, + 0.011284378357231617, + 0.03336435183882713, + 0.019860681146383286, + 0.013543112203478813, + 0.013708190992474556, + 0.009701567701995373, + 0.011838128790259361, + 0.010203891433775425, + 0.008788831532001495, + 0.03005939908325672, + 0.011142830364406109, + 0.007707628421485424, + 0.0052628484554588795, + 0.006825111340731382, + 0.012721138074994087, + 0.00555784348398447, + 0.005353695712983608, + 0.0048486883752048016, + 0.007932024076581001, + 0.020084910094738007, + 0.006582243833690882, + 0.006910714786499739, + 0.006833960767835379, + 0.004109437111765146, + 0.005525459069758654, + 0.015695840120315552, + 0.004685107618570328, + 0.00597582571208477, + 0.0052767409943044186, + 0.004376975819468498, + 0.01070853229612112, + 0.006396772805601358, + 0.007348863873630762, + 0.004577760584652424, + 0.00709872180595994, + 0.014530867338180542, + 0.006876039318740368, + 0.006580841261893511, + 0.008902301080524921, + 0.0059044999070465565, + 0.006801882293075323, + 0.010802443139255047, + 0.005547341890633106, + 0.003692260477691889, + 0.0036426265724003315, + 0.005484460387378931, + 0.004865109454840422, + 0.004253572318702936, + 0.0036922984290868044, + 0.0037478571757674217 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 4 + ], + "selected_reward": -0.38735127449035645 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 focused at near-eye level showing the face of a sleeping young woman?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene illuminated by soft candlelight and warm afternoon sun filtering through silk curtains?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the woman have pale pink eyeshadow and thick dark brown eyelashes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the woman's eyes snap open with dilated light brown pupils showing sudden surprise?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a multi-layered cake with sugar roses, fresh berries, and gold foil visible in the foreground?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029834305867552757 + ], + "opposite_logprob_tokens": [ + -3.529834270477295 + ], + "logprob_reward": -0.029834305867552757, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029834305867552757, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1555524617433548, + 0.041370537132024765, + 0.050932154059410095, + 0.24306315183639526, + 0.12386675924062729, + 0.04731813073158264, + 0.08605826646089554, + 0.08979140967130661, + 0.06520507484674454, + 0.09684208780527115 + ], + "credit_map": [ + 0.020451745018363, + 0.022061370313167572, + 0.01724635437130928, + 0.016789009794592857, + 0.022602714598178864, + 0.011932151392102242, + 0.016861578449606895, + 0.02760753035545349, + 0.004897372331470251, + 0.005112680606544018, + 0.0042602200992405415, + 0.00416643125936389, + 0.00670873187482357, + 0.003768258960917592, + 0.005586495157331228, + 0.0068703461438417435, + 0.009061037562787533, + 0.007230247370898724, + 0.006085122004151344, + 0.004830753430724144, + 0.008555943146348, + 0.005332238506525755, + 0.003569911466911435, + 0.006266898475587368, + 0.04057605192065239, + 0.029470980167388916, + 0.03014659509062767, + 0.02252292074263096, + 0.030713273212313652, + 0.023118406534194946, + 0.0171652864664793, + 0.04934961348772049, + 0.01973102055490017, + 0.014071248471736908, + 0.014832210727036, + 0.010229459963738918, + 0.014028343372046947, + 0.010701777413487434, + 0.009135507047176361, + 0.03113718144595623, + 0.006114338990300894, + 0.00835019163787365, + 0.0035787143278867006, + 0.003233636962249875, + 0.009755772538483143, + 0.0038672711234539747, + 0.004376011434942484, + 0.00804219115525484, + 0.014320393092930317, + 0.012477541342377663, + 0.005601970478892326, + 0.007196169812232256, + 0.017035041004419327, + 0.0066045960411429405, + 0.007657975424081087, + 0.015164566226303577, + 0.010308283381164074, + 0.02695046178996563, + 0.005006756167858839, + 0.00610137777402997, + 0.008274950087070465, + 0.005733736790716648, + 0.006536284927278757, + 0.020879551768302917, + 0.007786296308040619, + 0.017750579863786697, + 0.003000064520165324, + 0.0037295774091035128, + 0.0100630521774292, + 0.005144285969436169, + 0.006202921271324158, + 0.011528289876878262, + 0.010526529513299465, + 0.026232672855257988, + 0.005485415458679199, + 0.006303172092884779, + 0.01601841114461422, + 0.007462877780199051, + 0.006819028407335281, + 0.01799396425485611 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 4 + ], + "selected_reward": -0.029834305867552757 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 focused at near-eye level showing the face of a sleeping young woman?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene illuminated by soft candlelight and warm afternoon sun filtering through silk curtains?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the woman have pale pink eyeshadow and thick dark brown eyelashes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the woman's eyes snap open with dilated light brown pupils showing sudden surprise?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a multi-layered cake with sugar roses, fresh berries, and gold foil visible in the foreground?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06207651644945145 + ], + "opposite_logprob_tokens": [ + -2.8120765686035156 + ], + "logprob_reward": -0.06207651644945145, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06207651644945145, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04679734632372856, + 0.03299374505877495, + 0.022407744079828262, + 0.08885063976049423, + 0.05262041836977005, + 0.14258618652820587, + 0.1339614987373352, + 0.05939703807234764, + 0.23960496485233307, + 0.18078039586544037 + ], + "credit_map": [ + 0.009676818735897541, + 0.005794179625809193, + 0.003915860783308744, + 0.004155037458986044, + 0.00938976090401411, + 0.005382067058235407, + 0.004674885421991348, + 0.003808737499639392, + 0.003972781356424093, + 0.003917274530977011, + 0.003426636103540659, + 0.003768454771488905, + 0.005640653893351555, + 0.0037902933545410633, + 0.0040787747129797935, + 0.004398875404149294, + 0.004022094886749983, + 0.0026932249311357737, + 0.001836926443502307, + 0.0013912111753597856, + 0.004847269505262375, + 0.002790137194097042, + 0.0015547550283372402, + 0.003272125031799078, + 0.017078375443816185, + 0.01125236414372921, + 0.008667982183396816, + 0.006442733574658632, + 0.012634474784135818, + 0.01035038661211729, + 0.0067583369091153145, + 0.015665991231799126, + 0.009956100955605507, + 0.00631737569347024, + 0.006200856063514948, + 0.0039147683419287205, + 0.007328993175178766, + 0.004458825569599867, + 0.003779600141569972, + 0.010663899593055248, + 0.03248537704348564, + 0.02901574969291687, + 0.009008503518998623, + 0.007741506211459637, + 0.019656002521514893, + 0.009286642074584961, + 0.010516531765460968, + 0.024875856935977936, + 0.02597406879067421, + 0.025506634265184402, + 0.010070937685668468, + 0.01897951029241085, + 0.02417517639696598, + 0.008370928466320038, + 0.010623699985444546, + 0.01026054099202156, + 0.00762613071128726, + 0.008788676001131535, + 0.0038140667602419853, + 0.004014717880636454, + 0.014983227476477623, + 0.00921296514570713, + 0.005403170827776194, + 0.005554081406444311, + 0.031194718554615974, + 0.06479115784168243, + 0.012289229780435562, + 0.013879228383302689, + 0.035175081342458725, + 0.018362365663051605, + 0.019931741058826447, + 0.04398144409060478, + 0.01740262657403946, + 0.04050344601273537, + 0.01343502290546894, + 0.017517173662781715, + 0.0359613262116909, + 0.013502560555934906, + 0.015257949940860271, + 0.02720029279589653 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 5 + ], + "selected_reward": -0.06207651644945145 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 focused at near-eye level showing the face of a sleeping young woman?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene illuminated by soft candlelight and warm afternoon sun filtering through silk curtains?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the woman have pale pink eyeshadow and thick dark brown eyelashes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the woman's eyes snap open with dilated light brown pupils showing sudden surprise?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a multi-layered cake with sugar roses, fresh berries, and gold foil visible in the foreground?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0125464191660285 + ], + "opposite_logprob_tokens": [ + -4.387546539306641 + ], + "logprob_reward": -0.0125464191660285, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.0125464191660285, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4665689766407013, + 0.02713029272854328, + 0.029077809303998947, + 0.08366304636001587, + 0.08046969026327133, + 0.15750522911548615, + 0.0435987189412117, + 0.031091172248125076, + 0.04731851443648338, + 0.03357667848467827 + ], + "credit_map": [ + 0.05691162496805191, + 0.0836009755730629, + 0.037315696477890015, + 0.03752788156270981, + 0.0555441714823246, + 0.037915777415037155, + 0.05257860943675041, + 0.10517416149377823, + 0.003949666861444712, + 0.0045394753105938435, + 0.0023085640277713537, + 0.0021195607259869576, + 0.0045824977569282055, + 0.0021865726448595524, + 0.003427187679335475, + 0.004016764461994171, + 0.005182361230254173, + 0.0037743770517408848, + 0.002817536471411586, + 0.0023417517077177763, + 0.00489215599372983, + 0.0034093454014509916, + 0.0024589027743786573, + 0.00420137494802475, + 0.01369919627904892, + 0.010277247056365013, + 0.009554333053529263, + 0.006223438307642937, + 0.011940492317080498, + 0.008489223197102547, + 0.006549980957061052, + 0.016929123550653458, + 0.0143587626516819, + 0.00859851948916912, + 0.009788060560822487, + 0.006878592073917389, + 0.009429641999304295, + 0.007000143174082041, + 0.005582797806710005, + 0.018833158537745476, + 0.025698132812976837, + 0.03297160193324089, + 0.010505425743758678, + 0.010339372791349888, + 0.016939381137490273, + 0.01079088356345892, + 0.016605393961071968, + 0.033655013889074326, + 0.006989865098148584, + 0.00605141231790185, + 0.002996418857946992, + 0.005933673121035099, + 0.009347930550575256, + 0.0039982870221138, + 0.0034696778748184443, + 0.004811447579413652, + 0.0026976922526955605, + 0.008067631162703037, + 0.00141248875297606, + 0.0014784433878958225, + 0.004911511205136776, + 0.002465534256771207, + 0.0032949585001915693, + 0.006762911099940538, + 0.005149508360773325, + 0.009967059828341007, + 0.00510155875235796, + 0.007321970071643591, + 0.00543211167678237, + 0.0039680213667452335, + 0.00391607778146863, + 0.00646219914779067, + 0.003812497016042471, + 0.0038099954836070538, + 0.003591854590922594, + 0.0055876863189041615, + 0.006777627859264612, + 0.003405919997021556, + 0.0035348122473806143, + 0.003056282876059413 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 3 + ], + "selected_reward": -0.0125464191660285 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "stunning, close-up, static shot, focused at near-eye level, captures the face of a sleeping young woman, the entire scene steeped in the opulence and softness of Rococo style. Soft candlelight intertwines with the warm afternoon sun filtering through silk curtains, casting a warm, three-dimensional light on the girl's delicate, fair skin. Her pale pink eyeshadow adds a touch of sweetness. Initially, the subtle flutter of her thick, butterfly-like, dark brown eyelashes shatters the stillness of the scene. Then, her eyes snap open, her clear, light brown pupils dilating with sudden surprise. The camera captures, with pinpoint accuracy, the perfect reflection of a delicate, multi-layered cake in the foreground, adorned with sugar roses, fresh berries, and gold foil. Even the sheen of the buttercream is clearly visible. The scene, dominated by a palette of ivory, pale pink, and gold, creates a dramatic atmosphere of innocence, dreaminess, and sudden awakening, like a magnificent painting in motion.", + "seed": 8464, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000026/group_15/index_1868_seed_8464_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": -0.06489285137504339, + "vlm_logprob": -0.06489285137504339, + "vlm_yes_no_margin": 3.975, + "vlm_token_credit": -0.06489285137504339, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 focused at near-eye level showing the face of a sleeping young woman?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene illuminated by soft candlelight and warm afternoon sun filtering through silk curtains?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the woman have pale pink eyeshadow and thick dark brown eyelashes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the woman's eyes snap open with dilated light brown pupils showing sudden surprise?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a multi-layered cake with sugar roses, fresh berries, and gold foil visible in the foreground?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01038268581032753 + ], + "opposite_logprob_tokens": [ + -5.010382652282715 + ], + "logprob_reward": -0.01038268581032753, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.01038268581032753, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1810203343629837, + 0.05633587762713432, + 0.06148834154009819, + 0.01979496330022812, + 0.030688416212797165, + 0.029018433764576912, + 0.06466130912303925, + 0.04455685615539551, + 0.45021405816078186, + 0.06222141534090042 + ], + "credit_map": [ + 0.02314074896275997, + 0.03088224306702614, + 0.010778973810374737, + 0.014027963392436504, + 0.02534494176506996, + 0.03195182979106903, + 0.01852102018892765, + 0.026372626423835754, + 0.004305906593799591, + 0.009712308645248413, + 0.003839383600279689, + 0.004289153032004833, + 0.008191497065126896, + 0.009612290188670158, + 0.007967685349285603, + 0.008417653851211071, + 0.00543829333037138, + 0.010673170909285545, + 0.003997796215116978, + 0.005145853850990534, + 0.009028210304677486, + 0.008342307060956955, + 0.0078270910307765, + 0.011035621166229248, + 0.0016653637867420912, + 0.004058043472468853, + 0.001094965497031808, + 0.0013216340448707342, + 0.003113269805908203, + 0.003723680740222335, + 0.0024920417927205563, + 0.0023259632289409637, + 0.0031852172687649727, + 0.004150293301790953, + 0.0026340242475271225, + 0.00286702043376863, + 0.003914826549589634, + 0.0040843007154762745, + 0.0059006307274103165, + 0.003952101804316044, + 0.0024374311324208975, + 0.004985432606190443, + 0.0017553981160745025, + 0.0019978065975010395, + 0.003499614307656884, + 0.004151052795350552, + 0.004563477821648121, + 0.005628219805657864, + 0.0038311888929456472, + 0.00751841627061367, + 0.004243351984769106, + 0.004813257604837418, + 0.008066292852163315, + 0.009853505529463291, + 0.012903906404972076, + 0.013431382365524769, + 0.004969673231244087, + 0.005654460750520229, + 0.003914406057447195, + 0.005335908383131027, + 0.00523916631937027, + 0.004761043004691601, + 0.007881401106715202, + 0.006800799630582333, + 0.03442682698369026, + 0.04903869703412056, + 0.03424634411931038, + 0.05948120728135109, + 0.06328334659337997, + 0.03920504450798035, + 0.1191326230764389, + 0.05139998346567154, + 0.004847383592277765, + 0.007456955965608358, + 0.005347838159650564, + 0.006535952910780907, + 0.008615262806415558, + 0.0069030518643558025, + 0.012630476616322994, + 0.009884492494165897 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 6 + ], + "selected_reward": -0.01038268581032753 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 focused at near-eye level showing the face of a sleeping young woman?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene illuminated by soft candlelight and warm afternoon sun filtering through silk curtains?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the woman have pale pink eyeshadow and thick dark brown eyelashes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the woman's eyes snap open with dilated light brown pupils showing sudden surprise?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a multi-layered cake with sugar roses, fresh berries, and gold foil visible in the foreground?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.281472772359848 + ], + "opposite_logprob_tokens": [ + -1.4064728021621704 + ], + "logprob_reward": -0.281472772359848, + "yes_no_margin_reward": 1.125, + "token_credit_reward": -0.281472772359848, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24367843568325043, + 0.13901059329509735, + 0.06543099135160446, + 0.07493797689676285, + 0.0728887990117073, + 0.057280704379081726, + 0.0971628800034523, + 0.09450163692235947, + 0.09352264553308487, + 0.061585426330566406 + ], + "credit_map": [ + 0.02507537603378296, + 0.03806649148464203, + 0.017568601295351982, + 0.022771261632442474, + 0.034885700792074203, + 0.03973919153213501, + 0.0332389697432518, + 0.03233281150460243, + 0.012416085228323936, + 0.021264299750328064, + 0.009287561289966106, + 0.01262965239584446, + 0.018474556505680084, + 0.025776786729693413, + 0.015213209204375744, + 0.023948419839143753, + 0.00629102298989892, + 0.010687945410609245, + 0.004187974613159895, + 0.006308276671916246, + 0.008168330416083336, + 0.009730367921292782, + 0.0076220836490392685, + 0.01243498083204031, + 0.004808101803064346, + 0.012347661890089512, + 0.005184795707464218, + 0.009110560640692711, + 0.010456720367074013, + 0.012243703939020634, + 0.010671091265976429, + 0.010115331970155239, + 0.006942938547581434, + 0.010343731380999088, + 0.00439154589548707, + 0.0067625343799591064, + 0.010372847318649292, + 0.01316847838461399, + 0.010769736021757126, + 0.010136984288692474, + 0.0053456551395356655, + 0.006757540628314018, + 0.004224428441375494, + 0.006245430558919907, + 0.006736384239047766, + 0.008762778714299202, + 0.01020381785929203, + 0.009004668332636356, + 0.01026001013815403, + 0.015691980719566345, + 0.00599686149507761, + 0.00860587414354086, + 0.014842096716165543, + 0.012223362922668457, + 0.01788969337940216, + 0.011652992106974125, + 0.012796224094927311, + 0.011177650652825832, + 0.007889371365308762, + 0.010094193741679192, + 0.011784455738961697, + 0.010530433617532253, + 0.017298368737101555, + 0.012930928729474545, + 0.008567116223275661, + 0.014935491606593132, + 0.006907940376549959, + 0.009720348753035069, + 0.013115990906953812, + 0.009699175134301186, + 0.019533613696694374, + 0.01104296650737524, + 0.006124455947428942, + 0.007058985065668821, + 0.005269113928079605, + 0.006408642046153545, + 0.009394104592502117, + 0.007292497903108597, + 0.013073871843516827, + 0.006963746156543493 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 6 + ], + "selected_reward": -0.281472772359848 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 focused at near-eye level showing the face of a sleeping young woman?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene illuminated by soft candlelight and warm afternoon sun filtering through silk curtains?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the woman have pale pink eyeshadow and thick dark brown eyelashes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the woman's eyes snap open with dilated light brown pupils showing sudden surprise?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a multi-layered cake with sugar roses, fresh berries, and gold foil visible in the foreground?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016079863533377647 + ], + "opposite_logprob_tokens": [ + -4.141079902648926 + ], + "logprob_reward": -0.016079863533377647, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016079863533377647, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2413208931684494, + 0.18271800875663757, + 0.04934091866016388, + 0.04094630107283592, + 0.08279705047607422, + 0.051115553826093674, + 0.0720508024096489, + 0.09219125658273697, + 0.05392741784453392, + 0.13359178602695465 + ], + "credit_map": [ + 0.020675040781497955, + 0.041905421763658524, + 0.01361168920993805, + 0.018572479486465454, + 0.03544282913208008, + 0.05021636188030243, + 0.029703153297305107, + 0.031193900853395462, + 0.015078794211149216, + 0.02908116579055786, + 0.009703301824629307, + 0.014114965684711933, + 0.025992000475525856, + 0.04035129398107529, + 0.018056567758321762, + 0.030339917168021202, + 0.0029926556162536144, + 0.008051112294197083, + 0.003445346374064684, + 0.005111243110150099, + 0.007035285234451294, + 0.008407194167375565, + 0.007938899099826813, + 0.006359185557812452, + 0.0029242243617773056, + 0.006833357270807028, + 0.0029363082721829414, + 0.004008201416581869, + 0.006539763882756233, + 0.007034693844616413, + 0.005489740986377001, + 0.005180011969059706, + 0.005630481522530317, + 0.013230138458311558, + 0.00534295616671443, + 0.010873549617826939, + 0.010149023495614529, + 0.011330300942063332, + 0.013602697290480137, + 0.012637906707823277, + 0.005119642708450556, + 0.011471299454569817, + 0.0034815636463463306, + 0.0038405475206673145, + 0.006539752706885338, + 0.006640473380684853, + 0.00657371012493968, + 0.0074485670775175095, + 0.0071185193955898285, + 0.008726335130631924, + 0.005173208191990852, + 0.007811710238456726, + 0.01096112746745348, + 0.007219931576400995, + 0.015347829088568687, + 0.00969214178621769, + 0.009253410622477531, + 0.013535181991755962, + 0.007131459657102823, + 0.008813769556581974, + 0.011081092059612274, + 0.010542415082454681, + 0.01814435049891472, + 0.01368957944214344, + 0.005279218312352896, + 0.008412512950599194, + 0.00418098783120513, + 0.005968519486486912, + 0.007505600806325674, + 0.004886451177299023, + 0.01119527593255043, + 0.006498854607343674, + 0.011293758638203144, + 0.01692533865571022, + 0.009628646075725555, + 0.014720013365149498, + 0.019381428137421608, + 0.01807316020131111, + 0.02847549505531788, + 0.01509394682943821 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 9 + ], + "selected_reward": -0.016079863533377647 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 focused at near-eye level showing the face of a sleeping young woman?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene illuminated by soft candlelight and warm afternoon sun filtering through silk curtains?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the woman have pale pink eyeshadow and thick dark brown eyelashes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the woman's eyes snap open with dilated light brown pupils showing sudden surprise?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a multi-layered cake with sugar roses, fresh berries, and gold foil visible in the foreground?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009790970012545586 + ], + "opposite_logprob_tokens": [ + -4.634790897369385 + ], + "logprob_reward": -0.009790970012545586, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009790970012545586, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08255726844072342, + 0.07050061970949173, + 0.08658956736326218, + 0.0376560352742672, + 0.05174626037478447, + 0.0701417550444603, + 0.07793379575014114, + 0.06816523522138596, + 0.27559036016464233, + 0.17911909520626068 + ], + "credit_map": [ + 0.008023466914892197, + 0.012431644834578037, + 0.0051270583644509315, + 0.006377058103680611, + 0.012598731555044651, + 0.012779325246810913, + 0.010692954063415527, + 0.014527034014463425, + 0.005116300191730261, + 0.013945874758064747, + 0.003985341172665358, + 0.005224274005740881, + 0.011682983487844467, + 0.008914513513445854, + 0.0091931764036417, + 0.012438159435987473, + 0.004974902607500553, + 0.012031502090394497, + 0.005694056395441294, + 0.006685537751764059, + 0.012398183345794678, + 0.012387186288833618, + 0.013642074540257454, + 0.01877613179385662, + 0.0023755738511681557, + 0.0074739777483046055, + 0.0023913851473480463, + 0.003080291673541069, + 0.005356356501579285, + 0.0059905569069087505, + 0.005925614852458239, + 0.005062280222773552, + 0.005701607093214989, + 0.00941658765077591, + 0.003671028185635805, + 0.0034918603487312794, + 0.00602179579436779, + 0.0075986674055457115, + 0.007694103755056858, + 0.008150608278810978, + 0.005655345041304827, + 0.013228731229901314, + 0.0037749370094388723, + 0.0055379378609359264, + 0.009659526869654655, + 0.010610430501401424, + 0.010611222125589848, + 0.01106362696737051, + 0.004710896406322718, + 0.009592660702764988, + 0.005214985925704241, + 0.005946851335465908, + 0.009291267953813076, + 0.015415040776133537, + 0.012642757035791874, + 0.015119335614144802, + 0.007957130670547485, + 0.007121421862393618, + 0.008573934435844421, + 0.005565909668803215, + 0.008727586828172207, + 0.008327420800924301, + 0.01271384209394455, + 0.009177994914352894, + 0.02259325608611107, + 0.059983354061841965, + 0.018069595098495483, + 0.027844326570630074, + 0.03543922305107117, + 0.027029266580939293, + 0.044910430908203125, + 0.03972092643380165, + 0.01795629970729351, + 0.026836996898055077, + 0.011380806565284729, + 0.011527552269399166, + 0.02669818140566349, + 0.02954072318971157, + 0.023436754941940308, + 0.03174180909991264 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 2 + ], + "selected_reward": -0.009790970012545586 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 focused at near-eye level showing the face of a sleeping young woman?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene illuminated by soft candlelight and warm afternoon sun filtering through silk curtains?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the woman have pale pink eyeshadow and thick dark brown eyelashes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the woman's eyes snap open with dilated light brown pupils showing sudden surprise?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a multi-layered cake with sugar roses, fresh berries, and gold foil visible in the foreground?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"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": [ + 0.08756640553474426, + 0.043648406863212585, + 0.038406021893024445, + 0.035952143371105194, + 0.07003568112850189, + 0.04751253500580788, + 0.06669805198907852, + 0.031863827258348465, + 0.18338799476623535, + 0.394928902387619 + ], + "credit_map": [ + 0.00643546786159277, + 0.010547363199293613, + 0.004503708332777023, + 0.006946119479835033, + 0.01371108740568161, + 0.011707653291523457, + 0.017813779413700104, + 0.015901226550340652, + 0.00426866952329874, + 0.006839019246399403, + 0.002002537017688155, + 0.0041427938267588615, + 0.006709784269332886, + 0.007190525531768799, + 0.006287424825131893, + 0.00620765145868063, + 0.0027249993290752172, + 0.0032670453656464815, + 0.0030744855757802725, + 0.005247049033641815, + 0.005281980149447918, + 0.0043169185519218445, + 0.008762394078075886, + 0.0057311514392495155, + 0.002466462552547455, + 0.004819451831281185, + 0.0025957096368074417, + 0.0044318330474197865, + 0.00589176919311285, + 0.0050114598125219345, + 0.006167616695165634, + 0.004567840602248907, + 0.00353078730404377, + 0.008181891404092312, + 0.0061490340158343315, + 0.011361046694219112, + 0.008541414514183998, + 0.010271837003529072, + 0.011846628040075302, + 0.01015304122120142, + 0.0029897226486355066, + 0.005013209767639637, + 0.00408086646348238, + 0.009273524396121502, + 0.00615962129086256, + 0.005886479280889034, + 0.008689109236001968, + 0.005420004017651081, + 0.006698873359709978, + 0.00811009667813778, + 0.004183597397059202, + 0.0062523153610527515, + 0.009427696466445923, + 0.008474760688841343, + 0.011150495149195194, + 0.012400218285620213, + 0.0038439917843788862, + 0.0033784634433686733, + 0.003367847064509988, + 0.004489654675126076, + 0.003753282129764557, + 0.0033801663666963577, + 0.00598951755091548, + 0.0036609049420803785, + 0.011092374101281166, + 0.017505135387182236, + 0.01688954420387745, + 0.020055381581187248, + 0.03083706647157669, + 0.011110249906778336, + 0.05215070769190788, + 0.023747529834508896, + 0.02781609259545803, + 0.03462724760174751, + 0.029788699001073837, + 0.04977559298276901, + 0.060411982238292694, + 0.03687943518161774, + 0.10212786495685577, + 0.053502004593610764 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 0 + ], + "selected_reward": -0.0067379651591181755 + } + ], + "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/0000026/group_2/index_tensor([4783, 4783])_group_2_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000026/group_2/index_tensor([4783, 4783])_group_2_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..f49369126cf5b62e0822c427cbca099820a39b53 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000026/group_2/index_tensor([4783, 4783])_group_2_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "static, medium-long shot, shot from a level angle and composed with a steady eye, focuses on a dimly lit, classically styled study. A wall of floor-to-ceiling mahogany bookshelves dominates the view, densely packed with hardbound volumes. Their leather covers and gilded spines gleam softly and deeply in the warm, flickering glow of the off-screen fireplace. With a low, almost infrasonic sound of mechanical friction and rolling stone, the vast wall of shelves begins to solemnly and slowly rotate inward, its contents and a small brass globe remaining incredibly stable throughout the movement. As the shelves rotate, a bottomless, square entrance of absolute blackness gradually reveals itself. This darkness, seemingly devouring all light, creates a stark and dramatic contrast with the warm, orange-toned light within the study, creating a tense atmosphere filled with mystery, suspense, and a sense of uncharted territory.", + "seed": 28852, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000026/group_2/index_4783_seed_28852_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": -0.02499551046639681, + "vlm_logprob": -0.02499551046639681, + "vlm_yes_no_margin": 4.25, + "vlm_token_credit": -0.02499551046639681, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 dimly lit, classically styled study with a wall of floor-to-ceiling mahogany bookshelves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the bookshelves densely packed with hardbound volumes featuring leather covers and gilded spines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the wall of bookshelves rotate inward slowly while a small brass globe remains stable?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a bottomless, square entrance of absolute blackness gradually reveal itself as the shelves rotate?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static, medium-long shot composed from a level angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008455547504127026 + ], + "opposite_logprob_tokens": [ + -6.008455753326416 + ], + "logprob_reward": -0.008455547504127026, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.008455547504127026, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.01818818412721157, + 0.021280916407704353, + 0.03264909237623215, + 0.0298165250569582, + 0.038734275847673416, + 0.021961519494652748, + 0.017615245655179024, + 0.028851967304944992, + 0.731096088886261, + 0.059806156903505325 + ], + "credit_map": [ + 0.00279622501693666, + 0.0033101856242865324, + 0.0025344821624457836, + 0.00221859198063612, + 0.0011172185186296701, + 0.002297306666150689, + 0.0023254146799445152, + 0.0015887600602582097, + 0.002704732120037079, + 0.0021488675847649574, + 0.002955016680061817, + 0.002481454983353615, + 0.001982314744964242, + 0.0033865482546389103, + 0.003120140638202429, + 0.002501842100173235, + 0.005221855361014605, + 0.004453796427696943, + 0.004839128814637661, + 0.0046799988485872746, + 0.002333129523321986, + 0.0027689943090081215, + 0.0042892936617136, + 0.0040628956630826, + 0.005860233213752508, + 0.0030939525458961725, + 0.005532790906727314, + 0.003189357230439782, + 0.0024483606684952974, + 0.002839655615389347, + 0.0026880560908466578, + 0.004164118319749832, + 0.003957774955779314, + 0.007295012939721346, + 0.004551175516098738, + 0.0059457700699567795, + 0.0038666860200464725, + 0.00587631156668067, + 0.003674493869766593, + 0.0035670518409460783, + 0.0015496517298743129, + 0.0013970820000395179, + 0.005153815262019634, + 0.0050485869869589806, + 0.0024030101485550404, + 0.0032166405580937862, + 0.001915456959977746, + 0.0012772766640409827, + 0.0030723826494067907, + 0.0016351948725059628, + 0.002870902419090271, + 0.001778586651198566, + 0.0022761081345379353, + 0.0012528871884569526, + 0.0015823794528841972, + 0.0031468041706830263, + 0.0017287058290094137, + 0.0026781915221363306, + 0.003485405817627907, + 0.002534847240895033, + 0.0028369410429149866, + 0.0031101014465093613, + 0.0023285711649805307, + 0.010149204172194004, + 0.029371822252869606, + 0.020374193787574768, + 0.026478243991732597, + 0.06700481474399567, + 0.049486421048641205, + 0.2631142735481262, + 0.23512721061706543, + 0.04013910889625549, + 0.0026346405502408743, + 0.003022159915417433, + 0.003438579849898815, + 0.021636852994561195, + 0.00235866941511631, + 0.003265766194090247, + 0.02025585062801838, + 0.0031936336308717728 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 4 + ], + "selected_reward": -0.008455547504127026 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 dimly lit, classically styled study with a wall of floor-to-ceiling mahogany bookshelves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the bookshelves densely packed with hardbound volumes featuring leather covers and gilded spines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the wall of bookshelves rotate inward slowly while a small brass globe remains stable?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a bottomless, square entrance of absolute blackness gradually reveal itself as the shelves rotate?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static, medium-long shot composed from a level angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02062002755701542 + ], + "opposite_logprob_tokens": [ + -3.895620107650757 + ], + "logprob_reward": -0.02062002755701542, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.02062002755701542, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04972408339381218, + 0.036970898509025574, + 0.10529153794050217, + 0.08221662044525146, + 0.2296999841928482, + 0.11816654354333878, + 0.03449729084968567, + 0.13099157810211182, + 0.1899465173482895, + 0.022494958713650703 + ], + "credit_map": [ + 0.005163671914488077, + 0.005276082083582878, + 0.006693860981613398, + 0.006633175071328878, + 0.003235017415136099, + 0.00937685277312994, + 0.005954820197075605, + 0.007390601094812155, + 0.005814343225210905, + 0.0034505827352404594, + 0.005336933769285679, + 0.004132030066102743, + 0.0026315422728657722, + 0.0040041678585112095, + 0.004919318947941065, + 0.006681978702545166, + 0.0241155494004488, + 0.012369449250400066, + 0.02345235086977482, + 0.008909537456929684, + 0.00552617060020566, + 0.007266561035066843, + 0.00901843886822462, + 0.014633488841354847, + 0.010688802227377892, + 0.009392502717673779, + 0.011697391048073769, + 0.011474207043647766, + 0.007043155375868082, + 0.009565368294715881, + 0.009225737303495407, + 0.013129453174769878, + 0.01935485191643238, + 0.03966231644153595, + 0.02017928659915924, + 0.04519952833652496, + 0.024202777072787285, + 0.01978110708296299, + 0.018036413937807083, + 0.043283700942993164, + 0.01200722437351942, + 0.006149671971797943, + 0.027505872771143913, + 0.037776198238134384, + 0.007438349071890116, + 0.006745098624378443, + 0.012826603837311268, + 0.00771752605214715, + 0.0036361003294587135, + 0.003889441257342696, + 0.0050901928916573524, + 0.0046816132962703705, + 0.005987877491861582, + 0.0032661331351846457, + 0.003427294082939625, + 0.004518638364970684, + 0.008733845315873623, + 0.015387176536023617, + 0.015591131523251534, + 0.014822104945778847, + 0.010753397829830647, + 0.013927817344665527, + 0.011247956193983555, + 0.04052815958857536, + 0.008176525123417377, + 0.0060095577500760555, + 0.007910899817943573, + 0.018144018948078156, + 0.012751196511089802, + 0.06558393687009811, + 0.06004777178168297, + 0.011322620324790478, + 0.002090683439746499, + 0.0020319046452641487, + 0.0029659506399184465, + 0.0038345318753272295, + 0.0015988951781764627, + 0.004002179019153118, + 0.003582500386983156, + 0.0023883129470050335 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 8, + 7 + ], + "selected_reward": -0.02062002755701542 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 dimly lit, classically styled study with a wall of floor-to-ceiling mahogany bookshelves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the bookshelves densely packed with hardbound volumes featuring leather covers and gilded spines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the wall of bookshelves rotate inward slowly while a small brass globe remains stable?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a bottomless, square entrance of absolute blackness gradually reveal itself as the shelves rotate?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static, medium-long shot composed from a level angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07003693282604218 + ], + "opposite_logprob_tokens": [ + -2.6950368881225586 + ], + "logprob_reward": -0.07003693282604218, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07003693282604218, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10770483314990997, + 0.03485725820064545, + 0.055664386600255966, + 0.08252625167369843, + 0.1503710299730301, + 0.053345389664173126, + 0.04544264078140259, + 0.3164939284324646, + 0.06715808063745499, + 0.08643615990877151 + ], + "credit_map": [ + 0.015048322267830372, + 0.014194938354194164, + 0.012039273977279663, + 0.014123130589723587, + 0.0056352089159190655, + 0.016547415405511856, + 0.009996398352086544, + 0.02012014389038086, + 0.006211027968674898, + 0.004943526349961758, + 0.0054473052732646465, + 0.0037384892348200083, + 0.0025496617890894413, + 0.0033347688149660826, + 0.004074013326317072, + 0.00455846730619669, + 0.011201534420251846, + 0.00803396012634039, + 0.010301364585757256, + 0.005414793733507395, + 0.00325145130045712, + 0.004319978412240744, + 0.005445837043225765, + 0.007695471867918968, + 0.009190204553306103, + 0.007905786857008934, + 0.009472768753767014, + 0.01112267654389143, + 0.007673195097595453, + 0.008864309638738632, + 0.016950156539678574, + 0.011347162537276745, + 0.013112885877490044, + 0.02590135484933853, + 0.015349633991718292, + 0.02107168547809124, + 0.019182097166776657, + 0.02258186973631382, + 0.016460612416267395, + 0.016710905358195305, + 0.0033420799300074577, + 0.0035922217648476362, + 0.010321422480046749, + 0.013592958450317383, + 0.00923452340066433, + 0.005457553546875715, + 0.004029546398669481, + 0.0037750855553895235, + 0.004948385525494814, + 0.004466732498258352, + 0.008922303095459938, + 0.005314890760928392, + 0.0065460787154734135, + 0.0032989690080285072, + 0.0035362951457500458, + 0.008408988825976849, + 0.018553204834461212, + 0.031244903802871704, + 0.03825197368860245, + 0.030986078083515167, + 0.026564786210656166, + 0.03060184232890606, + 0.027218006551265717, + 0.11307314783334732, + 0.0031353647354990244, + 0.005406946875154972, + 0.006091671995818615, + 0.006942112930119038, + 0.007996082305908203, + 0.016598425805568695, + 0.015652474015951157, + 0.005335010122507811, + 0.0029285941272974014, + 0.003522491082549095, + 0.0036924979649484158, + 0.03341807797551155, + 0.002796251093968749, + 0.00441407784819603, + 0.03198283165693283, + 0.003681353759020567 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 4, + 0 + ], + "selected_reward": -0.07003693282604218 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 dimly lit, classically styled study with a wall of floor-to-ceiling mahogany bookshelves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the bookshelves densely packed with hardbound volumes featuring leather covers and gilded spines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the wall of bookshelves rotate inward slowly while a small brass globe remains stable?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a bottomless, square entrance of absolute blackness gradually reveal itself as the shelves rotate?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static, medium-long shot composed from a level angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016080567613244057 + ], + "opposite_logprob_tokens": [ + -4.141080379486084 + ], + "logprob_reward": -0.016080567613244057, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016080567613244057, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03967249393463135, + 0.01724270172417164, + 0.01626267284154892, + 0.01641201227903366, + 0.028456222265958786, + 0.03126009926199913, + 0.03539884090423584, + 0.11242108792066574, + 0.6677277684211731, + 0.0351460836827755 + ], + "credit_map": [ + 0.004510622471570969, + 0.004191211424767971, + 0.004734074231237173, + 0.004017113242298365, + 0.0022139661014080048, + 0.007213743403553963, + 0.004130957648158073, + 0.008660802617669106, + 0.002463341224938631, + 0.0015495208790525794, + 0.0025499912444502115, + 0.002318058628588915, + 0.0013887046370655298, + 0.002268363954499364, + 0.0023006161209195852, + 0.002404104685410857, + 0.002073311945423484, + 0.002753057051450014, + 0.0021572650875896215, + 0.0018385278526693583, + 0.0012322041438892484, + 0.001271724235266447, + 0.0025423916522413492, + 0.002394189592450857, + 0.0017586450558155775, + 0.0021587065421044827, + 0.0019748948980122805, + 0.002957929391413927, + 0.0015416726237162948, + 0.001947229728102684, + 0.0020244948100298643, + 0.002048440044745803, + 0.002534066094085574, + 0.0033615650609135628, + 0.0028615978080779314, + 0.004898251965641975, + 0.0044102855026721954, + 0.004306836053729057, + 0.0031929456163197756, + 0.0028906725347042084, + 0.0038605215959250927, + 0.0035278070718050003, + 0.004738999996334314, + 0.003223692299798131, + 0.004785358905792236, + 0.005528279114514589, + 0.003120916662737727, + 0.0024745233822613955, + 0.003506474196910858, + 0.0035146246664226055, + 0.007693518418818712, + 0.003674495732411742, + 0.005266290623694658, + 0.003034304827451706, + 0.003275955095887184, + 0.005433178506791592, + 0.009940020740032196, + 0.017895808443427086, + 0.013953682966530323, + 0.016442598775029182, + 0.013740180991590023, + 0.015478485263884068, + 0.012011437676846981, + 0.01295886654406786, + 0.027051396667957306, + 0.019107861444354057, + 0.023947501555085182, + 0.060023944824934006, + 0.043599095195531845, + 0.24213945865631104, + 0.21525956690311432, + 0.03659896180033684, + 0.00187968579120934, + 0.0027308666612952948, + 0.003861205419525504, + 0.009326677769422531, + 0.0021629314869642258, + 0.003198831109330058, + 0.009278726764023304, + 0.002707159146666527 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 0 + ], + "selected_reward": -0.016080567613244057 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 dimly lit, classically styled study with a wall of floor-to-ceiling mahogany bookshelves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the bookshelves densely packed with hardbound volumes featuring leather covers and gilded spines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the wall of bookshelves rotate inward slowly while a small brass globe remains stable?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a bottomless, square entrance of absolute blackness gradually reveal itself as the shelves rotate?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static, medium-long shot composed from a level angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009784476831555367 + ], + "opposite_logprob_tokens": [ + -4.634784698486328 + ], + "logprob_reward": -0.009784476831555367, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009784476831555367, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.036459244787693024, + 0.015986431390047073, + 0.042781028896570206, + 0.04725809395313263, + 0.0443899929523468, + 0.026305316016077995, + 0.02355910651385784, + 0.03806474804878235, + 0.5987437963485718, + 0.1264522224664688 + ], + "credit_map": [ + 0.004743903409689665, + 0.003644406795501709, + 0.005275229457765818, + 0.005672750994563103, + 0.0023492255713790655, + 0.0058504375629127026, + 0.004480876959860325, + 0.0044424147345125675, + 0.001863997895270586, + 0.0018609598046168685, + 0.0027036203537136316, + 0.0022065527737140656, + 0.0011443330440670252, + 0.002275059698149562, + 0.0019287287723273039, + 0.002003177534788847, + 0.009527556598186493, + 0.005967121105641127, + 0.008952409029006958, + 0.003622894175350666, + 0.0025590250734239817, + 0.00269629736430943, + 0.0032013331074267626, + 0.006254391744732857, + 0.00979890488088131, + 0.0046705217100679874, + 0.00572347454726696, + 0.0056680492125451565, + 0.00278839492239058, + 0.0035937230568379164, + 0.008613398298621178, + 0.006401627324521542, + 0.004482376854866743, + 0.007383367512375116, + 0.004389760084450245, + 0.006454330403357744, + 0.00556700024753809, + 0.005690762307494879, + 0.004250228870660067, + 0.006172169931232929, + 0.0018999023595824838, + 0.0020847655832767487, + 0.005438055377453566, + 0.0057625602930784225, + 0.0026462641544640064, + 0.003821931080892682, + 0.0029525007121264935, + 0.0016993347089737654, + 0.004002482630312443, + 0.0022268134634941816, + 0.003525993088260293, + 0.0024331696331501007, + 0.0033190737012773752, + 0.001993542304262519, + 0.002201439579948783, + 0.0038565911818295717, + 0.002968985354527831, + 0.0060443757101893425, + 0.004845031071454287, + 0.004796060733497143, + 0.003918524831533432, + 0.004039750434458256, + 0.0035448435228317976, + 0.007907175458967686, + 0.023635944351553917, + 0.016734302043914795, + 0.021922778338193893, + 0.05439003184437752, + 0.03981032967567444, + 0.21692268550395966, + 0.19201849400997162, + 0.033309224992990494, + 0.004483666270971298, + 0.004649828188121319, + 0.00497092492878437, + 0.05030331015586853, + 0.003923059441149235, + 0.006642226595431566, + 0.045811597257852554, + 0.0056676072999835014 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 3 + ], + "selected_reward": -0.009784476831555367 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, medium-long shot, shot from a level angle and composed with a steady eye, focuses on a dimly lit, classically styled study. A wall of floor-to-ceiling mahogany bookshelves dominates the view, densely packed with hardbound volumes. Their leather covers and gilded spines gleam softly and deeply in the warm, flickering glow of the off-screen fireplace. With a low, almost infrasonic sound of mechanical friction and rolling stone, the vast wall of shelves begins to solemnly and slowly rotate inward, its contents and a small brass globe remaining incredibly stable throughout the movement. As the shelves rotate, a bottomless, square entrance of absolute blackness gradually reveals itself. This darkness, seemingly devouring all light, creates a stark and dramatic contrast with the warm, orange-toned light within the study, creating a tense atmosphere filled with mystery, suspense, and a sense of uncharted territory.", + "seed": 28853, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000026/group_2/index_4783_seed_28853_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": -0.016499914042651654, + "vlm_logprob": -0.016499914042651654, + "vlm_yes_no_margin": 4.4, + "vlm_token_credit": -0.016499914042651654, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 dimly lit, classically styled study with a wall of floor-to-ceiling mahogany bookshelves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the bookshelves densely packed with hardbound volumes featuring leather covers and gilded spines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the wall of bookshelves rotate inward slowly while a small brass globe remains stable?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a bottomless, square entrance of absolute blackness gradually reveal itself as the shelves rotate?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static, medium-long shot composed from a level angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008462758734822273 + ], + "opposite_logprob_tokens": [ + -5.758462905883789 + ], + "logprob_reward": -0.008462758734822273, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.008462758734822273, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2834702134132385, + 0.08542729914188385, + 0.047773607075214386, + 0.08971600979566574, + 0.04938289150595665, + 0.06128205358982086, + 0.10764379799365997, + 0.044418465346097946, + 0.08155333250761032, + 0.14933227002620697 + ], + "credit_map": [ + 0.02465907484292984, + 0.0427483394742012, + 0.031125610694289207, + 0.03222488611936569, + 0.036063190549612045, + 0.04707527533173561, + 0.037808820605278015, + 0.03176503628492355, + 0.010113716125488281, + 0.009140674024820328, + 0.01009007915854454, + 0.011649458669126034, + 0.009226307272911072, + 0.011732048355042934, + 0.014271493069827557, + 0.009203532710671425, + 0.006629208568483591, + 0.0042273481376469135, + 0.005157019477337599, + 0.011070532724261284, + 0.0030423197895288467, + 0.0033105018083006144, + 0.007542734034359455, + 0.006793947424739599, + 0.011165618896484375, + 0.009193492121994495, + 0.011986959725618362, + 0.014720631763339043, + 0.0072961244732141495, + 0.009384294040501118, + 0.010915231890976429, + 0.015053668059408665, + 0.004816357512027025, + 0.005288901273161173, + 0.00764387147501111, + 0.012115568853914738, + 0.003339930670335889, + 0.004315501544624567, + 0.004607916809618473, + 0.007254844531416893, + 0.00760022085160017, + 0.007801947649568319, + 0.010280460119247437, + 0.010341274552047253, + 0.005293670576065779, + 0.006144774612039328, + 0.00737605569884181, + 0.006443653721362352, + 0.013629657216370106, + 0.008462906815111637, + 0.022493701428174973, + 0.020622387528419495, + 0.007816066965460777, + 0.00754530681297183, + 0.01412869431078434, + 0.012945082038640976, + 0.00496086897328496, + 0.003166083013638854, + 0.010522175580263138, + 0.004565950948745012, + 0.005822270642966032, + 0.0033393423072993755, + 0.006208224222064018, + 0.005833554081618786, + 0.007710345089435577, + 0.010006064549088478, + 0.018938494846224785, + 0.007981492206454277, + 0.007438856177031994, + 0.008082179352641106, + 0.012930181808769703, + 0.008465724065899849, + 0.013219884596765041, + 0.01299995742738247, + 0.02903088741004467, + 0.014499884098768234, + 0.019139863550662994, + 0.02409917302429676, + 0.021352069452404976, + 0.014990555122494698 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 6 + ], + "selected_reward": -0.008462758734822273 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 dimly lit, classically styled study with a wall of floor-to-ceiling mahogany bookshelves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the bookshelves densely packed with hardbound volumes featuring leather covers and gilded spines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the wall of bookshelves rotate inward slowly while a small brass globe remains stable?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a bottomless, square entrance of absolute blackness gradually reveal itself as the shelves rotate?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static, medium-long shot composed from a level angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020608698949217796 + ], + "opposite_logprob_tokens": [ + -3.89560866355896 + ], + "logprob_reward": -0.020608698949217796, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020608698949217796, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1781594306230545, + 0.0463910736143589, + 0.0972651019692421, + 0.089093416929245, + 0.06272345781326294, + 0.08109933882951736, + 0.14142636954784393, + 0.09609439969062805, + 0.06661371141672134, + 0.14113369584083557 + ], + "credit_map": [ + 0.014905581250786781, + 0.046704404056072235, + 0.015228310599923134, + 0.01875528320670128, + 0.0186441820114851, + 0.024215707555413246, + 0.023202337324619293, + 0.016503630205988884, + 0.006258832290768623, + 0.00640331394970417, + 0.005552926100790501, + 0.0057730055414140224, + 0.0046231565065681934, + 0.005837627686560154, + 0.006500022020190954, + 0.005442188587039709, + 0.013167515397071838, + 0.010455473326146603, + 0.011720728129148483, + 0.014822405762970448, + 0.0092057129368186, + 0.008033760823309422, + 0.012704254128038883, + 0.017155250534415245, + 0.01010818686336279, + 0.009040615521371365, + 0.012851253151893616, + 0.012440411373972893, + 0.007964861579239368, + 0.009740428999066353, + 0.014340857975184917, + 0.01260679867118597, + 0.007177385967224836, + 0.005008414853364229, + 0.012408330105245113, + 0.011881064623594284, + 0.005444370675832033, + 0.006640233099460602, + 0.00878951232880354, + 0.005374143365770578, + 0.008905594237148762, + 0.007502655498683453, + 0.01416672021150589, + 0.017177188768982887, + 0.006970610003918409, + 0.008792009204626083, + 0.006758527830243111, + 0.01082603819668293, + 0.020365172997117043, + 0.020631154999136925, + 0.02171679399907589, + 0.021021747961640358, + 0.01151745580136776, + 0.015425427816808224, + 0.013067116029560566, + 0.017681511119008064, + 0.012091288343071938, + 0.010969484224915504, + 0.013773022219538689, + 0.015111926011741161, + 0.011232457123696804, + 0.011543932370841503, + 0.009375547058880329, + 0.011996740475296974, + 0.007501377258449793, + 0.006545405834913254, + 0.013412836007773876, + 0.007571771740913391, + 0.005743670742958784, + 0.006450801156461239, + 0.012516111135482788, + 0.006871738936752081, + 0.012575040571391582, + 0.012714726850390434, + 0.021412495523691177, + 0.011232235468924046, + 0.021680496633052826, + 0.02768803760409355, + 0.01728443056344986, + 0.016546227037906647 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 9 + ], + "selected_reward": -0.020608698949217796 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 dimly lit, classically styled study with a wall of floor-to-ceiling mahogany bookshelves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the bookshelves densely packed with hardbound volumes featuring leather covers and gilded spines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the wall of bookshelves rotate inward slowly while a small brass globe remains stable?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a bottomless, square entrance of absolute blackness gradually reveal itself as the shelves rotate?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static, medium-long shot composed from a level angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029809892177581787 + ], + "opposite_logprob_tokens": [ + -3.5298099517822266 + ], + "logprob_reward": -0.029809892177581787, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029809892177581787, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2105989307165146, + 0.09040237963199615, + 0.1293061524629593, + 0.0702388808131218, + 0.1488795429468155, + 0.11334557086229324, + 0.09381277859210968, + 0.06470891088247299, + 0.04101785272359848, + 0.037689000368118286 + ], + "credit_map": [ + 0.016453886404633522, + 0.06410795450210571, + 0.017655149102211, + 0.028218727558851242, + 0.0254970695823431, + 0.020902441814541817, + 0.01996597833931446, + 0.017797715961933136, + 0.011664335615932941, + 0.00782269798219204, + 0.01966959424316883, + 0.009312358684837818, + 0.006936758290976286, + 0.019557639956474304, + 0.008938906714320183, + 0.006500089541077614, + 0.014881870709359646, + 0.01354469358921051, + 0.014819862321019173, + 0.02965688332915306, + 0.008535962551832199, + 0.01193949207663536, + 0.02238275110721588, + 0.013544633984565735, + 0.007121936418116093, + 0.008361628279089928, + 0.011851748451590538, + 0.016253873705863953, + 0.006421813741326332, + 0.005016816779971123, + 0.009202470071613789, + 0.006008598022162914, + 0.01996839977800846, + 0.020953316241502762, + 0.025785230100154877, + 0.033965837210416794, + 0.013258997350931168, + 0.012143708765506744, + 0.00807013176381588, + 0.014733925461769104, + 0.011189139448106289, + 0.009279862977564335, + 0.020882532000541687, + 0.0148278484120965, + 0.01654869318008423, + 0.012234463356435299, + 0.014432218857109547, + 0.013950815424323082, + 0.020461855456233025, + 0.009073681198060513, + 0.011916965246200562, + 0.01606242172420025, + 0.007228875067085028, + 0.006647174712270498, + 0.010423853062093258, + 0.011997956782579422, + 0.007349872030317783, + 0.004355601966381073, + 0.014566079713404179, + 0.008286092430353165, + 0.005130678415298462, + 0.0056091672740876675, + 0.009900066070258617, + 0.009511354379355907, + 0.006330288480967283, + 0.004857154097408056, + 0.007123855408281088, + 0.004567964468151331, + 0.002942581893876195, + 0.002340430859476328, + 0.00863627064973116, + 0.004219301976263523, + 0.004024790599942207, + 0.004279413726180792, + 0.005823660641908646, + 0.0040550571866333485, + 0.0040892912074923515, + 0.00559079647064209, + 0.004427438136190176, + 0.005398550070822239 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 2 + ], + "selected_reward": -0.029809892177581787 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a dimly lit, classically styled study with a wall of floor-to-ceiling mahogany bookshelves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the bookshelves densely packed with hardbound volumes featuring leather covers and gilded spines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the wall of bookshelves rotate inward slowly while a small brass globe remains stable?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a bottomless, square entrance of absolute blackness gradually reveal itself as the shelves rotate?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static, medium-long shot composed from a level angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012541710399091244 + ], + "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": [ + 0.11974239349365234, + 0.061470527201890945, + 0.0590561218559742, + 0.053191978484392166, + 0.13634979724884033, + 0.09767220914363861, + 0.0950629785656929, + 0.26520517468452454, + 0.052651546895504, + 0.059597279876470566 + ], + "credit_map": [ + 0.01388818584382534, + 0.021107472479343414, + 0.019001055508852005, + 0.01062450185418129, + 0.0138414790853858, + 0.016557486727833748, + 0.012176945805549622, + 0.012545269913971424, + 0.008075514808297157, + 0.005938335787504911, + 0.011593599803745747, + 0.007813237607479095, + 0.004315224010497332, + 0.008298489265143871, + 0.009588819928467274, + 0.0058473083190619946, + 0.007320392411202192, + 0.0055530257523059845, + 0.0073129418306052685, + 0.009881937876343727, + 0.007585007697343826, + 0.004132536705583334, + 0.008158981800079346, + 0.009111300110816956, + 0.005582516081631184, + 0.004663077648729086, + 0.008524016477167606, + 0.014146928675472736, + 0.0033082831650972366, + 0.004628458060324192, + 0.006295100785791874, + 0.006043597590178251, + 0.01671125367283821, + 0.014557763002812862, + 0.02563534490764141, + 0.04089934378862381, + 0.008533085696399212, + 0.009548784233629704, + 0.008005507290363312, + 0.01245870441198349, + 0.013530373573303223, + 0.010708032175898552, + 0.017498809844255447, + 0.019269149750471115, + 0.006772322580218315, + 0.01028617937117815, + 0.008762668818235397, + 0.010844670236110687, + 0.009491704404354095, + 0.009031706489622593, + 0.016079703345894814, + 0.01699269376695156, + 0.006755284499377012, + 0.009836663492023945, + 0.0099377715960145, + 0.016937440261244774, + 0.03296178951859474, + 0.018889224156737328, + 0.044700801372528076, + 0.03261316195130348, + 0.02510819397866726, + 0.021432669833302498, + 0.03914182260632515, + 0.05035750940442085, + 0.0058666798286139965, + 0.005993946455419064, + 0.008991059847176075, + 0.006560755893588066, + 0.004130921792238951, + 0.004793974105268717, + 0.009418378584086895, + 0.006895824801176786, + 0.006105959415435791, + 0.00493374140933156, + 0.011510265991091728, + 0.006280341651290655, + 0.006357979960739613, + 0.009098564274609089, + 0.008707032538950443, + 0.006603392772376537 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 4, + 0 + ], + "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 scene a dimly lit, classically styled study with a wall of floor-to-ceiling mahogany bookshelves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the bookshelves densely packed with hardbound volumes featuring leather covers and gilded spines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the wall of bookshelves rotate inward slowly while a small brass globe remains stable?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a bottomless, square entrance of absolute blackness gradually reveal itself as the shelves rotate?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static, medium-long shot composed from a level angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011076509952545166 + ], + "opposite_logprob_tokens": [ + -4.5110764503479 + ], + "logprob_reward": -0.011076509952545166, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011076509952545166, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10772519558668137, + 0.09295687079429626, + 0.09039261192083359, + 0.058998215943574905, + 0.08650685846805573, + 0.07796472311019897, + 0.18518052995204926, + 0.08190034329891205, + 0.0719408318400383, + 0.14643386006355286 + ], + "credit_map": [ + 0.013196132145822048, + 0.013916432857513428, + 0.017316000536084175, + 0.009622866287827492, + 0.011577618308365345, + 0.010969446040689945, + 0.018354589119553566, + 0.012772101908922195, + 0.00942852534353733, + 0.013401219621300697, + 0.015540681779384613, + 0.008713019080460072, + 0.009614662267267704, + 0.014432603493332863, + 0.013622133061289787, + 0.008204028941690922, + 0.010583311319351196, + 0.007256648503243923, + 0.00986555777490139, + 0.023495623841881752, + 0.0044718035496771336, + 0.006454526446759701, + 0.01899847202003002, + 0.009266662411391735, + 0.006213155575096607, + 0.007087056525051594, + 0.008606212213635445, + 0.011570503935217857, + 0.004985867999494076, + 0.007211458403617144, + 0.007041366305202246, + 0.0062825907953083515, + 0.011005834676325321, + 0.01284136064350605, + 0.013296146877110004, + 0.016469907015562057, + 0.007492735981941223, + 0.009061499498784542, + 0.007981685921549797, + 0.008357685059309006, + 0.010777330957353115, + 0.011010302230715752, + 0.012986592948436737, + 0.014582909643650055, + 0.004615258425474167, + 0.008239089511334896, + 0.0070409709587693214, + 0.008712268434464931, + 0.023652315139770508, + 0.016882378607988358, + 0.02577904425561428, + 0.04337933287024498, + 0.013547043316066265, + 0.011791971512138844, + 0.022640105336904526, + 0.027508322149515152, + 0.009220853447914124, + 0.005240350961685181, + 0.023103434592485428, + 0.006606120616197586, + 0.010333997197449207, + 0.005930024199187756, + 0.012688463553786278, + 0.008777100592851639, + 0.00919406209141016, + 0.004234619438648224, + 0.018569065257906914, + 0.005061238072812557, + 0.0043413820676505566, + 0.004373225383460522, + 0.017237059772014618, + 0.008930176496505737, + 0.012064525857567787, + 0.009630943648517132, + 0.02787352353334427, + 0.01445926632732153, + 0.020508384332060814, + 0.025367330759763718, + 0.02140098251402378, + 0.015128900296986103 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 9, + 0 + ], + "selected_reward": -0.011076509952545166 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, medium-long shot, shot from a level angle and composed with a steady eye, focuses on a dimly lit, classically styled study. A wall of floor-to-ceiling mahogany bookshelves dominates the view, densely packed with hardbound volumes. Their leather covers and gilded spines gleam softly and deeply in the warm, flickering glow of the off-screen fireplace. With a low, almost infrasonic sound of mechanical friction and rolling stone, the vast wall of shelves begins to solemnly and slowly rotate inward, its contents and a small brass globe remaining incredibly stable throughout the movement. As the shelves rotate, a bottomless, square entrance of absolute blackness gradually reveals itself. This darkness, seemingly devouring all light, creates a stark and dramatic contrast with the warm, orange-toned light within the study, creating a tense atmosphere filled with mystery, suspense, and a sense of uncharted territory.", + "seed": 28854, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000026/group_2/index_4783_seed_28854_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": -0.03099134359508753, + "vlm_logprob": -0.03099134359508753, + "vlm_yes_no_margin": 4.15, + "vlm_token_credit": -0.03099134359508753, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 dimly lit, classically styled study with a wall of floor-to-ceiling mahogany bookshelves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the bookshelves densely packed with hardbound volumes featuring leather covers and gilded spines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the wall of bookshelves rotate inward slowly while a small brass globe remains stable?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a bottomless, square entrance of absolute blackness gradually reveal itself as the shelves rotate?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static, medium-long shot composed from a level angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00889576692134142 + ], + "opposite_logprob_tokens": [ + -5.2588958740234375 + ], + "logprob_reward": -0.00889576692134142, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.00889576692134142, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1762847602367401, + 0.055287253111600876, + 0.02368933893740177, + 0.048244815319776535, + 0.03627890720963478, + 0.059136081486940384, + 0.42519184947013855, + 0.023743456229567528, + 0.053963035345077515, + 0.09818045794963837 + ], + "credit_map": [ + 0.02805008925497532, + 0.013950498774647713, + 0.0217747800052166, + 0.028407417237758636, + 0.010146742686629295, + 0.0181731004267931, + 0.030969930812716484, + 0.02481221593916416, + 0.008408677764236927, + 0.004853328224271536, + 0.007137888576835394, + 0.00781659223139286, + 0.005373326130211353, + 0.008329623378813267, + 0.00857391208410263, + 0.004793902859091759, + 0.0029485912527889013, + 0.00292055681347847, + 0.003989218734204769, + 0.003469855058938265, + 0.002740089548751712, + 0.0026269948575645685, + 0.0025703394785523415, + 0.0024236927274614573, + 0.008515724912285805, + 0.005472974386066198, + 0.005472837947309017, + 0.00357118621468544, + 0.010936960577964783, + 0.007261613383889198, + 0.003650116501376033, + 0.0033634034916758537, + 0.0039218515157699585, + 0.003932482562959194, + 0.005145528819411993, + 0.005629208870232105, + 0.0040814997628331184, + 0.005660719703882933, + 0.0040594469755887985, + 0.0038481689989566803, + 0.00817626342177391, + 0.005907109938561916, + 0.007401449605822563, + 0.009515397250652313, + 0.009687569923698902, + 0.004774583503603935, + 0.007159881759434938, + 0.006513824686408043, + 0.16018837690353394, + 0.16690921783447266, + 0.01702228933572769, + 0.02391311712563038, + 0.023760925978422165, + 0.015124213881790638, + 0.00923418253660202, + 0.009039538912475109, + 0.0037013839464634657, + 0.003598777111619711, + 0.003913422115147114, + 0.0026404215022921562, + 0.0018611830892041326, + 0.0019846793729811907, + 0.002930469112470746, + 0.0031131193973124027, + 0.008723380044102669, + 0.006819851230829954, + 0.005558101460337639, + 0.005297200288623571, + 0.006276002153754234, + 0.0067464690655469894, + 0.007996653206646442, + 0.006545378360897303, + 0.01752917282283306, + 0.010131112299859524, + 0.008996191434562206, + 0.011573861353099346, + 0.01002831943333149, + 0.013565524481236935, + 0.016469338908791542, + 0.009886940941214561 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 9 + ], + "selected_reward": -0.00889576692134142 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 dimly lit, classically styled study with a wall of floor-to-ceiling mahogany bookshelves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the bookshelves densely packed with hardbound volumes featuring leather covers and gilded spines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the wall of bookshelves rotate inward slowly while a small brass globe remains stable?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a bottomless, square entrance of absolute blackness gradually reveal itself as the shelves rotate?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static, medium-long shot composed from a level angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04310690239071846 + ], + "opposite_logprob_tokens": [ + -3.1681067943573 + ], + "logprob_reward": -0.04310690239071846, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04310690239071846, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4991145133972168, + 0.022257721051573753, + 0.023908620700240135, + 0.019048789516091347, + 0.029491042718291283, + 0.028904989361763, + 0.11094492673873901, + 0.011589961126446724, + 0.04756821319460869, + 0.2071712613105774 + ], + "credit_map": [ + 0.09005772322416306, + 0.03475601598620415, + 0.048359233886003494, + 0.10976504534482956, + 0.029028523713350296, + 0.052940722554922104, + 0.09949272871017456, + 0.03471451625227928, + 0.0030411162879318, + 0.0021172985434532166, + 0.0030204395297914743, + 0.0026734578423202038, + 0.0020258051808923483, + 0.004570353776216507, + 0.002880003536120057, + 0.0019292475190013647, + 0.003072402672842145, + 0.0034063002094626427, + 0.004162720404565334, + 0.0029750738758593798, + 0.002859764266759157, + 0.002775483066216111, + 0.002225665608420968, + 0.002431210596114397, + 0.002508502919226885, + 0.002942379331216216, + 0.0018866078462451696, + 0.0030233573634177446, + 0.003377912798896432, + 0.002196222310885787, + 0.0017820550128817558, + 0.0013317519333213568, + 0.0038497059140354395, + 0.0025465290527790785, + 0.002722284523770213, + 0.004506229888647795, + 0.006727089639753103, + 0.002870529657229781, + 0.003601966891437769, + 0.0026667064521461725, + 0.0084046870470047, + 0.003502096049487591, + 0.0027333220932632685, + 0.0028673841152340174, + 0.0016330836806446314, + 0.005947495810687542, + 0.0018370203906670213, + 0.001979901222512126, + 0.04446927830576897, + 0.04456935450434685, + 0.00414248974993825, + 0.004766248632222414, + 0.006290778052061796, + 0.003068026388064027, + 0.0020156227983534336, + 0.001623125746846199, + 0.0016201970865949988, + 0.0017227950738742948, + 0.0016190680908039212, + 0.0013992531457915902, + 0.0011752083664759994, + 0.0009986041113734245, + 0.001647192519158125, + 0.0014076426159590483, + 0.007128744386136532, + 0.005344097036868334, + 0.004512749146670103, + 0.0042295209132134914, + 0.006180743686854839, + 0.009132212027907372, + 0.005240464583039284, + 0.005799685604870319, + 0.028705162927508354, + 0.020098088309168816, + 0.017527084797620773, + 0.01418436598032713, + 0.017392074689269066, + 0.04474370926618576, + 0.03577108681201935, + 0.02874968759715557 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 6 + ], + "selected_reward": -0.04310690239071846 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 dimly lit, classically styled study with a wall of floor-to-ceiling mahogany bookshelves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the bookshelves densely packed with hardbound volumes featuring leather covers and gilded spines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the wall of bookshelves rotate inward slowly while a small brass globe remains stable?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a bottomless, square entrance of absolute blackness gradually reveal itself as the shelves rotate?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static, medium-long shot composed from a level angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08905156701803207 + ], + "opposite_logprob_tokens": [ + -2.4640514850616455 + ], + "logprob_reward": -0.08905156701803207, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08905156701803207, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19119343161582947, + 0.044750966131687164, + 0.03934597596526146, + 0.06889297813177109, + 0.02701549045741558, + 0.07412997633218765, + 0.02495654672384262, + 0.05999572202563286, + 0.10147605091333389, + 0.3682428300380707 + ], + "credit_map": [ + 0.03904911130666733, + 0.012058227322995663, + 0.025106651708483696, + 0.02473936416208744, + 0.00846623070538044, + 0.02204018458724022, + 0.027293309569358826, + 0.03244036063551903, + 0.0062785400077700615, + 0.003169469768181443, + 0.005334015935659409, + 0.005129741970449686, + 0.0038310110103338957, + 0.008617396466434002, + 0.008124956861138344, + 0.004265833180397749, + 0.005656407680362463, + 0.004393486771732569, + 0.007257089018821716, + 0.005055976565927267, + 0.0032533903140574694, + 0.0051404135301709175, + 0.0035624259617179632, + 0.005026789382100105, + 0.012883263640105724, + 0.007289994973689318, + 0.005000562407076359, + 0.004947919864207506, + 0.021100537851452827, + 0.008579986169934273, + 0.005150349345058203, + 0.003940361551940441, + 0.0027878156397491693, + 0.003120889188721776, + 0.003192123956978321, + 0.004769092425704002, + 0.004608534276485443, + 0.002284070709720254, + 0.003514515934512019, + 0.002738449489697814, + 0.01986030489206314, + 0.007835609838366508, + 0.00643958942964673, + 0.008098305203020573, + 0.0040377359837293625, + 0.015999790281057358, + 0.004382541868835688, + 0.007476102560758591, + 0.005393540486693382, + 0.0045278677716851234, + 0.0023080431856215, + 0.0018162274500355124, + 0.001884021214209497, + 0.0038161552511155605, + 0.0027017078828066587, + 0.002508981851860881, + 0.010943648405373096, + 0.010498682968318462, + 0.007404368370771408, + 0.005629613995552063, + 0.0045874337665736675, + 0.00523902103304863, + 0.005724961403757334, + 0.009967992082238197, + 0.015178598463535309, + 0.012139392085373402, + 0.008880289271473885, + 0.008919927291572094, + 0.01047143992036581, + 0.016575923189520836, + 0.016345586627721786, + 0.01296489592641592, + 0.05024629086256027, + 0.04034792631864548, + 0.03470084071159363, + 0.03021489828824997, + 0.0329478494822979, + 0.06951240450143814, + 0.05776198208332062, + 0.05251066014170647 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 8 + ], + "selected_reward": -0.08905156701803207 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 dimly lit, classically styled study with a wall of floor-to-ceiling mahogany bookshelves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the bookshelves densely packed with hardbound volumes featuring leather covers and gilded spines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the wall of bookshelves rotate inward slowly while a small brass globe remains stable?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a bottomless, square entrance of absolute blackness gradually reveal itself as the shelves rotate?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static, medium-long shot composed from a level angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00525664072483778 + ], + "opposite_logprob_tokens": [ + -5.255256652832031 + ], + "logprob_reward": -0.00525664072483778, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.00525664072483778, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.018276583403348923, + 0.01608770526945591, + 0.009231200441718102, + 0.007388224359601736, + 0.010778351686894894, + 0.02696051448583603, + 0.015891414135694504, + 0.008755705319344997, + 0.011288805864751339, + 0.8753415942192078 + ], + "credit_map": [ + 0.0025789341889321804, + 0.0015754102496430278, + 0.002611896488815546, + 0.0026015632320195436, + 0.001132229226641357, + 0.0027580419555306435, + 0.0023935509379953146, + 0.002624955028295517, + 0.002201702445745468, + 0.0011846717679873109, + 0.002632301300764084, + 0.0017874183831736445, + 0.001363090006634593, + 0.002858088817447424, + 0.00231413752771914, + 0.0017462936230003834, + 0.0012040683068335056, + 0.000979333883151412, + 0.0015222763177007437, + 0.001207487192004919, + 0.0007149488083086908, + 0.0015307534486055374, + 0.001066291006281972, + 0.0010060406057164073, + 0.0011767522664740682, + 0.0008090339251793921, + 0.0006364857545122504, + 0.0008979401900433004, + 0.001686405623331666, + 0.0008689386886544526, + 0.0005236508441157639, + 0.0007890161359682679, + 0.0014649590011686087, + 0.0010635893559083343, + 0.0010136232012882829, + 0.0013696093810722232, + 0.003090708516538143, + 0.0009271313901990652, + 0.001104454742744565, + 0.0007442752248607576, + 0.003074972191825509, + 0.002939847530797124, + 0.0034241732209920883, + 0.005085955839604139, + 0.00358613021671772, + 0.004291278775781393, + 0.0025123939849436283, + 0.0020457589998841286, + 0.004632945638149977, + 0.004319766070693731, + 0.0010622889967635274, + 0.0014607247430831194, + 0.001142958295531571, + 0.0016786027699708939, + 0.0007338480791077018, + 0.000860275817103684, + 0.0015120062744244933, + 0.0014929574681445956, + 0.0011984144803136587, + 0.0009434252278879285, + 0.0005824319669045508, + 0.0006742980913259089, + 0.0009875977411866188, + 0.0013645733706653118, + 0.0016830350505188107, + 0.0009674663888290524, + 0.0010135308839380741, + 0.0008347007678821683, + 0.0015541411703452468, + 0.001962232869118452, + 0.0019904030486941338, + 0.001283294172026217, + 0.11982546001672745, + 0.08195986598730087, + 0.06631547957658768, + 0.0560186430811882, + 0.07620237022638321, + 0.188667893409729, + 0.16280485689640045, + 0.12354690581560135 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 5, + 0 + ], + "selected_reward": -0.00525664072483778 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 dimly lit, classically styled study with a wall of floor-to-ceiling mahogany bookshelves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the bookshelves densely packed with hardbound volumes featuring leather covers and gilded spines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the wall of bookshelves rotate inward slowly while a small brass globe remains stable?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a bottomless, square entrance of absolute blackness gradually reveal itself as the shelves rotate?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static, medium-long shot composed from a level angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008645840920507908 + ], + "opposite_logprob_tokens": [ + -4.758646011352539 + ], + "logprob_reward": -0.008645840920507908, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008645840920507908, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19393301010131836, + 0.044816456735134125, + 0.02409091591835022, + 0.0302304457873106, + 0.03331775963306427, + 0.03991422429680824, + 0.10677175968885422, + 0.0601036436855793, + 0.1212737113237381, + 0.3455480933189392 + ], + "credit_map": [ + 0.03797423094511032, + 0.01402309164404869, + 0.01665770635008812, + 0.036543335765600204, + 0.012361710891127586, + 0.026909906417131424, + 0.03467943146824837, + 0.014783601276576519, + 0.006104560568928719, + 0.0037936284206807613, + 0.008391829207539558, + 0.004840636160224676, + 0.004211237654089928, + 0.007366082165390253, + 0.005370318423956633, + 0.004738163203001022, + 0.0037186210975050926, + 0.002278924221172929, + 0.004721282981336117, + 0.0028115203604102135, + 0.00233690207824111, + 0.003332301042973995, + 0.002391242887824774, + 0.00250012194737792, + 0.00492411432787776, + 0.0034833275713026524, + 0.002808788325637579, + 0.0019783126190304756, + 0.007722211070358753, + 0.0035556554794311523, + 0.0034785966854542494, + 0.0022794415708631277, + 0.0036092051304876804, + 0.002729841275140643, + 0.0039548552595078945, + 0.005653775297105312, + 0.005438680294901133, + 0.0030998331494629383, + 0.004154651891440153, + 0.0046769180335104465, + 0.004791674204170704, + 0.0047437516041100025, + 0.005876427516341209, + 0.0077607594430446625, + 0.004753386601805687, + 0.004571415949612856, + 0.0038189829792827368, + 0.003597827861085534, + 0.04157821089029312, + 0.041738323867321014, + 0.00412470567971468, + 0.005005732178688049, + 0.005807023961097002, + 0.0036585174966603518, + 0.0026688193902373314, + 0.0021904336754232645, + 0.012296579778194427, + 0.009215746074914932, + 0.008910927921533585, + 0.004602213390171528, + 0.006331233773380518, + 0.005370337516069412, + 0.004948134534060955, + 0.008428470231592655, + 0.019176311790943146, + 0.01377489510923624, + 0.01229871902614832, + 0.009938343428075314, + 0.014063311740756035, + 0.02080019749701023, + 0.016888171434402466, + 0.014333763159811497, + 0.04185936227440834, + 0.036711983382701874, + 0.027463629841804504, + 0.024699822068214417, + 0.034658972173929214, + 0.06647644937038422, + 0.06205069646239281, + 0.05162717029452324 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 8 + ], + "selected_reward": -0.008645840920507908 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, medium-long shot, shot from a level angle and composed with a steady eye, focuses on a dimly lit, classically styled study. A wall of floor-to-ceiling mahogany bookshelves dominates the view, densely packed with hardbound volumes. Their leather covers and gilded spines gleam softly and deeply in the warm, flickering glow of the off-screen fireplace. With a low, almost infrasonic sound of mechanical friction and rolling stone, the vast wall of shelves begins to solemnly and slowly rotate inward, its contents and a small brass globe remaining incredibly stable throughout the movement. As the shelves rotate, a bottomless, square entrance of absolute blackness gradually reveals itself. This darkness, seemingly devouring all light, creates a stark and dramatic contrast with the warm, orange-toned light within the study, creating a tense atmosphere filled with mystery, suspense, and a sense of uncharted territory.", + "seed": 28855, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000026/group_2/index_4783_seed_28855_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": -0.050407090410590175, + "vlm_logprob": -0.050407090410590175, + "vlm_yes_no_margin": 3.65, + "vlm_token_credit": -0.050407090410590175, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 dimly lit, classically styled study with a wall of floor-to-ceiling mahogany bookshelves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the bookshelves densely packed with hardbound volumes featuring leather covers and gilded spines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the wall of bookshelves rotate inward slowly while a small brass globe remains stable?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a bottomless, square entrance of absolute blackness gradually reveal itself as the shelves rotate?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static, medium-long shot composed from a level angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.024740388616919518 + ], + "opposite_logprob_tokens": [ + -4.149740219116211 + ], + "logprob_reward": -0.024740388616919518, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.024740388616919518, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15249496698379517, + 0.08535776287317276, + 0.059289224445819855, + 0.081584133207798, + 0.05465240404009819, + 0.19961777329444885, + 0.05753064155578613, + 0.061879292130470276, + 0.06904623657464981, + 0.17854756116867065 + ], + "credit_map": [ + 0.008488675579428673, + 0.01745452731847763, + 0.01774442009627819, + 0.021630382165312767, + 0.02326379530131817, + 0.017073292285203934, + 0.028864039108157158, + 0.017975836992263794, + 0.007356375455856323, + 0.007654584012925625, + 0.009224584326148033, + 0.014501448720693588, + 0.008535144850611687, + 0.0089032631367445, + 0.015834761783480644, + 0.013347595930099487, + 0.007182438857853413, + 0.004895183723419905, + 0.006358764134347439, + 0.009539552964270115, + 0.00700498279184103, + 0.007384670898318291, + 0.007181623484939337, + 0.009742005728185177, + 0.006398674100637436, + 0.008350590243935585, + 0.006853834260255098, + 0.009845566004514694, + 0.015179932117462158, + 0.009050236083567142, + 0.008613836951553822, + 0.017291458323597908, + 0.0060513997450470924, + 0.006253544241189957, + 0.005289511755108833, + 0.008027552627027035, + 0.006689511705189943, + 0.00954475812613964, + 0.00502063799649477, + 0.0077754901722073555, + 0.019816914573311806, + 0.011970271356403828, + 0.0138253690674901, + 0.015688737854361534, + 0.03004581294953823, + 0.05081324651837349, + 0.018093179911375046, + 0.03936424478888512, + 0.00576672051101923, + 0.003781559644266963, + 0.006710395682603121, + 0.0031300184782594442, + 0.01329960860311985, + 0.012404624372720718, + 0.005891589913517237, + 0.006546123884618282, + 0.005588100291788578, + 0.007920129224658012, + 0.008851677179336548, + 0.005273755639791489, + 0.0119983134791255, + 0.006105578504502773, + 0.008280032314360142, + 0.007861709222197533, + 0.008160601370036602, + 0.008960498496890068, + 0.005727874580770731, + 0.006796941161155701, + 0.007481793407350779, + 0.016069559380412102, + 0.006287341471761465, + 0.009561629965901375, + 0.027290944010019302, + 0.020941708236932755, + 0.02754225581884384, + 0.021892212331295013, + 0.021755119785666466, + 0.018857868388295174, + 0.01614564098417759, + 0.024121824651956558 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 9, + 0 + ], + "selected_reward": -0.024740388616919518 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 dimly lit, classically styled study with a wall of floor-to-ceiling mahogany bookshelves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the bookshelves densely packed with hardbound volumes featuring leather covers and gilded spines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the wall of bookshelves rotate inward slowly while a small brass globe remains stable?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a bottomless, square entrance of absolute blackness gradually reveal itself as the shelves rotate?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static, medium-long shot composed from a level angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012594802305102348 + ], + "opposite_logprob_tokens": [ + -4.387594699859619 + ], + "logprob_reward": -0.012594802305102348, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012594802305102348, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10628549009561539, + 0.06234758719801903, + 0.06816870719194412, + 0.47428759932518005, + 0.060661181807518005, + 0.02811390906572342, + 0.023576291278004646, + 0.03985641151666641, + 0.03899678960442543, + 0.09770605713129044 + ], + "credit_map": [ + 0.010780011303722858, + 0.012400887906551361, + 0.012191747315227985, + 0.01446795929223299, + 0.013578370213508606, + 0.011546655558049679, + 0.019514549523591995, + 0.011805313639342785, + 0.006422995124012232, + 0.006383505649864674, + 0.008916017599403858, + 0.011022573336958885, + 0.006450044922530651, + 0.005146228242665529, + 0.01084507443010807, + 0.007161142770200968, + 0.007989995181560516, + 0.0055381967686116695, + 0.00802705530077219, + 0.009322854690253735, + 0.007361188065260649, + 0.008967272937297821, + 0.009753566235303879, + 0.011208580806851387, + 0.039771243929862976, + 0.04559257626533508, + 0.045559242367744446, + 0.08400256931781769, + 0.04571130871772766, + 0.07781894505023956, + 0.06118718907237053, + 0.07464452087879181, + 0.0053833830170333385, + 0.00585217634215951, + 0.0057303449138998985, + 0.0038686050102114677, + 0.006781465839594603, + 0.014030812308192253, + 0.007834040559828281, + 0.011180352419614792, + 0.002536014188081026, + 0.002231014659628272, + 0.0020202212035655975, + 0.0021598455496132374, + 0.004982572980225086, + 0.0064370157197117805, + 0.0020326594822108746, + 0.005714563652873039, + 0.0027334652841091156, + 0.0014813251327723265, + 0.002182750729843974, + 0.0013657648814842105, + 0.003379101399332285, + 0.005084020551294088, + 0.0030886875465512276, + 0.004261175636202097, + 0.003410561475902796, + 0.003966460004448891, + 0.006393002346158028, + 0.0033254632726311684, + 0.0053551411256194115, + 0.0033664600923657417, + 0.009191513992846012, + 0.004847806878387928, + 0.005002595949918032, + 0.0037497258745133877, + 0.0029310088139027357, + 0.0033650060649961233, + 0.003634118242189288, + 0.009819955565035343, + 0.003163674846291542, + 0.007330705877393484, + 0.013110760599374771, + 0.010185042396187782, + 0.013241149485111237, + 0.012146142311394215, + 0.019043173640966415, + 0.010139591060578823, + 0.009463870897889137, + 0.010376321151852608 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 9 + ], + "selected_reward": -0.012594802305102348 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 dimly lit, classically styled study with a wall of floor-to-ceiling mahogany bookshelves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the bookshelves densely packed with hardbound volumes featuring leather covers and gilded spines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the wall of bookshelves rotate inward slowly while a small brass globe remains stable?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a bottomless, square entrance of absolute blackness gradually reveal itself as the shelves rotate?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static, medium-long shot composed from a level angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.17990727722644806 + ], + "opposite_logprob_tokens": [ + -1.8049073219299316 + ], + "logprob_reward": -0.17990727722644806, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.17990727722644806, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.039650894701480865, + 0.05143594369292259, + 0.027777189388871193, + 0.055112581700086594, + 0.028275176882743835, + 0.12740562856197357, + 0.03475203365087509, + 0.0414838008582592, + 0.17089207470417023, + 0.4232146739959717 + ], + "credit_map": [ + 0.0034240323584526777, + 0.0042370762676000595, + 0.0054076965898275375, + 0.005802381318062544, + 0.006065930239856243, + 0.005487792659550905, + 0.0046235742047429085, + 0.004602412227541208, + 0.005794330965727568, + 0.005240779835730791, + 0.006895785219967365, + 0.009491732344031334, + 0.005798885133117437, + 0.004655098542571068, + 0.008659048937261105, + 0.004900286439806223, + 0.0031997880432754755, + 0.002219092333689332, + 0.002754488494247198, + 0.0028371296357363462, + 0.0030759640503674746, + 0.00607238570228219, + 0.002576046157628298, + 0.005042295902967453, + 0.003912580199539661, + 0.004162763711065054, + 0.005472064949572086, + 0.004480513744056225, + 0.01179389376193285, + 0.011558630503714085, + 0.005596028175204992, + 0.00813610665500164, + 0.0024639905896037817, + 0.0025313510559499264, + 0.00238509732298553, + 0.0032630593050271273, + 0.0027560468297451735, + 0.007340645883232355, + 0.00276098120957613, + 0.004774005152285099, + 0.012921343557536602, + 0.007717138156294823, + 0.008385391905903816, + 0.009829770773649216, + 0.019460590556263924, + 0.03304152190685272, + 0.010959692299365997, + 0.025090187788009644, + 0.0036909847985953093, + 0.0030043807346373796, + 0.002733682980760932, + 0.002200212562456727, + 0.005462998524308205, + 0.009119954891502857, + 0.0033665806986391544, + 0.005173238459974527, + 0.004991029389202595, + 0.0035708826035261154, + 0.004823584109544754, + 0.0033676642924547195, + 0.0045145233161747456, + 0.004436119459569454, + 0.0068415505811572075, + 0.008938443847000599, + 0.02391038089990616, + 0.018348654732108116, + 0.016094157472252846, + 0.017251115292310715, + 0.015336118638515472, + 0.033212821930646896, + 0.016324589028954506, + 0.030414249747991562, + 0.055782366544008255, + 0.04383796080946922, + 0.05761023238301277, + 0.053675778210163116, + 0.08153053373098373, + 0.04351973906159401, + 0.04183268919587135, + 0.04542537406086922 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 5 + ], + "selected_reward": -0.17990727722644806 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 dimly lit, classically styled study with a wall of floor-to-ceiling mahogany bookshelves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the bookshelves densely packed with hardbound volumes featuring leather covers and gilded spines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the wall of bookshelves rotate inward slowly while a small brass globe remains stable?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a bottomless, square entrance of absolute blackness gradually reveal itself as the shelves rotate?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static, medium-long shot composed from a level angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014207523316144943 + ], + "opposite_logprob_tokens": [ + -4.264207363128662 + ], + "logprob_reward": -0.014207523316144943, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014207523316144943, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.057822275906801224, + 0.06815723329782486, + 0.028403397649526596, + 0.059896983206272125, + 0.04987470060586929, + 0.08019024133682251, + 0.0450255461037159, + 0.06398099660873413, + 0.2505623996257782, + 0.29608625173568726 + ], + "credit_map": [ + 0.0066433739848434925, + 0.005265612620860338, + 0.007565436419099569, + 0.008565299212932587, + 0.006948651745915413, + 0.00575614720582962, + 0.010117508471012115, + 0.006960246246308088, + 0.007534203585237265, + 0.006882564164698124, + 0.0128179294988513, + 0.009666246362030506, + 0.008159727789461613, + 0.005820724181830883, + 0.010756837204098701, + 0.006518996320664883, + 0.0029317494481801987, + 0.002814369974657893, + 0.0026485479902476072, + 0.004336565732955933, + 0.0039498088881373405, + 0.004263292066752911, + 0.0031515718437731266, + 0.004307491239160299, + 0.004783863667398691, + 0.005231066606938839, + 0.0062171053141355515, + 0.009155411273241043, + 0.006088251248002052, + 0.009844493120908737, + 0.009051566943526268, + 0.009525224566459656, + 0.005145107861608267, + 0.004631614778190851, + 0.005087460856884718, + 0.005062874872237444, + 0.005298304837197065, + 0.012877831235527992, + 0.005251995753496885, + 0.006519508548080921, + 0.009398197755217552, + 0.006068678572773933, + 0.005810234230011702, + 0.007136262487620115, + 0.01292707584798336, + 0.017674023285508156, + 0.0076441336423158646, + 0.013531642034649849, + 0.003768791211768985, + 0.004700985737144947, + 0.0055596064776182175, + 0.005652421619743109, + 0.004926750436425209, + 0.006176539231091738, + 0.008365923538804054, + 0.00587452994659543, + 0.007533712778240442, + 0.005539676174521446, + 0.010055619291961193, + 0.005242697894573212, + 0.005124233663082123, + 0.005761763546615839, + 0.012865723110735416, + 0.011857567355036736, + 0.034447867423295975, + 0.023958634585142136, + 0.023427462205290794, + 0.02512764185667038, + 0.02154122292995453, + 0.05172628536820412, + 0.023649774491786957, + 0.04668351262807846, + 0.03968917950987816, + 0.031111054122447968, + 0.03942684084177017, + 0.03747217357158661, + 0.05761658027768135, + 0.030717279762029648, + 0.02907409518957138, + 0.03097904659807682 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 5 + ], + "selected_reward": -0.014207523316144943 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 dimly lit, classically styled study with a wall of floor-to-ceiling mahogany bookshelves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the bookshelves densely packed with hardbound volumes featuring leather covers and gilded spines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the wall of bookshelves rotate inward slowly while a small brass globe remains stable?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a bottomless, square entrance of absolute blackness gradually reveal itself as the shelves rotate?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static, medium-long shot composed from a level angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02058546058833599 + ], + "opposite_logprob_tokens": [ + -3.895585536956787 + ], + "logprob_reward": -0.02058546058833599, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.02058546058833599, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08593108505010605, + 0.11270198225975037, + 0.025786666199564934, + 0.11932695657014847, + 0.04171019792556763, + 0.3147129714488983, + 0.025690769776701927, + 0.02347910776734352, + 0.18107512593269348, + 0.06958521157503128 + ], + "credit_map": [ + 0.007167392410337925, + 0.01064394786953926, + 0.008306167088449001, + 0.01148721668869257, + 0.012110563926398754, + 0.01141880638897419, + 0.014494987204670906, + 0.010302002541720867, + 0.01284036785364151, + 0.00913218129426241, + 0.022589679807424545, + 0.015170742757618427, + 0.01649962179362774, + 0.011773176491260529, + 0.01464912761002779, + 0.010047086514532566, + 0.0028167052660137415, + 0.0022085208911448717, + 0.0029719166923314333, + 0.004074194468557835, + 0.0027208090759813786, + 0.0038229317869991064, + 0.0031540978234261274, + 0.0040174867026507854, + 0.0098502729088068, + 0.01122504286468029, + 0.009231853298842907, + 0.009708389639854431, + 0.024817613884806633, + 0.026063116267323494, + 0.010530306957662106, + 0.01790035143494606, + 0.0036083690356463194, + 0.003504377556964755, + 0.0044481693767011166, + 0.00487876683473587, + 0.005575717892497778, + 0.008997754193842411, + 0.005060939118266106, + 0.005636102519929409, + 0.03279606252908707, + 0.019718198105692863, + 0.022268816828727722, + 0.025586457923054695, + 0.0468941293656826, + 0.07744795829057693, + 0.029126012697815895, + 0.060875311493873596, + 0.0025533591397106647, + 0.0019043739885091782, + 0.0022763595916330814, + 0.002003807108849287, + 0.0029709399677813053, + 0.004297020845115185, + 0.0054970113560557365, + 0.004187894985079765, + 0.0025274092331528664, + 0.0033954859245568514, + 0.0021491048391908407, + 0.002218506531789899, + 0.003598902141675353, + 0.002393196104094386, + 0.0028664579149335623, + 0.004330042749643326, + 0.025472475215792656, + 0.012759982608258724, + 0.014362996444106102, + 0.017392855137586594, + 0.019559545442461967, + 0.02907484956085682, + 0.017942603677511215, + 0.04450978338718414, + 0.00780970323830843, + 0.006208430044353008, + 0.00815634336322546, + 0.007492356933653355, + 0.014045034535229206, + 0.007838760502636433, + 0.007161407731473446, + 0.010873174294829369 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 8, + 3 + ], + "selected_reward": -0.02058546058833599 + } + ], + "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/0000026/group_5/index_tensor([2446, 2446])_group_5_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000026/group_5/index_tensor([2446, 2446])_group_5_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..22cb619ccbfc2901debf3361aee8df6adb410e0a --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000026/group_5/index_tensor([2446, 2446])_group_5_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "An extreme close-up shot from a slightly elevated angle remains still, intensely focused on a small, round, white pill, perched precariously in the center of a trembling hand. The scene is bathed in dramatic, low-key lighting, with a single, harsh light source from the side carving the hand out from the deep, ambiguous darkness of the background. This stark contrast of black and white creates a stark contrast, casting long, deep shadows within the creases of the palm while capturing sharp highlights on the pale skin of the knuckles, the edges of the nails, and the smooth, chalky surface of the pill. The fingers, curled slightly inward, emit an uncontrollable, tiny vibration, noticeably accentuated by the play of light and shadow. Their unsteady posture suggests a profound inner struggle. A cool warmth permeates the entire image, heightening the tense, somber, and deeply personal atmosphere of quiet despair and momentous decision.", + "seed": 28430, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000026/group_5/index_2446_seed_28430_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": -0.14323142080102116, + "vlm_logprob": -0.14323142080102116, + "vlm_yes_no_margin": 3.875, + "vlm_token_credit": -0.14323142080102116, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 small, round, white pill?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the pill perched in the center of a trembling hand with fingers curled slightly inward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene lit by a single, harsh light source from the side creating high contrast?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background deep and ambiguous darkness with long, deep shadows in the palm creases?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera angle slightly elevated and completely still throughout the 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": [ + -0.010881843976676464 + ], + "opposite_logprob_tokens": [ + -5.5108819007873535 + ], + "logprob_reward": -0.010881843976676464, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.010881843976676464, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21210967004299164, + 0.08011490106582642, + 0.05136353522539139, + 0.05222350358963013, + 0.07578295469284058, + 0.05860103666782379, + 0.07813670486211777, + 0.07539600133895874, + 0.13706400990486145, + 0.17920763790607452 + ], + "credit_map": [ + 0.01816423609852791, + 0.02121148630976677, + 0.013791457749903202, + 0.023955995216965675, + 0.025560816749930382, + 0.027982043102383614, + 0.04823654890060425, + 0.03320710361003876, + 0.009889285080134869, + 0.006901768036186695, + 0.005760170519351959, + 0.011278136633336544, + 0.00918499194085598, + 0.009704695083200932, + 0.018804164603352547, + 0.008591693826019764, + 0.004788786172866821, + 0.00654575414955616, + 0.004509622696787119, + 0.00660876976326108, + 0.00743149546906352, + 0.007560526020824909, + 0.009070292115211487, + 0.004848294425755739, + 0.0039359163492918015, + 0.006271583493798971, + 0.004039733204990625, + 0.009323296137154102, + 0.005870030727237463, + 0.007214253768324852, + 0.008708992972970009, + 0.006859702989459038, + 0.006322863977402449, + 0.00804104097187519, + 0.004871175158768892, + 0.011384519748389721, + 0.007577441167086363, + 0.009467530995607376, + 0.01705949753522873, + 0.011058897711336613, + 0.005447094794362783, + 0.006251070648431778, + 0.0042838542722165585, + 0.010074256919324398, + 0.006685612723231316, + 0.007092494983226061, + 0.0100813964381814, + 0.008685259148478508, + 0.010233390145003796, + 0.0066452836617827415, + 0.005201032385230064, + 0.008710591122508049, + 0.0063113062642514706, + 0.011956153437495232, + 0.018377196043729782, + 0.01070176251232624, + 0.01267909724265337, + 0.009118339978158474, + 0.005090363789349794, + 0.008643965236842632, + 0.008625194430351257, + 0.011041439138352871, + 0.009656327776610851, + 0.010541279800236225, + 0.010976461693644524, + 0.013673338107764721, + 0.007963930256664753, + 0.015085470862686634, + 0.014377236366271973, + 0.019394123926758766, + 0.04125208035111427, + 0.014341384172439575, + 0.02867751568555832, + 0.014548525214195251, + 0.018092574551701546, + 0.02144424058496952, + 0.014789415523409843, + 0.02646494284272194, + 0.033705808222293854, + 0.02148464135825634 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 8 + ], + "selected_reward": -0.010881843976676464 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 small, round, white pill?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the pill perched in the center of a trembling hand with fingers curled slightly inward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene lit by a single, harsh light source from the side creating high contrast?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background deep and ambiguous darkness with long, deep shadows in the palm creases?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera angle slightly elevated and completely still throughout the 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": [ + -0.0036264387890696526 + ], + "opposite_logprob_tokens": [ + -5.628626346588135 + ], + "logprob_reward": -0.0036264387890696526, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.0036264387890696526, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12098909914493561, + 0.07491365820169449, + 0.03681072220206261, + 0.028643539175391197, + 0.05271269381046295, + 0.025492409244179726, + 0.04755435511469841, + 0.15156807005405426, + 0.312623530626297, + 0.14869190752506256 + ], + "credit_map": [ + 0.012896468862891197, + 0.01693600043654442, + 0.008469752967357635, + 0.01691582053899765, + 0.018474668264389038, + 0.012911546975374222, + 0.017773238942027092, + 0.016611594706773758, + 0.00626315688714385, + 0.007936383597552776, + 0.00546222971752286, + 0.013892811723053455, + 0.0065080695785582066, + 0.009762544184923172, + 0.011869275011122227, + 0.013219188898801804, + 0.004829654470086098, + 0.004100229125469923, + 0.003289170330390334, + 0.004029214847832918, + 0.0040224171243608, + 0.005438263062387705, + 0.006319636013358831, + 0.004782140254974365, + 0.0027186735533177853, + 0.002905006054788828, + 0.0018496038392186165, + 0.00420024897903204, + 0.003740172367542982, + 0.0037829983048141003, + 0.005488540045917034, + 0.003958294168114662, + 0.004946048837155104, + 0.00516916811466217, + 0.0034568989649415016, + 0.0077588739804923534, + 0.007268107030540705, + 0.006226710509508848, + 0.009263266809284687, + 0.008623615838587284, + 0.002696502488106489, + 0.0032499898225069046, + 0.0016194864874705672, + 0.0039673782885074615, + 0.0025646239519119263, + 0.0029559931717813015, + 0.004727434366941452, + 0.0037110012490302324, + 0.0036877936217933893, + 0.006122751627117395, + 0.003173376666381955, + 0.007076993118971586, + 0.004714055918157101, + 0.007332008332014084, + 0.010227635502815247, + 0.00521974079310894, + 0.01081063412129879, + 0.008511090651154518, + 0.011084573343396187, + 0.023300260305404663, + 0.009576826356351376, + 0.02507094107568264, + 0.045465607196092606, + 0.017748145386576653, + 0.02339182049036026, + 0.025485312566161156, + 0.022854279726743698, + 0.03390482813119888, + 0.02487185411155224, + 0.04614567756652832, + 0.09229732304811478, + 0.043672457337379456, + 0.013071020133793354, + 0.013196448795497417, + 0.01640118844807148, + 0.02054758369922638, + 0.0139771094545722, + 0.023778898641467094, + 0.027190376073122025, + 0.020529288798570633 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 9 + ], + "selected_reward": -0.0036264387890696526 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 small, round, white pill?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the pill perched in the center of a trembling hand with fingers curled slightly inward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene lit by a single, harsh light source from the side creating high contrast?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background deep and ambiguous darkness with long, deep shadows in the palm creases?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera angle slightly elevated and completely still throughout the 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": [ + -0.22561755776405334 + ], + "opposite_logprob_tokens": [ + -1.600617527961731 + ], + "logprob_reward": -0.22561755776405334, + "yes_no_margin_reward": 1.375, + "token_credit_reward": -0.22561755776405334, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14247290790081024, + 0.05507706478238106, + 0.04139525443315506, + 0.016298841685056686, + 0.02764333412051201, + 0.027729718014597893, + 0.035482387989759445, + 0.16251245141029358, + 0.24942739307880402, + 0.24196062982082367 + ], + "credit_map": [ + 0.01453733816742897, + 0.014698795042932034, + 0.007875715382397175, + 0.021741202101111412, + 0.02109912969172001, + 0.017462853342294693, + 0.02707783505320549, + 0.017980031669139862, + 0.008579549379646778, + 0.007005738560110331, + 0.004215365741401911, + 0.008296094834804535, + 0.004203035030514002, + 0.007580648176372051, + 0.007426131516695023, + 0.00777050107717514, + 0.004926185589283705, + 0.004918497987091541, + 0.0038065395783632994, + 0.006298177409917116, + 0.00479546282440424, + 0.005531713832169771, + 0.005698627792298794, + 0.005420051980763674, + 0.0017491485923528671, + 0.0017578911501914263, + 0.0013973464956507087, + 0.0018662401707842946, + 0.0025008281227201223, + 0.0018763297703117132, + 0.0029490538872778416, + 0.0022020037285983562, + 0.002607925795018673, + 0.002948177745565772, + 0.002408792730420828, + 0.0038422795478254557, + 0.00362992100417614, + 0.004072338808327913, + 0.004739981610327959, + 0.00339391827583313, + 0.0029636730905622244, + 0.00320040644146502, + 0.0030134839471429586, + 0.004485075827687979, + 0.0023587695322930813, + 0.0037670580204576254, + 0.00355811626650393, + 0.004383135586977005, + 0.003778995480388403, + 0.0043382784351706505, + 0.0029511204920709133, + 0.003853129455819726, + 0.00413941964507103, + 0.0054416349157691, + 0.007507560774683952, + 0.003472249023616314, + 0.01799568347632885, + 0.01714406907558441, + 0.013194107450544834, + 0.023399047553539276, + 0.01814674772322178, + 0.026497140526771545, + 0.02567741461098194, + 0.020458240061998367, + 0.021981123834848404, + 0.03022356890141964, + 0.020871810615062714, + 0.033517587929964066, + 0.026330025866627693, + 0.03623760864138603, + 0.03789498284459114, + 0.04237069562077522, + 0.028653690591454506, + 0.023520901799201965, + 0.01912105642259121, + 0.033241696655750275, + 0.02280360832810402, + 0.03720839321613312, + 0.0452759824693203, + 0.032135311514139175 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 7 + ], + "selected_reward": -0.22561755776405334 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 small, round, white pill?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the pill perched in the center of a trembling hand with fingers curled slightly inward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene lit by a single, harsh light source from the side creating high contrast?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background deep and ambiguous darkness with long, deep shadows in the palm creases?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera angle slightly elevated and completely still throughout the 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": [ + -0.0017144280718639493 + ], + "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": [ + 0.16583725810050964, + 0.0529346764087677, + 0.03925386443734169, + 0.03337514027953148, + 0.047105029225349426, + 0.05293815955519676, + 0.04296693578362465, + 0.16426067054271698, + 0.191063791513443, + 0.2102644294500351 + ], + "credit_map": [ + 0.015886398032307625, + 0.021043460816144943, + 0.008484682068228722, + 0.02664441056549549, + 0.026397699490189552, + 0.020881909877061844, + 0.0252773929387331, + 0.02122132107615471, + 0.0077734352089464664, + 0.007062577176839113, + 0.004085632506757975, + 0.007623227778822184, + 0.004730491898953915, + 0.006828824058175087, + 0.007730509154498577, + 0.007099987473338842, + 0.0048135616816580296, + 0.005001791287213564, + 0.0034535638988018036, + 0.005372608546167612, + 0.004320048727095127, + 0.005548525135964155, + 0.006190694868564606, + 0.004553072154521942, + 0.003787529654800892, + 0.004472279455512762, + 0.002613632706925273, + 0.0047109127044677734, + 0.0050377254374325275, + 0.004318762104958296, + 0.003974663559347391, + 0.004459637217223644, + 0.004084951709955931, + 0.004974365234375, + 0.004298023413866758, + 0.007378404028713703, + 0.006553533021360636, + 0.006327101960778236, + 0.007249636575579643, + 0.00623902166262269, + 0.004561177454888821, + 0.004629936069250107, + 0.004195645451545715, + 0.008802497759461403, + 0.004647182300686836, + 0.008085346780717373, + 0.010749722830951214, + 0.007266651373356581, + 0.006510766223073006, + 0.004324021749198437, + 0.005074754822999239, + 0.005834508221596479, + 0.003685642033815384, + 0.006649702321738005, + 0.006272811442613602, + 0.00461473548784852, + 0.01939493417739868, + 0.019245684146881104, + 0.013984544202685356, + 0.021149538457393646, + 0.01862453669309616, + 0.02768494375050068, + 0.022564072161912918, + 0.02161243185400963, + 0.019691063091158867, + 0.013029752299189568, + 0.012095379643142223, + 0.020374812185764313, + 0.013423909433186054, + 0.028921140357851982, + 0.055275555700063705, + 0.02825220301747322, + 0.030048783868551254, + 0.014724993146955967, + 0.02267337404191494, + 0.028491564095020294, + 0.017978042364120483, + 0.03279885649681091, + 0.03239918872714043, + 0.031149646267294884 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 0 + ], + "selected_reward": -0.0017144280718639493 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a small, round, white pill?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the pill perched in the center of a trembling hand with fingers curled slightly inward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene lit by a single, harsh light source from the side creating high contrast?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background deep and ambiguous darkness with long, deep shadows in the palm creases?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera angle slightly elevated and completely still throughout the 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": [ + -0.4743168354034424 + ], + "opposite_logprob_tokens": [ + -0.9743168354034424 + ], + "logprob_reward": -0.4743168354034424, + "yes_no_margin_reward": 0.5, + "token_credit_reward": -0.4743168354034424, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11966698616743088, + 0.07539468258619308, + 0.040032047778367996, + 0.035228390246629715, + 0.0273983683437109, + 0.03985734283924103, + 0.041022900491952896, + 0.16939599812030792, + 0.30682551860809326, + 0.14517776668071747 + ], + "credit_map": [ + 0.011999577283859253, + 0.014736002311110497, + 0.008768734522163868, + 0.015500372275710106, + 0.018840190023183823, + 0.0160874892026186, + 0.018288226798176765, + 0.01544638630002737, + 0.005894279573112726, + 0.006622487213462591, + 0.005112363491207361, + 0.014175943098962307, + 0.00852114800363779, + 0.009931433945894241, + 0.012850826606154442, + 0.012286195531487465, + 0.0052132210694253445, + 0.0033748557325452566, + 0.0034808251075446606, + 0.0055521572940051556, + 0.0040190196596086025, + 0.006027323193848133, + 0.006979890167713165, + 0.0053847553208470345, + 0.0029055397026240826, + 0.003945499192923307, + 0.002278720261529088, + 0.005457203835248947, + 0.0045474013313651085, + 0.004307935945689678, + 0.006213824264705181, + 0.005572265014052391, + 0.002193235559388995, + 0.0025434899143874645, + 0.001714137033559382, + 0.0046439459547400475, + 0.0030649923719465733, + 0.0033545708283782005, + 0.0054611447267234325, + 0.004422852303832769, + 0.003423987189307809, + 0.003745956113561988, + 0.002984128426760435, + 0.005946441553533077, + 0.003345303703099489, + 0.0059180548414587975, + 0.008573035709559917, + 0.005920434836298227, + 0.003752085380256176, + 0.005420485977083445, + 0.0033602856565266848, + 0.005396762862801552, + 0.003948417957872152, + 0.006016452331095934, + 0.008493276312947273, + 0.0046351319178938866, + 0.01613110676407814, + 0.01873856596648693, + 0.016383275389671326, + 0.025592226535081863, + 0.016390595585107803, + 0.02628602832555771, + 0.02771066129207611, + 0.022163528949022293, + 0.03506225347518921, + 0.024967338889837265, + 0.022890081629157066, + 0.03318364918231964, + 0.028651420027017593, + 0.04312256723642349, + 0.06791669130325317, + 0.05103153735399246, + 0.02463875524699688, + 0.01212749257683754, + 0.008303959853947163, + 0.017547380179166794, + 0.017348838970065117, + 0.019573472440242767, + 0.027234988287091255, + 0.018402887508273125 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 9 + ], + "selected_reward": -0.4743168354034424 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An extreme close-up shot from a slightly elevated angle remains still, intensely focused on a small, round, white pill, perched precariously in the center of a trembling hand. The scene is bathed in dramatic, low-key lighting, with a single, harsh light source from the side carving the hand out from the deep, ambiguous darkness of the background. This stark contrast of black and white creates a stark contrast, casting long, deep shadows within the creases of the palm while capturing sharp highlights on the pale skin of the knuckles, the edges of the nails, and the smooth, chalky surface of the pill. The fingers, curled slightly inward, emit an uncontrollable, tiny vibration, noticeably accentuated by the play of light and shadow. Their unsteady posture suggests a profound inner struggle. A cool warmth permeates the entire image, heightening the tense, somber, and deeply personal atmosphere of quiet despair and momentous decision.", + "seed": 28431, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000026/group_5/index_2446_seed_28431_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": -0.0704661157913506, + "vlm_logprob": -0.0704661157913506, + "vlm_yes_no_margin": 3.65, + "vlm_token_credit": -0.0704661157913506, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 small, round, white pill?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the pill perched in the center of a trembling hand with fingers curled slightly inward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene lit by a single, harsh light source from the side creating high contrast?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background deep and ambiguous darkness with long, deep shadows in the palm creases?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera angle slightly elevated and completely still throughout the 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": [ + -0.04597953334450722 + ], + "opposite_logprob_tokens": [ + -3.1709794998168945 + ], + "logprob_reward": -0.04597953334450722, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04597953334450722, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16563613712787628, + 0.11483363062143326, + 0.11342024803161621, + 0.04248327016830444, + 0.030282258987426758, + 0.22801968455314636, + 0.06600736826658249, + 0.06651952117681503, + 0.033804751932621, + 0.13899315893650055 + ], + "credit_map": [ + 0.02048575133085251, + 0.010065586306154728, + 0.023971162736415863, + 0.019605077803134918, + 0.008134531788527966, + 0.017342275008559227, + 0.018301960080862045, + 0.04772977530956268, + 0.013474161736667156, + 0.009210252203047276, + 0.01663549616932869, + 0.014839970506727695, + 0.008202208206057549, + 0.015479987487196922, + 0.01420579943805933, + 0.022785747423768044, + 0.011511199176311493, + 0.00710672652348876, + 0.020451046526432037, + 0.01712711714208126, + 0.0067366622388362885, + 0.013237313367426395, + 0.01670468971133232, + 0.020545480772852898, + 0.005187498405575752, + 0.0037858986761420965, + 0.005525005981326103, + 0.006324923597276211, + 0.0027767173014581203, + 0.0044945920817554, + 0.005183514207601547, + 0.009205116890370846, + 0.003057322232052684, + 0.0024356802459806204, + 0.003636028850451112, + 0.005053467582911253, + 0.0018211446003988385, + 0.0027232710272073746, + 0.004148731008172035, + 0.0074066114611923695, + 0.027465609833598137, + 0.01800045557320118, + 0.030530981719493866, + 0.049846623092889786, + 0.014183162711560726, + 0.01767575368285179, + 0.032727811485528946, + 0.03758927062153816, + 0.007224200759083033, + 0.004863369278609753, + 0.010436154901981354, + 0.012855419889092445, + 0.004589172080159187, + 0.007290179841220379, + 0.0066373310983181, + 0.012111530639231205, + 0.007387101650238037, + 0.004866831004619598, + 0.008659282699227333, + 0.010457579977810383, + 0.004966774024069309, + 0.006657965946942568, + 0.00950637087225914, + 0.01401760894805193, + 0.004089339170604944, + 0.003981008660048246, + 0.005061638541519642, + 0.007203672546893358, + 0.0018753678305074573, + 0.0027812730986624956, + 0.004495482426136732, + 0.004316967912018299, + 0.0162952970713377, + 0.017197301611304283, + 0.019827181473374367, + 0.017812535166740417, + 0.013390809297561646, + 0.015813367441296577, + 0.013092118315398693, + 0.02556455321609974 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 9 + ], + "selected_reward": -0.04597953334450722 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 small, round, white pill?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the pill perched in the center of a trembling hand with fingers curled slightly inward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene lit by a single, harsh light source from the side creating high contrast?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background deep and ambiguous darkness with long, deep shadows in the palm creases?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera angle slightly elevated and completely still throughout the 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": [ + -0.012564901262521744 + ], + "opposite_logprob_tokens": [ + -4.3875651359558105 + ], + "logprob_reward": -0.012564901262521744, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012564901262521744, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2205718755722046, + 0.12516731023788452, + 0.13137415051460266, + 0.045885708183050156, + 0.03072032891213894, + 0.16930431127548218, + 0.07521496713161469, + 0.061904747039079666, + 0.05382726714015007, + 0.08602936565876007 + ], + "credit_map": [ + 0.024723146110773087, + 0.016736237332224846, + 0.029878171160817146, + 0.026449991390109062, + 0.012825224548578262, + 0.019287867471575737, + 0.027041785418987274, + 0.06362944096326828, + 0.014644673094153404, + 0.009512964636087418, + 0.016349032521247864, + 0.01918543502688408, + 0.008603320457041264, + 0.01637064479291439, + 0.016891904175281525, + 0.02360934391617775, + 0.014034591615200043, + 0.00789946410804987, + 0.023581208661198616, + 0.020115600898861885, + 0.007900058291852474, + 0.01608683541417122, + 0.01820036582648754, + 0.02355603314936161, + 0.005240435712039471, + 0.004067661706358194, + 0.005789023358374834, + 0.00798128917813301, + 0.0031564864329993725, + 0.00545695424079895, + 0.0058507355861365795, + 0.008343123830854893, + 0.003397574182599783, + 0.0024333989713340998, + 0.0038674278184771538, + 0.0046425978653132915, + 0.0019667469896376133, + 0.002740586642175913, + 0.0047028688713908195, + 0.0069691273383796215, + 0.020939696580171585, + 0.013378717936575413, + 0.021995285525918007, + 0.0360802486538887, + 0.011254949495196342, + 0.014059551060199738, + 0.023420589044690132, + 0.028175268322229385, + 0.008672032505273819, + 0.004798361100256443, + 0.010059889405965805, + 0.012066211551427841, + 0.004852097947150469, + 0.009472977370023727, + 0.008876392617821693, + 0.01641700603067875, + 0.006824002135545015, + 0.004126511048525572, + 0.008838693611323833, + 0.010062064975500107, + 0.004043765366077423, + 0.006283164955675602, + 0.008140621706843376, + 0.013585923239588737, + 0.006915274541825056, + 0.006416122429072857, + 0.0064269243739545345, + 0.00763245252892375, + 0.004705087281763554, + 0.005907333921641111, + 0.00756946112960577, + 0.008254611864686012, + 0.009517590515315533, + 0.01118418574333191, + 0.01081684697419405, + 0.00968457106500864, + 0.008222648873925209, + 0.010464892722666264, + 0.00885999295860529, + 0.01727863773703575 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 2 + ], + "selected_reward": -0.012564901262521744 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 small, round, white pill?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the pill perched in the center of a trembling hand with fingers curled slightly inward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene lit by a single, harsh light source from the side creating high contrast?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background deep and ambiguous darkness with long, deep shadows in the palm creases?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera angle slightly elevated and completely still throughout the 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": [ + -0.11294373869895935 + ], + "opposite_logprob_tokens": [ + -2.237943649291992 + ], + "logprob_reward": -0.11294373869895935, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11294373869895935, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2085263431072235, + 0.14792680740356445, + 0.13491098582744598, + 0.05167486146092415, + 0.02383863925933838, + 0.06753887236118317, + 0.1791674643754959, + 0.06425842642784119, + 0.09593287855386734, + 0.02622472308576107 + ], + "credit_map": [ + 0.0467086099088192, + 0.021344464272260666, + 0.035873401910066605, + 0.02187696471810341, + 0.016033006832003593, + 0.016771584749221802, + 0.022874288260936737, + 0.027044018730521202, + 0.032024335116147995, + 0.01088442001491785, + 0.027598842978477478, + 0.015406963415443897, + 0.014108267612755299, + 0.016024071723222733, + 0.013387853279709816, + 0.018492061644792557, + 0.028056448325514793, + 0.011221633292734623, + 0.022192394360899925, + 0.012728933244943619, + 0.014513732865452766, + 0.011953720822930336, + 0.013841005973517895, + 0.020403115078806877, + 0.008702733553946018, + 0.004335380159318447, + 0.006656918674707413, + 0.007343896199017763, + 0.0034752357751131058, + 0.00528379064053297, + 0.006365967448800802, + 0.009510934352874756, + 0.0032462230883538723, + 0.001698242500424385, + 0.0032030020374804735, + 0.004582681227475405, + 0.0014741979539394379, + 0.0020144954323768616, + 0.0039634620770812035, + 0.0036563349422067404, + 0.010611023753881454, + 0.006886322051286697, + 0.008224005810916424, + 0.011000283062458038, + 0.0055098338052630424, + 0.005508861504495144, + 0.009744462557137012, + 0.01005407702177763, + 0.02210877649486065, + 0.00972068216651678, + 0.023670537397265434, + 0.033737801015377045, + 0.015820706263184547, + 0.0246560238301754, + 0.016382230445742607, + 0.03307070583105087, + 0.011182763613760471, + 0.005078515503555536, + 0.01149957999587059, + 0.011553065851330757, + 0.002728373510763049, + 0.004551331978291273, + 0.0058076102286577225, + 0.01185718271881342, + 0.015524953603744507, + 0.007905415259301662, + 0.013953985646367073, + 0.014087295159697533, + 0.008038503117859364, + 0.00838767271488905, + 0.014269274659454823, + 0.013765784911811352, + 0.004201267380267382, + 0.0028036576695740223, + 0.0038984306156635284, + 0.004604949615895748, + 0.0014415710465982556, + 0.0015574457356706262, + 0.003586638020351529, + 0.004130764398723841 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 1 + ], + "selected_reward": -0.11294373869895935 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 small, round, white pill?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the pill perched in the center of a trembling hand with fingers curled slightly inward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene lit by a single, harsh light source from the side creating high contrast?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background deep and ambiguous darkness with long, deep shadows in the palm creases?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera angle slightly elevated and completely still throughout the 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": [ + -0.000922969076782465 + ], + "opposite_logprob_tokens": [ + -7.000923156738281 + ], + "logprob_reward": -0.000922969076782465, + "yes_no_margin_reward": 7.0, + "token_credit_reward": -0.000922969076782465, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17947037518024445, + 0.09095530956983566, + 0.052013322710990906, + 0.0859452337026596, + 0.05944470688700676, + 0.13203652203083038, + 0.18340139091014862, + 0.10160645097494125, + 0.056583382189273834, + 0.05854331701993942 + ], + "credit_map": [ + 0.026229925453662872, + 0.01696624606847763, + 0.026877643540501595, + 0.025415077805519104, + 0.015524791553616524, + 0.015533051453530788, + 0.026449695229530334, + 0.026473956182599068, + 0.01293940283358097, + 0.007777529302984476, + 0.013214057311415672, + 0.01183730736374855, + 0.00854631420224905, + 0.008824380114674568, + 0.011283455416560173, + 0.016532864421606064, + 0.007171158213168383, + 0.0032705964986234903, + 0.008140793070197105, + 0.008257264271378517, + 0.003330629551783204, + 0.004909641109406948, + 0.008015500381588936, + 0.008917740546166897, + 0.012913830578327179, + 0.00862516276538372, + 0.010254262015223503, + 0.011172333732247353, + 0.00509532680734992, + 0.009376609697937965, + 0.013599192723631859, + 0.014908520504832268, + 0.008734292350709438, + 0.004468468017876148, + 0.008677280507981777, + 0.009673785418272018, + 0.002816680585965514, + 0.00394250825047493, + 0.010129988193511963, + 0.01100170612335205, + 0.01388858538120985, + 0.008373499847948551, + 0.016290605068206787, + 0.024833587929606438, + 0.010026280768215656, + 0.011601308360695839, + 0.02200913056731224, + 0.025013532489538193, + 0.023943684995174408, + 0.011544531211256981, + 0.03079758957028389, + 0.033743180334568024, + 0.012078416533768177, + 0.015357131138443947, + 0.019670795649290085, + 0.03626605495810509, + 0.010937273502349854, + 0.009305234998464584, + 0.014374847523868084, + 0.019978273659944534, + 0.00690765306353569, + 0.010619963519275188, + 0.010752707719802856, + 0.01873050071299076, + 0.0074058412574231625, + 0.006179723888635635, + 0.007759592495858669, + 0.008753211237490177, + 0.004675737116485834, + 0.006018412299454212, + 0.00841569621115923, + 0.0073751709423959255, + 0.009456321597099304, + 0.007265125401318073, + 0.00851796381175518, + 0.007113539148122072, + 0.004603683948516846, + 0.005635824985802174, + 0.006214672699570656, + 0.009736184030771255 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 5 + ], + "selected_reward": -0.000922969076782465 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 small, round, white pill?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the pill perched in the center of a trembling hand with fingers curled slightly inward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene lit by a single, harsh light source from the side creating high contrast?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background deep and ambiguous darkness with long, deep shadows in the palm creases?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera angle slightly elevated and completely still throughout the 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": [ + -0.17991943657398224 + ], + "opposite_logprob_tokens": [ + -1.8049194812774658 + ], + "logprob_reward": -0.17991943657398224, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.17991943657398224, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09439673274755478, + 0.041513293981552124, + 0.042949479073286057, + 0.06938090175390244, + 0.04266020283102989, + 0.08152076601982117, + 0.27092674374580383, + 0.23796625435352325, + 0.07744234800338745, + 0.041243262588977814 + ], + "credit_map": [ + 0.013858589343726635, + 0.007453585043549538, + 0.014012513682246208, + 0.009800441563129425, + 0.004852595739066601, + 0.007637124042958021, + 0.008648167364299297, + 0.028133712708950043, + 0.0059122564271092415, + 0.0020693810656666756, + 0.006329168099910021, + 0.004467817954719067, + 0.0024095354601740837, + 0.004575316794216633, + 0.00406953739002347, + 0.01168027799576521, + 0.004897992126643658, + 0.0028403657488524914, + 0.006632791832089424, + 0.004911565221846104, + 0.002693962538614869, + 0.005228198133409023, + 0.006485562771558762, + 0.009259040467441082, + 0.0070314533077180386, + 0.005566168110817671, + 0.008554674685001373, + 0.011287791654467583, + 0.005567032378166914, + 0.007301702629774809, + 0.013194497674703598, + 0.010877584107220173, + 0.005922558717429638, + 0.0032219651620835066, + 0.00538151990622282, + 0.007246793247759342, + 0.0027683216612786055, + 0.004197068978101015, + 0.007290325127542019, + 0.006631648167967796, + 0.011384568177163601, + 0.006439007353037596, + 0.012070061638951302, + 0.010448972694575787, + 0.005542512517422438, + 0.009459707885980606, + 0.010200447402894497, + 0.01597549021244049, + 0.029166309162974358, + 0.01876714453101158, + 0.04295267164707184, + 0.04958679527044296, + 0.018106132745742798, + 0.024523427709937096, + 0.03356406092643738, + 0.05426020547747612, + 0.02638780139386654, + 0.020125042647123337, + 0.034119896590709686, + 0.043450213968753815, + 0.01757913827896118, + 0.027392050251364708, + 0.024982063099741936, + 0.0439300499856472, + 0.0112899886444211, + 0.007908312603831291, + 0.011309642344713211, + 0.01160056795924902, + 0.0061204456724226475, + 0.008183276280760765, + 0.01187887042760849, + 0.009151243604719639, + 0.005288355518132448, + 0.004485354758799076, + 0.005235370248556137, + 0.005278008058667183, + 0.005090306047350168, + 0.0038115354254841805, + 0.005371043924242258, + 0.0066832927986979485 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 7, + 0 + ], + "selected_reward": -0.17991943657398224 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An extreme close-up shot from a slightly elevated angle remains still, intensely focused on a small, round, white pill, perched precariously in the center of a trembling hand. The scene is bathed in dramatic, low-key lighting, with a single, harsh light source from the side carving the hand out from the deep, ambiguous darkness of the background. This stark contrast of black and white creates a stark contrast, casting long, deep shadows within the creases of the palm while capturing sharp highlights on the pale skin of the knuckles, the edges of the nails, and the smooth, chalky surface of the pill. The fingers, curled slightly inward, emit an uncontrollable, tiny vibration, noticeably accentuated by the play of light and shadow. Their unsteady posture suggests a profound inner struggle. A cool warmth permeates the entire image, heightening the tense, somber, and deeply personal atmosphere of quiet despair and momentous decision.", + "seed": 28432, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000026/group_5/index_2446_seed_28432_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": -0.06886574502568692, + "vlm_logprob": -0.06886574502568692, + "vlm_yes_no_margin": 3.975, + "vlm_token_credit": -0.06886574502568692, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 small, round, white pill?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the pill perched in the center of a trembling hand with fingers curled slightly inward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene lit by a single, harsh light source from the side creating high contrast?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background deep and ambiguous darkness with long, deep shadows in the palm creases?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera angle slightly elevated and completely still throughout the 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": [ + -0.013021906837821007 + ], + "opposite_logprob_tokens": [ + -4.638021945953369 + ], + "logprob_reward": -0.013021906837821007, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.013021906837821007, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11849072575569153, + 0.0851387158036232, + 0.10631180554628372, + 0.20948849618434906, + 0.07768530398607254, + 0.18621082603931427, + 0.09679817408323288, + 0.05476285517215729, + 0.05115712806582451, + 0.013955920934677124 + ], + "credit_map": [ + 0.01190629880875349, + 0.009832370094954967, + 0.009114192798733711, + 0.01652040146291256, + 0.016632279381155968, + 0.019852226600050926, + 0.01911921240389347, + 0.01551374513655901, + 0.006208003032952547, + 0.0074755409732460976, + 0.005390031263232231, + 0.011236939579248428, + 0.01525269914418459, + 0.017196407541632652, + 0.01082844752818346, + 0.011550655588507652, + 0.008938114158809185, + 0.010134536772966385, + 0.008091507479548454, + 0.01627478003501892, + 0.015624510124325752, + 0.01915578730404377, + 0.01440754160284996, + 0.013685042969882488, + 0.015860790386795998, + 0.012809512205421925, + 0.017354752868413925, + 0.02266370691359043, + 0.02882491797208786, + 0.04236110299825668, + 0.022052699699997902, + 0.047561030834913254, + 0.005660983268171549, + 0.0074585312977433205, + 0.005881286226212978, + 0.008240561001002789, + 0.014201867394149303, + 0.011131873354315758, + 0.014613062143325806, + 0.010497147217392921, + 0.0133562833070755, + 0.01038033701479435, + 0.015157808549702168, + 0.01286919042468071, + 0.0465710386633873, + 0.044027477502822876, + 0.013081023469567299, + 0.03076768107712269, + 0.006096437573432922, + 0.0034665733110159636, + 0.006507336627691984, + 0.004546094685792923, + 0.01098554115742445, + 0.02804066799581051, + 0.0055993362329900265, + 0.03155619278550148, + 0.004574470221996307, + 0.004699404817074537, + 0.00443356903269887, + 0.008699663914740086, + 0.009893879294395447, + 0.008790234103798866, + 0.006937861442565918, + 0.006733776535838842, + 0.004347086884081364, + 0.00423835776746273, + 0.004774110857397318, + 0.007750159595161676, + 0.008984915912151337, + 0.00816282257437706, + 0.006637896411120892, + 0.006261784583330154, + 0.0012037220876663923, + 0.0013405790086835623, + 0.0011968578910455108, + 0.002031064825132489, + 0.0022641122341156006, + 0.0023743899073451757, + 0.0015964574413374066, + 0.001948738587088883 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 5, + 0 + ], + "selected_reward": -0.013021906837821007 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 small, round, white pill?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the pill perched in the center of a trembling hand with fingers curled slightly inward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene lit by a single, harsh light source from the side creating high contrast?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background deep and ambiguous darkness with long, deep shadows in the palm creases?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera angle slightly elevated and completely still throughout the 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": [ + -0.007644092198461294 + ], + "opposite_logprob_tokens": [ + -4.882644176483154 + ], + "logprob_reward": -0.007644092198461294, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007644092198461294, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2302767038345337, + 0.1795371174812317, + 0.09390296787023544, + 0.1601765900850296, + 0.08578693121671677, + 0.040246374905109406, + 0.05504414439201355, + 0.06012165546417236, + 0.061402566730976105, + 0.03350501507520676 + ], + "credit_map": [ + 0.017273398116230965, + 0.02561976946890354, + 0.020899465307593346, + 0.030057203024625778, + 0.03587079048156738, + 0.03771210089325905, + 0.03141229227185249, + 0.031431667506694794, + 0.01592262275516987, + 0.014126201160252094, + 0.015187823213636875, + 0.021361317485570908, + 0.032528117299079895, + 0.031526122242212296, + 0.023301774635910988, + 0.025583114475011826, + 0.007791409734636545, + 0.007058829069137573, + 0.007518331054598093, + 0.010583152994513512, + 0.015352007932960987, + 0.021479157730937004, + 0.0123238954693079, + 0.011796184815466404, + 0.012042086571455002, + 0.00885050930082798, + 0.01132140401750803, + 0.014196506701409817, + 0.018754422664642334, + 0.03449036180973053, + 0.018528450280427933, + 0.04199283942580223, + 0.0079866424202919, + 0.005863866303116083, + 0.009220041334629059, + 0.009549322538077831, + 0.015923673287034035, + 0.012147200293838978, + 0.012039457447826862, + 0.013056721538305283, + 0.0022121290676295757, + 0.002971880603581667, + 0.003380562411621213, + 0.0035776973236352205, + 0.008839936926960945, + 0.007848268374800682, + 0.004602501634508371, + 0.00681339344009757, + 0.004483248572796583, + 0.00486399931833148, + 0.0071726213209331036, + 0.00597818847745657, + 0.008939211256802082, + 0.00799931213259697, + 0.007759321480989456, + 0.007848239503800869, + 0.0056466879323124886, + 0.005399998743087053, + 0.006030289456248283, + 0.008703265339136124, + 0.010152474977076054, + 0.01041026134043932, + 0.007276956923305988, + 0.0065017156302928925, + 0.004755547735840082, + 0.005124139599502087, + 0.006178978364914656, + 0.009967400692403316, + 0.010139976628124714, + 0.00919042993336916, + 0.008785557001829147, + 0.007260533981025219, + 0.0029319049790501595, + 0.0033793915063142776, + 0.0019469105172902346, + 0.0035052674356848, + 0.005664722062647343, + 0.006621881388127804, + 0.004553426057100296, + 0.004901512060314417 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 3 + ], + "selected_reward": -0.007644092198461294 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 small, round, white pill?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the pill perched in the center of a trembling hand with fingers curled slightly inward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene lit by a single, harsh light source from the side creating high contrast?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background deep and ambiguous darkness with long, deep shadows in the palm creases?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera angle slightly elevated and completely still throughout the 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": [ + -0.07003648579120636 + ], + "opposite_logprob_tokens": [ + -2.6950364112854004 + ], + "logprob_reward": -0.07003648579120636, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07003648579120636, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10285726934671402, + 0.046011783182621, + 0.030956916511058807, + 0.09580785036087036, + 0.41861608624458313, + 0.11507269740104675, + 0.13544823229312897, + 0.02911866270005703, + 0.016637170687317848, + 0.00947339367121458 + ], + "credit_map": [ + 0.009240771643817425, + 0.01169864647090435, + 0.007980524562299252, + 0.013007257133722305, + 0.015426959842443466, + 0.016596810892224312, + 0.0141074750572443, + 0.014798808842897415, + 0.004111606627702713, + 0.005573485977947712, + 0.0029638241976499557, + 0.007045935373753309, + 0.006547302007675171, + 0.007844564504921436, + 0.005446950905025005, + 0.006478107534348965, + 0.002289664698764682, + 0.0028517921455204487, + 0.002012687735259533, + 0.004709049127995968, + 0.005708653945475817, + 0.0049673994071781635, + 0.004166168160736561, + 0.00425149779766798, + 0.006433245725929737, + 0.0070951152592897415, + 0.007629472762346268, + 0.011380970478057861, + 0.011152185499668121, + 0.01909071020781994, + 0.01219241227954626, + 0.020833726972341537, + 0.039254046976566315, + 0.026715964078903198, + 0.036281194537878036, + 0.05016544088721275, + 0.09561644494533539, + 0.06118212267756462, + 0.051171328872442245, + 0.0582294762134552, + 0.010160686448216438, + 0.008293257094919682, + 0.01345005165785551, + 0.010606663301587105, + 0.023469917476177216, + 0.024215111508965492, + 0.007140288129448891, + 0.017736706882715225, + 0.009833421558141708, + 0.006756088696420193, + 0.011264342814683914, + 0.00998444203287363, + 0.00948931835591793, + 0.035449400544166565, + 0.011690294370055199, + 0.040980905294418335, + 0.0023374303709715605, + 0.0025140540674328804, + 0.002148366067558527, + 0.004243719857186079, + 0.005123966373503208, + 0.004761841613799334, + 0.00412680534645915, + 0.003862476907670498, + 0.001606540521606803, + 0.0016062005888670683, + 0.0014600820140913129, + 0.00197485089302063, + 0.002777191810309887, + 0.0029431800357997417, + 0.002199812326580286, + 0.002069310750812292, + 0.0005367333069443703, + 0.0007601565448567271, + 0.0007913640583865345, + 0.0016763812163844705, + 0.0012229522690176964, + 0.0014354134909808636, + 0.001269156695343554, + 0.0017812346341088414 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 6, + 5 + ], + "selected_reward": -0.07003648579120636 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 small, round, white pill?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the pill perched in the center of a trembling hand with fingers curled slightly inward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene lit by a single, harsh light source from the side creating high contrast?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background deep and ambiguous darkness with long, deep shadows in the palm creases?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera angle slightly elevated and completely still throughout the 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": [ + -0.001515788841061294 + ], + "opposite_logprob_tokens": [ + -6.501515865325928 + ], + "logprob_reward": -0.001515788841061294, + "yes_no_margin_reward": 6.5, + "token_credit_reward": -0.001515788841061294, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1123085618019104, + 0.09056445211172104, + 0.028382405638694763, + 0.1497531533241272, + 0.21782396733760834, + 0.1338888704776764, + 0.17258146405220032, + 0.03800158575177193, + 0.025979861617088318, + 0.030715664848685265 + ], + "credit_map": [ + 0.009150856174528599, + 0.013398479670286179, + 0.008894353173673153, + 0.015995686873793602, + 0.014916491694748402, + 0.016040528193116188, + 0.018839465454220772, + 0.015072700567543507, + 0.008354906924068928, + 0.00990793202072382, + 0.006155959330499172, + 0.008660227060317993, + 0.015837542712688446, + 0.015328519977629185, + 0.013286205939948559, + 0.01303315069526434, + 0.0023036657366901636, + 0.0033190050162374973, + 0.001720028230920434, + 0.003972548991441727, + 0.0049791112542152405, + 0.005085056181997061, + 0.0026290612295269966, + 0.004373928532004356, + 0.009460686706006527, + 0.01142543088644743, + 0.010954434983432293, + 0.015390726737678051, + 0.0187041275203228, + 0.03135889768600464, + 0.017759744077920914, + 0.03469910845160484, + 0.019257429987192154, + 0.01459287665784359, + 0.021785613149404526, + 0.02285301685333252, + 0.05376327037811279, + 0.03191472962498665, + 0.02358625829219818, + 0.030070776119828224, + 0.012484760954976082, + 0.01066210214048624, + 0.01573784649372101, + 0.012342484667897224, + 0.020178327336907387, + 0.025125902146100998, + 0.016211774200201035, + 0.021145671606063843, + 0.012615710496902466, + 0.007328260689973831, + 0.016112929210066795, + 0.011736901476979256, + 0.0124169010668993, + 0.04534902796149254, + 0.014276000671088696, + 0.05274572968482971, + 0.004227042198181152, + 0.0031175222247838974, + 0.0038358436431735754, + 0.004113722126930952, + 0.007154528517276049, + 0.006381405051797628, + 0.003753666067495942, + 0.005417854990810156, + 0.0022857803851366043, + 0.002866673981770873, + 0.0023766059894114733, + 0.003611832158640027, + 0.0033710359130054712, + 0.004379233345389366, + 0.0034243741538375616, + 0.0036643254570662975, + 0.0027221343480050564, + 0.0026318144518882036, + 0.0026245680637657642, + 0.004508923273533583, + 0.004673401825129986, + 0.004991929046809673, + 0.004203130956739187, + 0.004359763115644455 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 6, + 3 + ], + "selected_reward": -0.001515788841061294 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 small, round, white pill?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the pill perched in the center of a trembling hand with fingers curled slightly inward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene lit by a single, harsh light source from the side creating high contrast?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background deep and ambiguous darkness with long, deep shadows in the palm creases?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera angle slightly elevated and completely still throughout the 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": [ + -0.25211045145988464 + ], + "opposite_logprob_tokens": [ + -1.502110481262207 + ], + "logprob_reward": -0.25211045145988464, + "yes_no_margin_reward": 1.25, + "token_credit_reward": -0.25211045145988464, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09186651557683945, + 0.03633805364370346, + 0.04984939843416214, + 0.2043713629245758, + 0.18743528425693512, + 0.2264789193868637, + 0.09946756809949875, + 0.029810717329382896, + 0.052296463400125504, + 0.022085711359977722 + ], + "credit_map": [ + 0.0071642217226326466, + 0.006988279987126589, + 0.006434318143874407, + 0.012401844374835491, + 0.015632914379239082, + 0.018458891659975052, + 0.011523433029651642, + 0.013262614607810974, + 0.0028041619807481766, + 0.00307782762683928, + 0.002409484004601836, + 0.003797338344156742, + 0.0052470192313194275, + 0.008424154482781887, + 0.0051149651408195496, + 0.005463105160742998, + 0.0037564763333648443, + 0.004280037246644497, + 0.003421599743887782, + 0.005723812151700258, + 0.008365604095160961, + 0.009831034578382969, + 0.0076721664518117905, + 0.006798666436225176, + 0.015185533091425896, + 0.013745730742812157, + 0.017058338969945908, + 0.026376718655228615, + 0.023084202781319618, + 0.03985368832945824, + 0.024576455354690552, + 0.044490691274404526, + 0.017174651846289635, + 0.013337739743292332, + 0.016963083297014236, + 0.02063174359500408, + 0.03603223338723183, + 0.026957770809531212, + 0.02725670114159584, + 0.029081357643008232, + 0.014737481251358986, + 0.013696501962840557, + 0.017705747857689857, + 0.014898794703185558, + 0.042249489575624466, + 0.062042057514190674, + 0.01690598390996456, + 0.04424286633729935, + 0.005670599639415741, + 0.004542797803878784, + 0.007570657413452864, + 0.0061115603893995285, + 0.008917419239878654, + 0.029377581551671028, + 0.005078145302832127, + 0.03219881281256676, + 0.0027296303305774927, + 0.0023851559963077307, + 0.0028960471972823143, + 0.0051312814466655254, + 0.004418328870087862, + 0.004197960719466209, + 0.0048417868092656136, + 0.003210527589544654, + 0.005245533771812916, + 0.00460180826485157, + 0.005141773726791143, + 0.007068917620927095, + 0.00740869902074337, + 0.009569836780428886, + 0.006824595853686333, + 0.006435296032577753, + 0.0020582014694809914, + 0.0014520405093207955, + 0.0023658673744648695, + 0.00423628743737936, + 0.002700581680983305, + 0.0027294002939015627, + 0.003881629556417465, + 0.002661702921614051 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 3, + 4 + ], + "selected_reward": -0.25211045145988464 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An extreme close-up shot from a slightly elevated angle remains still, intensely focused on a small, round, white pill, perched precariously in the center of a trembling hand. The scene is bathed in dramatic, low-key lighting, with a single, harsh light source from the side carving the hand out from the deep, ambiguous darkness of the background. This stark contrast of black and white creates a stark contrast, casting long, deep shadows within the creases of the palm while capturing sharp highlights on the pale skin of the knuckles, the edges of the nails, and the smooth, chalky surface of the pill. The fingers, curled slightly inward, emit an uncontrollable, tiny vibration, noticeably accentuated by the play of light and shadow. Their unsteady posture suggests a profound inner struggle. A cool warmth permeates the entire image, heightening the tense, somber, and deeply personal atmosphere of quiet despair and momentous decision.", + "seed": 28433, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000026/group_5/index_2446_seed_28433_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": -0.1948461970780045, + "vlm_logprob": -0.1948461970780045, + "vlm_yes_no_margin": 3.675, + "vlm_token_credit": -0.1948461970780045, + "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 of a small, round, white pill?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the pill perched in the center of a trembling hand with fingers curled slightly inward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene lit by a single, harsh light source from the side creating high contrast?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background deep and ambiguous darkness with long, deep shadows in the palm creases?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera angle slightly elevated and completely still throughout the 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": [ + -0.0073251002468168736 + ], + "opposite_logprob_tokens": [ + -5.757325172424316 + ], + "logprob_reward": -0.0073251002468168736, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.0073251002468168736, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11102525144815445, + 0.15678423643112183, + 0.03788815066218376, + 0.12427535653114319, + 0.2548392713069916, + 0.05657554790377617, + 0.11196578294038773, + 0.08958514034748077, + 0.03674144297838211, + 0.020319782197475433 + ], + "credit_map": [ + 0.009759425185620785, + 0.00756909791380167, + 0.004915793426334858, + 0.007559136021882296, + 0.014446268789470196, + 0.02418804168701172, + 0.010435580275952816, + 0.03215191140770912, + 0.015528729185461998, + 0.00635936064645648, + 0.004951195791363716, + 0.004349668975919485, + 0.008390498347580433, + 0.05273137986660004, + 0.007037255447357893, + 0.05743614211678505, + 0.0050945524126291275, + 0.004357029218226671, + 0.00200819643214345, + 0.0017718872986733913, + 0.009992220439016819, + 0.006324707064777613, + 0.004503471776843071, + 0.0038360811304301023, + 0.031617823988199234, + 0.006963212974369526, + 0.005142271053045988, + 0.0047804489731788635, + 0.007646559271961451, + 0.04781761020421982, + 0.005057508125901222, + 0.015249921008944511, + 0.06563927978277206, + 0.011404314078390598, + 0.010228431783616543, + 0.005389427300542593, + 0.01849169097840786, + 0.10377640277147293, + 0.008394446223974228, + 0.03151528909802437, + 0.011580469086766243, + 0.003925904631614685, + 0.002832205733284354, + 0.0018515348201617599, + 0.008509126491844654, + 0.015387078747153282, + 0.004928641952574253, + 0.007560587488114834, + 0.029573434963822365, + 0.005292598158121109, + 0.0048265159130096436, + 0.003480437910184264, + 0.006056877784430981, + 0.04518430307507515, + 0.006068991497159004, + 0.011482622474431992, + 0.010401397943496704, + 0.008468221873044968, + 0.0044782524928450584, + 0.006093362346291542, + 0.024498090147972107, + 0.015359148383140564, + 0.012392401695251465, + 0.00789426825940609, + 0.004897462669759989, + 0.002974982839077711, + 0.0022850933019071817, + 0.004981569014489651, + 0.00620950385928154, + 0.005020677112042904, + 0.005604817532002926, + 0.004767334554344416, + 0.002920580329373479, + 0.0018246350809931755, + 0.0013455458683893085, + 0.0018521274905651808, + 0.0026911329478025436, + 0.003767241258174181, + 0.0028176845517009497, + 0.0031008359510451555 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 1, + 3 + ], + "selected_reward": -0.0073251002468168736 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 small, round, white pill?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the pill perched in the center of a trembling hand with fingers curled slightly inward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene lit by a single, harsh light source from the side creating high contrast?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background deep and ambiguous darkness with long, deep shadows in the palm creases?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera angle slightly elevated and completely still throughout the 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": [ + -0.0046546440571546555 + ], + "opposite_logprob_tokens": [ + -5.379654407501221 + ], + "logprob_reward": -0.0046546440571546555, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.0046546440571546555, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06603525578975677, + 0.11370508372783661, + 0.021462148055434227, + 0.1610700488090515, + 0.32325252890586853, + 0.1061592698097229, + 0.024977145716547966, + 0.03220806643366814, + 0.1323317140340805, + 0.01879872754216194 + ], + "credit_map": [ + 0.007573002018034458, + 0.007875146344304085, + 0.004185661673545837, + 0.006443197838962078, + 0.009470593184232712, + 0.011584751307964325, + 0.00857549998909235, + 0.010327409021556377, + 0.015194442123174667, + 0.006774216424673796, + 0.004760629963129759, + 0.005432359408587217, + 0.015871809795498848, + 0.03512788191437721, + 0.008517404086887836, + 0.022026345133781433, + 0.004752669483423233, + 0.0012642709771171212, + 0.0008856749045662582, + 0.0011432425817474723, + 0.0014449235750362277, + 0.007969584316015244, + 0.001387258991599083, + 0.0026145256124436855, + 0.041060492396354675, + 0.008402220904827118, + 0.007137670647352934, + 0.0071456460282206535, + 0.009072997607290745, + 0.06361151486635208, + 0.006783926393836737, + 0.01785559393465519, + 0.08299509435892105, + 0.018633777275681496, + 0.014861535280942917, + 0.0076192026026546955, + 0.019987596198916435, + 0.1165812686085701, + 0.016729658469557762, + 0.045844390988349915, + 0.025215541943907738, + 0.0059012253768742085, + 0.003999721724539995, + 0.0028464861679822206, + 0.007872776128351688, + 0.043142545968294144, + 0.004949909634888172, + 0.0122310696169734, + 0.003116287523880601, + 0.0023182351142168045, + 0.001801252132281661, + 0.0021291610319167376, + 0.0023664673790335655, + 0.005230308044701815, + 0.002965729683637619, + 0.005049706902354956, + 0.003696079133078456, + 0.0030126848723739386, + 0.0017347028478980064, + 0.0020605321042239666, + 0.008682209998369217, + 0.006264731287956238, + 0.0041033122688531876, + 0.0026538167148828506, + 0.030505456030368805, + 0.008387911133468151, + 0.007153244689106941, + 0.007773547898977995, + 0.011136387474834919, + 0.04394698515534401, + 0.010343722999095917, + 0.013084473088383675, + 0.002381059806793928, + 0.0020466481801122427, + 0.0013101216172799468, + 0.001739477040246129, + 0.002492570085451007, + 0.0035250047221779823, + 0.002101966179907322, + 0.0032018814235925674 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 3, + 8 + ], + "selected_reward": -0.0046546440571546555 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 small, round, white pill?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the pill perched in the center of a trembling hand with fingers curled slightly inward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene lit by a single, harsh light source from the side creating high contrast?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background deep and ambiguous darkness with long, deep shadows in the palm creases?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera angle slightly elevated and completely still throughout the 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": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.7578229904174805 + ], + "opposite_logprob_tokens": [ + -0.6328229904174805 + ], + "logprob_reward": -0.7578229904174805, + "yes_no_margin_reward": -0.125, + "token_credit_reward": -0.7578229904174805, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.06550820916891098, + 0.19936418533325195, + 0.01693706400692463, + 0.07576070725917816, + 0.32871580123901367, + 0.029693691059947014, + 0.03500913456082344, + 0.15501844882965088, + 0.08513133227825165, + 0.008861429058015347 + ], + "credit_map": [ + 0.006250040140002966, + 0.0051179490983486176, + 0.0037322852294892073, + 0.007202285807579756, + 0.013401622883975506, + 0.00961890909820795, + 0.008963530883193016, + 0.011221583932638168, + 0.031956031918525696, + 0.00789003074169159, + 0.0060872239992022514, + 0.0038912531454116106, + 0.008647222071886063, + 0.08502990007400513, + 0.004973257426172495, + 0.0508892647922039, + 0.0021141550969332457, + 0.0015191823476925492, + 0.000982599682174623, + 0.000934951298404485, + 0.004106287844479084, + 0.0028027084190398455, + 0.0015908913919702172, + 0.0028862874023616314, + 0.017603736370801926, + 0.004952790215611458, + 0.0032712342217564583, + 0.0035815280862152576, + 0.005110916681587696, + 0.027351966127753258, + 0.003869252745062113, + 0.010019281879067421, + 0.08606043457984924, + 0.014773383736610413, + 0.012102703563869, + 0.009920096956193447, + 0.014227296225726604, + 0.13789300620555878, + 0.011274329386651516, + 0.042464546859264374, + 0.00734616769477725, + 0.0015583571512252092, + 0.001429605414159596, + 0.0007155042258091271, + 0.0026275552809238434, + 0.011048478074371815, + 0.00148325739428401, + 0.0034847641363739967, + 0.008115126751363277, + 0.001990261720493436, + 0.0015875461976975203, + 0.0018891642102971673, + 0.002126164734363556, + 0.01299747359007597, + 0.002666783519089222, + 0.003636610694229603, + 0.015419581905007362, + 0.012675076723098755, + 0.007921854965388775, + 0.009471781551837921, + 0.0465502105653286, + 0.0263723935931921, + 0.017953136935830116, + 0.01865440234541893, + 0.019225239753723145, + 0.005352059844881296, + 0.0043719979003071785, + 0.0029575226362794638, + 0.013745701871812344, + 0.0277396272867918, + 0.0041450010612607, + 0.007594180293381214, + 0.0016553431050851941, + 0.00060858647339046, + 0.0005799793289043009, + 0.0006500904564745724, + 0.0011567813344299793, + 0.0018893411615863442, + 0.0009076540591195226, + 0.0014136526733636856 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 1, + 7 + ], + "selected_reward": -0.7578229904174805 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 small, round, white pill?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the pill perched in the center of a trembling hand with fingers curled slightly inward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene lit by a single, harsh light source from the side creating high contrast?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background deep and ambiguous darkness with long, deep shadows in the palm creases?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera angle slightly elevated and completely still throughout the 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": [ + -0.0028212766628712416 + ], + "opposite_logprob_tokens": [ + -5.877821445465088 + ], + "logprob_reward": -0.0028212766628712416, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.0028212766628712416, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13420967757701874, + 0.07735560834407806, + 0.009988222271203995, + 0.03977608308196068, + 0.20071619749069214, + 0.04597582668066025, + 0.06474802643060684, + 0.10852795094251633, + 0.3038022220134735, + 0.014900182373821735 + ], + "credit_map": [ + 0.011521929875016212, + 0.006557538639754057, + 0.005974275059998035, + 0.008599928580224514, + 0.01902701146900654, + 0.027026597410440445, + 0.013851866126060486, + 0.041650522500276566, + 0.006744622252881527, + 0.0031102250795811415, + 0.002577963750809431, + 0.0037088445387780666, + 0.006990992464125156, + 0.02477332018315792, + 0.004531728569418192, + 0.024917909875512123, + 0.0013768334174528718, + 0.000951060326769948, + 0.0009753736667335033, + 0.0007853818242438138, + 0.0009263532119803131, + 0.0014923368580639362, + 0.0011412891326472163, + 0.002339594066143036, + 0.004697345197200775, + 0.005056357942521572, + 0.002437352202832699, + 0.002230672398582101, + 0.008877326734364033, + 0.007196406833827496, + 0.005415766499936581, + 0.003864854108542204, + 0.04802114889025688, + 0.010634570382535458, + 0.010449028573930264, + 0.0061576832085847855, + 0.01677989773452282, + 0.06628815829753876, + 0.010396200232207775, + 0.03198951482772827, + 0.008021068759262562, + 0.0027232642751187086, + 0.002038518665358424, + 0.002327133435755968, + 0.005160411354154348, + 0.014300831593573093, + 0.005078874062746763, + 0.006325722672045231, + 0.016662783920764923, + 0.0029003056697547436, + 0.0025396174751222134, + 0.002219666726887226, + 0.002952245296910405, + 0.025389252230525017, + 0.0031832358799874783, + 0.008900917135179043, + 0.010705703869462013, + 0.009592804126441479, + 0.0070296525955200195, + 0.007615985814481974, + 0.03103215992450714, + 0.01625446416437626, + 0.012676671147346497, + 0.01362051535397768, + 0.07099630683660507, + 0.01783253438770771, + 0.015395009890198708, + 0.011416364461183548, + 0.045144349336624146, + 0.10354255884885788, + 0.014861376024782658, + 0.024613739922642708, + 0.0019959681667387486, + 0.001852344605140388, + 0.00069217081181705, + 0.0010340851731598377, + 0.003149785567075014, + 0.00194209860637784, + 0.0021325605921447277, + 0.00210116826929152 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 4, + 0 + ], + "selected_reward": -0.0028212766628712416 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 small, round, white pill?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the pill perched in the center of a trembling hand with fingers curled slightly inward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene lit by a single, harsh light source from the side creating high contrast?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background deep and ambiguous darkness with long, deep shadows in the palm creases?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera angle slightly elevated and completely still throughout the 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": [ + -0.20160697400569916 + ], + "opposite_logprob_tokens": [ + -1.7016069889068604 + ], + "logprob_reward": -0.20160697400569916, + "yes_no_margin_reward": 1.5, + "token_credit_reward": -0.20160697400569916, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10281594842672348, + 0.06436609476804733, + 0.03901248425245285, + 0.08141002804040909, + 0.19286364316940308, + 0.1005842536687851, + 0.12388395518064499, + 0.01908545196056366, + 0.26264896988868713, + 0.013329179026186466 + ], + "credit_map": [ + 0.01192106306552887, + 0.010640142485499382, + 0.005703188478946686, + 0.010730461217463017, + 0.015303258784115314, + 0.01740782894194126, + 0.014233309775590897, + 0.016876697540283203, + 0.013998490758240223, + 0.005515867844223976, + 0.004249539691954851, + 0.003976117819547653, + 0.007746982388198376, + 0.012169931083917618, + 0.004365168511867523, + 0.012343998067080975, + 0.005980781279504299, + 0.002858694875612855, + 0.0017267909133806825, + 0.0028168046846985817, + 0.007398183457553387, + 0.01013578288257122, + 0.004022960551083088, + 0.004072484094649553, + 0.010349889285862446, + 0.009982967749238014, + 0.005636225920170546, + 0.005914514418691397, + 0.016662264242768288, + 0.013968822546303272, + 0.010502241551876068, + 0.008393101394176483, + 0.051086004823446274, + 0.009339177049696445, + 0.008214994333684444, + 0.004474857356399298, + 0.010685289278626442, + 0.07665283232927322, + 0.007231681142002344, + 0.025178806856274605, + 0.011991203762590885, + 0.009989271871745586, + 0.003092992352321744, + 0.005389623809605837, + 0.028261248022317886, + 0.018269900232553482, + 0.014649894088506699, + 0.008940114639699459, + 0.030446495860815048, + 0.006816650740802288, + 0.006015670485794544, + 0.00499371625483036, + 0.009146989323198795, + 0.04511621966958046, + 0.008095412515103817, + 0.013252809643745422, + 0.002465886762365699, + 0.0024299388751387596, + 0.0015645693056285381, + 0.0015508558135479689, + 0.003383408300578594, + 0.0033415157813578844, + 0.0018792659975588322, + 0.002470011357218027, + 0.06013071909546852, + 0.0156792551279068, + 0.014402514323592186, + 0.010838217101991177, + 0.04079213738441467, + 0.0860290676355362, + 0.013700373470783234, + 0.02107667736709118, + 0.002142018638551235, + 0.0011515889782458544, + 0.001280896714888513, + 0.0010624679271131754, + 0.0015479064313694835, + 0.002972352085635066, + 0.001339936861768365, + 0.0018320110393688083 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 4, + 6 + ], + "selected_reward": -0.20160697400569916 + } + ], + "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/0000026/group_6/index_tensor([6973, 6973])_group_6_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000026/group_6/index_tensor([6973, 6973])_group_6_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..e1e140deb5a1a4a9768e6aa02b1063ff08c8207e --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000026/group_6/index_tensor([6973, 6973])_group_6_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "In hyper-realistic CGI, a dynamic close-up tracking shot follows a fully ripe lemon as it dangles violently from its gnarled, sun-drenched branch. The camera remains locked on the lemon as it decisively separates and begins to tumble downward through the air, its oval shape slowly rotating. Intense, direct sunlight gleams through the microscopic pores of its bumpy, vibrant yellow rind, momentarily illuminating its waxy texture. As it falls, the background transforms into a soft, cinematic blur, the deep green of the leaves and the speckled light filtering through them blending into a beautiful, out-of-focus bokeh that enhances the sharp focus on the tumbling fruit. The movement is both abrupt and graceful, a momentary natural force captured with surreal detail and a palpable sense of weight and momentum.", + "seed": 12049, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000026/group_6/index_6973_seed_12049_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": -0.25656825352925805, + "vlm_logprob": -0.25656825352925805, + "vlm_yes_no_margin": 3.525, + "vlm_token_credit": -0.25656825352925805, + "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 fully ripe lemon with a vibrant yellow, bumpy rind?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the lemon start by dangling from a gnarled, sun-drenched branch before separating and tumbling downward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the lemon's oval shape visible rotating slowly as it falls through the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video feature a dynamic close-up tracking shot that remains locked on the lemon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a soft, out-of-focus bokeh of deep green leaves and speckled 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": [ + -0.004538592416793108 + ], + "opposite_logprob_tokens": [ + -6.379538536071777 + ], + "logprob_reward": -0.004538592416793108, + "yes_no_margin_reward": 6.375, + "token_credit_reward": -0.004538592416793108, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3955884873867035, + 0.30133622884750366, + 0.02541419118642807, + 0.018725771456956863, + 0.05128741264343262, + 0.05832892656326294, + 0.05653030425310135, + 0.0244899932295084, + 0.03894107788801193, + 0.029357578605413437 + ], + "credit_map": [ + 0.02738678641617298, + 0.006593627855181694, + 0.15830105543136597, + 0.016226302832365036, + 0.006368153728544712, + 0.019689591601490974, + 0.012073716148734093, + 0.1489492952823639, + 0.016175027936697006, + 0.00506465882062912, + 0.12160775065422058, + 0.011706002056598663, + 0.005148455500602722, + 0.014683583751320839, + 0.0087317805737257, + 0.11821895837783813, + 0.0032141038682311773, + 0.0024608697276562452, + 0.004074057564139366, + 0.003668523393571377, + 0.002186474157497287, + 0.003255440155044198, + 0.003576158545911312, + 0.0029785630758851767, + 0.0016322382725775242, + 0.0013605430722236633, + 0.0028340185526758432, + 0.0024536647833883762, + 0.0016010416438803077, + 0.002838586922734976, + 0.002956563141196966, + 0.0030491140205413103, + 0.0038283204194158316, + 0.004024642053991556, + 0.014354396611452103, + 0.003160573309287429, + 0.003073191037401557, + 0.005912263877689838, + 0.004835639148950577, + 0.012098386883735657, + 0.007015013135969639, + 0.008281925693154335, + 0.0091516999527812, + 0.010128623805940151, + 0.004477042704820633, + 0.008085510693490505, + 0.005587243940681219, + 0.005601865239441395, + 0.0050043025985360146, + 0.006165160797536373, + 0.009791379794478416, + 0.0071462299674749374, + 0.0063964007422327995, + 0.010684911161661148, + 0.006488828454166651, + 0.004853087943047285, + 0.0030449004843831062, + 0.0030975560657680035, + 0.0038185613229870796, + 0.003436170518398285, + 0.0021110509987920523, + 0.002843753667548299, + 0.0032071762252599, + 0.002930824412032962, + 0.0034606691915541887, + 0.005834563635289669, + 0.00670189643278718, + 0.004709345288574696, + 0.0036326455883681774, + 0.006745276506990194, + 0.004815215244889259, + 0.003041468095034361, + 0.002291196258738637, + 0.0033935904502868652, + 0.004902747459709644, + 0.004336697049438953, + 0.0027381093241274357, + 0.005944321397691965, + 0.0036371960304677486, + 0.0021137213334441185 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 5 + ], + "selected_reward": -0.004538592416793108 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a fully ripe lemon with a vibrant yellow, bumpy rind?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the lemon start by dangling from a gnarled, sun-drenched branch before separating and tumbling downward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the lemon's oval shape visible rotating slowly as it falls through the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video feature a dynamic close-up tracking shot that remains locked on the lemon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a soft, out-of-focus bokeh of deep green leaves and speckled 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": [ + -0.026351304724812508 + ], + "opposite_logprob_tokens": [ + -3.6513512134552 + ], + "logprob_reward": -0.026351304724812508, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026351304724812508, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2638412415981293, + 0.6443725824356079, + 0.02100389450788498, + 0.006782324053347111, + 0.01049153320491314, + 0.01355206873267889, + 0.00970142800360918, + 0.004780653398483992, + 0.0207633338868618, + 0.004710972309112549 + ], + "credit_map": [ + 0.015376358292996883, + 0.003775888355448842, + 0.1097094863653183, + 0.00840866006910801, + 0.002983092563226819, + 0.011159717105329037, + 0.006086844950914383, + 0.1063411682844162, + 0.035075172781944275, + 0.008560530841350555, + 0.27099230885505676, + 0.019849659875035286, + 0.007058108691126108, + 0.025913052260875702, + 0.013847081921994686, + 0.26307666301727295, + 0.0029472163878381252, + 0.0017254078993573785, + 0.0033493181690573692, + 0.002360873157158494, + 0.0015285784611478448, + 0.0028618923388421535, + 0.0022970112040638924, + 0.003933595027774572, + 0.0006463452591560781, + 0.0007598461816087365, + 0.0009089577943086624, + 0.0009825399611145258, + 0.0006860269349999726, + 0.0010329682845622301, + 0.0009013193193823099, + 0.0008643201435916126, + 0.0007500979118049145, + 0.0005724231014028192, + 0.0034155810717493296, + 0.0004985214327462018, + 0.00043841227306984365, + 0.0008520213887095451, + 0.0008217013091780245, + 0.003142774570733309, + 0.0015848048496991396, + 0.001363108167424798, + 0.002180708572268486, + 0.0022273946087807417, + 0.001291210064664483, + 0.0016151661984622478, + 0.002077556913718581, + 0.0012121177278459072, + 0.0009656777256168425, + 0.0009887574706226587, + 0.0012759745586663485, + 0.0011576976394280791, + 0.0010482852812856436, + 0.0023341914638876915, + 0.0011260699247941375, + 0.0008047737064771354, + 0.0004992200993001461, + 0.0005674617132171988, + 0.0008543986477889121, + 0.0007682509021833539, + 0.0003792000061366707, + 0.0006479351432062685, + 0.0005568424239754677, + 0.0005073444917798042, + 0.001680136425420642, + 0.003769458504393697, + 0.0031874682754278183, + 0.002341136336326599, + 0.0017501618713140488, + 0.003580213524401188, + 0.0027002247516065836, + 0.0017545310547575355, + 0.0003493216063361615, + 0.0005596705595962703, + 0.0009596225572749972, + 0.000779840862378478, + 0.0004248358600307256, + 0.0007302780286408961, + 0.0005689768586307764, + 0.0003384253941476345 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 2 + ], + "selected_reward": -0.026351304724812508 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a fully ripe lemon with a vibrant yellow, bumpy rind?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the lemon start by dangling from a gnarled, sun-drenched branch before separating and tumbling downward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the lemon's oval shape visible rotating slowly as it falls through the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video feature a dynamic close-up tracking shot that remains locked on the lemon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a soft, out-of-focus bokeh of deep green leaves and speckled 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": [ + -0.11294437944889069 + ], + "opposite_logprob_tokens": [ + -2.2379443645477295 + ], + "logprob_reward": -0.11294437944889069, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11294437944889069, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.029336361214518547, + 0.8828295469284058, + 0.02863641083240509, + 0.0071893297135829926, + 0.005281839054077864, + 0.00928948912769556, + 0.008279366418719292, + 0.0038033563178032637, + 0.022102896124124527, + 0.003251501591876149 + ], + "credit_map": [ + 0.0035467694979161024, + 0.0029082375112921, + 0.0037031129468232393, + 0.005618435796350241, + 0.0017117905663326383, + 0.004004376009106636, + 0.004052117001265287, + 0.003791520372033119, + 0.04922536388039589, + 0.01198501419275999, + 0.37076684832572937, + 0.02705107070505619, + 0.00944727286696434, + 0.0353691466152668, + 0.018684105947613716, + 0.36030063033103943, + 0.0037710180040448904, + 0.0023051684256643057, + 0.004437346011400223, + 0.0035722856409847736, + 0.0024307365529239178, + 0.004291572608053684, + 0.0030704683158546686, + 0.0047578150406479836, + 0.0004730319487862289, + 0.0004940317594446242, + 0.0008082209387794137, + 0.0007553565665148199, + 0.000514731858856976, + 0.0008180888253264129, + 0.0018287274288013577, + 0.0014971391065046191, + 0.0005183856701478362, + 0.0007909747073426843, + 0.0006603207439184189, + 0.0006272952305153012, + 0.0005881820106878877, + 0.0005115655949339271, + 0.0009902832098305225, + 0.0005948318284936249, + 0.00116480712313205, + 0.0010458369506523013, + 0.0013815220445394516, + 0.0009846327593550086, + 0.00113580166362226, + 0.00148441840428859, + 0.0010222896235063672, + 0.0010701795108616352, + 0.0009493674733676016, + 0.0008978726109489799, + 0.0019084915984421968, + 0.0013461294583976269, + 0.0005367855192162097, + 0.0011177619453519583, + 0.0007288646884262562, + 0.0007940924842841923, + 0.0004552232858259231, + 0.0005313350702635944, + 0.0005257829907350242, + 0.0005211098468862474, + 0.0004399501485750079, + 0.0005299179465509951, + 0.0005019822856411338, + 0.00029805424856022, + 0.0018043436575680971, + 0.003305261954665184, + 0.004100431222468615, + 0.0025074740406125784, + 0.0018768432782962918, + 0.003605717094615102, + 0.0029983024578541517, + 0.0019045195076614618, + 0.00029015837935730815, + 0.00043276918586343527, + 0.00044993290794081986, + 0.0004692482762038708, + 0.0003541376499924809, + 0.0006197553011588752, + 0.0003965690266340971, + 0.00023893047182355076 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 2 + ], + "selected_reward": -0.11294437944889069 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a fully ripe lemon with a vibrant yellow, bumpy rind?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the lemon start by dangling from a gnarled, sun-drenched branch before separating and tumbling downward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the lemon's oval shape visible rotating slowly as it falls through the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video feature a dynamic close-up tracking shot that remains locked on the lemon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a soft, out-of-focus bokeh of deep green leaves and speckled light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.1370640993118286 + ], + "opposite_logprob_tokens": [ + -0.3870640993118286 + ], + "logprob_reward": -1.1370640993118286, + "yes_no_margin_reward": -0.75, + "token_credit_reward": -1.1370640993118286, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.036502655595541, + 0.7692899107933044, + 0.056678786873817444, + 0.02058332972228527, + 0.019660858437418938, + 0.042067207396030426, + 0.02902115508913994, + 0.010586672462522984, + 0.009477908723056316, + 0.006131613161414862 + ], + "credit_map": [ + 0.006286463234573603, + 0.0031386485788971186, + 0.00555112911388278, + 0.00304355937987566, + 0.0019255821825936437, + 0.006855164188891649, + 0.003314066445454955, + 0.00638803793117404, + 0.04343174770474434, + 0.010559038259088993, + 0.32260140776634216, + 0.02337431162595749, + 0.00819336250424385, + 0.031024983152747154, + 0.016379712149500847, + 0.3137252628803253, + 0.00619346322491765, + 0.0028535681776702404, + 0.014627541415393353, + 0.003641619114205241, + 0.003836191026493907, + 0.007862565107643604, + 0.004169953055679798, + 0.013493875041604042, + 0.00241530267521739, + 0.001196172903291881, + 0.0019437646260485053, + 0.001988897332921624, + 0.0012633496662601829, + 0.002839406719431281, + 0.004393768031150103, + 0.004542666021734476, + 0.0014246205100789666, + 0.0013137554051354527, + 0.005985981319099665, + 0.0009916278067976236, + 0.0010333890095353127, + 0.0018983829068019986, + 0.0011955552035942674, + 0.005817546043545008, + 0.0033755514305084944, + 0.004466441925615072, + 0.008481364697217941, + 0.007592258043587208, + 0.004010593052953482, + 0.005290979519486427, + 0.005139719694852829, + 0.0037102955393493176, + 0.002825214061886072, + 0.0023870684672147036, + 0.0061790235340595245, + 0.00441198144108057, + 0.002710919827222824, + 0.00528907822445035, + 0.0026630989741533995, + 0.0025547670666128397, + 0.0010467790998518467, + 0.0013721603900194168, + 0.0016641670372337103, + 0.0013669254258275032, + 0.0011613199021667242, + 0.0015090497909113765, + 0.0013678256655111909, + 0.0010984438704326749, + 0.0014186346670612693, + 0.0014808793785050511, + 0.0017839843640103936, + 0.0007747070048935711, + 0.000752882391680032, + 0.0011060734977945685, + 0.0010851618135347962, + 0.001075585139915347, + 0.0006327525479719043, + 0.0008359042112715542, + 0.0008553517400287092, + 0.0007075773319229484, + 0.0006799033144488931, + 0.0011522488202899694, + 0.0007378771551884711, + 0.000529997400008142 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 2, + 5 + ], + "selected_reward": -1.1370640993118286 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a fully ripe lemon with a vibrant yellow, bumpy rind?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the lemon start by dangling from a gnarled, sun-drenched branch before separating and tumbling downward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the lemon's oval shape visible rotating slowly as it falls through the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video feature a dynamic close-up tracking shot that remains locked on the lemon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a soft, out-of-focus bokeh of deep green leaves and speckled 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": [ + -0.001942891743965447 + ], + "opposite_logprob_tokens": [ + -6.251943111419678 + ], + "logprob_reward": -0.001942891743965447, + "yes_no_margin_reward": 6.25, + "token_credit_reward": -0.001942891743965447, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2209809273481369, + 0.3556472659111023, + 0.06133177503943443, + 0.020103218033909798, + 0.035785458981990814, + 0.05246196687221527, + 0.038728147745132446, + 0.03870956227183342, + 0.15478113293647766, + 0.021470554172992706 + ], + "credit_map": [ + 0.013130638748407364, + 0.004393317271023989, + 0.0856633186340332, + 0.008283190429210663, + 0.0040527526289224625, + 0.014009692706167698, + 0.006997285410761833, + 0.08445074409246445, + 0.01994631439447403, + 0.005097975954413414, + 0.1484888195991516, + 0.012243162840604782, + 0.004560941364616156, + 0.014012299478054047, + 0.008230304345488548, + 0.14306744933128357, + 0.006152106449007988, + 0.003585238242521882, + 0.014432921074330807, + 0.005199480336159468, + 0.004810566548258066, + 0.007674433756619692, + 0.004627629648894072, + 0.01484939455986023, + 0.001978806685656309, + 0.001261615427210927, + 0.0023578868713229895, + 0.0019756583496928215, + 0.0018809252651408315, + 0.003213508054614067, + 0.0039313845336437225, + 0.0035034343600273132, + 0.002463500713929534, + 0.003503437852486968, + 0.00920940563082695, + 0.001760682207532227, + 0.002510417252779007, + 0.004477069713175297, + 0.002635464072227478, + 0.009225482121109962, + 0.004629276692867279, + 0.004134111572057009, + 0.011826465837657452, + 0.009355702437460423, + 0.003874216228723526, + 0.006403257604688406, + 0.00618300074711442, + 0.006055937614291906, + 0.0037531480193138123, + 0.004770217463374138, + 0.006563690025359392, + 0.006271666847169399, + 0.0034371016081422567, + 0.005957686807960272, + 0.004776280838996172, + 0.003198357066139579, + 0.003344160970300436, + 0.005752011202275753, + 0.005544384475797415, + 0.004570608027279377, + 0.004743918310850859, + 0.00674755685031414, + 0.0047894748859107494, + 0.0032174468506127596, + 0.012211926281452179, + 0.027524717152118683, + 0.02412881702184677, + 0.017209747806191444, + 0.013191904872655869, + 0.026965077966451645, + 0.020528078079223633, + 0.013020860031247139, + 0.0020888817962259054, + 0.002910939045250416, + 0.0033004239667207003, + 0.003509530797600746, + 0.0020798174664378166, + 0.0033844145946204662, + 0.0024728369899094105, + 0.0017237093998119235 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 8 + ], + "selected_reward": -0.001942891743965447 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "No", + "Yes" + ] + }, + { + "prompt": "In hyper-realistic CGI, a dynamic close-up tracking shot follows a fully ripe lemon as it dangles violently from its gnarled, sun-drenched branch. The camera remains locked on the lemon as it decisively separates and begins to tumble downward through the air, its oval shape slowly rotating. Intense, direct sunlight gleams through the microscopic pores of its bumpy, vibrant yellow rind, momentarily illuminating its waxy texture. As it falls, the background transforms into a soft, cinematic blur, the deep green of the leaves and the speckled light filtering through them blending into a beautiful, out-of-focus bokeh that enhances the sharp focus on the tumbling fruit. The movement is both abrupt and graceful, a momentary natural force captured with surreal detail and a palpable sense of weight and momentum.", + "seed": 12050, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000026/group_6/index_6973_seed_12050_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": -0.03525351455900818, + "vlm_logprob": -0.03525351455900818, + "vlm_yes_no_margin": 4.45, + "vlm_token_credit": -0.03525351455900818, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a fully ripe lemon with a vibrant yellow, bumpy rind?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the lemon start by dangling from a gnarled, sun-drenched branch before separating and tumbling downward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the lemon's oval shape visible rotating slowly as it falls through the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video feature a dynamic close-up tracking shot that remains locked on the lemon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a soft, out-of-focus bokeh of deep green leaves and speckled 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": [ + -0.006916272919625044 + ], + "opposite_logprob_tokens": [ + -5.506916046142578 + ], + "logprob_reward": -0.006916272919625044, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.006916272919625044, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.203008234500885, + 0.05136176198720932, + 0.08597590029239655, + 0.20095334947109222, + 0.05780632048845291, + 0.19972635805606842, + 0.048118412494659424, + 0.05668914318084717, + 0.053007952868938446, + 0.04335261508822441 + ], + "credit_map": [ + 0.027929721400141716, + 0.015130937099456787, + 0.026640109717845917, + 0.029552357271313667, + 0.0212249718606472, + 0.025127889588475227, + 0.03719786927103996, + 0.02020437642931938, + 0.009116802364587784, + 0.004754229448735714, + 0.006954185664653778, + 0.005000269040465355, + 0.006509355269372463, + 0.00611809641122818, + 0.007756843231618404, + 0.005151981022208929, + 0.007172903046011925, + 0.006847410928457975, + 0.009995114989578724, + 0.017474014312028885, + 0.009421016089618206, + 0.010328597389161587, + 0.015642885118722916, + 0.00909395981580019, + 0.019776783883571625, + 0.017749212682247162, + 0.023299366235733032, + 0.030462438240647316, + 0.037145812064409256, + 0.022719882428646088, + 0.033904943615198135, + 0.015894917771220207, + 0.007604815997183323, + 0.00576391164213419, + 0.008972102776169777, + 0.0071067120879888535, + 0.005526701919734478, + 0.007812729105353355, + 0.007389433681964874, + 0.007629909552633762, + 0.01835016906261444, + 0.02603289857506752, + 0.031197229400277138, + 0.025350358337163925, + 0.020606067031621933, + 0.030242156237363815, + 0.028798608109354973, + 0.019148875027894974, + 0.004596574231982231, + 0.005615641362965107, + 0.00564524857327342, + 0.006461287848651409, + 0.010321798734366894, + 0.004602152854204178, + 0.006347413640469313, + 0.0045282961800694466, + 0.0038016552571207285, + 0.005577852018177509, + 0.00813230685889721, + 0.007655476685613394, + 0.0052058021537959576, + 0.007670155260711908, + 0.009865335188806057, + 0.00878056138753891, + 0.008926181122660637, + 0.006072606425732374, + 0.007628125138580799, + 0.004846186842769384, + 0.004499587696045637, + 0.005824410356581211, + 0.008012625388801098, + 0.007198230363428593, + 0.005120307672768831, + 0.005239278078079224, + 0.005630552303045988, + 0.004511237144470215, + 0.004338027909398079, + 0.0038380210753530264, + 0.006810739636421204, + 0.007864447310566902 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 5 + ], + "selected_reward": -0.006916272919625044 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a fully ripe lemon with a vibrant yellow, bumpy rind?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the lemon start by dangling from a gnarled, sun-drenched branch before separating and tumbling downward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the lemon's oval shape visible rotating slowly as it falls through the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video feature a dynamic close-up tracking shot that remains locked on the lemon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a soft, out-of-focus bokeh of deep green leaves and speckled 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": [ + -0.012546301819384098 + ], + "opposite_logprob_tokens": [ + -4.387546539306641 + ], + "logprob_reward": -0.012546301819384098, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012546301819384098, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17765319347381592, + 0.027689632028341293, + 0.032142508774995804, + 0.02490811049938202, + 0.028552193194627762, + 0.03647300601005554, + 0.1460927128791809, + 0.07789725065231323, + 0.285465806722641, + 0.16312570869922638 + ], + "credit_map": [ + 0.021111706271767616, + 0.013612465001642704, + 0.025854935869574547, + 0.02851761132478714, + 0.01755368709564209, + 0.0183450598269701, + 0.03535445034503937, + 0.017303260043263435, + 0.0031937728635966778, + 0.0024943521711975336, + 0.003729880088940263, + 0.005005403887480497, + 0.003074578009545803, + 0.0029044447001069784, + 0.004168746527284384, + 0.0031184509862214327, + 0.007050514221191406, + 0.002695863600820303, + 0.0037511608097702265, + 0.004335141275078058, + 0.003193131648004055, + 0.0034097093157470226, + 0.0046818130649626255, + 0.0030251701828092337, + 0.0024649775587022305, + 0.0024757429491728544, + 0.0029052800964564085, + 0.0026088212616741657, + 0.0028943074867129326, + 0.004310643300414085, + 0.003328475868329406, + 0.003919860348105431, + 0.004835197236388922, + 0.002283772686496377, + 0.0037256216164678335, + 0.003013127250596881, + 0.0037316461093723774, + 0.0038493687752634287, + 0.0028510368429124355, + 0.004262419417500496, + 0.003032534848898649, + 0.002824339084327221, + 0.005178615916520357, + 0.0045771789737045765, + 0.003156563499942422, + 0.00629028957337141, + 0.004912490025162697, + 0.006500991526991129, + 0.012636633589863777, + 0.016171684488654137, + 0.01599758304655552, + 0.01695520058274269, + 0.039233073592185974, + 0.013486490584909916, + 0.021261559799313545, + 0.010350470431149006, + 0.006360393948853016, + 0.0077279917895793915, + 0.009150192141532898, + 0.01103311125189066, + 0.006853086873888969, + 0.012635277584195137, + 0.011435110121965408, + 0.012702074833214283, + 0.02646111696958542, + 0.03143227472901344, + 0.038307685405015945, + 0.03405332565307617, + 0.03766176849603653, + 0.029086774215102196, + 0.06376045942306519, + 0.024702351540327072, + 0.028865618631243706, + 0.02123405411839485, + 0.021645359694957733, + 0.015398402698338032, + 0.01958460547029972, + 0.018027599900960922, + 0.01832185685634613, + 0.020048193633556366 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 9 + ], + "selected_reward": -0.012546301819384098 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a fully ripe lemon with a vibrant yellow, bumpy rind?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the lemon start by dangling from a gnarled, sun-drenched branch before separating and tumbling downward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the lemon's oval shape visible rotating slowly as it falls through the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video feature a dynamic close-up tracking shot that remains locked on the lemon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a soft, out-of-focus bokeh of deep green leaves and speckled 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": [ + -0.04307127743959427 + ], + "opposite_logprob_tokens": [ + -3.1680712699890137 + ], + "logprob_reward": -0.04307127743959427, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04307127743959427, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.31747621297836304, + 0.027368098497390747, + 0.037106748670339584, + 0.1211385726928711, + 0.11063403636217117, + 0.024621479213237762, + 0.13781583309173584, + 0.05135847628116608, + 0.09657810628414154, + 0.07590249925851822 + ], + "credit_map": [ + 0.038798294961452484, + 0.024046508595347404, + 0.04555530101060867, + 0.053328175097703934, + 0.03139570727944374, + 0.032267212867736816, + 0.06472676992416382, + 0.027358224615454674, + 0.003830182831734419, + 0.0016120218206197023, + 0.004500993527472019, + 0.003779786406084895, + 0.002927639754489064, + 0.0033590556122362614, + 0.004396044183522463, + 0.0029623694717884064, + 0.004359791520982981, + 0.0034389100037515163, + 0.005404266528785229, + 0.005031433422118425, + 0.004170541185885668, + 0.0040748040191829205, + 0.00628249067813158, + 0.004344508983194828, + 0.012162109836935997, + 0.010629130527377129, + 0.015049444511532784, + 0.01472549606114626, + 0.027821600437164307, + 0.010255075991153717, + 0.020559702068567276, + 0.00993600208312273, + 0.012146196328103542, + 0.012126113288104534, + 0.01583682745695114, + 0.018877122551202774, + 0.008691161870956421, + 0.013212807476520538, + 0.015573794953525066, + 0.014170000329613686, + 0.002576197497546673, + 0.002335704630240798, + 0.0031409412622451782, + 0.0036285328678786755, + 0.0019056502496823668, + 0.0036786748096346855, + 0.0035133562050759792, + 0.0038424215745180845, + 0.013630113564431667, + 0.015883170068264008, + 0.014788425527513027, + 0.01656801998615265, + 0.034633710980415344, + 0.014802937395870686, + 0.017791617661714554, + 0.009717834182083607, + 0.0055726696737110615, + 0.006832218263298273, + 0.008152785710990429, + 0.004778943490236998, + 0.005152198486030102, + 0.005452749785035849, + 0.009764510206878185, + 0.0056523969396948814, + 0.013712186366319656, + 0.012575758621096611, + 0.011809042654931545, + 0.010765152983367443, + 0.011299319565296173, + 0.012699722312390804, + 0.012921443209052086, + 0.0107954740524292, + 0.010801995173096657, + 0.008718186989426613, + 0.012727968394756317, + 0.009880268946290016, + 0.00821736827492714, + 0.007007287349551916, + 0.009123828262090683, + 0.009425587020814419 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 3 + ], + "selected_reward": -0.04307127743959427 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a fully ripe lemon with a vibrant yellow, bumpy rind?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the lemon start by dangling from a gnarled, sun-drenched branch before separating and tumbling downward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the lemon's oval shape visible rotating slowly as it falls through the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video feature a dynamic close-up tracking shot that remains locked on the lemon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a soft, out-of-focus bokeh of deep green leaves and speckled 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": [ + -0.11292020976543427 + ], + "opposite_logprob_tokens": [ + -2.2379202842712402 + ], + "logprob_reward": -0.11292020976543427, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11292020976543427, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3548191785812378, + 0.044284749776124954, + 0.042743898928165436, + 0.060574669390916824, + 0.04317708685994148, + 0.027618272230029106, + 0.1446906179189682, + 0.07247792929410934, + 0.14016659557819366, + 0.0694471076130867 + ], + "credit_map": [ + 0.04330235347151756, + 0.025512341409921646, + 0.05162440985441208, + 0.05898385867476463, + 0.03559013828635216, + 0.03645186871290207, + 0.0722227618098259, + 0.03113142028450966, + 0.00812619086354971, + 0.0041495724581182, + 0.005914604756981134, + 0.003233077470213175, + 0.004911921918392181, + 0.008262413553893566, + 0.005209716036915779, + 0.004477249924093485, + 0.0051149772480130196, + 0.003191292518749833, + 0.006667018868029118, + 0.00515928165987134, + 0.0033650645054876804, + 0.006751906592398882, + 0.006128406617790461, + 0.006365950684994459, + 0.005622108466923237, + 0.004968942608684301, + 0.007199050858616829, + 0.007621278986334801, + 0.012827775441110134, + 0.005807951558381319, + 0.010373231954872608, + 0.006154327653348446, + 0.005985582247376442, + 0.003977376502007246, + 0.006005274131894112, + 0.008018404245376587, + 0.003950535319745541, + 0.005057319533079863, + 0.005454621277749538, + 0.004727973137050867, + 0.0041188765317201614, + 0.003053459571674466, + 0.0033049429766833782, + 0.004382650833576918, + 0.002131849527359009, + 0.0032167555764317513, + 0.003953958861529827, + 0.0034557790495455265, + 0.01755559630692005, + 0.014291196130216122, + 0.01598806120455265, + 0.015800759196281433, + 0.030169561505317688, + 0.019819872453808784, + 0.01779787428677082, + 0.013267687521874905, + 0.011301513761281967, + 0.008481783792376518, + 0.011518768034875393, + 0.004303231369704008, + 0.006943815853446722, + 0.013820488005876541, + 0.00791496317833662, + 0.008193365298211575, + 0.018412720412015915, + 0.021831713616847992, + 0.01857946068048477, + 0.014150720089673996, + 0.022143051028251648, + 0.01450779102742672, + 0.019988352432847023, + 0.010552781634032726, + 0.013148580677807331, + 0.009056868962943554, + 0.010886730626225471, + 0.007028298452496529, + 0.007652125786989927, + 0.004862836562097073, + 0.008477125316858292, + 0.008334539830684662 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 8 + ], + "selected_reward": -0.11292020976543427 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a fully ripe lemon with a vibrant yellow, bumpy rind?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the lemon start by dangling from a gnarled, sun-drenched branch before separating and tumbling downward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the lemon's oval shape visible rotating slowly as it falls through the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video feature a dynamic close-up tracking shot that remains locked on the lemon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a soft, out-of-focus bokeh of deep green leaves and speckled 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": [ + -0.0008135108510032296 + ], + "opposite_logprob_tokens": [ + -7.1258134841918945 + ], + "logprob_reward": -0.0008135108510032296, + "yes_no_margin_reward": 7.125, + "token_credit_reward": -0.0008135108510032296, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06372828036546707, + 0.13873744010925293, + 0.025257136672735214, + 0.05751747637987137, + 0.07304823398590088, + 0.0428902767598629, + 0.12338294088840485, + 0.06793244928121567, + 0.2677273452281952, + 0.13977845013141632 + ], + "credit_map": [ + 0.010288461111485958, + 0.006198331248015165, + 0.00854518637061119, + 0.005985068157315254, + 0.008381935767829418, + 0.009248018264770508, + 0.007547635585069656, + 0.007533645257353783, + 0.028680451214313507, + 0.012359345331788063, + 0.01848597824573517, + 0.013627356849610806, + 0.015105425380170345, + 0.01976332813501358, + 0.017080117017030716, + 0.013635437935590744, + 0.0034693710040301085, + 0.00167724653147161, + 0.00337881944142282, + 0.0038052864838391542, + 0.0023928123991936445, + 0.0034385949838906527, + 0.004180140793323517, + 0.002914865268394351, + 0.005822652019560337, + 0.004501005169004202, + 0.006673775147646666, + 0.005631086882203817, + 0.01264907792210579, + 0.008684320375323296, + 0.007759250700473785, + 0.005796308629214764, + 0.007913266308605671, + 0.006101804319769144, + 0.01042621023952961, + 0.013060078024864197, + 0.006252775900065899, + 0.010230684652924538, + 0.0089487349614501, + 0.01011468656361103, + 0.006888879928737879, + 0.0033988712821155787, + 0.006001986563205719, + 0.005461596883833408, + 0.004040715284645557, + 0.005982339382171631, + 0.004388227127492428, + 0.0067276619374752045, + 0.01303543895483017, + 0.013043210841715336, + 0.01324666291475296, + 0.013465565629303455, + 0.03135794401168823, + 0.013025710359215736, + 0.01685137301683426, + 0.009357035160064697, + 0.005515896249562502, + 0.008587301708757877, + 0.011323705315589905, + 0.006100829690694809, + 0.006725323852151632, + 0.008685847744345665, + 0.010445510037243366, + 0.010548037476837635, + 0.027839211747050285, + 0.03454187884926796, + 0.03231549635529518, + 0.027349162846803665, + 0.03464286029338837, + 0.03369947895407677, + 0.04140614718198776, + 0.03593308851122856, + 0.020504973828792572, + 0.01963873766362667, + 0.021216362714767456, + 0.015398390591144562, + 0.012262657284736633, + 0.015022998675704002, + 0.018335405737161636, + 0.017398929223418236 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 1 + ], + "selected_reward": -0.0008135108510032296 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In hyper-realistic CGI, a dynamic close-up tracking shot follows a fully ripe lemon as it dangles violently from its gnarled, sun-drenched branch. The camera remains locked on the lemon as it decisively separates and begins to tumble downward through the air, its oval shape slowly rotating. Intense, direct sunlight gleams through the microscopic pores of its bumpy, vibrant yellow rind, momentarily illuminating its waxy texture. As it falls, the background transforms into a soft, cinematic blur, the deep green of the leaves and the speckled light filtering through them blending into a beautiful, out-of-focus bokeh that enhances the sharp focus on the tumbling fruit. The movement is both abrupt and graceful, a momentary natural force captured with surreal detail and a palpable sense of weight and momentum.", + "seed": 12051, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000026/group_6/index_6973_seed_12051_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": -0.02205111035145819, + "vlm_logprob": -0.02205111035145819, + "vlm_yes_no_margin": 4.7, + "vlm_token_credit": -0.02205111035145819, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a fully ripe lemon with a vibrant yellow, bumpy rind?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the lemon start by dangling from a gnarled, sun-drenched branch before separating and tumbling downward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the lemon's oval shape visible rotating slowly as it falls through the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video feature a dynamic close-up tracking shot that remains locked on the lemon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a soft, out-of-focus bokeh of deep green leaves and speckled 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": [ + -0.0074604712426662445 + ], + "opposite_logprob_tokens": [ + -5.257460594177246 + ], + "logprob_reward": -0.0074604712426662445, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.0074604712426662445, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07976055890321732, + 0.16444267332553864, + 0.19597868621349335, + 0.0681518018245697, + 0.05239356309175491, + 0.054456014186143875, + 0.1468334197998047, + 0.06074080243706703, + 0.0276335459202528, + 0.14960895478725433 + ], + "credit_map": [ + 0.0064760735258460045, + 0.010555566288530827, + 0.011547363363206387, + 0.010675429366528988, + 0.009521856904029846, + 0.01249459944665432, + 0.007047981955111027, + 0.01144168060272932, + 0.01038007065653801, + 0.05042290315032005, + 0.012655400671064854, + 0.007473500911146402, + 0.013157141394913197, + 0.01162969134747982, + 0.007530728820711374, + 0.0511932335793972, + 0.017308205366134644, + 0.031343791633844376, + 0.022779695689678192, + 0.012622423470020294, + 0.032179079949855804, + 0.023561712354421616, + 0.029261544346809387, + 0.02692222222685814, + 0.005345802288502455, + 0.008181675337255001, + 0.006710936315357685, + 0.003502183360978961, + 0.009777745231986046, + 0.006705139297991991, + 0.011110927909612656, + 0.01681738905608654, + 0.003990335389971733, + 0.005180425476282835, + 0.007753504440188408, + 0.006172908004373312, + 0.0074559724889695644, + 0.005437575746327639, + 0.004919645842164755, + 0.011483194306492805, + 0.004536451306194067, + 0.0034267788287252188, + 0.011861947365105152, + 0.0031134947203099728, + 0.003065466647967696, + 0.007911065593361855, + 0.003877362236380577, + 0.0166634451597929, + 0.012565279379487038, + 0.011403711512684822, + 0.019464829936623573, + 0.017528902739286423, + 0.012546618469059467, + 0.01637190580368042, + 0.02622343972325325, + 0.03072873130440712, + 0.008445512503385544, + 0.0034371076617389917, + 0.00736656179651618, + 0.007480275351554155, + 0.00326272239908576, + 0.011160974390804768, + 0.008070489391684532, + 0.011517155915498734, + 0.005016034934669733, + 0.0019123939564451575, + 0.003668375313282013, + 0.004527411889284849, + 0.0021384304855018854, + 0.0039016008377075195, + 0.0030702350195497274, + 0.0033990636002272367, + 0.012013803236186504, + 0.009842527098953724, + 0.026068584993481636, + 0.018889615312218666, + 0.011103572323918343, + 0.01648005284368992, + 0.0206968542188406, + 0.0345139354467392 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 1, + 9 + ], + "selected_reward": -0.0074604712426662445 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a fully ripe lemon with a vibrant yellow, bumpy rind?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the lemon start by dangling from a gnarled, sun-drenched branch before separating and tumbling downward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the lemon's oval shape visible rotating slowly as it falls through the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video feature a dynamic close-up tracking shot that remains locked on the lemon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a soft, out-of-focus bokeh of deep green leaves and speckled 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": [ + -0.07002515345811844 + ], + "opposite_logprob_tokens": [ + -2.6950252056121826 + ], + "logprob_reward": -0.07002515345811844, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07002515345811844, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16317670047283173, + 0.14300782978534698, + 0.2194143384695053, + 0.039307232946157455, + 0.051824767142534256, + 0.07562355697154999, + 0.17589692771434784, + 0.07055293768644333, + 0.042910002171993256, + 0.0182857196778059 + ], + "credit_map": [ + 0.01803174987435341, + 0.017102031037211418, + 0.017018333077430725, + 0.011825424619019032, + 0.02331347018480301, + 0.029525335878133774, + 0.018185673281550407, + 0.02817467600107193, + 0.013756302185356617, + 0.04164775833487511, + 0.010521938093006611, + 0.009237990714609623, + 0.023795928806066513, + 0.021591436117887497, + 0.0067429612390697, + 0.015713520348072052, + 0.01872837170958519, + 0.04654887691140175, + 0.019121356308460236, + 0.01481781154870987, + 0.028251515701413155, + 0.024745063856244087, + 0.03176380693912506, + 0.03543752804398537, + 0.006312351208180189, + 0.0049199871718883514, + 0.005077187437564135, + 0.004538869485259056, + 0.005067695397883654, + 0.0034176912158727646, + 0.0035764877684414387, + 0.006396962329745293, + 0.003039843635633588, + 0.004817964043468237, + 0.007293395698070526, + 0.007869905792176723, + 0.00698804622516036, + 0.006924530956894159, + 0.0036294241435825825, + 0.011261659674346447, + 0.011841165833175182, + 0.007033398374915123, + 0.012222240678966045, + 0.0077095599845051765, + 0.006563133094459772, + 0.01016069296747446, + 0.008187259547412395, + 0.011906103231012821, + 0.013112416490912437, + 0.00832412764430046, + 0.03885004296898842, + 0.013101914897561073, + 0.008938138373196125, + 0.018754130229353905, + 0.019217781722545624, + 0.05559837445616722, + 0.010239806957542896, + 0.00407989788800478, + 0.00884322915226221, + 0.008695084601640701, + 0.003889409126713872, + 0.011349018663167953, + 0.00962501484900713, + 0.01383147481828928, + 0.004720673430711031, + 0.0023272668477147818, + 0.007417580112814903, + 0.004345690365880728, + 0.0035710223019123077, + 0.006324718240648508, + 0.004867200274020433, + 0.009335853159427643, + 0.00212086527608335, + 0.0010752155212685466, + 0.0027469173073768616, + 0.0031870699021965265, + 0.0015966164646670222, + 0.0020012082532048225, + 0.002811602083966136, + 0.0027462253347039223 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 6, + 0 + ], + "selected_reward": -0.07002515345811844 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a fully ripe lemon with a vibrant yellow, bumpy rind?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the lemon start by dangling from a gnarled, sun-drenched branch before separating and tumbling downward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the lemon's oval shape visible rotating slowly as it falls through the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video feature a dynamic close-up tracking shot that remains locked on the lemon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a soft, out-of-focus bokeh of deep green leaves and speckled 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": [ + -0.016070479527115822 + ], + "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": [ + 0.07303644716739655, + 0.1657101809978485, + 0.11753001809120178, + 0.039502423256635666, + 0.04103462025523186, + 0.02763167954981327, + 0.40315333008766174, + 0.0814017504453659, + 0.02343781106173992, + 0.02756178006529808 + ], + "credit_map": [ + 0.009829842485487461, + 0.006877826992422342, + 0.011534432880580425, + 0.005662722513079643, + 0.01017901860177517, + 0.009084269404411316, + 0.008480265736579895, + 0.011388065293431282, + 0.011371643282473087, + 0.06556851416826248, + 0.014901931397616863, + 0.007718615233898163, + 0.012933837249875069, + 0.02157011814415455, + 0.005149055738002062, + 0.026496464386582375, + 0.008935622870922089, + 0.03283323720097542, + 0.011696098372340202, + 0.0081010228022933, + 0.01274096593260765, + 0.01482793502509594, + 0.017097849398851395, + 0.011297270655632019, + 0.0056454818695783615, + 0.005043996963649988, + 0.004657280631363392, + 0.002750582527369261, + 0.005247240886092186, + 0.0037610267754644156, + 0.004498796071857214, + 0.007898014038801193, + 0.003617680398747325, + 0.004443140234798193, + 0.007251403760164976, + 0.004253176040947437, + 0.005236080847680569, + 0.004538383334875107, + 0.0036070067435503006, + 0.008087745867669582, + 0.004437659401446581, + 0.002768588252365589, + 0.00375260878354311, + 0.00221514655277133, + 0.0030605043284595013, + 0.003908076789230108, + 0.003490346483886242, + 0.0039987461641430855, + 0.03429644927382469, + 0.026080043986439705, + 0.07288837432861328, + 0.03889927268028259, + 0.02766280807554722, + 0.043600279837846756, + 0.05659852921962738, + 0.10312754660844803, + 0.008050937205553055, + 0.004553087987005711, + 0.012493878602981567, + 0.010163232684135437, + 0.00472955871373415, + 0.011207923293113708, + 0.010430163703858852, + 0.019772961735725403, + 0.0017265740316361189, + 0.0013138367794454098, + 0.003538752207532525, + 0.002341747749596834, + 0.0017681765602901578, + 0.003673550207167864, + 0.002726789563894272, + 0.006348384078592062, + 0.0030024568550288677, + 0.0013714204542338848, + 0.006448149681091309, + 0.0025215886998921633, + 0.00183784868568182, + 0.003001395147293806, + 0.004333254415541887, + 0.005045663565397263 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 1, + 2 + ], + "selected_reward": -0.016070479527115822 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a fully ripe lemon with a vibrant yellow, bumpy rind?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the lemon start by dangling from a gnarled, sun-drenched branch before separating and tumbling downward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the lemon's oval shape visible rotating slowly as it falls through the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video feature a dynamic close-up tracking shot that remains locked on the lemon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a soft, out-of-focus bokeh of deep green leaves and speckled 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": [ + -0.016064144670963287 + ], + "opposite_logprob_tokens": [ + -4.141064167022705 + ], + "logprob_reward": -0.016064144670963287, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016064144670963287, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08086928725242615, + 0.05803752690553665, + 0.39495012164115906, + 0.08316359668970108, + 0.03587067872285843, + 0.10740421712398529, + 0.10572389513254166, + 0.0320882648229599, + 0.08345910161733627, + 0.0184333398938179 + ], + "credit_map": [ + 0.007397525478154421, + 0.012724189087748528, + 0.0082487091422081, + 0.006697417702525854, + 0.013160709291696548, + 0.011734705418348312, + 0.010513757355511189, + 0.010392269119620323, + 0.004417758900672197, + 0.008529341779649258, + 0.009831789880990982, + 0.0032377284951508045, + 0.006836172193288803, + 0.005963804200291634, + 0.0036289487034082413, + 0.015591982752084732, + 0.03263753280043602, + 0.10210680961608887, + 0.03219126537442207, + 0.02487538754940033, + 0.05181568115949631, + 0.04831966012716293, + 0.059368349611759186, + 0.04363543912768364, + 0.007198150269687176, + 0.00886547565460205, + 0.008322831243276596, + 0.005454421043395996, + 0.013055798597633839, + 0.008076961152255535, + 0.014323951676487923, + 0.01786600612103939, + 0.0027770563028752804, + 0.004240874666720629, + 0.005402432754635811, + 0.004639945458620787, + 0.0046126400120556355, + 0.00572897307574749, + 0.0029094924684613943, + 0.005559262819588184, + 0.017206339165568352, + 0.011941653676331043, + 0.017520302906632423, + 0.008672177791595459, + 0.006315229460597038, + 0.015040993690490723, + 0.009026252664625645, + 0.021681267768144608, + 0.009928993880748749, + 0.006950646638870239, + 0.015429486520588398, + 0.012104956433176994, + 0.008487262763082981, + 0.012802192009985447, + 0.017613422125577927, + 0.022406933829188347, + 0.00566658703610301, + 0.0017866642447188497, + 0.00453862315043807, + 0.0035594776272773743, + 0.0021506091579794884, + 0.0054317861795425415, + 0.0032796207815408707, + 0.005674897693097591, + 0.007825804874300957, + 0.004940591752529144, + 0.013178270310163498, + 0.009240257553756237, + 0.006283901631832123, + 0.011401304975152016, + 0.01094396784901619, + 0.019644999876618385, + 0.0019609362352639437, + 0.0012471589725464582, + 0.002731921849772334, + 0.0024915076792240143, + 0.0015008591581135988, + 0.0035791194532066584, + 0.002563231624662876, + 0.0023586053866893053 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 5, + 6 + ], + "selected_reward": -0.016064144670963287 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a fully ripe lemon with a vibrant yellow, bumpy rind?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the lemon start by dangling from a gnarled, sun-drenched branch before separating and tumbling downward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the lemon's oval shape visible rotating slowly as it falls through the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video feature a dynamic close-up tracking shot that remains locked on the lemon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a soft, out-of-focus bokeh of deep green leaves and speckled 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": [ + -0.0006353028584271669 + ], + "opposite_logprob_tokens": [ + -7.375635147094727 + ], + "logprob_reward": -0.0006353028584271669, + "yes_no_margin_reward": 7.375, + "token_credit_reward": -0.0006353028584271669, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1059146448969841, + 0.09489773213863373, + 0.2580377757549286, + 0.08942043781280518, + 0.04619646444916725, + 0.14147335290908813, + 0.10380633175373077, + 0.05124706029891968, + 0.03625505045056343, + 0.07275116443634033 + ], + "credit_map": [ + 0.00945298746228218, + 0.014469863846898079, + 0.011773223988711834, + 0.008308538235723972, + 0.012360743246972561, + 0.01846020296216011, + 0.010955814272165298, + 0.020133275538682938, + 0.006879129912704229, + 0.03154611587524414, + 0.004192824941128492, + 0.0035715503618121147, + 0.010266193188726902, + 0.009281373582780361, + 0.004281777422875166, + 0.024878762662410736, + 0.020233891904354095, + 0.07674305140972137, + 0.017427897080779076, + 0.01310420036315918, + 0.02672390267252922, + 0.03251799941062927, + 0.04066544771194458, + 0.030621368438005447, + 0.009508038870990276, + 0.009527631103992462, + 0.014361342415213585, + 0.009418394416570663, + 0.011399517767131329, + 0.010383118875324726, + 0.0119243822991848, + 0.012898004613816738, + 0.004275903571397066, + 0.00508560985326767, + 0.007408871315419674, + 0.003946206532418728, + 0.006796520669013262, + 0.007879982702434063, + 0.003435524646192789, + 0.007367847487330437, + 0.028838202357292175, + 0.009732301346957684, + 0.02686670981347561, + 0.010042564943432808, + 0.00737904105335474, + 0.022013526409864426, + 0.013580898754298687, + 0.023020101711153984, + 0.00903905089944601, + 0.005102337803691626, + 0.024781186133623123, + 0.007803266867995262, + 0.005096223670989275, + 0.011240279302001, + 0.008281290531158447, + 0.032462697476148605, + 0.007935946807265282, + 0.0029274174012243748, + 0.008422588929533958, + 0.005502777174115181, + 0.004172398243099451, + 0.008431493304669857, + 0.004842786118388176, + 0.009011647664010525, + 0.002665169071406126, + 0.0022675690706819296, + 0.005177604965865612, + 0.005405230447649956, + 0.0036978560965508223, + 0.004347707610577345, + 0.004526134114712477, + 0.008167780004441738, + 0.006664291024208069, + 0.0032751569524407387, + 0.013940015807747841, + 0.005382811650633812, + 0.004214428830891848, + 0.01196037232875824, + 0.009351207874715328, + 0.017962882295250893 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 5, + 0 + ], + "selected_reward": -0.0006353028584271669 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In hyper-realistic CGI, a dynamic close-up tracking shot follows a fully ripe lemon as it dangles violently from its gnarled, sun-drenched branch. The camera remains locked on the lemon as it decisively separates and begins to tumble downward through the air, its oval shape slowly rotating. Intense, direct sunlight gleams through the microscopic pores of its bumpy, vibrant yellow rind, momentarily illuminating its waxy texture. As it falls, the background transforms into a soft, cinematic blur, the deep green of the leaves and the speckled light filtering through them blending into a beautiful, out-of-focus bokeh that enhances the sharp focus on the tumbling fruit. The movement is both abrupt and graceful, a momentary natural force captured with surreal detail and a palpable sense of weight and momentum.", + "seed": 12052, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000026/group_6/index_6973_seed_12052_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": -0.043435284274164584, + "vlm_logprob": -0.043435284274164584, + "vlm_yes_no_margin": 4.325, + "vlm_token_credit": -0.043435284274164584, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a fully ripe lemon with a vibrant yellow, bumpy rind?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the lemon start by dangling from a gnarled, sun-drenched branch before separating and tumbling downward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the lemon's oval shape visible rotating slowly as it falls through the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video feature a dynamic close-up tracking shot that remains locked on the lemon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a soft, out-of-focus bokeh of deep green leaves and speckled 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": [ + -0.004435344599187374 + ], + "opposite_logprob_tokens": [ + -6.2544355392456055 + ], + "logprob_reward": -0.004435344599187374, + "yes_no_margin_reward": 6.25, + "token_credit_reward": -0.004435344599187374, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08006493747234344, + 0.021252820268273354, + 0.020498856902122498, + 0.015392277389764786, + 0.03401191532611847, + 0.21129365265369415, + 0.05819261446595192, + 0.025304468348622322, + 0.017085308209061623, + 0.5169031620025635 + ], + "credit_map": [ + 0.007455979008227587, + 0.004778611473739147, + 0.009637993760406971, + 0.0037550379056483507, + 0.009354598820209503, + 0.012090308591723442, + 0.007992858067154884, + 0.024999534711241722, + 0.0021507900673896074, + 0.0021624157670885324, + 0.0031624273397028446, + 0.002492062048986554, + 0.002413245616480708, + 0.0038066913839429617, + 0.0027253383304923773, + 0.0023398459888994694, + 0.002055420773103833, + 0.0009058286086656153, + 0.0027469496708363295, + 0.001398871187120676, + 0.004054128658026457, + 0.0033684680238366127, + 0.0027439030818641186, + 0.0032252834644168615, + 0.0018978876760229468, + 0.0009963156189769506, + 0.0027084497269243, + 0.0012977065052837133, + 0.0014707776717841625, + 0.00292256660759449, + 0.0019760888535529375, + 0.00212248251773417, + 0.005616507958620787, + 0.002499269088730216, + 0.005324790719896555, + 0.005262947641313076, + 0.004373730160295963, + 0.003567334031686187, + 0.004121360834687948, + 0.0032459693029522896, + 0.020152708515524864, + 0.00952268484979868, + 0.013587473891675472, + 0.014560149982571602, + 0.017131496220827103, + 0.04497061297297478, + 0.02534165233373642, + 0.06602685153484344, + 0.0038778192829340696, + 0.007105020806193352, + 0.0052672154270112514, + 0.007544544525444508, + 0.019186319783329964, + 0.004558582324534655, + 0.0034952631685882807, + 0.007157838903367519, + 0.002871431177482009, + 0.0017500062240287662, + 0.0030872460920363665, + 0.0027317858766764402, + 0.0031268673483282328, + 0.005238755606114864, + 0.0028996265027672052, + 0.003598747542127967, + 0.002343556145206094, + 0.0017479440430179238, + 0.002291424199938774, + 0.002272191224619746, + 0.0018442691070958972, + 0.002356133423745632, + 0.0022592844907194376, + 0.001970501383766532, + 0.06972891837358475, + 0.010977613739669323, + 0.20552103221416473, + 0.039267394691705704, + 0.03514455631375313, + 0.047254886478185654, + 0.02487150952219963, + 0.08413723111152649 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 5, + 0 + ], + "selected_reward": -0.004435344599187374 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a fully ripe lemon with a vibrant yellow, bumpy rind?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the lemon start by dangling from a gnarled, sun-drenched branch before separating and tumbling downward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the lemon's oval shape visible rotating slowly as it falls through the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video feature a dynamic close-up tracking shot that remains locked on the lemon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a soft, out-of-focus bokeh of deep green leaves and speckled 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": [ + -0.02980387583374977 + ], + "opposite_logprob_tokens": [ + -3.529803991317749 + ], + "logprob_reward": -0.02980387583374977, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.02980387583374977, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.225518137216568, + 0.09760306030511856, + 0.05120484158396721, + 0.05202865228056908, + 0.08918227255344391, + 0.2019178867340088, + 0.04618706554174423, + 0.033325761556625366, + 0.048366159200668335, + 0.15466612577438354 + ], + "credit_map": [ + 0.029342973604798317, + 0.011305566877126694, + 0.03608138859272003, + 0.010722212493419647, + 0.026213614270091057, + 0.03387463837862015, + 0.019087478518486023, + 0.05889028310775757, + 0.01261636521667242, + 0.00822999607771635, + 0.019098250195384026, + 0.007745278999209404, + 0.00989020150154829, + 0.016940947622060776, + 0.00852309912443161, + 0.01455893088132143, + 0.006032150704413652, + 0.0028239539824426174, + 0.008629056625068188, + 0.0043598017655313015, + 0.005944507662206888, + 0.008755946531891823, + 0.006512409076094627, + 0.008147019892930984, + 0.004577335435897112, + 0.0030259843915700912, + 0.012140778824687004, + 0.0029472140595316887, + 0.003345992648974061, + 0.007312980946153402, + 0.0041246493346989155, + 0.014553720131516457, + 0.014858556911349297, + 0.006952059920877218, + 0.014367293566465378, + 0.011439581401646137, + 0.008933261036872864, + 0.011224735528230667, + 0.009920461103320122, + 0.011486324481666088, + 0.02057138830423355, + 0.021426338702440262, + 0.020401379093527794, + 0.025437042117118835, + 0.02735978364944458, + 0.03263213485479355, + 0.02625603787600994, + 0.027833783999085426, + 0.0044064512476325035, + 0.003966950345784426, + 0.0055534797720611095, + 0.009223336353898048, + 0.005424373783171177, + 0.006299273110926151, + 0.0060847108252346516, + 0.00522849103435874, + 0.004067162983119488, + 0.0028518415056169033, + 0.005261239595711231, + 0.002983557526022196, + 0.005531465169042349, + 0.005029771942645311, + 0.003541767131537199, + 0.004058957565575838, + 0.0062369052320718765, + 0.006292799022048712, + 0.005492736119776964, + 0.00682624988257885, + 0.005491890944540501, + 0.00597961712628603, + 0.007080004084855318, + 0.004965959116816521, + 0.020525915548205376, + 0.0036473844666033983, + 0.06113457307219505, + 0.012240360490977764, + 0.010228386148810387, + 0.014140861108899117, + 0.007340229582041502, + 0.02540842816233635 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 9 + ], + "selected_reward": -0.02980387583374977 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a fully ripe lemon with a vibrant yellow, bumpy rind?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the lemon start by dangling from a gnarled, sun-drenched branch before separating and tumbling downward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the lemon's oval shape visible rotating slowly as it falls through the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video feature a dynamic close-up tracking shot that remains locked on the lemon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a soft, out-of-focus bokeh of deep green leaves and speckled 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": [ + -0.05497494712471962 + ], + "opposite_logprob_tokens": [ + -2.9299750328063965 + ], + "logprob_reward": -0.05497494712471962, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05497494712471962, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09800421446561813, + 0.04851819574832916, + 0.025252876803278923, + 0.05891126021742821, + 0.083698570728302, + 0.19894662499427795, + 0.17429830133914948, + 0.035632625222206116, + 0.055949531495571136, + 0.22078774869441986 + ], + "credit_map": [ + 0.014391393400728703, + 0.007819345220923424, + 0.011989918537437916, + 0.0050398665480315685, + 0.011187509633600712, + 0.014590276405215263, + 0.0076341754756867886, + 0.02535172551870346, + 0.00413772277534008, + 0.0036853905767202377, + 0.009674553759396076, + 0.0045815142802894115, + 0.005103954579681158, + 0.008822669275105, + 0.005737032275646925, + 0.006775359157472849, + 0.003081921488046646, + 0.0013273208169266582, + 0.003926945384591818, + 0.002162112155929208, + 0.0027951642405241728, + 0.005060768686234951, + 0.0028009084053337574, + 0.004097737371921539, + 0.005782624706625938, + 0.002111641690135002, + 0.015620660036802292, + 0.0036414964124560356, + 0.0036601948086172342, + 0.005771283991634846, + 0.0051809195429086685, + 0.017142441123723984, + 0.011082314886152744, + 0.00886981375515461, + 0.012082095257937908, + 0.01240193285048008, + 0.007866800762712955, + 0.008885441347956657, + 0.013977100141346455, + 0.00853306706994772, + 0.013846649788320065, + 0.01627730205655098, + 0.01541314460337162, + 0.011719688773155212, + 0.02404981479048729, + 0.034968115389347076, + 0.02952510304749012, + 0.05314679816365242, + 0.021044228225946426, + 0.01653715968132019, + 0.020143376663327217, + 0.03668659180402756, + 0.02735239826142788, + 0.021618574857711792, + 0.016521388664841652, + 0.014394576661288738, + 0.00443221814930439, + 0.0035016858018934727, + 0.005036226473748684, + 0.003269190201535821, + 0.005484140478074551, + 0.0050403643399477005, + 0.004876218736171722, + 0.003992580808699131, + 0.008085835725069046, + 0.006893826182931662, + 0.006720810197293758, + 0.00685638515278697, + 0.007015268784016371, + 0.006225816905498505, + 0.008580878376960754, + 0.005570706911385059, + 0.02966739609837532, + 0.004851259756833315, + 0.08755472302436829, + 0.01701640710234642, + 0.01576879806816578, + 0.019708693027496338, + 0.010625354014337063, + 0.035595137625932693 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 5, + 6 + ], + "selected_reward": -0.05497494712471962 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a fully ripe lemon with a vibrant yellow, bumpy rind?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the lemon start by dangling from a gnarled, sun-drenched branch before separating and tumbling downward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the lemon's oval shape visible rotating slowly as it falls through the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video feature a dynamic close-up tracking shot that remains locked on the lemon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a soft, out-of-focus bokeh of deep green leaves and speckled 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": [ + -0.12704059481620789 + ], + "opposite_logprob_tokens": [ + -2.1270406246185303 + ], + "logprob_reward": -0.12704059481620789, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.12704059481620789, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.02118576131761074, + 0.010404136963188648, + 0.009233511984348297, + 0.022568099200725555, + 0.0170847587287426, + 0.7471222877502441, + 0.07350121438503265, + 0.023947980254888535, + 0.03349090740084648, + 0.04146136716008186 + ], + "credit_map": [ + 0.0024791774339973927, + 0.0023330289404839277, + 0.002366354689002037, + 0.00161367270629853, + 0.0021509940270334482, + 0.003101498121395707, + 0.0018242723308503628, + 0.005316765047609806, + 0.001055571949109435, + 0.0009132602717727423, + 0.002257850021123886, + 0.001012154039926827, + 0.001242281636223197, + 0.0013843867927789688, + 0.0013150438899174333, + 0.0012235885951668024, + 0.0012897064443677664, + 0.0007215907098725438, + 0.001077672466635704, + 0.001137163140811026, + 0.001100098597817123, + 0.001306543592363596, + 0.0014644311740994453, + 0.001136306207627058, + 0.0018325613345950842, + 0.0008774130255915225, + 0.006579173728823662, + 0.0008831843151710927, + 0.001349232392385602, + 0.002120314398780465, + 0.001406845636665821, + 0.007519375067204237, + 0.0024549446534365416, + 0.0020388804841786623, + 0.0022361709270626307, + 0.0022261578124016523, + 0.001534854294732213, + 0.0022462119814008474, + 0.0021954302210360765, + 0.0021521104499697685, + 0.06515642255544662, + 0.03945060819387436, + 0.044500820338726044, + 0.0491669736802578, + 0.06639602035284042, + 0.159270778298378, + 0.0860360637307167, + 0.2371445596218109, + 0.005600167904049158, + 0.008173475041985512, + 0.00633340934291482, + 0.011703528463840485, + 0.021661752834916115, + 0.006428690627217293, + 0.00535994116216898, + 0.008240253664553165, + 0.002619177335873246, + 0.001762576401233673, + 0.00419966597110033, + 0.0022935548331588507, + 0.0026814397424459457, + 0.002989639528095722, + 0.0029796918388456106, + 0.0044222380965948105, + 0.004324199631810188, + 0.004031817894428968, + 0.004078639205545187, + 0.0038857548497617245, + 0.004659067373722792, + 0.0038516325876116753, + 0.005214848555624485, + 0.0034449510276317596, + 0.005550192203372717, + 0.0009089751401916146, + 0.016293974593281746, + 0.00313216308131814, + 0.0027766097337007523, + 0.0038384380750358105, + 0.002033821539953351, + 0.0069271959364414215 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 6, + 9 + ], + "selected_reward": -0.12704059481620789 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a fully ripe lemon with a vibrant yellow, bumpy rind?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the lemon start by dangling from a gnarled, sun-drenched branch before separating and tumbling downward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the lemon's oval shape visible rotating slowly as it falls through the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video feature a dynamic close-up tracking shot that remains locked on the lemon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a soft, out-of-focus bokeh of deep green leaves and speckled 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": [ + -0.0009216589969582856 + ], + "opposite_logprob_tokens": [ + -7.000921726226807 + ], + "logprob_reward": -0.0009216589969582856, + "yes_no_margin_reward": 7.0, + "token_credit_reward": -0.0009216589969582856, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.036803774535655975, + 0.022939687594771385, + 0.028316082432866096, + 0.03276919573545456, + 0.015918362885713577, + 0.640305757522583, + 0.04015730693936348, + 0.020378902554512024, + 0.017157141119241714, + 0.14525386691093445 + ], + "credit_map": [ + 0.004982564598321915, + 0.0020878459326922894, + 0.006205615121871233, + 0.002264000941067934, + 0.004158833995461464, + 0.005184900015592575, + 0.0029002390801906586, + 0.009019769728183746, + 0.0021981699392199516, + 0.0015495361294597387, + 0.00501031382009387, + 0.002256605075672269, + 0.0029046726413071156, + 0.0028387238271534443, + 0.002542247297242284, + 0.0036394151393324137, + 0.003417747328057885, + 0.0014397847699001431, + 0.005320931319147348, + 0.0032334793359041214, + 0.0026397351175546646, + 0.004393626470118761, + 0.0034363737795501947, + 0.004434402100741863, + 0.003906959667801857, + 0.001592734712176025, + 0.006628946401178837, + 0.00250458437949419, + 0.0021793607156723738, + 0.006146419793367386, + 0.0035575772635638714, + 0.00625260965898633, + 0.002139490097761154, + 0.0006454710382968187, + 0.0028351396322250366, + 0.0010610843310132623, + 0.000887792615685612, + 0.0045445142313838005, + 0.0012751062167808414, + 0.0025297612883150578, + 0.05816008150577545, + 0.03333143889904022, + 0.038133326917886734, + 0.041912421584129333, + 0.057329464703798294, + 0.13541226089000702, + 0.0718444362282753, + 0.20418228209018707, + 0.004076572135090828, + 0.0040386756882071495, + 0.004605588968843222, + 0.007153133396059275, + 0.008433081209659576, + 0.004464388359338045, + 0.003171981777995825, + 0.004213877487927675, + 0.002962494269013405, + 0.0017496193759143353, + 0.002402064623311162, + 0.002428860403597355, + 0.0032566054724156857, + 0.0038294820114970207, + 0.0016443816712126136, + 0.002105392748489976, + 0.0020976329687982798, + 0.0018692075973376632, + 0.002133917296305299, + 0.0019956110045313835, + 0.0024400614202022552, + 0.002167175756767392, + 0.0025588797871023417, + 0.0018946516793221235, + 0.01946597918868065, + 0.003337848000228405, + 0.05720563232898712, + 0.011302194558084011, + 0.009801754727959633, + 0.013735827058553696, + 0.006999434903264046, + 0.0234051663428545 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 9, + 6 + ], + "selected_reward": -0.0009216589969582856 + } + ], + "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/0000026/group_8/index_tensor([5807, 5807])_group_8_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000026/group_8/index_tensor([5807, 5807])_group_8_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..938f5f69456ce8dc67272409881c76e9647b8efb --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000026/group_8/index_tensor([5807, 5807])_group_8_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "Told from a dreamlike first-person perspective, the camera glides forward in an incredibly smooth, floating motion across the vast, sun-drenched halls of the Baroque palace. Golden, three-dimensional sunlight pours through the soaring arched windows, illuminating the air with a visible divine glow and dancing across the polished marble floors. As the camera pans forward, the ornate gilded walls, covered in intricate Rococo-style gold leaf patterns, begin to gracefully deconstruct; golden plaster calmly dissolves at the edges, transforming into swirling cascades of shimmering, ethereal dust. Above, the vast frescoed ceiling, depicting a celestial scene of cherubs and clouds in soft, muted hues, undergoes a similarly magical weathering. The painted figures and sky gently condense, releasing their pigments into shimmering particles of light, joining in a floating dance. These magical dust particles, glowing with an inner glow, catch the brilliant sunlight and create a breathtaking, shimmering cloud that fills the vast space. A slow and beautiful process of magical weathering creates a tranquil and transcendent atmosphere, as the solid architecture melts into pure, floating magic.", + "seed": 1830, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000026/group_8/index_5807_seed_1830_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": -0.06124546593055129, + "vlm_logprob": -0.06124546593055129, + "vlm_yes_no_margin": 3.249999976158142, + "vlm_token_credit": -0.06124546593055129, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 vast Baroque palace with soaring arched windows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the ornate gilded walls feature intricate Rococo-style gold leaf patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the golden plaster on the walls dissolve and transform into swirling cascades of shimmering dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the frescoed ceiling depicting cherubs and clouds condense and release pigments into floating particles of light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera glide forward in a smooth, floating motion across the hall?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07744806259870529 + ], + "opposite_logprob_tokens": [ + -2.7024481296539307 + ], + "logprob_reward": -0.07744806259870529, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07744806259870529, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21981729567050934, + 0.20092108845710754, + 0.07763245701789856, + 0.0883733257651329, + 0.03360060602426529, + 0.18059390783309937, + 0.053654201328754425, + 0.09120194613933563, + 0.015553135424852371, + 0.03865215182304382 + ], + "credit_map": [ + 0.01787213608622551, + 0.05726727098226547, + 0.07331147789955139, + 0.02423590049147606, + 0.013032762333750725, + 0.008190548978745937, + 0.014977376908063889, + 0.010929788462817669, + 0.014680791646242142, + 0.06506432592868805, + 0.06584678590297699, + 0.019411839544773102, + 0.00863313302397728, + 0.007596905808895826, + 0.012329022400081158, + 0.007358286529779434, + 0.00819074921309948, + 0.01236689742654562, + 0.014884667471051216, + 0.009803537279367447, + 0.007565655279904604, + 0.00664426339790225, + 0.006669390946626663, + 0.011507292278110981, + 0.007748893462121487, + 0.016696657985448837, + 0.018675418570637703, + 0.012551823630928993, + 0.011801005341112614, + 0.00609208969399333, + 0.007241889834403992, + 0.007565537467598915, + 0.004629716742783785, + 0.006333956960588694, + 0.004748718813061714, + 0.005288530606776476, + 0.004937604069709778, + 0.0027829082682728767, + 0.002416453091427684, + 0.0024627132806926966, + 0.0633307546377182, + 0.016387201845645905, + 0.062203481793403625, + 0.013334380462765694, + 0.010188962332904339, + 0.00414956733584404, + 0.004271093290299177, + 0.006728438660502434, + 0.008940019644796848, + 0.007084055803716183, + 0.009663690812885761, + 0.006775117013603449, + 0.006093516480177641, + 0.004994383547455072, + 0.004660314880311489, + 0.005443098023533821, + 0.010946608148515224, + 0.01625259779393673, + 0.009392281994223595, + 0.009230771102011204, + 0.010989419184625149, + 0.018555104732513428, + 0.00739041343331337, + 0.008444741368293762, + 0.002544692950323224, + 0.0021627384703606367, + 0.001753459800966084, + 0.001897500827908516, + 0.0018722517415881157, + 0.0013912193244323134, + 0.002301726257428527, + 0.0016295441891998053, + 0.004135036841034889, + 0.008520594798028469, + 0.0035744919441640377, + 0.004202147014439106, + 0.00572377722710371, + 0.0059949965216219425, + 0.0033674335572868586, + 0.003133672522380948 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 5 + ], + "selected_reward": -0.07744806259870529 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 vast Baroque palace with soaring arched windows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the ornate gilded walls feature intricate Rococo-style gold leaf patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the golden plaster on the walls dissolve and transform into swirling cascades of shimmering dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the frescoed ceiling depicting cherubs and clouds condense and release pigments into floating particles of light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera glide forward in a smooth, floating motion across the hall?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.062127042561769485 + ], + "opposite_logprob_tokens": [ + -2.812127113342285 + ], + "logprob_reward": -0.062127042561769485, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.062127042561769485, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14961925148963928, + 0.17723149061203003, + 0.10119085013866425, + 0.08207967877388, + 0.02835727110505104, + 0.177534282207489, + 0.12591896951198578, + 0.07036163657903671, + 0.025550303980708122, + 0.06215634196996689 + ], + "credit_map": [ + 0.014173440635204315, + 0.03301982954144478, + 0.0455232672393322, + 0.018529009073972702, + 0.011019689030945301, + 0.00832857470959425, + 0.01069140899926424, + 0.00833402294665575, + 0.013595296069979668, + 0.05227385833859444, + 0.05266721919178963, + 0.020466258749365807, + 0.00825177039951086, + 0.007996662519872189, + 0.014390402473509312, + 0.007590008899569511, + 0.013803000561892986, + 0.01659504324197769, + 0.01946713961660862, + 0.010083179920911789, + 0.009981341660022736, + 0.007891662418842316, + 0.007694068364799023, + 0.01567540317773819, + 0.009205312468111515, + 0.015100877732038498, + 0.011735630221664906, + 0.013937221840023994, + 0.010469410568475723, + 0.006754544097930193, + 0.007200125139206648, + 0.007676548790186644, + 0.00406850827857852, + 0.004896547179669142, + 0.0038464567624032497, + 0.003938470035791397, + 0.0038597872480750084, + 0.00357407471165061, + 0.0021708253771066666, + 0.002002599649131298, + 0.05719427019357681, + 0.017108077183365822, + 0.05495522543787956, + 0.014100277796387672, + 0.010892855003476143, + 0.0061500356532633305, + 0.00551252905279398, + 0.011620991863310337, + 0.03756388649344444, + 0.010907510295510292, + 0.03735949844121933, + 0.011110447347164154, + 0.007751212455332279, + 0.007162928581237793, + 0.0076915062963962555, + 0.006371970288455486, + 0.006561108399182558, + 0.012406849302351475, + 0.006935938727110624, + 0.006292213220149279, + 0.008123461157083511, + 0.016014719381928444, + 0.006270783022046089, + 0.007756559643894434, + 0.004055399913340807, + 0.0037669995799660683, + 0.003043998032808304, + 0.003144049784168601, + 0.0027186768129467964, + 0.0026240707375109196, + 0.003089950419962406, + 0.003107155906036496, + 0.006569152232259512, + 0.01205423753708601, + 0.006095158867537975, + 0.008005976676940918, + 0.008243296295404434, + 0.00802415981888771, + 0.005895393434911966, + 0.007268961984664202 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 1, + 0 + ], + "selected_reward": -0.062127042561769485 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 vast Baroque palace with soaring arched windows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the ornate gilded walls feature intricate Rococo-style gold leaf patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the golden plaster on the walls dissolve and transform into swirling cascades of shimmering dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the frescoed ceiling depicting cherubs and clouds condense and release pigments into floating particles of light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera glide forward in a smooth, floating motion across the hall?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007639241870492697 + ], + "opposite_logprob_tokens": [ + -4.882639408111572 + ], + "logprob_reward": -0.007639241870492697, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007639241870492697, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0757003128528595, + 0.10622915625572205, + 0.08030355721712112, + 0.05587006360292435, + 0.037356920540332794, + 0.27926918864250183, + 0.16362734138965607, + 0.0948178768157959, + 0.037356216460466385, + 0.06946931779384613 + ], + "credit_map": [ + 0.008106970228254795, + 0.014900678768754005, + 0.020724322646856308, + 0.011487371288239956, + 0.005426920484751463, + 0.00627306941896677, + 0.004741991870105267, + 0.004038996063172817, + 0.008891330100595951, + 0.035714007914066315, + 0.037494439631700516, + 0.00603445153683424, + 0.004953095689415932, + 0.0036982696037739515, + 0.005243574734777212, + 0.004199992399662733, + 0.011888733133673668, + 0.01326340064406395, + 0.015397921204566956, + 0.006506901700049639, + 0.008182132616639137, + 0.007439056411385536, + 0.005287517327815294, + 0.012337898835539818, + 0.005503722932189703, + 0.013844577595591545, + 0.007952593266963959, + 0.011482935398817062, + 0.006254164036363363, + 0.0036158559378236532, + 0.0036023252177983522, + 0.003613889217376709, + 0.006106814835220575, + 0.0072807990945875645, + 0.004647504072636366, + 0.007482015993446112, + 0.004543975926935673, + 0.002673554001376033, + 0.002104364801198244, + 0.0025178962387144566, + 0.09212938696146011, + 0.03340456634759903, + 0.08804004639387131, + 0.021052207797765732, + 0.017739182338118553, + 0.006378925871104002, + 0.007602941244840622, + 0.012921939603984356, + 0.05304674804210663, + 0.0117514468729496, + 0.05285707488656044, + 0.011418571695685387, + 0.010845012962818146, + 0.006935689598321915, + 0.009036476723849773, + 0.007736337836831808, + 0.014241798780858517, + 0.015947921201586723, + 0.00925054308027029, + 0.009734654799103737, + 0.011290183290839195, + 0.018220186233520508, + 0.007796997204422951, + 0.008335600607097149, + 0.005017178598791361, + 0.004196994472295046, + 0.004540804773569107, + 0.0070154424756765366, + 0.00426647812128067, + 0.003655318869277835, + 0.00515458220615983, + 0.003509420668706298, + 0.008682959713041782, + 0.013057246804237366, + 0.007246330846101046, + 0.011862090788781643, + 0.006252871826291084, + 0.007800930179655552, + 0.006153582129627466, + 0.008413310162723064 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 6, + 1 + ], + "selected_reward": -0.007639241870492697 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 vast Baroque palace with soaring arched windows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the ornate gilded walls feature intricate Rococo-style gold leaf patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the golden plaster on the walls dissolve and transform into swirling cascades of shimmering dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the frescoed ceiling depicting cherubs and clouds condense and release pigments into floating particles of light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera glide forward in a smooth, floating motion across the hall?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.14294801652431488 + ], + "opposite_logprob_tokens": [ + -2.0179479122161865 + ], + "logprob_reward": -0.14294801652431488, + "yes_no_margin_reward": 1.8749998807907104, + "token_credit_reward": -0.14294801652431488, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0724736824631691, + 0.36360371112823486, + 0.04457133263349533, + 0.06311843544244766, + 0.025532033294439316, + 0.13861536979675293, + 0.14045217633247375, + 0.08114077895879745, + 0.02387228049337864, + 0.0466202087700367 + ], + "credit_map": [ + 0.0058626290410757065, + 0.018473036587238312, + 0.021165622398257256, + 0.007803894113749266, + 0.004623730666935444, + 0.0033900474663823843, + 0.006709029898047447, + 0.004445691127330065, + 0.027988776564598083, + 0.13852916657924652, + 0.14249275624752045, + 0.014031547121703625, + 0.013825633563101292, + 0.009692061692476273, + 0.008560148999094963, + 0.008483633399009705, + 0.003838211763650179, + 0.006268379744142294, + 0.00699974037706852, + 0.004675234667956829, + 0.007213750388473272, + 0.00820518471300602, + 0.0031437550205737352, + 0.004227076657116413, + 0.007382584735751152, + 0.016254281625151634, + 0.008357222191989422, + 0.012530221603810787, + 0.006167047191411257, + 0.004108837805688381, + 0.00441759591922164, + 0.003900643903762102, + 0.0035538875963538885, + 0.005414064973592758, + 0.00352388434112072, + 0.0032566897571086884, + 0.003763501299545169, + 0.0026015769690275192, + 0.00154294993262738, + 0.0018754792399704456, + 0.05002056434750557, + 0.01036179531365633, + 0.048995692282915115, + 0.010559736751019955, + 0.007237169425934553, + 0.003268458414822817, + 0.0029080486856400967, + 0.005263902712613344, + 0.048054955899715424, + 0.009138410910964012, + 0.04937995597720146, + 0.010736576281487942, + 0.006886741146445274, + 0.004408310633152723, + 0.006879101973026991, + 0.004968129564076662, + 0.0074638803489506245, + 0.01596974954009056, + 0.010084628127515316, + 0.00898150447756052, + 0.00991274043917656, + 0.014293747954070568, + 0.006474160589277744, + 0.0079603660851717, + 0.0045514958910644054, + 0.0029020949732512236, + 0.0029029964935034513, + 0.0032452335581183434, + 0.0019710459746420383, + 0.002760958392173052, + 0.0024576475843787193, + 0.003080809721723199, + 0.004749916028231382, + 0.009059455245733261, + 0.004577119369059801, + 0.005509327165782452, + 0.006041419226676226, + 0.006849500350654125, + 0.004294269718229771, + 0.005539202596992254 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 6, + 5 + ], + "selected_reward": -0.14294801652431488 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 vast Baroque palace with soaring arched windows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the ornate gilded walls feature intricate Rococo-style gold leaf patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the golden plaster on the walls dissolve and transform into swirling cascades of shimmering dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the frescoed ceiling depicting cherubs and clouds condense and release pigments into floating particles of light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera glide forward in a smooth, floating motion across the hall?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016064966097474098 + ], + "opposite_logprob_tokens": [ + -4.1410651206970215 + ], + "logprob_reward": -0.016064966097474098, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016064966097474098, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08962079137563705, + 0.555376410484314, + 0.046345122158527374, + 0.05040566250681877, + 0.019473055377602577, + 0.08712522685527802, + 0.06823206692934036, + 0.029691416770219803, + 0.031682178378105164, + 0.022048119455575943 + ], + "credit_map": [ + 0.009416047483682632, + 0.021048130467534065, + 0.022956307977437973, + 0.008192349225282669, + 0.007370644714683294, + 0.006329444237053394, + 0.007569297216832638, + 0.006738560739904642, + 0.042542289942502975, + 0.21060964465141296, + 0.21624122560024261, + 0.02275603637099266, + 0.021452557295560837, + 0.014884034171700478, + 0.013444487005472183, + 0.013446060940623283, + 0.005018503870815039, + 0.007957388646900654, + 0.0077015794813632965, + 0.0062823183834552765, + 0.005634724628180265, + 0.00585070950910449, + 0.00423237681388855, + 0.0036675145383924246, + 0.005549371242523193, + 0.012410776689648628, + 0.011093799024820328, + 0.008060593158006668, + 0.004520806483924389, + 0.0030258712358772755, + 0.003105402924120426, + 0.002639035228639841, + 0.0030409670434892178, + 0.005851162131875753, + 0.0019824898336082697, + 0.003110436024144292, + 0.0022903182543814182, + 0.0011423800606280565, + 0.0010170845780521631, + 0.001038214541040361, + 0.024058209732174873, + 0.011474132537841797, + 0.024685684591531754, + 0.006805392447859049, + 0.007088480982929468, + 0.003361969254910946, + 0.0024935572873800993, + 0.007157792337238789, + 0.018248634412884712, + 0.00640623876824975, + 0.016920870169997215, + 0.00492350198328495, + 0.006216449197381735, + 0.005639158189296722, + 0.004987659398466349, + 0.004889552481472492, + 0.002825329313054681, + 0.005363557022064924, + 0.0041525703854858875, + 0.0035289516672492027, + 0.003723947796970606, + 0.004590913653373718, + 0.00308990734629333, + 0.0024162358604371548, + 0.0036647836677730083, + 0.005103420931845903, + 0.004425020422786474, + 0.006153278052806854, + 0.0034208204597234726, + 0.0025374609977006912, + 0.003660102840512991, + 0.0027172917034476995, + 0.003288588020950556, + 0.00338281923905015, + 0.002603680593892932, + 0.0030426352750509977, + 0.0024283388629555702, + 0.0028174843173474073, + 0.002414425602182746, + 0.0020701454486697912 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 5 + ], + "selected_reward": -0.016064966097474098 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "Told from a dreamlike first-person perspective, the camera glides forward in an incredibly smooth, floating motion across the vast, sun-drenched halls of the Baroque palace. Golden, three-dimensional sunlight pours through the soaring arched windows, illuminating the air with a visible divine glow and dancing across the polished marble floors. As the camera pans forward, the ornate gilded walls, covered in intricate Rococo-style gold leaf patterns, begin to gracefully deconstruct; golden plaster calmly dissolves at the edges, transforming into swirling cascades of shimmering, ethereal dust. Above, the vast frescoed ceiling, depicting a celestial scene of cherubs and clouds in soft, muted hues, undergoes a similarly magical weathering. The painted figures and sky gently condense, releasing their pigments into shimmering particles of light, joining in a floating dance. These magical dust particles, glowing with an inner glow, catch the brilliant sunlight and create a breathtaking, shimmering cloud that fills the vast space. A slow and beautiful process of magical weathering creates a tranquil and transcendent atmosphere, as the solid architecture melts into pure, floating magic.", + "seed": 1831, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000026/group_8/index_5807_seed_1831_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": -0.012988507188856602, + "vlm_logprob": -0.012988507188856602, + "vlm_yes_no_margin": 4.575, + "vlm_token_credit": -0.012988507188856602, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 vast Baroque palace with soaring arched windows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the ornate gilded walls feature intricate Rococo-style gold leaf patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the golden plaster on the walls dissolve and transform into swirling cascades of shimmering dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the frescoed ceiling depicting cherubs and clouds condense and release pigments into floating particles of light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera glide forward in a smooth, floating motion across the hall?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007185568567365408 + ], + "opposite_logprob_tokens": [ + -5.257185459136963 + ], + "logprob_reward": -0.007185568567365408, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.007185568567365408, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.36192816495895386, + 0.06105465069413185, + 0.12901973724365234, + 0.06680069118738174, + 0.05753733217716217, + 0.10963619500398636, + 0.06426909565925598, + 0.0670703649520874, + 0.04900212958455086, + 0.03368169069290161 + ], + "credit_map": [ + 0.04337335750460625, + 0.0873333290219307, + 0.09131451696157455, + 0.0403987355530262, + 0.02378673106431961, + 0.019035717472434044, + 0.02919713966548443, + 0.02748860977590084, + 0.010374638251960278, + 0.009218970313668251, + 0.00816491711884737, + 0.006350804585963488, + 0.006422128994017839, + 0.006025868933647871, + 0.00796740222722292, + 0.006529915612190962, + 0.033565450459718704, + 0.014185740612447262, + 0.010790684260427952, + 0.012220022268593311, + 0.00947448518127203, + 0.00746920146048069, + 0.031522709876298904, + 0.009791434742510319, + 0.011672932654619217, + 0.009351932443678379, + 0.008622675202786922, + 0.0056130425073206425, + 0.0070449295453727245, + 0.007419436704367399, + 0.006576641462743282, + 0.010499094612896442, + 0.007443287409842014, + 0.007939309813082218, + 0.009739633649587631, + 0.008214722387492657, + 0.004770178813487291, + 0.00878159049898386, + 0.005762944929301739, + 0.004885661415755749, + 0.010274812579154968, + 0.01401529274880886, + 0.03728475049138069, + 0.009500126354396343, + 0.009430731646716595, + 0.00820426270365715, + 0.008287887088954449, + 0.012638323940336704, + 0.0047995299100875854, + 0.014335988089442253, + 0.007451469544321299, + 0.008372372947633266, + 0.009521194733679295, + 0.005211051553487778, + 0.006775516550987959, + 0.007801967207342386, + 0.005409323610365391, + 0.00895706843584776, + 0.013207798823714256, + 0.0085389269515872, + 0.007940933108329773, + 0.007161471992731094, + 0.009820712730288506, + 0.006034126505255699, + 0.004721003584563732, + 0.006671398412436247, + 0.004777081776410341, + 0.007236870937049389, + 0.005133415572345257, + 0.006108853034675121, + 0.005126029718667269, + 0.00922747515141964, + 0.0029037543572485447, + 0.003985119052231312, + 0.0036264685913920403, + 0.007220450788736343, + 0.0035821772180497646, + 0.0058601838536560535, + 0.0032956332433968782, + 0.0032079005613923073 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 5 + ], + "selected_reward": -0.007185568567365408 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 vast Baroque palace with soaring arched windows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the ornate gilded walls feature intricate Rococo-style gold leaf patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the golden plaster on the walls dissolve and transform into swirling cascades of shimmering dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the frescoed ceiling depicting cherubs and clouds condense and release pigments into floating particles of light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera glide forward in a smooth, floating motion across the hall?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014216925017535686 + ], + "opposite_logprob_tokens": [ + -4.264216899871826 + ], + "logprob_reward": -0.014216925017535686, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014216925017535686, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23625417053699493, + 0.07307185977697372, + 0.10159400850534439, + 0.0860305055975914, + 0.05341004952788353, + 0.06104021519422531, + 0.13435277342796326, + 0.06768185645341873, + 0.1420758068561554, + 0.04448878765106201 + ], + "credit_map": [ + 0.02457060106098652, + 0.06326426565647125, + 0.051298435777425766, + 0.026151929050683975, + 0.01914067566394806, + 0.013544702902436256, + 0.02010885812342167, + 0.018174685537815094, + 0.007382941897958517, + 0.0128379100933671, + 0.008360709995031357, + 0.010448555462062359, + 0.009008549153804779, + 0.006219914648681879, + 0.009285654872655869, + 0.009527621790766716, + 0.028522614389657974, + 0.009411284700036049, + 0.008152562193572521, + 0.008903649635612965, + 0.005995017010718584, + 0.005877260118722916, + 0.02775595337152481, + 0.00697567081078887, + 0.013514203950762749, + 0.010786795988678932, + 0.010266712866723537, + 0.007304752711206675, + 0.009653083048760891, + 0.012403390370309353, + 0.008907249197363853, + 0.013194315135478973, + 0.006326720584183931, + 0.006588018033653498, + 0.007112086284905672, + 0.010565949603915215, + 0.00480826897546649, + 0.005320041906088591, + 0.0077184345573186874, + 0.004970534238964319, + 0.007385936100035906, + 0.01021080557256937, + 0.007455840241163969, + 0.008417002856731415, + 0.009674138389527798, + 0.006867325399070978, + 0.007022195961326361, + 0.004006973467767239, + 0.00899839773774147, + 0.028773007914423943, + 0.014560566283762455, + 0.017431143671274185, + 0.022604988887906075, + 0.01603912189602852, + 0.011376670561730862, + 0.014568870887160301, + 0.005881197284907103, + 0.006887892726808786, + 0.014056982472538948, + 0.008773363195359707, + 0.005926190409809351, + 0.009146715514361858, + 0.00910855457186699, + 0.007900958880782127, + 0.02342744544148445, + 0.018345648422837257, + 0.013872814364731312, + 0.014246425591409206, + 0.02339431457221508, + 0.023503292351961136, + 0.011748060584068298, + 0.013537811115384102, + 0.004542395006865263, + 0.0056601460091769695, + 0.0043972693383693695, + 0.008561911061406136, + 0.004929346032440662, + 0.007390149869024754, + 0.004302657674998045, + 0.004704911261796951 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 6 + ], + "selected_reward": -0.014216925017535686 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 vast Baroque palace with soaring arched windows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the ornate gilded walls feature intricate Rococo-style gold leaf patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the golden plaster on the walls dissolve and transform into swirling cascades of shimmering dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the frescoed ceiling depicting cherubs and clouds condense and release pigments into floating particles of light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera glide forward in a smooth, floating motion across the hall?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00464135454967618 + ], + "opposite_logprob_tokens": [ + -5.379641532897949 + ], + "logprob_reward": -0.00464135454967618, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.00464135454967618, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09937520325183868, + 0.09679586440324783, + 0.17018583416938782, + 0.04727596044540405, + 0.06473623961210251, + 0.19358700513839722, + 0.11655761301517487, + 0.0831153616309166, + 0.09773026406764984, + 0.03064059466123581 + ], + "credit_map": [ + 0.012718407437205315, + 0.014689539559185505, + 0.014456799253821373, + 0.012792818248271942, + 0.009918843396008015, + 0.006150336004793644, + 0.017045732587575912, + 0.011602723971009254, + 0.009889178909361362, + 0.01552582997828722, + 0.01119875069707632, + 0.014124562963843346, + 0.010143718682229519, + 0.00846939068287611, + 0.012372531928122044, + 0.01507190428674221, + 0.049552902579307556, + 0.013941755518317223, + 0.009023357182741165, + 0.012640390545129776, + 0.010114191100001335, + 0.009293876588344574, + 0.056949418038129807, + 0.008669929578900337, + 0.007181755267083645, + 0.006185394246131182, + 0.005829483270645142, + 0.0055924984626472, + 0.005893561989068985, + 0.005634077358990908, + 0.004991145338863134, + 0.005968041718006134, + 0.010255394503474236, + 0.00869205966591835, + 0.00947022344917059, + 0.00944590475410223, + 0.004850866738706827, + 0.008740616030991077, + 0.0069471546448767185, + 0.006334019359201193, + 0.014025445096194744, + 0.030311772599816322, + 0.04704868420958519, + 0.021132951602339745, + 0.02424550987780094, + 0.016407473012804985, + 0.022515425458550453, + 0.017899740487337112, + 0.011448645032942295, + 0.020386215299367905, + 0.014843129552900791, + 0.010106422938406467, + 0.021779488772153854, + 0.015933237969875336, + 0.011167069897055626, + 0.010893408209085464, + 0.008101210929453373, + 0.010761207900941372, + 0.012468818575143814, + 0.011413010768592358, + 0.009530787356197834, + 0.010096773505210876, + 0.012188494205474854, + 0.008555062115192413, + 0.013299456797540188, + 0.012884491123259068, + 0.010156082920730114, + 0.013920966535806656, + 0.012166080065071583, + 0.01522455271333456, + 0.0106447609141469, + 0.00943388044834137, + 0.0029628570191562176, + 0.004511995241045952, + 0.002762318355962634, + 0.005157726816833019, + 0.004039968363940716, + 0.00457925396040082, + 0.003442357527092099, + 0.00318411854095757 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 2, + 6 + ], + "selected_reward": -0.00464135454967618 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 vast Baroque palace with soaring arched windows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the ornate gilded walls feature intricate Rococo-style gold leaf patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the golden plaster on the walls dissolve and transform into swirling cascades of shimmering dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the frescoed ceiling depicting cherubs and clouds condense and release pigments into floating particles of light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera glide forward in a smooth, floating motion across the hall?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02636663056910038 + ], + "opposite_logprob_tokens": [ + -3.651366710662842 + ], + "logprob_reward": -0.02636663056910038, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.02636663056910038, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11700480431318283, + 0.10924454033374786, + 0.12791933119297028, + 0.058782078325748444, + 0.05832694098353386, + 0.15226924419403076, + 0.1457134634256363, + 0.13088449835777283, + 0.05868781730532646, + 0.04116722568869591 + ], + "credit_map": [ + 0.013467662036418915, + 0.02160540781915188, + 0.021458175033330917, + 0.014103163965046406, + 0.010476247407495975, + 0.007356794085353613, + 0.01633208990097046, + 0.012205269187688828, + 0.013172462582588196, + 0.015417645685374737, + 0.01404672209173441, + 0.0148233687505126, + 0.011459930799901485, + 0.010730027221143246, + 0.01327640749514103, + 0.01631798781454563, + 0.03786385431885719, + 0.00989750400185585, + 0.006165511906147003, + 0.009459668770432472, + 0.007493085693567991, + 0.006796592380851507, + 0.04352787137031555, + 0.006715245079249144, + 0.0096234530210495, + 0.0069414833560585976, + 0.007408872712403536, + 0.007388677913695574, + 0.006596777588129044, + 0.007242331746965647, + 0.005415080580860376, + 0.00816540140658617, + 0.007369755767285824, + 0.008082481101155281, + 0.008160257712006569, + 0.008079995401203632, + 0.007346917875111103, + 0.008296739310026169, + 0.006443424616008997, + 0.004547369200736284, + 0.011576329357922077, + 0.022724155336618423, + 0.04971778765320778, + 0.013687501661479473, + 0.015487761236727238, + 0.011001421138644218, + 0.011463470757007599, + 0.016610827296972275, + 0.011680237017571926, + 0.030574532225728035, + 0.013684166595339775, + 0.015830758959054947, + 0.030905574560165405, + 0.018737051635980606, + 0.011548152193427086, + 0.012752990238368511, + 0.01550783310085535, + 0.016994701698422432, + 0.018438221886754036, + 0.016673577949404716, + 0.014081605710089207, + 0.020983673632144928, + 0.014467833563685417, + 0.013737034052610397, + 0.007401464972645044, + 0.010330159217119217, + 0.0053685931488871574, + 0.006121227517724037, + 0.008970250375568867, + 0.007288369350135326, + 0.005778791848570108, + 0.007428955752402544, + 0.003761052852496505, + 0.005228116177022457, + 0.004139616619795561, + 0.006616548169404268, + 0.005243220366537571, + 0.0080771055072546, + 0.0038763645570725203, + 0.004225202836096287 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 6, + 7 + ], + "selected_reward": -0.02636663056910038 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 vast Baroque palace with soaring arched windows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the ornate gilded walls feature intricate Rococo-style gold leaf patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the golden plaster on the walls dissolve and transform into swirling cascades of shimmering dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the frescoed ceiling depicting cherubs and clouds condense and release pigments into floating particles of light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera glide forward in a smooth, floating motion across the hall?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012532057240605354 + ], + "opposite_logprob_tokens": [ + -4.3875322341918945 + ], + "logprob_reward": -0.012532057240605354, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012532057240605354, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0775180459022522, + 0.4702407419681549, + 0.07439326494932175, + 0.08001028001308441, + 0.06960213929414749, + 0.04819232597947121, + 0.04432310163974762, + 0.03974815085530281, + 0.06334256380796432, + 0.03262937068939209 + ], + "credit_map": [ + 0.009180473163723946, + 0.011501148343086243, + 0.013014929369091988, + 0.012621196918189526, + 0.00734538771212101, + 0.007013671565800905, + 0.0072876447811722755, + 0.00955360010266304, + 0.0453651063144207, + 0.09854009747505188, + 0.05302067846059799, + 0.0761193260550499, + 0.040327463299036026, + 0.035349030047655106, + 0.05063692480325699, + 0.07088211923837662, + 0.01634182780981064, + 0.010558645240962505, + 0.007914233952760696, + 0.007432444486767054, + 0.0047296485863626, + 0.005488586612045765, + 0.01643131859600544, + 0.005496560130268335, + 0.016741253435611725, + 0.010326305404305458, + 0.010554625652730465, + 0.008335696533322334, + 0.006959676276892424, + 0.00971025601029396, + 0.008657119236886501, + 0.008725359104573727, + 0.009638674557209015, + 0.010605642572045326, + 0.009381253272294998, + 0.011719537898898125, + 0.006139695178717375, + 0.0056955814361572266, + 0.010159485973417759, + 0.00626227306202054, + 0.00433615455403924, + 0.008254357613623142, + 0.007479899562895298, + 0.005275518633425236, + 0.009216437116265297, + 0.005053533241152763, + 0.004761172458529472, + 0.003815255593508482, + 0.00343776261433959, + 0.007509688846766949, + 0.007219034247100353, + 0.0034491815604269505, + 0.006603570189327002, + 0.005142477340996265, + 0.004123294726014137, + 0.006838093977421522, + 0.004220121074467897, + 0.004792561288923025, + 0.007955512031912804, + 0.004563003778457642, + 0.004341116175055504, + 0.0048862602561712265, + 0.004185094498097897, + 0.0048044826835393906, + 0.010382584296166897, + 0.008760252967476845, + 0.005604958161711693, + 0.006659908685833216, + 0.009450828656554222, + 0.009726416319608688, + 0.005715267267078161, + 0.007042355369776487, + 0.002566824434325099, + 0.005238999612629414, + 0.003334708046168089, + 0.005754775833338499, + 0.0045204926282167435, + 0.004493154119700193, + 0.004068027716130018, + 0.0026523913256824017 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 3, + 0 + ], + "selected_reward": -0.012532057240605354 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "Told from a dreamlike first-person perspective, the camera glides forward in an incredibly smooth, floating motion across the vast, sun-drenched halls of the Baroque palace. Golden, three-dimensional sunlight pours through the soaring arched windows, illuminating the air with a visible divine glow and dancing across the polished marble floors. As the camera pans forward, the ornate gilded walls, covered in intricate Rococo-style gold leaf patterns, begin to gracefully deconstruct; golden plaster calmly dissolves at the edges, transforming into swirling cascades of shimmering, ethereal dust. Above, the vast frescoed ceiling, depicting a celestial scene of cherubs and clouds in soft, muted hues, undergoes a similarly magical weathering. The painted figures and sky gently condense, releasing their pigments into shimmering particles of light, joining in a floating dance. These magical dust particles, glowing with an inner glow, catch the brilliant sunlight and create a breathtaking, shimmering cloud that fills the vast space. A slow and beautiful process of magical weathering creates a tranquil and transcendent atmosphere, as the solid architecture melts into pure, floating magic.", + "seed": 1832, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000026/group_8/index_5807_seed_1832_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": -0.03600840177386999, + "vlm_logprob": -0.03600840177386999, + "vlm_yes_no_margin": 3.525, + "vlm_token_credit": -0.03600840177386999, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 vast Baroque palace with soaring arched windows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the ornate gilded walls feature intricate Rococo-style gold leaf patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the golden plaster on the walls dissolve and transform into swirling cascades of shimmering dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the frescoed ceiling depicting cherubs and clouds condense and release pigments into floating particles of light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera glide forward in a smooth, floating motion across the hall?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.024648625403642654 + ], + "opposite_logprob_tokens": [ + -3.899648666381836 + ], + "logprob_reward": -0.024648625403642654, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.024648625403642654, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.41900834441185, + 0.058737196028232574, + 0.051463522017002106, + 0.21775594353675842, + 0.05326985567808151, + 0.059727992862463, + 0.04496239870786667, + 0.024482451379299164, + 0.02617081068456173, + 0.044421400874853134 + ], + "credit_map": [ + 0.14630663394927979, + 0.1556229442358017, + 0.028767365962266922, + 0.025746552273631096, + 0.025324780493974686, + 0.009666170924901962, + 0.014748355373740196, + 0.01282559335231781, + 0.016105521470308304, + 0.008596560917794704, + 0.008043140172958374, + 0.008485177531838417, + 0.005630168132483959, + 0.0035522293765097857, + 0.004600803833454847, + 0.0037236062344163656, + 0.009881961159408092, + 0.006789840292185545, + 0.006830339320003986, + 0.006681353319436312, + 0.008554106578230858, + 0.0022379199508577585, + 0.005503178108483553, + 0.004984829109162092, + 0.03162685036659241, + 0.033829957246780396, + 0.08120618015527725, + 0.017402106896042824, + 0.011628721840679646, + 0.005704480689018965, + 0.027174759656190872, + 0.009182915091514587, + 0.00806159246712923, + 0.013691502623260021, + 0.0042894077487289906, + 0.00782069843262434, + 0.006224021315574646, + 0.0026136545930057764, + 0.005407740827649832, + 0.005161243956536055, + 0.008304689079523087, + 0.01175826694816351, + 0.005982633680105209, + 0.009208079427480698, + 0.006694966461509466, + 0.005404973868280649, + 0.004364438354969025, + 0.008009953424334526, + 0.0037344510201364756, + 0.006119379773736, + 0.0059067681431770325, + 0.004071089439094067, + 0.003633741522207856, + 0.004963153973221779, + 0.006963802967220545, + 0.009570018388330936, + 0.0036136063281446695, + 0.0032374269794672728, + 0.0061721582897007465, + 0.002396183554083109, + 0.002382508246228099, + 0.0017971581546589732, + 0.0026508939918130636, + 0.0022325199097394943, + 0.004695717711001635, + 0.005477357655763626, + 0.0037591103464365005, + 0.00289604882709682, + 0.0036337729543447495, + 0.0016279545379802585, + 0.0019579166546463966, + 0.0021229363046586514, + 0.007532143034040928, + 0.009167816489934921, + 0.007619250100106001, + 0.006814396474510431, + 0.005558504723012447, + 0.0017003845423460007, + 0.0029328525997698307, + 0.0030960598960518837 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 5 + ], + "selected_reward": -0.024648625403642654 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 vast Baroque palace with soaring arched windows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the ornate gilded walls feature intricate Rococo-style gold leaf patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the golden plaster on the walls dissolve and transform into swirling cascades of shimmering dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the frescoed ceiling depicting cherubs and clouds condense and release pigments into floating particles of light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera glide forward in a smooth, floating motion across the hall?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.043109528720378876 + ], + "opposite_logprob_tokens": [ + -3.16810941696167 + ], + "logprob_reward": -0.043109528720378876, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.043109528720378876, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.26231271028518677, + 0.061758752912282944, + 0.05721313878893852, + 0.3270082473754883, + 0.06352636218070984, + 0.056498847901821136, + 0.059705402702093124, + 0.04791514575481415, + 0.023540502414107323, + 0.040520936250686646 + ], + "credit_map": [ + 0.0947750061750412, + 0.10146423429250717, + 0.01529491227120161, + 0.012870922684669495, + 0.01584351435303688, + 0.0065030972473323345, + 0.008267815224826336, + 0.00729318568482995, + 0.018549945205450058, + 0.011623837053775787, + 0.007211936172097921, + 0.00733168376609683, + 0.006216277368366718, + 0.003039577277377248, + 0.004270968958735466, + 0.0035145271103829145, + 0.01168428547680378, + 0.008923943154513836, + 0.00801069289445877, + 0.005561851896345615, + 0.00937205832451582, + 0.0027936059050261974, + 0.0057851099409163, + 0.005081588868051767, + 0.0591115765273571, + 0.07989532500505447, + 0.09466153383255005, + 0.023405417799949646, + 0.022532111033797264, + 0.009207435883581638, + 0.022732125595211983, + 0.015462711453437805, + 0.009683444164693356, + 0.013449487276375294, + 0.006665022578090429, + 0.00874602422118187, + 0.007632907945662737, + 0.0028619214426726103, + 0.007057093549519777, + 0.007430455181747675, + 0.008758693933486938, + 0.00958772748708725, + 0.009464649483561516, + 0.006005590781569481, + 0.008049744181334972, + 0.004387708846479654, + 0.004874556325376034, + 0.005370175465941429, + 0.006091527175158262, + 0.006604187656193972, + 0.009429432451725006, + 0.00465758889913559, + 0.005242858547717333, + 0.0059277708642184734, + 0.008252108469605446, + 0.013499926775693893, + 0.006638793274760246, + 0.006728835869580507, + 0.011802596971392632, + 0.004949222318828106, + 0.0038734825793653727, + 0.003401436610147357, + 0.005192400887608528, + 0.005328370723873377, + 0.006029769778251648, + 0.004440739285200834, + 0.0036096067633479834, + 0.002613555872812867, + 0.0025949289556592703, + 0.0012739638332277536, + 0.001434885198250413, + 0.001543052843771875, + 0.006706360261887312, + 0.007443662267178297, + 0.007445400580763817, + 0.006536198779940605, + 0.004290003329515457, + 0.001408133888617158, + 0.002565839095041156, + 0.004125333856791258 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 4 + ], + "selected_reward": -0.043109528720378876 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 vast Baroque palace with soaring arched windows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the ornate gilded walls feature intricate Rococo-style gold leaf patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the golden plaster on the walls dissolve and transform into swirling cascades of shimmering dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the frescoed ceiling depicting cherubs and clouds condense and release pigments into floating particles of light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera glide forward in a smooth, floating motion across the hall?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0486576184630394 + ], + "opposite_logprob_tokens": [ + -3.0486576557159424 + ], + "logprob_reward": -0.0486576184630394, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.0486576184630394, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.42565998435020447, + 0.06349238753318787, + 0.03042036108672619, + 0.11877067387104034, + 0.16194571554660797, + 0.07541099935770035, + 0.06937068700790405, + 0.02533835731446743, + 0.014070545323193073, + 0.015520348213613033 + ], + "credit_map": [ + 0.14618819952011108, + 0.1651148647069931, + 0.0297442227602005, + 0.01915343850851059, + 0.02794583886861801, + 0.011009309440851212, + 0.01290143746882677, + 0.013602661900222301, + 0.015115613117814064, + 0.010290390811860561, + 0.00894956011325121, + 0.00696444371715188, + 0.0067920684814453125, + 0.004925406072288752, + 0.006535576656460762, + 0.003919330891221762, + 0.006925289984792471, + 0.003922460600733757, + 0.003285423619672656, + 0.004054064862430096, + 0.004897499457001686, + 0.0018814948853105307, + 0.0029729108791798353, + 0.002481216797605157, + 0.02166014164686203, + 0.01825101301074028, + 0.041672248393297195, + 0.008350108750164509, + 0.007797970902174711, + 0.0032330143731087446, + 0.012719163671135902, + 0.005087018944323063, + 0.02825067937374115, + 0.03450756147503853, + 0.020305663347244263, + 0.022438138723373413, + 0.014742329716682434, + 0.006491375155746937, + 0.01650419272482395, + 0.018705787137150764, + 0.010989130474627018, + 0.013096805661916733, + 0.011096777394413948, + 0.013987143523991108, + 0.00740324379876256, + 0.0051678866147994995, + 0.006718270946294069, + 0.0069517414085567, + 0.008055916987359524, + 0.012410406954586506, + 0.012105193920433521, + 0.008029943332076073, + 0.005933358799666166, + 0.004263173323124647, + 0.007152927573770285, + 0.011419766582548618, + 0.0033648423850536346, + 0.004124362021684647, + 0.007255049422383308, + 0.0023382706567645073, + 0.0015860229032114148, + 0.0018175530713051558, + 0.002543651731684804, + 0.0023086033761501312, + 0.002584298839792609, + 0.0019823512993752956, + 0.0037093269638717175, + 0.001618897425942123, + 0.0016312877414748073, + 0.0006311513716354966, + 0.0011449548183009028, + 0.0007682761643081903, + 0.002828977769240737, + 0.002013531746342778, + 0.002396330237388611, + 0.003737415885552764, + 0.001350514474324882, + 0.0006357709062285721, + 0.0013249196344986558, + 0.0012328877346590161 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 3 + ], + "selected_reward": -0.0486576184630394 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set inside a vast Baroque palace with soaring arched windows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the ornate gilded walls feature intricate Rococo-style gold leaf patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the golden plaster on the walls dissolve and transform into swirling cascades of shimmering dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the frescoed ceiling depicting cherubs and clouds condense and release pigments into floating particles of light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera glide forward in a smooth, floating motion across the hall?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.054989613592624664 + ], + "opposite_logprob_tokens": [ + -2.9299895763397217 + ], + "logprob_reward": -0.054989613592624664, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.054989613592624664, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2782971262931824, + 0.05085376650094986, + 0.0890466645359993, + 0.15802372992038727, + 0.16255660355091095, + 0.037887319922447205, + 0.07479612529277802, + 0.05104421451687813, + 0.04019317403435707, + 0.05730125680565834 + ], + "credit_map": [ + 0.10331875085830688, + 0.09505949169397354, + 0.026157258078455925, + 0.012913071550428867, + 0.015640677884221077, + 0.006407337263226509, + 0.009178642183542252, + 0.00962191540747881, + 0.013766189105808735, + 0.007027561776340008, + 0.006034309044480324, + 0.005803237669169903, + 0.0063177356496453285, + 0.0033196040894836187, + 0.004897143226116896, + 0.003687982214614749, + 0.01895042136311531, + 0.012317760847508907, + 0.010773904621601105, + 0.009680432267487049, + 0.016770360991358757, + 0.0054366709664464, + 0.00750947929918766, + 0.0076076313853263855, + 0.022739281877875328, + 0.0560031495988369, + 0.022639907896518707, + 0.008486189879477024, + 0.014505088329315186, + 0.0055028838105499744, + 0.01968429423868656, + 0.008462931960821152, + 0.02882857248187065, + 0.037122942507267, + 0.017376618459820747, + 0.015258364379405975, + 0.02131170593202114, + 0.007537294179201126, + 0.01837664656341076, + 0.016744457185268402, + 0.006207797210663557, + 0.008025337010622025, + 0.005459900945425034, + 0.003475317033007741, + 0.005240503698587418, + 0.0026018796488642693, + 0.003261565463617444, + 0.003615019144490361, + 0.010604222305119038, + 0.00882643274962902, + 0.013073577545583248, + 0.006079866550862789, + 0.005412286147475243, + 0.006324172485619783, + 0.009561245329678059, + 0.014914325438439846, + 0.0067775617353618145, + 0.007867912761867046, + 0.01029371190816164, + 0.006068670656532049, + 0.00573197053745389, + 0.0050538587383925915, + 0.005184580106288195, + 0.004065948072820902, + 0.007464928086847067, + 0.008037165738642216, + 0.0075013646855950356, + 0.004186092875897884, + 0.004917106591165066, + 0.0021255628671497107, + 0.0028096851892769337, + 0.0031512670684605837, + 0.010199818760156631, + 0.0106568718329072, + 0.010106890462338924, + 0.010724371299147606, + 0.005359459202736616, + 0.0020597223192453384, + 0.003730545286089182, + 0.004463574383407831 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 3 + ], + "selected_reward": -0.054989613592624664 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 vast Baroque palace with soaring arched windows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the ornate gilded walls feature intricate Rococo-style gold leaf patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the golden plaster on the walls dissolve and transform into swirling cascades of shimmering dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the frescoed ceiling depicting cherubs and clouds condense and release pigments into floating particles of light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera glide forward in a smooth, floating motion across the hall?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008636622689664364 + ], + "opposite_logprob_tokens": [ + -4.758636474609375 + ], + "logprob_reward": -0.008636622689664364, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008636622689664364, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.45943066477775574, + 0.08407548815011978, + 0.03425808623433113, + 0.1580212414264679, + 0.06973941624164581, + 0.07691186666488647, + 0.05177031457424164, + 0.025370610877871513, + 0.019823789596557617, + 0.0205985140055418 + ], + "credit_map": [ + 0.1592148393392563, + 0.18010887503623962, + 0.027634993195533752, + 0.01808849163353443, + 0.027750341221690178, + 0.01272521074861288, + 0.01684720069169998, + 0.017060691490769386, + 0.01544870063662529, + 0.014164288528263569, + 0.008473615162074566, + 0.013254189863801003, + 0.00918814167380333, + 0.005926941055804491, + 0.010622763074934483, + 0.0069968500174582005, + 0.0062661441043019295, + 0.0043646665289998055, + 0.002924517262727022, + 0.004240942187607288, + 0.005847024731338024, + 0.002758870366960764, + 0.003831649199128151, + 0.004024272318929434, + 0.015338758938014507, + 0.04277670755982399, + 0.0645861029624939, + 0.007339249365031719, + 0.007348258513957262, + 0.003208331996574998, + 0.012262288480997086, + 0.0051615433767437935, + 0.010819652117788792, + 0.014641692861914635, + 0.010123935528099537, + 0.009385243058204651, + 0.008183164522051811, + 0.003266554791480303, + 0.006895864848047495, + 0.00642330851405859, + 0.009372495114803314, + 0.0137286726385355, + 0.014232002198696136, + 0.010300561785697937, + 0.008013078942894936, + 0.0064372653141617775, + 0.007164109032601118, + 0.007663681637495756, + 0.009030584245920181, + 0.009676299057900906, + 0.006734147202223539, + 0.004782599397003651, + 0.006366490852087736, + 0.0033639874309301376, + 0.005828112363815308, + 0.005988099612295628, + 0.0030137307476252317, + 0.0033170797396451235, + 0.004007535986602306, + 0.0026913967449218035, + 0.0021807977464050055, + 0.003440620843321085, + 0.0037061895709484816, + 0.003013260429725051, + 0.003033657791092992, + 0.0033966663759201765, + 0.004226280841976404, + 0.002429074142128229, + 0.0019543340895324945, + 0.0011795901227742434, + 0.0019465794321149588, + 0.001657606684602797, + 0.004114942625164986, + 0.004185735248029232, + 0.0035299567971378565, + 0.002681409940123558, + 0.0023657011333853006, + 0.0008875924395397305, + 0.0014605597825720906, + 0.0013726148754358292 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 1 + ], + "selected_reward": -0.008636622689664364 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "Told from a dreamlike first-person perspective, the camera glides forward in an incredibly smooth, floating motion across the vast, sun-drenched halls of the Baroque palace. Golden, three-dimensional sunlight pours through the soaring arched windows, illuminating the air with a visible divine glow and dancing across the polished marble floors. As the camera pans forward, the ornate gilded walls, covered in intricate Rococo-style gold leaf patterns, begin to gracefully deconstruct; golden plaster calmly dissolves at the edges, transforming into swirling cascades of shimmering, ethereal dust. Above, the vast frescoed ceiling, depicting a celestial scene of cherubs and clouds in soft, muted hues, undergoes a similarly magical weathering. The painted figures and sky gently condense, releasing their pigments into shimmering particles of light, joining in a floating dance. These magical dust particles, glowing with an inner glow, catch the brilliant sunlight and create a breathtaking, shimmering cloud that fills the vast space. A slow and beautiful process of magical weathering creates a tranquil and transcendent atmosphere, as the solid architecture melts into pure, floating magic.", + "seed": 1833, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000026/group_8/index_5807_seed_1833_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": -0.007890761364251375, + "vlm_logprob": -0.007890761364251375, + "vlm_yes_no_margin": 4.975, + "vlm_token_credit": -0.007890761364251375, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 vast Baroque palace with soaring arched windows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the ornate gilded walls feature intricate Rococo-style gold leaf patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the golden plaster on the walls dissolve and transform into swirling cascades of shimmering dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the frescoed ceiling depicting cherubs and clouds condense and release pigments into floating particles of light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera glide forward in a smooth, floating motion across the hall?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008062437176704407 + ], + "opposite_logprob_tokens": [ + -5.258062362670898 + ], + "logprob_reward": -0.008062437176704407, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.008062437176704407, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06698354333639145, + 0.04955350235104561, + 0.02727608196437359, + 0.036029618233442307, + 0.037520669400691986, + 0.0669354721903801, + 0.09102460741996765, + 0.37403932213783264, + 0.03379221260547638, + 0.21684496104717255 + ], + "credit_map": [ + 0.00926146935671568, + 0.009000590071082115, + 0.005290106870234013, + 0.006961111910641193, + 0.00932516809552908, + 0.010356486774981022, + 0.011088790372014046, + 0.0056998226791620255, + 0.005307355895638466, + 0.004021512810140848, + 0.00427096476778388, + 0.004874400794506073, + 0.0060796234756708145, + 0.007066006306558847, + 0.0071614086627960205, + 0.010772228240966797, + 0.002510106423869729, + 0.003935234621167183, + 0.002869692863896489, + 0.0027744004037231207, + 0.003459949977695942, + 0.0028468563687056303, + 0.006366359535604715, + 0.00251348246820271, + 0.004952068440616131, + 0.004067548085004091, + 0.00476762605831027, + 0.006136333104223013, + 0.00444904575124383, + 0.0030219778418540955, + 0.004879390355199575, + 0.00375563045963645, + 0.003586712060496211, + 0.005947798490524292, + 0.0043366956524550915, + 0.003659762442111969, + 0.004181978292763233, + 0.004696385003626347, + 0.006890625227242708, + 0.004220712464302778, + 0.006557127460837364, + 0.007144934963434935, + 0.008175431750714779, + 0.005976383108645678, + 0.005726697389036417, + 0.00854598917067051, + 0.01153527107089758, + 0.013273635879158974, + 0.007288808468729258, + 0.011477814055979252, + 0.008218558505177498, + 0.010627311654388905, + 0.01230667158961296, + 0.01605224609375, + 0.013378375209867954, + 0.011674817651510239, + 0.024451086297631264, + 0.0472152978181839, + 0.01530668418854475, + 0.02328534424304962, + 0.11471118777990341, + 0.04650668054819107, + 0.05240688845515251, + 0.05015616491436958, + 0.004573458805680275, + 0.004560396540910006, + 0.003909460734575987, + 0.0037830581422895193, + 0.003934559877961874, + 0.004325407091528177, + 0.005309564061462879, + 0.0033963066525757313, + 0.016237368807196617, + 0.026445157825946808, + 0.015738220885396004, + 0.02625100128352642, + 0.025962620973587036, + 0.050221461802721024, + 0.01972806267440319, + 0.03626108914613724 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 9, + 6 + ], + "selected_reward": -0.008062437176704407 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 vast Baroque palace with soaring arched windows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the ornate gilded walls feature intricate Rococo-style gold leaf patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the golden plaster on the walls dissolve and transform into swirling cascades of shimmering dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the frescoed ceiling depicting cherubs and clouds condense and release pigments into floating particles of light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera glide forward in a smooth, floating motion across the hall?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009801947511732578 + ], + "opposite_logprob_tokens": [ + -4.634801864624023 + ], + "logprob_reward": -0.009801947511732578, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009801947511732578, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.027806147933006287, + 0.0312499962747097, + 0.037842702120542526, + 0.026382863521575928, + 0.020189765840768814, + 0.03616701066493988, + 0.041766781359910965, + 0.4128834903240204, + 0.02997514419257641, + 0.3357361853122711 + ], + "credit_map": [ + 0.0026885345578193665, + 0.0032028958667069674, + 0.00284375436604023, + 0.0027782812248915434, + 0.004038739483803511, + 0.005429039243608713, + 0.00366630544885993, + 0.0031585944816470146, + 0.004139852710068226, + 0.0034379870630800724, + 0.0031044497154653072, + 0.0027702394872903824, + 0.005105553660541773, + 0.004120407626032829, + 0.0032047913409769535, + 0.005366712808609009, + 0.002697547199204564, + 0.007228932809084654, + 0.0032411213032901287, + 0.004107552580535412, + 0.009367767721414566, + 0.002911490388214588, + 0.005899740383028984, + 0.0023885478731244802, + 0.0034878780134022236, + 0.0028671743348240852, + 0.004487389698624611, + 0.0033930279314517975, + 0.003340828465297818, + 0.002669524634256959, + 0.003897574730217457, + 0.002239460591226816, + 0.0022714745718985796, + 0.003009542590007186, + 0.0028536892496049404, + 0.002187849720939994, + 0.0023155915550887585, + 0.0018253755988553166, + 0.0030651711858808994, + 0.002661069855093956, + 0.004206123296171427, + 0.0044136932119727135, + 0.00478433258831501, + 0.003815319621935487, + 0.0032800580374896526, + 0.005122701171785593, + 0.004455373622477055, + 0.00608940701931715, + 0.0034966813400387764, + 0.006746167317032814, + 0.0028776084072887897, + 0.002992379479110241, + 0.012305359356105328, + 0.004249561112374067, + 0.006331394426524639, + 0.0027676206082105637, + 0.024175314232707024, + 0.05569116026163101, + 0.01478472352027893, + 0.023521864786744118, + 0.15270595252513885, + 0.038117945194244385, + 0.05993383750319481, + 0.04395262151956558, + 0.002922876039519906, + 0.004989100620150566, + 0.0022098380140960217, + 0.003788607893511653, + 0.003938449081033468, + 0.0038523913826793432, + 0.005856366828083992, + 0.00241751316934824, + 0.027088504284620285, + 0.04904552176594734, + 0.02426844835281372, + 0.05067058280110359, + 0.04819665104150772, + 0.05395811051130295, + 0.04166143760085106, + 0.04084689915180206 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 9, + 6 + ], + "selected_reward": -0.009801947511732578 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 vast Baroque palace with soaring arched windows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the ornate gilded walls feature intricate Rococo-style gold leaf patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the golden plaster on the walls dissolve and transform into swirling cascades of shimmering dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the frescoed ceiling depicting cherubs and clouds condense and release pigments into floating particles of light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera glide forward in a smooth, floating motion across the hall?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011078749783337116 + ], + "opposite_logprob_tokens": [ + -4.511078834533691 + ], + "logprob_reward": -0.011078749783337116, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011078749783337116, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06590931862592697, + 0.02191300503909588, + 0.10520856082439423, + 0.14220581948757172, + 0.035247378051280975, + 0.10415040701627731, + 0.07769941538572311, + 0.14813636243343353, + 0.10696809738874435, + 0.19256159663200378 + ], + "credit_map": [ + 0.008298104628920555, + 0.008432162925601006, + 0.007314853370189667, + 0.006352103315293789, + 0.007808479480445385, + 0.008469503372907639, + 0.009108565747737885, + 0.010125559754669666, + 0.0021566241048276424, + 0.002520122565329075, + 0.0023830346763134003, + 0.0027103552129119635, + 0.0024687270633876324, + 0.0026044147089123726, + 0.004492282401770353, + 0.002577445702627301, + 0.005750971846282482, + 0.014636396430432796, + 0.00478420639410615, + 0.005775077268481255, + 0.030602432787418365, + 0.008244650438427925, + 0.027887670323252678, + 0.007527167908847332, + 0.021282266825437546, + 0.014717992395162582, + 0.022979779168963432, + 0.014469916000962257, + 0.025584399700164795, + 0.015055306255817413, + 0.014191195368766785, + 0.01392496656626463, + 0.005067089572548866, + 0.005353545304387808, + 0.006120031233876944, + 0.0048618256114423275, + 0.00334722688421607, + 0.0029797907918691635, + 0.003158756298944354, + 0.00435911538079381, + 0.011720039881765842, + 0.010884648188948631, + 0.013854846358299255, + 0.008339830674231052, + 0.010177819058299065, + 0.01670031063258648, + 0.013049142435193062, + 0.019423775374889374, + 0.006177923176437616, + 0.01173938438296318, + 0.006914282217621803, + 0.007665663026273251, + 0.014574162662029266, + 0.01062945369631052, + 0.00939487386494875, + 0.010603674687445164, + 0.011953439563512802, + 0.017941949889063835, + 0.012716908939182758, + 0.011562522500753403, + 0.021367449313402176, + 0.03101121447980404, + 0.01575024239718914, + 0.0258326455950737, + 0.009667394682765007, + 0.012985802255570889, + 0.016864286735653877, + 0.015538904815912247, + 0.014951078221201897, + 0.011990086175501347, + 0.012113010510802269, + 0.012857555411756039, + 0.015719780698418617, + 0.020898746326565742, + 0.01508262101560831, + 0.03165271133184433, + 0.0210476815700531, + 0.04050789400935173, + 0.02346828021109104, + 0.02418390102684498 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 7, + 3 + ], + "selected_reward": -0.011078749783337116 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 vast Baroque palace with soaring arched windows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the ornate gilded walls feature intricate Rococo-style gold leaf patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the golden plaster on the walls dissolve and transform into swirling cascades of shimmering dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the frescoed ceiling depicting cherubs and clouds condense and release pigments into floating particles of light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera glide forward in a smooth, floating motion across the hall?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005260079633444548 + ], + "opposite_logprob_tokens": [ + -5.255259990692139 + ], + "logprob_reward": -0.005260079633444548, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005260079633444548, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.029864318668842316, + 0.015674058347940445, + 0.2798261344432831, + 0.012687242589890957, + 0.014410747215151787, + 0.012247915379703045, + 0.05397559329867363, + 0.3325551450252533, + 0.04423902556300163, + 0.20451980829238892 + ], + "credit_map": [ + 0.003622728632763028, + 0.003003074787557125, + 0.004121371079236269, + 0.002787810517475009, + 0.003358645597472787, + 0.003158574691042304, + 0.004460079129785299, + 0.0053520346991717815, + 0.0020846922416239977, + 0.0016357580898329616, + 0.0014385291142389178, + 0.0012773408088833094, + 0.002195363864302635, + 0.0024276969488710165, + 0.001994943246245384, + 0.002619734965264797, + 0.014515481889247894, + 0.04168805107474327, + 0.011694930493831635, + 0.014236638322472572, + 0.08742374926805496, + 0.021393615752458572, + 0.07099060714244843, + 0.01788308285176754, + 0.0016509245615452528, + 0.0016008944949135184, + 0.0018705615075305104, + 0.0014461551327258348, + 0.0016979348147287965, + 0.0011888289591297507, + 0.0017693669069558382, + 0.001462576794438064, + 0.002022722503170371, + 0.001872354419901967, + 0.002295723184943199, + 0.001614216947928071, + 0.0014716617297381163, + 0.0014551039785146713, + 0.001650384278036654, + 0.002028581453487277, + 0.0010696790413931012, + 0.0016903589712455869, + 0.0012232876615598798, + 0.001745652174577117, + 0.00136344856582582, + 0.001955375773832202, + 0.0016106307739391923, + 0.0015894828829914331, + 0.004270216915756464, + 0.006754724774509668, + 0.003679803106933832, + 0.005982710514217615, + 0.00915417168289423, + 0.007656658999621868, + 0.009690932929515839, + 0.006786379497498274, + 0.016708694398403168, + 0.04474921524524689, + 0.008903766050934792, + 0.01937304623425007, + 0.144594207406044, + 0.027724195271730423, + 0.04932452738285065, + 0.02117753028869629, + 0.004003094509243965, + 0.005083984229713678, + 0.0050519793294370174, + 0.008130005560815334, + 0.005242218263447285, + 0.005216630175709724, + 0.0066266548819839954, + 0.00488445907831192, + 0.010566149838268757, + 0.013672858476638794, + 0.012605544179677963, + 0.017034919932484627, + 0.016303321346640587, + 0.05528100207448006, + 0.020942451432347298, + 0.05811356380581856 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 2, + 9 + ], + "selected_reward": -0.005260079633444548 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 vast Baroque palace with soaring arched windows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the ornate gilded walls feature intricate Rococo-style gold leaf patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the golden plaster on the walls dissolve and transform into swirling cascades of shimmering dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the frescoed ceiling depicting cherubs and clouds condense and release pigments into floating particles of light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera glide forward in a smooth, floating motion across the hall?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005250592716038227 + ], + "opposite_logprob_tokens": [ + -5.255250453948975 + ], + "logprob_reward": -0.005250592716038227, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005250592716038227, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0466841422021389, + 0.04365820810198784, + 0.0367303304374218, + 0.06792829185724258, + 0.04035021737217903, + 0.11967988312244415, + 0.0928356721997261, + 0.4004061818122864, + 0.05410689488053322, + 0.09762018918991089 + ], + "credit_map": [ + 0.005957582034170628, + 0.005293292924761772, + 0.006972524803131819, + 0.003858284093439579, + 0.005611454602330923, + 0.0050088828429579735, + 0.007596900686621666, + 0.006385223474353552, + 0.004445348400622606, + 0.004878019914031029, + 0.004203110933303833, + 0.00583280436694622, + 0.0049981847405433655, + 0.005908680614084005, + 0.007069829851388931, + 0.006322228815406561, + 0.0025268341414630413, + 0.00528276152908802, + 0.0021879549603909254, + 0.002316653961315751, + 0.009832222945988178, + 0.002955171512439847, + 0.00907615851610899, + 0.00255257124081254, + 0.011021360754966736, + 0.006162898149341345, + 0.00815028976649046, + 0.007327524479478598, + 0.011840187944471836, + 0.006057679653167725, + 0.011228125542402267, + 0.006140224169939756, + 0.005370904225856066, + 0.005594645626842976, + 0.006780225317925215, + 0.005546624306589365, + 0.004522446542978287, + 0.003873971290886402, + 0.0038567378651350737, + 0.004804665222764015, + 0.012007779441773891, + 0.012687058188021183, + 0.015774313360452652, + 0.010723487474024296, + 0.009997537359595299, + 0.016312286257743835, + 0.018026242032647133, + 0.02415117435157299, + 0.007952149957418442, + 0.012924958020448685, + 0.006236382760107517, + 0.010016532614827156, + 0.012235351838171482, + 0.015217175707221031, + 0.014428744092583656, + 0.01382437814027071, + 0.019428422674536705, + 0.05816778168082237, + 0.012352509424090385, + 0.023974506184458733, + 0.1740366667509079, + 0.03761744871735573, + 0.0576665885746479, + 0.017162300646305084, + 0.005107397213578224, + 0.0055502308532595634, + 0.00874212384223938, + 0.00800272636115551, + 0.005095511209219694, + 0.006903825793415308, + 0.005935241933912039, + 0.008769838139414787, + 0.00703772995620966, + 0.011714128777384758, + 0.005619820207357407, + 0.011157378554344177, + 0.013493225909769535, + 0.015720827504992485, + 0.011770027689635754, + 0.021107051521539688 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 5, + 9 + ], + "selected_reward": -0.005250592716038227 + } + ], + "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/0000026/group_9/index_tensor([4175, 4175])_group_9_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000026/group_9/index_tensor([4175, 4175])_group_9_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..464d69cfb52c9bbf3dfad533e1ad3b0e4cc3eaa1 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000026/group_9/index_tensor([4175, 4175])_group_9_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "The film begins with an erratic handheld tracking shot, the frame punctuated by digital noise and flickering like signal interference. The focus shifts rapidly between foreground and background, closely following a cuddly Disney Pixar-esque spherical robot. Its smooth, egg-yellow shell, a pair of enormous, digital blue eyes flashing wildly with utter terror, moves on three spindly metal legs, frantically darting through the ruins of a dilapidated biocontrol demonstration area. This area seems like an isolated island amidst a vast, gray wasteland, the sky a sickly, dim yellow. Countless thick veins of mutant plants, emitting an eerie purple and green bioluminescence, surge, entwine, and grow at a visible speed, as if alive, through the cracked ground and rusted debris of equipment. Each pulse of these veins causes the fluorescent light on their surface to flicker. The robot clumsily dodges a glowing vine that suddenly breaks out from the ground under its feet and almost trips. The whole scene forms a strong visual contrast between the cute character's terrified escape and the invasion of terrifying plants. Combined with the unstable lens and digital glitch effects, it successfully creates a unique, unsettling and weird contrast atmosphere.", + "seed": 2676, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000026/group_9/index_4175_seed_2676_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": -0.07914303597062826, + "vlm_logprob": -0.07914303597062826, + "vlm_yes_no_margin": 3.775, + "vlm_token_credit": -0.07914303597062826, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main character a spherical robot with a smooth egg-yellow shell and enormous digital blue eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the robot move on three spindly metal legs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment the ruins of a dilapidated biocontrol demonstration area located on an isolated island in a gray wasteland?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are thick veins of mutant plants emitting purple and green bioluminescence visible growing rapidly through the ground and debris?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot erratic and handheld with digital noise and flickering signal interference effects?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003286320250481367 + ], + "opposite_logprob_tokens": [ + -6.253286361694336 + ], + "logprob_reward": -0.003286320250481367, + "yes_no_margin_reward": 6.25, + "token_credit_reward": -0.003286320250481367, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18719319999217987, + 0.08338160812854767, + 0.09172763675451279, + 0.06411417573690414, + 0.10458248108625412, + 0.07319352775812149, + 0.09388919174671173, + 0.17939209938049316, + 0.054866235703229904, + 0.06765981018543243 + ], + "credit_map": [ + 0.02558884583413601, + 0.015615605749189854, + 0.02247638627886772, + 0.03529190644621849, + 0.015876809135079384, + 0.022271396592259407, + 0.030320782214403152, + 0.01975146494805813, + 0.007337873335927725, + 0.009659881703555584, + 0.008660933934152126, + 0.010725609958171844, + 0.008494443260133266, + 0.011666728183627129, + 0.008960411883890629, + 0.017875714227557182, + 0.009632116183638573, + 0.008899361826479435, + 0.01643853262066841, + 0.007372320629656315, + 0.006306122988462448, + 0.014845339581370354, + 0.0055685704573988914, + 0.022665271535515785, + 0.011890892870724201, + 0.00537959486246109, + 0.010410130955278873, + 0.005969947669655085, + 0.003987360745668411, + 0.011828938499093056, + 0.006795626133680344, + 0.007851682603359222, + 0.009145358577370644, + 0.017627622932195663, + 0.010195458307862282, + 0.011304760351777077, + 0.02212635427713394, + 0.014201915822923183, + 0.01206727884709835, + 0.007913735695183277, + 0.007144513074308634, + 0.010412929579615593, + 0.005636366084218025, + 0.005876148119568825, + 0.00873099360615015, + 0.012710078619420528, + 0.017373710870742798, + 0.00530878733843565, + 0.00686269998550415, + 0.008950152434408665, + 0.013903656974434853, + 0.011049957945942879, + 0.013525297865271568, + 0.011775500141084194, + 0.014184036292135715, + 0.013637889176607132, + 0.011388566344976425, + 0.020707568153738976, + 0.017383309081196785, + 0.015213058330118656, + 0.036839086562395096, + 0.027514498680830002, + 0.015501344576478004, + 0.034844666719436646, + 0.008343166671693325, + 0.004560634959489107, + 0.008783707395195961, + 0.004255617503076792, + 0.0041803233325481415, + 0.013782155700027943, + 0.005017447285354137, + 0.005943182855844498, + 0.006459514610469341, + 0.004858306143432856, + 0.020202454179525375, + 0.004909547511488199, + 0.003747028298676014, + 0.007077174261212349, + 0.003623725613579154, + 0.016782056540250778 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 4 + ], + "selected_reward": -0.003286320250481367 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main character a spherical robot with a smooth egg-yellow shell and enormous digital blue eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the robot move on three spindly metal legs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment the ruins of a dilapidated biocontrol demonstration area located on an isolated island in a gray wasteland?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are thick veins of mutant plants emitting purple and green bioluminescence visible growing rapidly through the ground and debris?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot erratic and handheld with digital noise and flickering signal interference effects?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08906225860118866 + ], + "opposite_logprob_tokens": [ + -2.464062213897705 + ], + "logprob_reward": -0.08906225860118866, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08906225860118866, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10084454715251923, + 0.1079402044415474, + 0.12533700466156006, + 0.029291996732354164, + 0.1266554743051529, + 0.14504565298557281, + 0.0667463093996048, + 0.24222181737422943, + 0.02747821994125843, + 0.028438802808523178 + ], + "credit_map": [ + 0.014090166427195072, + 0.012320238165557384, + 0.012594381347298622, + 0.012901142239570618, + 0.009483424946665764, + 0.013146698474884033, + 0.015729373320937157, + 0.010579131543636322, + 0.010740529745817184, + 0.010364074259996414, + 0.013355699367821217, + 0.012731693685054779, + 0.011848833411931992, + 0.019068649038672447, + 0.012622860260307789, + 0.01720787212252617, + 0.008511095307767391, + 0.005585432518273592, + 0.01367542240768671, + 0.006788188125938177, + 0.0060666752979159355, + 0.0373675711452961, + 0.007049751002341509, + 0.04029286652803421, + 0.004075299948453903, + 0.002151682274416089, + 0.005114460363984108, + 0.00233744690194726, + 0.0017524780705571175, + 0.007145716808736324, + 0.0026538739912211895, + 0.00406103627756238, + 0.02504686266183853, + 0.014798741787672043, + 0.017743950709700584, + 0.01295468956232071, + 0.013853068463504314, + 0.01557091437280178, + 0.011456586420536041, + 0.015230660326778889, + 0.0146761704236269, + 0.005203413311392069, + 0.009485832415521145, + 0.015457517467439175, + 0.0066133481450378895, + 0.02821962907910347, + 0.055917222052812576, + 0.009472519159317017, + 0.0052299462258815765, + 0.007250308524817228, + 0.010857010260224342, + 0.0073097459971904755, + 0.009686487726867199, + 0.006818093825131655, + 0.010930848307907581, + 0.008663867600262165, + 0.01281280629336834, + 0.022994879633188248, + 0.029872041195631027, + 0.014908792451024055, + 0.052119359374046326, + 0.03157930448651314, + 0.015338328666985035, + 0.06259628385305405, + 0.0037772993091493845, + 0.0029682056047022343, + 0.005290711764246225, + 0.0026898919604718685, + 0.0020305190701037645, + 0.005254463292658329, + 0.002483188873156905, + 0.0029839409980922937, + 0.0028388642240315676, + 0.0020049354061484337, + 0.006437574978917837, + 0.002586434595286846, + 0.0015712276799604297, + 0.004568197298794985, + 0.002836792729794979, + 0.0055947755463421345 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 5, + 4 + ], + "selected_reward": -0.08906225860118866 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main character a spherical robot with a smooth egg-yellow shell and enormous digital blue eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the robot move on three spindly metal legs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment the ruins of a dilapidated biocontrol demonstration area located on an isolated island in a gray wasteland?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are thick veins of mutant plants emitting purple and green bioluminescence visible growing rapidly through the ground and debris?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot erratic and handheld with digital noise and flickering signal interference effects?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.25223255157470703 + ], + "opposite_logprob_tokens": [ + -1.502232551574707 + ], + "logprob_reward": -0.25223255157470703, + "yes_no_margin_reward": 1.25, + "token_credit_reward": -0.25223255157470703, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15136364102363586, + 0.09097132086753845, + 0.04519890621304512, + 0.0366964265704155, + 0.15373682975769043, + 0.2030618041753769, + 0.04428806155920029, + 0.08006422221660614, + 0.036526139825582504, + 0.15809263288974762 + ], + "credit_map": [ + 0.02606622502207756, + 0.017255278304219246, + 0.014464830048382282, + 0.022533399984240532, + 0.014505420811474323, + 0.020223191007971764, + 0.018787166103720665, + 0.017528114840388298, + 0.011361582204699516, + 0.006271795369684696, + 0.021632760763168335, + 0.009070364758372307, + 0.007870973087847233, + 0.01378551684319973, + 0.00809681136161089, + 0.01288150716573, + 0.0031139557249844074, + 0.003066933248192072, + 0.0036388023290783167, + 0.002619336359202862, + 0.0025447176303714514, + 0.01432172954082489, + 0.0029709143564105034, + 0.012922512367367744, + 0.004357115365564823, + 0.003444361500442028, + 0.005757184233516455, + 0.004371850285679102, + 0.0033494154922664165, + 0.007902354001998901, + 0.003499538404867053, + 0.004014607053250074, + 0.03373848646879196, + 0.014520227909088135, + 0.02345888875424862, + 0.0175398550927639, + 0.010852623730897903, + 0.019053706899285316, + 0.013655001297593117, + 0.020918043330311775, + 0.020405307412147522, + 0.005539715755730867, + 0.016568822786211967, + 0.02236921899020672, + 0.009232401847839355, + 0.03680340200662613, + 0.07916922867298126, + 0.01297371182590723, + 0.004328511655330658, + 0.004494675900787115, + 0.005734620615839958, + 0.006328299641609192, + 0.004526671953499317, + 0.0055195968598127365, + 0.009164043702185154, + 0.004191641695797443, + 0.008207689970731735, + 0.006943372078239918, + 0.011358865536749363, + 0.005477037746459246, + 0.016503505408763885, + 0.007514093071222305, + 0.005369789432734251, + 0.018689868971705437, + 0.006480870768427849, + 0.0032260327134281397, + 0.00684772152453661, + 0.002689594868570566, + 0.0027366008143872023, + 0.007137854117900133, + 0.003320395713672042, + 0.004087068606168032, + 0.013981720432639122, + 0.008867932483553886, + 0.04603606089949608, + 0.010086318477988243, + 0.00807760190218687, + 0.017327796667814255, + 0.014805088751018047, + 0.03891009837388992 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 9, + 4 + ], + "selected_reward": -0.25223255157470703 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main character a spherical robot with a smooth egg-yellow shell and enormous digital blue eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the robot move on three spindly metal legs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment the ruins of a dilapidated biocontrol demonstration area located on an isolated island in a gray wasteland?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are thick veins of mutant plants emitting purple and green bioluminescence visible growing rapidly through the ground and debris?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot erratic and handheld with digital noise and flickering signal interference effects?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0024903970770537853 + ], + "opposite_logprob_tokens": [ + -6.002490520477295 + ], + "logprob_reward": -0.0024903970770537853, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.0024903970770537853, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06085982918739319, + 0.041239477694034576, + 0.05316227301955223, + 0.05226542800664902, + 0.15463022887706757, + 0.25106561183929443, + 0.07040005177259445, + 0.2154000997543335, + 0.029126353561878204, + 0.07185070961713791 + ], + "credit_map": [ + 0.00818092841655016, + 0.007121740374714136, + 0.0057175010442733765, + 0.009589575231075287, + 0.005876730661839247, + 0.00648356880992651, + 0.009836779907345772, + 0.008052999153733253, + 0.0039254846051335335, + 0.0032431127037853003, + 0.008884768933057785, + 0.004201401956379414, + 0.0037410894874483347, + 0.00588684668764472, + 0.003279686439782381, + 0.008077085949480534, + 0.004441835917532444, + 0.003930971026420593, + 0.004961987491697073, + 0.00350415101274848, + 0.0035265563055872917, + 0.018874051049351692, + 0.0025797190610319376, + 0.011343000456690788, + 0.009337469935417175, + 0.004126929212361574, + 0.008276182226836681, + 0.0044717141427099705, + 0.0029180545825511217, + 0.011625311337411404, + 0.0041387989185750484, + 0.0073709674179553986, + 0.03305352106690407, + 0.016666563227772713, + 0.023061484098434448, + 0.016846591606736183, + 0.013766458258032799, + 0.01866498962044716, + 0.012727286666631699, + 0.019843319430947304, + 0.031581293791532516, + 0.008599434047937393, + 0.033953700214624405, + 0.023832568898797035, + 0.008778216317296028, + 0.045252520591020584, + 0.07903668284416199, + 0.02003117837011814, + 0.006316381506621838, + 0.007070870604366064, + 0.008447510190308094, + 0.008973187766969204, + 0.00895348098129034, + 0.010058705694973469, + 0.011349432170391083, + 0.009230483323335648, + 0.012516272254288197, + 0.02324879914522171, + 0.022039402276277542, + 0.015623400919139385, + 0.04850117862224579, + 0.029982926324009895, + 0.014803573489189148, + 0.048684533685445786, + 0.004392155446112156, + 0.0032211055513471365, + 0.005221960600465536, + 0.0029192452784627676, + 0.0017711763503029943, + 0.004540679976344109, + 0.0027073859237134457, + 0.004352642223238945, + 0.009256621822714806, + 0.0038445075042545795, + 0.022746721282601357, + 0.004142845049500465, + 0.0026537852827459574, + 0.007861549966037273, + 0.0029889284633100033, + 0.01835574209690094 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 7, + 4 + ], + "selected_reward": -0.0024903970770537853 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main character a spherical robot with a smooth egg-yellow shell and enormous digital blue eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the robot move on three spindly metal legs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment the ruins of a dilapidated biocontrol demonstration area located on an isolated island in a gray wasteland?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are thick veins of mutant plants emitting purple and green bioluminescence visible growing rapidly through the ground and debris?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot erratic and handheld with digital noise and flickering signal interference effects?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.048643652349710464 + ], + "opposite_logprob_tokens": [ + -3.0486435890197754 + ], + "logprob_reward": -0.048643652349710464, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.048643652349710464, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11641405522823334, + 0.047387391328811646, + 0.19293324649333954, + 0.07131195068359375, + 0.05200381949543953, + 0.06696400046348572, + 0.06454107910394669, + 0.28328630328178406, + 0.0416015163064003, + 0.06355661153793335 + ], + "credit_map": [ + 0.014576424844563007, + 0.012737007811665535, + 0.01405011210590601, + 0.018373366445302963, + 0.01109448540955782, + 0.015332943759858608, + 0.016572389751672745, + 0.013677328824996948, + 0.004623590037226677, + 0.004563185386359692, + 0.007138012908399105, + 0.00693341763690114, + 0.0034856919664889574, + 0.0067537338472902775, + 0.004474382847547531, + 0.009415377862751484, + 0.01329865213483572, + 0.011215833947062492, + 0.015337989665567875, + 0.008529983460903168, + 0.012594797648489475, + 0.06937910616397858, + 0.008081845007836819, + 0.05449504032731056, + 0.013867281377315521, + 0.006072660908102989, + 0.009634978137910366, + 0.005818268284201622, + 0.00494384765625, + 0.0154957827180624, + 0.004597363993525505, + 0.01088177040219307, + 0.00631074421107769, + 0.00517771951854229, + 0.006795378401875496, + 0.006977022625505924, + 0.006469838786870241, + 0.007080657873302698, + 0.007038506679236889, + 0.006153948605060577, + 0.009052307344973087, + 0.0051682936027646065, + 0.009015622548758984, + 0.0054972111247479916, + 0.005253015086054802, + 0.012712325900793076, + 0.0100490627810359, + 0.010216159746050835, + 0.006363396532833576, + 0.008248850703239441, + 0.007653810549527407, + 0.008503830060362816, + 0.008462666533887386, + 0.00692960387095809, + 0.011497827246785164, + 0.006881090812385082, + 0.01845623180270195, + 0.035168156027793884, + 0.02189772203564644, + 0.029920876026153564, + 0.05757690221071243, + 0.04957105964422226, + 0.02813592180609703, + 0.04255944862961769, + 0.004261859692633152, + 0.0036344658583402634, + 0.006783377379179001, + 0.005712450481951237, + 0.003120028181001544, + 0.008074693381786346, + 0.004843317903578281, + 0.005171324126422405, + 0.008694116957485676, + 0.004037783946841955, + 0.017414357513189316, + 0.004581696819514036, + 0.0032782582566142082, + 0.009637863375246525, + 0.002672686241567135, + 0.013239852152764797 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 2, + 0 + ], + "selected_reward": -0.048643652349710464 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "The film begins with an erratic handheld tracking shot, the frame punctuated by digital noise and flickering like signal interference. The focus shifts rapidly between foreground and background, closely following a cuddly Disney Pixar-esque spherical robot. Its smooth, egg-yellow shell, a pair of enormous, digital blue eyes flashing wildly with utter terror, moves on three spindly metal legs, frantically darting through the ruins of a dilapidated biocontrol demonstration area. This area seems like an isolated island amidst a vast, gray wasteland, the sky a sickly, dim yellow. Countless thick veins of mutant plants, emitting an eerie purple and green bioluminescence, surge, entwine, and grow at a visible speed, as if alive, through the cracked ground and rusted debris of equipment. Each pulse of these veins causes the fluorescent light on their surface to flicker. The robot clumsily dodges a glowing vine that suddenly breaks out from the ground under its feet and almost trips. The whole scene forms a strong visual contrast between the cute character's terrified escape and the invasion of terrifying plants. Combined with the unstable lens and digital glitch effects, it successfully creates a unique, unsettling and weird contrast atmosphere.", + "seed": 2677, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000026/group_9/index_4175_seed_2677_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": -0.037396639585494995, + "vlm_logprob": -0.037396639585494995, + "vlm_yes_no_margin": 4.225, + "vlm_token_credit": -0.037396639585494995, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main character a spherical robot with a smooth egg-yellow shell and enormous digital blue eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the robot move on three spindly metal legs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment the ruins of a dilapidated biocontrol demonstration area located on an isolated island in a gray wasteland?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are thick veins of mutant plants emitting purple and green bioluminescence visible growing rapidly through the ground and debris?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot erratic and handheld with digital noise and flickering signal interference effects?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003378281369805336 + ], + "opposite_logprob_tokens": [ + -6.128378391265869 + ], + "logprob_reward": -0.003378281369805336, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.003378281369805336, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.40992939472198486, + 0.057102832943201065, + 0.05610094591975212, + 0.0594097301363945, + 0.09759848564863205, + 0.06640083342790604, + 0.09855080395936966, + 0.03998724743723869, + 0.051668260246515274, + 0.06325144320726395 + ], + "credit_map": [ + 0.049055300652980804, + 0.0428524911403656, + 0.03950272873044014, + 0.03228408843278885, + 0.06979035586118698, + 0.057563040405511856, + 0.05494168400764465, + 0.06393972784280777, + 0.006741046905517578, + 0.006021467968821526, + 0.009630872868001461, + 0.006341906730085611, + 0.0062365154735744, + 0.008764533326029778, + 0.006617357023060322, + 0.006749132182449102, + 0.006227283738553524, + 0.006698182318359613, + 0.007578895892947912, + 0.0058165243826806545, + 0.004903968423604965, + 0.012496034614741802, + 0.005328053142875433, + 0.007052005268633366, + 0.005273832473903894, + 0.007922164164483547, + 0.009246157482266426, + 0.010911393910646439, + 0.0041962433606386185, + 0.0054606287740170956, + 0.007347108330577612, + 0.009052206762135029, + 0.008725816383957863, + 0.008092846721410751, + 0.011650349013507366, + 0.018155818805098534, + 0.006988325621932745, + 0.008549203164875507, + 0.01346697099506855, + 0.021969156339764595, + 0.009734596125781536, + 0.00513104721903801, + 0.01135952677577734, + 0.01266220211982727, + 0.0058839754201471806, + 0.004049542360007763, + 0.009472287259995937, + 0.008107651956379414, + 0.023280221968889236, + 0.01630970649421215, + 0.01273552980273962, + 0.015363698825240135, + 0.006276738364249468, + 0.004692526068538427, + 0.008866515010595322, + 0.011025873944163322, + 0.0044873966835439205, + 0.005163354333490133, + 0.0060730441473424435, + 0.006726154591888189, + 0.003409907454624772, + 0.0040665664710104465, + 0.004548994358628988, + 0.005511830560863018, + 0.005354437045753002, + 0.010471181944012642, + 0.005473115481436253, + 0.006393159739673138, + 0.007657077629119158, + 0.00550431152805686, + 0.005449492949992418, + 0.005365482997149229, + 0.006634560413658619, + 0.006100110709667206, + 0.007628594059497118, + 0.015592861920595169, + 0.0072697862051427364, + 0.006809181999415159, + 0.004835646599531174, + 0.008380703628063202 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 4 + ], + "selected_reward": -0.003378281369805336 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main character a spherical robot with a smooth egg-yellow shell and enormous digital blue eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the robot move on three spindly metal legs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment the ruins of a dilapidated biocontrol demonstration area located on an isolated island in a gray wasteland?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are thick veins of mutant plants emitting purple and green bioluminescence visible growing rapidly through the ground and debris?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot erratic and handheld with digital noise and flickering signal interference effects?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05496411398053169 + ], + "opposite_logprob_tokens": [ + -2.929964065551758 + ], + "logprob_reward": -0.05496411398053169, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05496411398053169, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23432452976703644, + 0.26360395550727844, + 0.09949595481157303, + 0.07136521488428116, + 0.10378330200910568, + 0.041405126452445984, + 0.052362602204084396, + 0.04260677844285965, + 0.03490709885954857, + 0.056145425885915756 + ], + "credit_map": [ + 0.027629340067505836, + 0.022324511781334877, + 0.02706255204975605, + 0.03591328486800194, + 0.026161573827266693, + 0.02695983275771141, + 0.0377969890832901, + 0.03047644905745983, + 0.029959123581647873, + 0.027130432426929474, + 0.031397588551044464, + 0.026368804275989532, + 0.01719571277499199, + 0.06507238000631332, + 0.020086180418729782, + 0.04639372602105141, + 0.008045673370361328, + 0.010685420595109463, + 0.01170169934630394, + 0.01168894674628973, + 0.00709457416087389, + 0.020608432590961456, + 0.010781247168779373, + 0.018889961764216423, + 0.005922787357121706, + 0.007723769173026085, + 0.013288586400449276, + 0.011963489465415478, + 0.004563314840197563, + 0.006746232509613037, + 0.006552668288350105, + 0.014604364521801472, + 0.011116609908640385, + 0.012095270678400993, + 0.015869546681642532, + 0.023977551609277725, + 0.006845570169389248, + 0.007542012725025415, + 0.011498303152620792, + 0.01483843568712473, + 0.004385155159980059, + 0.00470727076753974, + 0.007669419515877962, + 0.007402661256492138, + 0.0026798227336257696, + 0.0029375399462878704, + 0.005619002040475607, + 0.006004254333674908, + 0.010569374077022076, + 0.006085705943405628, + 0.00610085716471076, + 0.008300294168293476, + 0.005676319822669029, + 0.0036307647824287415, + 0.005944901145994663, + 0.006054387893527746, + 0.007894570007920265, + 0.004503817763179541, + 0.0072023868560791016, + 0.00510869175195694, + 0.004643352702260017, + 0.0038627430330961943, + 0.005533116869628429, + 0.003858099225908518, + 0.004554489627480507, + 0.0035572585184127092, + 0.004782741889357567, + 0.006593677680939436, + 0.0035252096131443977, + 0.0036021156702190638, + 0.004451033193618059, + 0.003840572899207473, + 0.005777199286967516, + 0.006563679780811071, + 0.01200510747730732, + 0.011033311486244202, + 0.006113459821790457, + 0.005940036382526159, + 0.003698506625369191, + 0.005014125723391771 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 4 + ], + "selected_reward": -0.05496411398053169 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main character a spherical robot with a smooth egg-yellow shell and enormous digital blue eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the robot move on three spindly metal legs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment the ruins of a dilapidated biocontrol demonstration area located on an isolated island in a gray wasteland?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are thick veins of mutant plants emitting purple and green bioluminescence visible growing rapidly through the ground and debris?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot erratic and handheld with digital noise and flickering signal interference effects?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10036002844572067 + ], + "opposite_logprob_tokens": [ + -2.3503599166870117 + ], + "logprob_reward": -0.10036002844572067, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10036002844572067, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1964394897222519, + 0.11097443848848343, + 0.08560322970151901, + 0.07348468899726868, + 0.0443260632455349, + 0.1943763792514801, + 0.08505868166685104, + 0.09211135655641556, + 0.04496300593018532, + 0.07266271859407425 + ], + "credit_map": [ + 0.01908116787672043, + 0.023732956498861313, + 0.025455204769968987, + 0.018773922696709633, + 0.02483544312417507, + 0.03106299787759781, + 0.02498534508049488, + 0.028512435033917427, + 0.011013680137693882, + 0.013062527403235435, + 0.019286993891000748, + 0.021303867921233177, + 0.0067324512638151646, + 0.015062515623867512, + 0.0139998197555542, + 0.010512572713196278, + 0.008554891683161259, + 0.008381219580769539, + 0.017621802166104317, + 0.01068496610969305, + 0.004663479048758745, + 0.01646152138710022, + 0.005934203043580055, + 0.013301153667271137, + 0.004980850033462048, + 0.005276833660900593, + 0.011697079055011272, + 0.013542844913899899, + 0.0035792605485767126, + 0.0053013768047094345, + 0.009175636805593967, + 0.019930806010961533, + 0.00552614638581872, + 0.005833137780427933, + 0.0055479928851127625, + 0.006919792853295803, + 0.003267258172854781, + 0.0034871867392212152, + 0.005762028973549604, + 0.00798251572996378, + 0.04844426363706589, + 0.01637105457484722, + 0.040901102125644684, + 0.04141521453857422, + 0.008971805684268475, + 0.009232872165739536, + 0.015223963186144829, + 0.01381609681993723, + 0.017929023131728172, + 0.015118777751922607, + 0.012240960262715816, + 0.009862697683274746, + 0.007466408424079418, + 0.005299882497638464, + 0.008227271027863026, + 0.008913666941225529, + 0.008607353083789349, + 0.009663543663918972, + 0.015553812496364117, + 0.014953969977796078, + 0.00826025940477848, + 0.008388692513108253, + 0.013317306526005268, + 0.013366412371397018, + 0.004632246680557728, + 0.009665253572165966, + 0.005670438054949045, + 0.004924894776195288, + 0.005763808265328407, + 0.006241900380700827, + 0.00421769917011261, + 0.0038467629346996546, + 0.009690940380096436, + 0.0073359389789402485, + 0.01371984463185072, + 0.009943347424268723, + 0.00961366854608059, + 0.006063248496502638, + 0.005858141928911209, + 0.010437584482133389 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 1 + ], + "selected_reward": -0.10036002844572067 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main character a spherical robot with a smooth egg-yellow shell and enormous digital blue eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the robot move on three spindly metal legs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment the ruins of a dilapidated biocontrol demonstration area located on an isolated island in a gray wasteland?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are thick veins of mutant plants emitting purple and green bioluminescence visible growing rapidly through the ground and debris?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot erratic and handheld with digital noise and flickering signal interference effects?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.001940036192536354 + ], + "opposite_logprob_tokens": [ + -6.2519402503967285 + ], + "logprob_reward": -0.001940036192536354, + "yes_no_margin_reward": 6.25, + "token_credit_reward": -0.001940036192536354, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16581985354423523, + 0.10342832654714584, + 0.0925416648387909, + 0.05113157257437706, + 0.09191080182790756, + 0.0803494080901146, + 0.06475859135389328, + 0.1747535616159439, + 0.08547916263341904, + 0.0898270308971405 + ], + "credit_map": [ + 0.01809404417872429, + 0.015023249201476574, + 0.02233557403087616, + 0.01814967393875122, + 0.02035534754395485, + 0.028476031497120857, + 0.02185996063053608, + 0.021525967866182327, + 0.012816756032407284, + 0.0160312969237566, + 0.010864286683499813, + 0.01383384969085455, + 0.008005911484360695, + 0.016892943531274796, + 0.011885476298630238, + 0.013097800314426422, + 0.008589963428676128, + 0.01162030827254057, + 0.01098733302205801, + 0.01088173408061266, + 0.004715866409242153, + 0.02548176422715187, + 0.006329227704554796, + 0.013935461640357971, + 0.0044319224543869495, + 0.0058824047446250916, + 0.00972751434892416, + 0.009472430683672428, + 0.003994364757090807, + 0.004094487056136131, + 0.006627076305449009, + 0.006901375483721495, + 0.009700832888484001, + 0.00856176670640707, + 0.011327502317726612, + 0.01588398404419422, + 0.0074225221760571, + 0.008112695068120956, + 0.011472401209175587, + 0.019429100677371025, + 0.01045297086238861, + 0.010529821738600731, + 0.012919261120259762, + 0.014056331478059292, + 0.006995994132012129, + 0.00610613776370883, + 0.00880146212875843, + 0.010487428866326809, + 0.008568931370973587, + 0.00771813839673996, + 0.007523931097239256, + 0.008396996185183525, + 0.010950707830488682, + 0.007749538403004408, + 0.008078943006694317, + 0.005771401803940535, + 0.018526718020439148, + 0.016133030876517296, + 0.017788395285606384, + 0.024898217990994453, + 0.02201681397855282, + 0.026565002277493477, + 0.022574562579393387, + 0.02625081315636635, + 0.007938006892800331, + 0.019411752000451088, + 0.009324206970632076, + 0.012113692238926888, + 0.008810841478407383, + 0.008029816672205925, + 0.011259027756750584, + 0.008591820485889912, + 0.011779450811445713, + 0.009810474701225758, + 0.012440584599971771, + 0.012699712067842484, + 0.012356353923678398, + 0.011861673556268215, + 0.007613665424287319, + 0.011265111155807972 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 1 + ], + "selected_reward": -0.001940036192536354 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main character a spherical robot with a smooth egg-yellow shell and enormous digital blue eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the robot move on three spindly metal legs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment the ruins of a dilapidated biocontrol demonstration area located on an isolated island in a gray wasteland?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are thick veins of mutant plants emitting purple and green bioluminescence visible growing rapidly through the ground and debris?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot erratic and handheld with digital noise and flickering signal interference effects?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02634073793888092 + ], + "opposite_logprob_tokens": [ + -3.6513407230377197 + ], + "logprob_reward": -0.02634073793888092, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.02634073793888092, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08423688262701035, + 0.060326408594846725, + 0.05030883476138115, + 0.03861956298351288, + 0.06857770681381226, + 0.2379600554704666, + 0.19040395319461823, + 0.046812549233436584, + 0.06346866488456726, + 0.15928538143634796 + ], + "credit_map": [ + 0.009381777606904507, + 0.007891569286584854, + 0.014190011657774448, + 0.010253304615616798, + 0.008175646886229515, + 0.012417560443282127, + 0.008524232544004917, + 0.013402783311903477, + 0.006280296016484499, + 0.004613530822098255, + 0.008744681254029274, + 0.00567040778696537, + 0.004962103441357613, + 0.013353140093386173, + 0.005960206035524607, + 0.010742043145000935, + 0.004240938927978277, + 0.005851606838405132, + 0.007046228740364313, + 0.00701550068333745, + 0.004419500008225441, + 0.009275948628783226, + 0.004887049552053213, + 0.0075720613822340965, + 0.0028123222291469574, + 0.004299231339246035, + 0.005524301901459694, + 0.007793337106704712, + 0.003730380441993475, + 0.003344772383570671, + 0.00568393524736166, + 0.005431281868368387, + 0.006675174459815025, + 0.007099101319909096, + 0.007981927134096622, + 0.011577208526432514, + 0.0051778992637991905, + 0.005864947568625212, + 0.01016896404325962, + 0.014032483100891113, + 0.03639790043234825, + 0.02172774262726307, + 0.056829798966646194, + 0.061653345823287964, + 0.011087910272181034, + 0.009047980420291424, + 0.023789405822753906, + 0.01742597110569477, + 0.03847609460353851, + 0.030134333297610283, + 0.024796582758426666, + 0.030950387939810753, + 0.012809515930712223, + 0.01277617085725069, + 0.0178777314722538, + 0.022583112120628357, + 0.006071994546800852, + 0.003992675803601742, + 0.006521057803183794, + 0.006276983767747879, + 0.005227439571171999, + 0.005948233883827925, + 0.006297928746789694, + 0.006476235575973988, + 0.006096545606851578, + 0.006865520961582661, + 0.006746460217982531, + 0.011463495902717113, + 0.008772586472332478, + 0.0064196158200502396, + 0.006905013229697943, + 0.010199424810707569, + 0.01863042823970318, + 0.01518284436315298, + 0.03496554493904114, + 0.03146006911993027, + 0.019740581512451172, + 0.016754502430558205, + 0.010964130982756615, + 0.011587277054786682 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 6, + 9 + ], + "selected_reward": -0.02634073793888092 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "The film begins with an erratic handheld tracking shot, the frame punctuated by digital noise and flickering like signal interference. The focus shifts rapidly between foreground and background, closely following a cuddly Disney Pixar-esque spherical robot. Its smooth, egg-yellow shell, a pair of enormous, digital blue eyes flashing wildly with utter terror, moves on three spindly metal legs, frantically darting through the ruins of a dilapidated biocontrol demonstration area. This area seems like an isolated island amidst a vast, gray wasteland, the sky a sickly, dim yellow. Countless thick veins of mutant plants, emitting an eerie purple and green bioluminescence, surge, entwine, and grow at a visible speed, as if alive, through the cracked ground and rusted debris of equipment. Each pulse of these veins causes the fluorescent light on their surface to flicker. The robot clumsily dodges a glowing vine that suddenly breaks out from the ground under its feet and almost trips. The whole scene forms a strong visual contrast between the cute character's terrified escape and the invasion of terrifying plants. Combined with the unstable lens and digital glitch effects, it successfully creates a unique, unsettling and weird contrast atmosphere.", + "seed": 2678, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000026/group_9/index_4175_seed_2678_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": -0.04993648314848542, + "vlm_logprob": -0.04993648314848542, + "vlm_yes_no_margin": 3.625, + "vlm_token_credit": -0.04993648314848542, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main character a spherical robot with a smooth egg-yellow shell and enormous digital blue eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the robot move on three spindly metal legs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment the ruins of a dilapidated biocontrol demonstration area located on an isolated island in a gray wasteland?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are thick veins of mutant plants emitting purple and green bioluminescence visible growing rapidly through the ground and debris?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot erratic and handheld with digital noise and flickering signal interference effects?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.015016373246908188 + ], + "opposite_logprob_tokens": [ + -4.265016555786133 + ], + "logprob_reward": -0.015016373246908188, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.015016373246908188, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1200551986694336, + 0.06547757983207703, + 0.3386658728122711, + 0.03899865597486496, + 0.2074456661939621, + 0.01329199317842722, + 0.12187129259109497, + 0.030620234087109566, + 0.008715475909411907, + 0.05485804006457329 + ], + "credit_map": [ + 0.01703290082514286, + 0.015445997007191181, + 0.014081294648349285, + 0.02064668759703636, + 0.0070031736977398396, + 0.020488696172833443, + 0.016161344945430756, + 0.009195102378726006, + 0.00825345516204834, + 0.006162514444440603, + 0.013407602906227112, + 0.01595575548708439, + 0.003776335157454014, + 0.005712749436497688, + 0.006747236475348473, + 0.005461933556944132, + 0.029069526121020317, + 0.09103819727897644, + 0.02163243480026722, + 0.061783336102962494, + 0.030668217688798904, + 0.03284754231572151, + 0.0413106344640255, + 0.030315963551402092, + 0.0036475216038525105, + 0.004584831185638905, + 0.00236521870829165, + 0.0029996312223374844, + 0.007168315351009369, + 0.007867232896387577, + 0.0041403258219361305, + 0.0062255775555968285, + 0.02084592543542385, + 0.004999075550585985, + 0.012641893699765205, + 0.023117881268262863, + 0.022328441962599754, + 0.028147738426923752, + 0.08221632242202759, + 0.01314839068800211, + 0.002611661795526743, + 0.001438530394807458, + 0.0018457560800015926, + 0.0014711299445480108, + 0.0011135705281049013, + 0.0015231776051223278, + 0.0018149134702980518, + 0.001473252777941525, + 0.010520240291953087, + 0.007212186697870493, + 0.013826118782162666, + 0.017740273848176003, + 0.009777904488146305, + 0.013114592991769314, + 0.03081880696117878, + 0.01886116713285446, + 0.003245127387344837, + 0.0031658851075917482, + 0.004826242569833994, + 0.005168332252651453, + 0.0029560180846601725, + 0.003454274497926235, + 0.004113936331123114, + 0.003690418554469943, + 0.0009823297150433064, + 0.0015480557922273874, + 0.000834052509162575, + 0.0012440511491149664, + 0.001096548279747367, + 0.0008087096503004432, + 0.001071654842235148, + 0.001130073331296444, + 0.0051591177470982075, + 0.011145961470901966, + 0.005774354562163353, + 0.009744995273649693, + 0.005215836688876152, + 0.003825406776741147, + 0.008084598928689957, + 0.005907769780606031 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 4, + 6 + ], + "selected_reward": -0.015016373246908188 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main character a spherical robot with a smooth egg-yellow shell and enormous digital blue eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the robot move on three spindly metal legs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment the ruins of a dilapidated biocontrol demonstration area located on an isolated island in a gray wasteland?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are thick veins of mutant plants emitting purple and green bioluminescence visible growing rapidly through the ground and debris?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot erratic and handheld with digital noise and flickering signal interference effects?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10033608973026276 + ], + "opposite_logprob_tokens": [ + -2.3503360748291016 + ], + "logprob_reward": -0.10033608973026276, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10033608973026276, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15785512328147888, + 0.07021134346723557, + 0.04330376535654068, + 0.054392147809267044, + 0.4679054319858551, + 0.06330043822526932, + 0.07615961134433746, + 0.01911892369389534, + 0.013150941580533981, + 0.03460223972797394 + ], + "credit_map": [ + 0.02610873244702816, + 0.016187259927392006, + 0.018238235265016556, + 0.028323139995336533, + 0.008144441992044449, + 0.027245692908763885, + 0.021894458681344986, + 0.011713152751326561, + 0.012222936376929283, + 0.0051854816265404224, + 0.018833596259355545, + 0.013073621317744255, + 0.003821037942543626, + 0.003818332450464368, + 0.008585221134126186, + 0.004671120550483465, + 0.004265666473656893, + 0.004849204793572426, + 0.0036837479565292597, + 0.005883913952857256, + 0.004346441477537155, + 0.004762158263474703, + 0.009121107868850231, + 0.006391526665538549, + 0.006368185393512249, + 0.006080657243728638, + 0.00540272006765008, + 0.007139611057937145, + 0.005663216579705477, + 0.009483147412538528, + 0.006344104651361704, + 0.007910503074526787, + 0.047422491014003754, + 0.011725971475243568, + 0.02466236986219883, + 0.05408673360943794, + 0.04946545884013176, + 0.07206102460622787, + 0.17996613681316376, + 0.028515253216028214, + 0.013699563220143318, + 0.0057472288608551025, + 0.008635306730866432, + 0.00562850059941411, + 0.004594694823026657, + 0.008221542462706566, + 0.006537145469337702, + 0.010236447677016258, + 0.007246818859130144, + 0.005304468795657158, + 0.007934657856822014, + 0.010166743770241737, + 0.0050077252089977264, + 0.009762072004377842, + 0.018123425543308258, + 0.012613696977496147, + 0.0029025354888290167, + 0.0014863257529214025, + 0.0021884264424443245, + 0.002287948038429022, + 0.0011920029064640403, + 0.0033389409072697163, + 0.002545831957831979, + 0.0031769098713994026, + 0.002285514259710908, + 0.0014382232911884785, + 0.0019840202294290066, + 0.0014041434042155743, + 0.0008648372604511678, + 0.0013741771690547466, + 0.002003438537940383, + 0.0017965877195820212, + 0.004276041407138109, + 0.005767177324742079, + 0.0028082225471735, + 0.0062515148892998695, + 0.00333785661496222, + 0.0027252689469605684, + 0.005994703620672226, + 0.0034414553083479404 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 6 + ], + "selected_reward": -0.10033608973026276 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main character a spherical robot with a smooth egg-yellow shell and enormous digital blue eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the robot move on three spindly metal legs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment the ruins of a dilapidated biocontrol demonstration area located on an isolated island in a gray wasteland?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are thick veins of mutant plants emitting purple and green bioluminescence visible growing rapidly through the ground and debris?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot erratic and handheld with digital noise and flickering signal interference effects?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08907785266637802 + ], + "opposite_logprob_tokens": [ + -2.464077949523926 + ], + "logprob_reward": -0.08907785266637802, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08907785266637802, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24099496006965637, + 0.0869750827550888, + 0.2688809037208557, + 0.03972760587930679, + 0.23277124762535095, + 0.03455258533358574, + 0.03846215829253197, + 0.026425138115882874, + 0.01178060844540596, + 0.01942971721291542 + ], + "credit_map": [ + 0.033397283405065536, + 0.027660643681883812, + 0.029105406254529953, + 0.0503777451813221, + 0.0172916017472744, + 0.0342218242585659, + 0.02744617685675621, + 0.02149428427219391, + 0.011290503665804863, + 0.012185543775558472, + 0.010642002336680889, + 0.014140632934868336, + 0.006834462285041809, + 0.009340697899460793, + 0.012167680077254772, + 0.010373562574386597, + 0.02490348555147648, + 0.08360538631677628, + 0.017985614016652107, + 0.05920898914337158, + 0.027476904913783073, + 0.0263157207518816, + 0.014645470306277275, + 0.014739332720637321, + 0.0036695082671940327, + 0.005499660968780518, + 0.005931312218308449, + 0.0038965868297964334, + 0.005875795613974333, + 0.00561657827347517, + 0.004640880972146988, + 0.004597281571477652, + 0.023542417213320732, + 0.006269088946282864, + 0.012579021044075489, + 0.027368959039449692, + 0.02422880195081234, + 0.034594275057315826, + 0.08969401568174362, + 0.014494677074253559, + 0.00431208498775959, + 0.0030205645598471165, + 0.004335630685091019, + 0.004141922574490309, + 0.00289440737105906, + 0.0034128790721297264, + 0.004484147299081087, + 0.007950949482619762, + 0.003443717025220394, + 0.003646101802587509, + 0.005010874010622501, + 0.0033495742827653885, + 0.003657884430140257, + 0.006483904086053371, + 0.004411075729876757, + 0.008459025993943214, + 0.0027760500088334084, + 0.0016522271325811744, + 0.0032800380140542984, + 0.0022489007096737623, + 0.0020332378335297108, + 0.004740775562822819, + 0.005731288809329271, + 0.003962620161473751, + 0.0012468363856896758, + 0.0013473419239744544, + 0.0017745395889505744, + 0.0014434270560741425, + 0.0011684551136568189, + 0.0012522460892796516, + 0.0018752296455204487, + 0.0016725321765989065, + 0.0018784436397254467, + 0.0021511015947908163, + 0.003908661659806967, + 0.003112869104370475, + 0.0014624698087573051, + 0.0021935184486210346, + 0.0020822372753173113, + 0.0026404138188809156 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 4 + ], + "selected_reward": -0.08907785266637802 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main character a spherical robot with a smooth egg-yellow shell and enormous digital blue eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the robot move on three spindly metal legs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment the ruins of a dilapidated biocontrol demonstration area located on an isolated island in a gray wasteland?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are thick veins of mutant plants emitting purple and green bioluminescence visible growing rapidly through the ground and debris?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot erratic and handheld with digital noise and flickering signal interference effects?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002197947818785906 + ], + "opposite_logprob_tokens": [ + -6.127197742462158 + ], + "logprob_reward": -0.002197947818785906, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.002197947818785906, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13142916560173035, + 0.05306171998381615, + 0.3226177990436554, + 0.056393321603536606, + 0.22906458377838135, + 0.03233177959918976, + 0.06617105007171631, + 0.03795934095978737, + 0.014734620228409767, + 0.056236617267131805 + ], + "credit_map": [ + 0.015765316784381866, + 0.015775980427861214, + 0.015726421028375626, + 0.025345169007778168, + 0.011191410943865776, + 0.013645162805914879, + 0.02266533672809601, + 0.011314371600747108, + 0.00479570496827364, + 0.006174195557832718, + 0.008962982334196568, + 0.011050608940422535, + 0.006023494526743889, + 0.004282153677195311, + 0.005826354026794434, + 0.005946221761405468, + 0.02962508797645569, + 0.10208772122859955, + 0.02068881317973137, + 0.0703960582613945, + 0.033292677253484726, + 0.031914275139570236, + 0.01622973196208477, + 0.018383419141173363, + 0.007661120966076851, + 0.0070791905745863914, + 0.008989199995994568, + 0.00617729127407074, + 0.005293407943099737, + 0.008628254756331444, + 0.005854207556694746, + 0.006710642483085394, + 0.02308487892150879, + 0.0069611165672540665, + 0.014345704577863216, + 0.02599509060382843, + 0.024437470361590385, + 0.032444994896650314, + 0.08777163922786713, + 0.014023683033883572, + 0.004438995849341154, + 0.003005305537953973, + 0.00479007326066494, + 0.0033882104326039553, + 0.0036789048463106155, + 0.004224976059049368, + 0.005173659883439541, + 0.0036316525656729937, + 0.008493252098560333, + 0.005215875804424286, + 0.007453265134245157, + 0.00723580177873373, + 0.00499314907938242, + 0.016078101471066475, + 0.007332819513976574, + 0.009368788450956345, + 0.003554718568921089, + 0.0026915078051388264, + 0.005051978398114443, + 0.0030569778755307198, + 0.00300653581507504, + 0.008593134582042694, + 0.0049798195250332355, + 0.007024666760116816, + 0.0015412532957270741, + 0.0019277742831036448, + 0.0016750748036429286, + 0.0015791557962074876, + 0.001771358191035688, + 0.0020710364915430546, + 0.0018591757398098707, + 0.002309790812432766, + 0.006084369495511055, + 0.013281412422657013, + 0.004692462272942066, + 0.007206670008599758, + 0.00617359671741724, + 0.005375802982598543, + 0.00748980138450861, + 0.005932498723268509 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 4, + 0 + ], + "selected_reward": -0.002197947818785906 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main character a spherical robot with a smooth egg-yellow shell and enormous digital blue eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the robot move on three spindly metal legs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment the ruins of a dilapidated biocontrol demonstration area located on an isolated island in a gray wasteland?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are thick veins of mutant plants emitting purple and green bioluminescence visible growing rapidly through the ground and debris?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot erratic and handheld with digital noise and flickering signal interference effects?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04305415228009224 + ], + "opposite_logprob_tokens": [ + -3.1680541038513184 + ], + "logprob_reward": -0.04305415228009224, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04305415228009224, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14301863312721252, + 0.04247410595417023, + 0.504303514957428, + 0.09850922226905823, + 0.08048513531684875, + 0.04968268796801567, + 0.027871066704392433, + 0.016266925260424614, + 0.01545603759586811, + 0.021932702511548996 + ], + "credit_map": [ + 0.017565513029694557, + 0.021121134981513023, + 0.017651332542300224, + 0.022676823660731316, + 0.009821946732699871, + 0.028119919821619987, + 0.013839752413332462, + 0.01222220528870821, + 0.007303277496248484, + 0.004239912144839764, + 0.006024294998496771, + 0.010226888582110405, + 0.002430247375741601, + 0.0028556769248098135, + 0.005043389275670052, + 0.004350416827946901, + 0.049168623983860016, + 0.15580204129219055, + 0.03386779874563217, + 0.11709369719028473, + 0.050413645803928375, + 0.04776560515165329, + 0.024535302072763443, + 0.025656770914793015, + 0.009041045792400837, + 0.014175950549542904, + 0.014536462724208832, + 0.008875987492501736, + 0.018153492361307144, + 0.011082793585956097, + 0.010618781670928001, + 0.01202470064163208, + 0.008052446879446507, + 0.0031342003494501114, + 0.005414009094238281, + 0.010180015116930008, + 0.007957200519740582, + 0.010581159964203835, + 0.029523395001888275, + 0.005642710253596306, + 0.009793554432690144, + 0.004832816775888205, + 0.0065023344941437244, + 0.004116600379347801, + 0.005328563041985035, + 0.006844162009656429, + 0.006002838723361492, + 0.006261818576604128, + 0.004230978898704052, + 0.0025346793700009584, + 0.0038805250078439713, + 0.003443773603066802, + 0.002055268967524171, + 0.00361328711733222, + 0.0038161908742040396, + 0.004296363331377506, + 0.0014229579828679562, + 0.0020250838715583086, + 0.0020054676569998264, + 0.0024608930107206106, + 0.0016342410817742348, + 0.002072787145152688, + 0.0027297777123749256, + 0.0019157155184075236, + 0.0019284383160993457, + 0.0013992430176585913, + 0.0016651060432195663, + 0.0019347475608810782, + 0.0015485159819945693, + 0.0017260455060750246, + 0.0033714191522449255, + 0.0018825234146788716, + 0.0021654090378433466, + 0.002550346776843071, + 0.002647539135068655, + 0.0030920160934329033, + 0.002323402091860771, + 0.0025170878507196903, + 0.004157865419983864, + 0.0024790368042886257 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 3 + ], + "selected_reward": -0.04305415228009224 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "The film begins with an erratic handheld tracking shot, the frame punctuated by digital noise and flickering like signal interference. The focus shifts rapidly between foreground and background, closely following a cuddly Disney Pixar-esque spherical robot. Its smooth, egg-yellow shell, a pair of enormous, digital blue eyes flashing wildly with utter terror, moves on three spindly metal legs, frantically darting through the ruins of a dilapidated biocontrol demonstration area. This area seems like an isolated island amidst a vast, gray wasteland, the sky a sickly, dim yellow. Countless thick veins of mutant plants, emitting an eerie purple and green bioluminescence, surge, entwine, and grow at a visible speed, as if alive, through the cracked ground and rusted debris of equipment. Each pulse of these veins causes the fluorescent light on their surface to flicker. The robot clumsily dodges a glowing vine that suddenly breaks out from the ground under its feet and almost trips. The whole scene forms a strong visual contrast between the cute character's terrified escape and the invasion of terrifying plants. Combined with the unstable lens and digital glitch effects, it successfully creates a unique, unsettling and weird contrast atmosphere.", + "seed": 2679, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000026/group_9/index_4175_seed_2679_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": -0.03182434826157987, + "vlm_logprob": -0.03182434826157987, + "vlm_yes_no_margin": 4.025, + "vlm_token_credit": -0.03182434826157987, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main character a spherical robot with a smooth egg-yellow shell and enormous digital blue eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the robot move on three spindly metal legs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment the ruins of a dilapidated biocontrol demonstration area located on an isolated island in a gray wasteland?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are thick veins of mutant plants emitting purple and green bioluminescence visible growing rapidly through the ground and debris?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot erratic and handheld with digital noise and flickering signal interference effects?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00519805820658803 + ], + "opposite_logprob_tokens": [ + -5.380198001861572 + ], + "logprob_reward": -0.00519805820658803, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.00519805820658803, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3120700716972351, + 0.05110439658164978, + 0.07709148526191711, + 0.04115403816103935, + 0.12452255934476852, + 0.08418010175228119, + 0.06930854916572571, + 0.10296662896871567, + 0.03689806908369064, + 0.10070401430130005 + ], + "credit_map": [ + 0.030501138418912888, + 0.042830243706703186, + 0.031200051307678223, + 0.03906376659870148, + 0.047419726848602295, + 0.030183224007487297, + 0.06434839218854904, + 0.02652355469763279, + 0.003945063799619675, + 0.004952710121870041, + 0.008562390692532063, + 0.007236999925225973, + 0.007904738187789917, + 0.005787464790046215, + 0.005185922607779503, + 0.007529112510383129, + 0.010273095220327377, + 0.011349253356456757, + 0.011789694428443909, + 0.007439486216753721, + 0.00785787496715784, + 0.011151782236993313, + 0.008270148187875748, + 0.008960157632827759, + 0.004198572598397732, + 0.008062324486672878, + 0.003937181551009417, + 0.00504312664270401, + 0.005464090965688229, + 0.004215907771140337, + 0.0060602943412959576, + 0.004172546323388815, + 0.008630126714706421, + 0.008800371550023556, + 0.0171336829662323, + 0.02725788950920105, + 0.01903909258544445, + 0.009691592305898666, + 0.018245112150907516, + 0.01572471112012863, + 0.007485013455152512, + 0.008217506110668182, + 0.012456727214157581, + 0.014502888545393944, + 0.008677824400365353, + 0.005103484261780977, + 0.020029019564390182, + 0.007707652635872364, + 0.006015080027282238, + 0.007335889618843794, + 0.015037192031741142, + 0.00845137145370245, + 0.00849140528589487, + 0.00893858727067709, + 0.008299333974719048, + 0.006739695090800524, + 0.007467986550182104, + 0.008322695270180702, + 0.010927734896540642, + 0.021323949098587036, + 0.009622069075703621, + 0.007603582926094532, + 0.018873779103159904, + 0.01882484182715416, + 0.0047385916113853455, + 0.0025294357910752296, + 0.003994836006313562, + 0.004225316923111677, + 0.0048224288038909435, + 0.0036996426060795784, + 0.008606850169599056, + 0.004280970431864262, + 0.012400565668940544, + 0.008278368040919304, + 0.022211559116840363, + 0.013324667699635029, + 0.007331198547035456, + 0.00604186300188303, + 0.00830919574946165, + 0.022806603461503983 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 7 + ], + "selected_reward": -0.00519805820658803 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main character a spherical robot with a smooth egg-yellow shell and enormous digital blue eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the robot move on three spindly metal legs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment the ruins of a dilapidated biocontrol demonstration area located on an isolated island in a gray wasteland?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are thick veins of mutant plants emitting purple and green bioluminescence visible growing rapidly through the ground and debris?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot erratic and handheld with digital noise and flickering signal interference effects?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.054968513548374176 + ], + "opposite_logprob_tokens": [ + -2.9299685955047607 + ], + "logprob_reward": -0.054968513548374176, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.054968513548374176, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21729080379009247, + 0.07870499789714813, + 0.1903066188097, + 0.11928356438875198, + 0.08025971055030823, + 0.06519531458616257, + 0.03985600918531418, + 0.06897973269224167, + 0.05752672255039215, + 0.08259651809930801 + ], + "credit_map": [ + 0.03744892030954361, + 0.03376060724258423, + 0.017654024064540863, + 0.01618466153740883, + 0.019877124577760696, + 0.058301717042922974, + 0.02183321863412857, + 0.012230523861944675, + 0.0066084410063922405, + 0.006760198622941971, + 0.01754765585064888, + 0.010054845362901688, + 0.009401284158229828, + 0.010765690356492996, + 0.0070380535908043385, + 0.010528827086091042, + 0.02826637402176857, + 0.03307810798287392, + 0.025329338386654854, + 0.019152363762259483, + 0.01655668392777443, + 0.0326499417424202, + 0.018588587641716003, + 0.01668522320687771, + 0.012357885017991066, + 0.028199220076203346, + 0.010313848033547401, + 0.013491670601069927, + 0.014102368615567684, + 0.011081794276833534, + 0.017055688425898552, + 0.012681091204285622, + 0.009459208697080612, + 0.01097509078681469, + 0.010391593910753727, + 0.009632494300603867, + 0.01066390611231327, + 0.009482404217123985, + 0.008769561536610126, + 0.0108854491263628, + 0.0053734988905489445, + 0.0061753131449222565, + 0.009398571215569973, + 0.012869703583419323, + 0.008686634711921215, + 0.005159843247383833, + 0.011356497183442116, + 0.0061752512119710445, + 0.003866884857416153, + 0.004301121924072504, + 0.006898549385368824, + 0.00522211380302906, + 0.004680689889937639, + 0.004644463304430246, + 0.00502932071685791, + 0.005212861578911543, + 0.005147331859916449, + 0.006040965672582388, + 0.008811854757368565, + 0.021675005555152893, + 0.005909901577979326, + 0.005540433339774609, + 0.008568253368139267, + 0.007285988889634609, + 0.005422922316938639, + 0.005649957340210676, + 0.005252080503851175, + 0.005995656829327345, + 0.009316357783973217, + 0.006048010196536779, + 0.012379361316561699, + 0.0074623762629926205, + 0.006761391181498766, + 0.007558593992143869, + 0.022606315091252327, + 0.009152146987617016, + 0.006109865382313728, + 0.006155949551612139, + 0.006904236972332001, + 0.017348019406199455 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 3 + ], + "selected_reward": -0.054968513548374176 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main character a spherical robot with a smooth egg-yellow shell and enormous digital blue eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the robot move on three spindly metal legs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment the ruins of a dilapidated biocontrol demonstration area located on an isolated island in a gray wasteland?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are thick veins of mutant plants emitting purple and green bioluminescence visible growing rapidly through the ground and debris?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot erratic and handheld with digital noise and flickering signal interference effects?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.062072932720184326 + ], + "opposite_logprob_tokens": [ + -2.812072992324829 + ], + "logprob_reward": -0.062072932720184326, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.062072932720184326, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16655924916267395, + 0.10427743941545486, + 0.13577976822853088, + 0.06804541498422623, + 0.1196557953953743, + 0.12714768946170807, + 0.07193537801504135, + 0.06000002101063728, + 0.06185321509838104, + 0.08474602550268173 + ], + "credit_map": [ + 0.028441043570637703, + 0.026240704581141472, + 0.017330048605799675, + 0.012060711160302162, + 0.014613804407417774, + 0.042626142501831055, + 0.011904719285666943, + 0.013342070393264294, + 0.009759948588907719, + 0.010460182093083858, + 0.01986525021493435, + 0.012463723309338093, + 0.012857181020081043, + 0.01632213406264782, + 0.011311129666864872, + 0.011237888596951962, + 0.018829314038157463, + 0.020771002396941185, + 0.015406642109155655, + 0.013057001866400242, + 0.014402315951883793, + 0.018114928156137466, + 0.015344381332397461, + 0.019854189828038216, + 0.006455879658460617, + 0.006790599785745144, + 0.007229918614029884, + 0.009597337804734707, + 0.008749352768063545, + 0.008474001660943031, + 0.011464976705610752, + 0.009283349849283695, + 0.010537789203226566, + 0.011896109208464622, + 0.021851522848010063, + 0.028063472360372543, + 0.009944157674908638, + 0.008222312666475773, + 0.013092627748847008, + 0.016047805547714233, + 0.011361469514667988, + 0.012273836880922318, + 0.017712827771902084, + 0.021271435543894768, + 0.014408884570002556, + 0.010686264373362064, + 0.024603169411420822, + 0.01482979767024517, + 0.004849105607718229, + 0.006870038341730833, + 0.010287811979651451, + 0.00887025985866785, + 0.013947181403636932, + 0.008232378400862217, + 0.010317315347492695, + 0.008561287075281143, + 0.01058869156986475, + 0.007574800401926041, + 0.00574364373460412, + 0.006659276317805052, + 0.005504745990037918, + 0.006604272406548262, + 0.007821213454008102, + 0.009503376670181751, + 0.00581996189430356, + 0.007142483722418547, + 0.005210683215409517, + 0.00584582332521677, + 0.00943869911134243, + 0.0068559483624994755, + 0.0129768755286932, + 0.008562739938497543, + 0.006671931594610214, + 0.010084063746035099, + 0.025514913722872734, + 0.005402295850217342, + 0.004735617898404598, + 0.004388259258121252, + 0.0037803405430167913, + 0.024168599396944046 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 5 + ], + "selected_reward": -0.062072932720184326 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main character a spherical robot with a smooth egg-yellow shell and enormous digital blue eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the robot move on three spindly metal legs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment the ruins of a dilapidated biocontrol demonstration area located on an isolated island in a gray wasteland?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are thick veins of mutant plants emitting purple and green bioluminescence visible growing rapidly through the ground and debris?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot erratic and handheld with digital noise and flickering signal interference effects?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0031919742468744516 + ], + "opposite_logprob_tokens": [ + -5.753191947937012 + ], + "logprob_reward": -0.0031919742468744516, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.0031919742468744516, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14405791461467743, + 0.053045693784952164, + 0.05551852658390999, + 0.03609443083405495, + 0.1271108090877533, + 0.1163351833820343, + 0.13008427619934082, + 0.13587334752082825, + 0.10133084654808044, + 0.10054895281791687 + ], + "credit_map": [ + 0.024100836366415024, + 0.02521536499261856, + 0.01445755921304226, + 0.01346680149435997, + 0.01052962988615036, + 0.03440140187740326, + 0.011306767351925373, + 0.010579569265246391, + 0.00440186494961381, + 0.004722759127616882, + 0.0067151696421206, + 0.009001201950013638, + 0.005191121716052294, + 0.004656463861465454, + 0.011569922789931297, + 0.006787190679460764, + 0.003977929707616568, + 0.006381216924637556, + 0.007828906178474426, + 0.005896962247788906, + 0.007922139950096607, + 0.008995478972792625, + 0.007261223159730434, + 0.007254674099385738, + 0.0030978552531450987, + 0.0038258375134319067, + 0.0029349951073527336, + 0.0053583416156470776, + 0.004362375941127539, + 0.003867429681122303, + 0.00782803539186716, + 0.004819564986974001, + 0.016925282776355743, + 0.015441551804542542, + 0.01538008637726307, + 0.017299117520451546, + 0.018038462847471237, + 0.014919418841600418, + 0.015110213309526443, + 0.013996686786413193, + 0.011103147640824318, + 0.014989652670919895, + 0.017005089670419693, + 0.027557170018553734, + 0.009982552379369736, + 0.009678593836724758, + 0.012278073467314243, + 0.013740915805101395, + 0.01364080049097538, + 0.011772754602134228, + 0.03534441813826561, + 0.01683000661432743, + 0.008209050633013248, + 0.0134658208116889, + 0.017541155219078064, + 0.013280285522341728, + 0.02025262638926506, + 0.02072197012603283, + 0.012929821386933327, + 0.020791910588741302, + 0.013828265480697155, + 0.012832914479076862, + 0.01706854999065399, + 0.017447305843234062, + 0.01459331251680851, + 0.009850354865193367, + 0.010126518085598946, + 0.011770817451179028, + 0.018459822982549667, + 0.012134985998272896, + 0.014781503938138485, + 0.009613546542823315, + 0.011025259271264076, + 0.015709800645709038, + 0.014145031571388245, + 0.013126016594469547, + 0.009161427617073059, + 0.010352037847042084, + 0.010015908628702164, + 0.01701347716152668 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 6 + ], + "selected_reward": -0.0031919742468744516 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main character a spherical robot with a smooth egg-yellow shell and enormous digital blue eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the robot move on three spindly metal legs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment the ruins of a dilapidated biocontrol demonstration area located on an isolated island in a gray wasteland?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are thick veins of mutant plants emitting purple and green bioluminescence visible growing rapidly through the ground and debris?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot erratic and handheld with digital noise and flickering signal interference effects?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"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": [ + 0.25862348079681396, + 0.06813646852970123, + 0.043044283986091614, + 0.06810493767261505, + 0.10435699671506882, + 0.1430666297674179, + 0.07648627460002899, + 0.06855227053165436, + 0.053682178258895874, + 0.1159464493393898 + ], + "credit_map": [ + 0.04839921370148659, + 0.03819571062922478, + 0.027473021298646927, + 0.023300794884562492, + 0.026337411254644394, + 0.04328175261616707, + 0.031604450196027756, + 0.020031146705150604, + 0.008489159867167473, + 0.006921735592186451, + 0.0111785177141428, + 0.008928784169256687, + 0.009582316502928734, + 0.00761062977835536, + 0.006993697956204414, + 0.008431627415120602, + 0.005611142609268427, + 0.008218502625823021, + 0.005509174428880215, + 0.004351879935711622, + 0.003744863672181964, + 0.005710605066269636, + 0.005827920977026224, + 0.004070192109793425, + 0.006175587419420481, + 0.00955507904291153, + 0.007308576721698046, + 0.007240489590913057, + 0.008203506469726562, + 0.005249325186014175, + 0.016902616247534752, + 0.007469756528735161, + 0.01198419276624918, + 0.014128080569207668, + 0.016384640708565712, + 0.012495704926550388, + 0.012135287746787071, + 0.014892717823386192, + 0.011692482978105545, + 0.01064387708902359, + 0.01538935024291277, + 0.022796325385570526, + 0.019135857000947, + 0.018937045708298683, + 0.014946067705750465, + 0.017150754109025, + 0.018934650346636772, + 0.015776580199599266, + 0.00944406259804964, + 0.008191249333322048, + 0.013300621882081032, + 0.00869785901159048, + 0.0067308624275028706, + 0.011874273419380188, + 0.009751778095960617, + 0.008495570160448551, + 0.011501733213663101, + 0.007160961627960205, + 0.0070397015661001205, + 0.007634499575942755, + 0.005358810070902109, + 0.009149918332695961, + 0.011063222773373127, + 0.009643420577049255, + 0.005888894200325012, + 0.004826537799090147, + 0.006146352272480726, + 0.004338065627962351, + 0.008617744781076908, + 0.007754833437502384, + 0.007512297015637159, + 0.00859745405614376, + 0.011850595474243164, + 0.009290882386267185, + 0.02256958745419979, + 0.019758477807044983, + 0.008737543597817421, + 0.008864326402544975, + 0.01171096321195364, + 0.023164071142673492 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 9 + ], + "selected_reward": -0.03369026258587837 + } + ], + "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/0000056/group_0/index_tensor([8869, 8869])_group_0_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000056/group_0/index_tensor([8869, 8869])_group_0_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..66c832b191fb397f0c907de6b263d643a98e6ede --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000056/group_0/index_tensor([8869, 8869])_group_0_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "surreal low-angle approaching shot tracks smoothly alongside a J-20 stealth fighter as it roars down a rain-slicked runway with a powerful takeoff roll. The camera maintains a constant speed with the accelerating aircraft, emphasizing its immense velocity. The fighter's dark, angular fuselage, coated in a matte, radar-absorbing material, gleams in the diffuse light of an overcast sky, every rivet, seam, and panel line rendered in crisp, lifelike detail. Water streams across its geometric surfaces and sleek, monolithic canopy. As the aircraft surges forward, its powerful twin engines generate a shimmering, translucent heat haze, violently vaporizing the layer of water on the dark asphalt, creating billowing clouds of white steam that blend with the plumes of spray kicked up by its landing gear. The entire scene, dominated by a sober, monochromatic palette of grays and blacks, creates a tense and breathtaking atmosphere of raw technological power unleashed.", + "seed": 19024, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000056/group_0/index_8869_seed_19024_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": -0.07131142746657133, + "vlm_logprob": -0.07131142746657133, + "vlm_yes_no_margin": 3.3, + "vlm_token_credit": -0.07131142746657133, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a J-20 stealth fighter aircraft?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the aircraft performing a takeoff roll on a rain-slicked runway?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera use a low-angle approach tracking smoothly alongside the accelerating aircraft?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are visible visual effects including water streams on the fuselage, a shimmering heat haze from the engines, and billowing white steam?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene dominated by a monochromatic palette of grays and blacks under an overcast sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.13526801764965057 + ], + "opposite_logprob_tokens": [ + -2.135267972946167 + ], + "logprob_reward": -0.13526801764965057, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.13526801764965057, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16070708632469177, + 0.2988240122795105, + 0.11531831324100494, + 0.15154944360256195, + 0.06496375799179077, + 0.041533198207616806, + 0.028259195387363434, + 0.020635560154914856, + 0.061041124165058136, + 0.057168252766132355 + ], + "credit_map": [ + 0.02387087792158127, + 0.020929187536239624, + 0.02272936888039112, + 0.01245522778481245, + 0.01678207330405712, + 0.02472548931837082, + 0.011483720503747463, + 0.027731139212846756, + 0.041684750467538834, + 0.05916818603873253, + 0.036076486110687256, + 0.03312854841351509, + 0.041253626346588135, + 0.040296778082847595, + 0.024762257933616638, + 0.02245338261127472, + 0.016671672463417053, + 0.013829919509589672, + 0.008584657683968544, + 0.011276447214186192, + 0.02533770352602005, + 0.018271371722221375, + 0.008614365942776203, + 0.012732173316180706, + 0.012041500769555569, + 0.03234096243977547, + 0.012929114513099194, + 0.01423044502735138, + 0.025889532640576363, + 0.027936214581131935, + 0.014337015338242054, + 0.011844671331346035, + 0.005524435080587864, + 0.009156325832009315, + 0.007969620637595654, + 0.0065168836154043674, + 0.01317581906914711, + 0.010646162554621696, + 0.005321883130818605, + 0.006652633659541607, + 0.005265498999506235, + 0.006946361158043146, + 0.0034672520123422146, + 0.00498763844370842, + 0.005213406402617693, + 0.005592897534370422, + 0.005685686133801937, + 0.004374452866613865, + 0.004988228436559439, + 0.0037282882258296013, + 0.002823484130203724, + 0.0033219947945326567, + 0.0034704420249909163, + 0.004315997939556837, + 0.003493088763207197, + 0.0021176720038056374, + 0.005725862458348274, + 0.0026252625975757837, + 0.002114651957526803, + 0.001914282445795834, + 0.0027514623943716288, + 0.002472295193001628, + 0.0015590169932693243, + 0.0014727270463481545, + 0.011625915765762329, + 0.005625167395919561, + 0.009276934899389744, + 0.007982871495187283, + 0.006322967819869518, + 0.006622949615120888, + 0.00684511661529541, + 0.006739196367561817, + 0.005495401099324226, + 0.004871945828199387, + 0.008797123096883297, + 0.018683049827814102, + 0.004666518419981003, + 0.004054334480315447, + 0.004631216172128916, + 0.005968658719211817 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 3 + ], + "selected_reward": -0.13526801764965057 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a J-20 stealth fighter aircraft?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the aircraft performing a takeoff roll on a rain-slicked runway?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera use a low-angle approach tracking smoothly alongside the accelerating aircraft?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are visible visual effects including water streams on the fuselage, a shimmering heat haze from the engines, and billowing white steam?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene dominated by a monochromatic palette of grays and blacks under an overcast sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.17991982400417328 + ], + "opposite_logprob_tokens": [ + -1.8049198389053345 + ], + "logprob_reward": -0.17991982400417328, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.17991982400417328, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15458406507968903, + 0.06592363119125366, + 0.135994091629982, + 0.061070531606674194, + 0.07649936527013779, + 0.05842160806059837, + 0.0982191264629364, + 0.19044269621372223, + 0.08205590397119522, + 0.0767890065908432 + ], + "credit_map": [ + 0.02111881785094738, + 0.019923949614167213, + 0.018262559548020363, + 0.012295772321522236, + 0.019958093762397766, + 0.022956063970923424, + 0.010657962411642075, + 0.0294108334928751, + 0.00921782199293375, + 0.012344499118626118, + 0.007593146059662104, + 0.004875672981142998, + 0.009663834236562252, + 0.010215542279183865, + 0.004505570977926254, + 0.00750754214823246, + 0.021325314417481422, + 0.018240049481391907, + 0.011009549722075462, + 0.013183807954192162, + 0.023260602727532387, + 0.021050168201327324, + 0.010382739827036858, + 0.017541855573654175, + 0.009215468540787697, + 0.007996431551873684, + 0.010363386012613773, + 0.009354900568723679, + 0.0053575714118778706, + 0.009358013980090618, + 0.0047653717920184135, + 0.00465938588604331, + 0.008202279917895794, + 0.012892285361886024, + 0.0077058300375938416, + 0.006037841085344553, + 0.01386738009750843, + 0.011770622804760933, + 0.008698424324393272, + 0.007324698381125927, + 0.011349363252520561, + 0.008364818058907986, + 0.004673948045819998, + 0.005955907981842756, + 0.006681304425001144, + 0.009278826415538788, + 0.004317167680710554, + 0.007800274062901735, + 0.018605004996061325, + 0.013556779362261295, + 0.008767366409301758, + 0.009182194247841835, + 0.010789129883050919, + 0.015376186929643154, + 0.013384215533733368, + 0.008558246307075024, + 0.044956400990486145, + 0.0344206877052784, + 0.01802683062851429, + 0.023217977955937386, + 0.022090332582592964, + 0.022909583523869514, + 0.014635209925472736, + 0.010185661725699902, + 0.013722950592637062, + 0.007720837835222483, + 0.010684562847018242, + 0.015756765380501747, + 0.007952076382935047, + 0.008059418760240078, + 0.008571930229663849, + 0.009587363339960575, + 0.008209753781557083, + 0.009443572722375393, + 0.012730208225548267, + 0.015536796301603317, + 0.006252513732761145, + 0.008259328082203865, + 0.00889196339994669, + 0.007464868947863579 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 2 + ], + "selected_reward": -0.17991982400417328 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a J-20 stealth fighter aircraft?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the aircraft performing a takeoff roll on a rain-slicked runway?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera use a low-angle approach tracking smoothly alongside the accelerating aircraft?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are visible visual effects including water streams on the fuselage, a shimmering heat haze from the engines, and billowing white steam?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene dominated by a monochromatic palette of grays and blacks under an overcast sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014207640662789345 + ], + "opposite_logprob_tokens": [ + -4.26420783996582 + ], + "logprob_reward": -0.014207640662789345, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014207640662789345, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18867409229278564, + 0.1362285166978836, + 0.04945516586303711, + 0.02909853868186474, + 0.05034405738115311, + 0.035339921712875366, + 0.06179928407073021, + 0.29495254158973694, + 0.06392001360654831, + 0.09018788486719131 + ], + "credit_map": [ + 0.02459223009645939, + 0.028811359778046608, + 0.02221292071044445, + 0.017485875636339188, + 0.023495906963944435, + 0.029075687751173973, + 0.012436846271157265, + 0.030563263222575188, + 0.01882089488208294, + 0.022370589897036552, + 0.017387939617037773, + 0.014652235433459282, + 0.01953365094959736, + 0.017151199281215668, + 0.0125685790553689, + 0.013743430376052856, + 0.007179593667387962, + 0.007101487368345261, + 0.0047939615324139595, + 0.0041457610204815865, + 0.0058395410887897015, + 0.00809885561466217, + 0.0035814926959574223, + 0.00871447566896677, + 0.0030036307871341705, + 0.005815551616251469, + 0.002439365955069661, + 0.002715134294703603, + 0.004699419718235731, + 0.005043352954089642, + 0.002868042327463627, + 0.002514039631932974, + 0.0038344881031662226, + 0.008307591080665588, + 0.0052680885419249535, + 0.004639041144400835, + 0.009717993438243866, + 0.00890649575740099, + 0.005253446288406849, + 0.004416914191097021, + 0.004119593650102615, + 0.00559454970061779, + 0.002847474068403244, + 0.0035132169723510742, + 0.0045393360778689384, + 0.005480786319822073, + 0.003768493654206395, + 0.005476471036672592, + 0.007252609822899103, + 0.010100322775542736, + 0.0073516457341611385, + 0.0060974773950874805, + 0.006777761038392782, + 0.008882089518010616, + 0.008077415637671947, + 0.007259959354996681, + 0.06946559250354767, + 0.04986698180437088, + 0.029794853180646896, + 0.031033983454108238, + 0.03926779702305794, + 0.03322148323059082, + 0.023403067141771317, + 0.01889875903725624, + 0.008776229806244373, + 0.005653652362525463, + 0.013313092291355133, + 0.007241731975227594, + 0.006265340372920036, + 0.006599329877644777, + 0.006341011729091406, + 0.009729630313813686, + 0.0075317020528018475, + 0.008296490646898746, + 0.01123348530381918, + 0.03244239091873169, + 0.006524758879095316, + 0.0057557751424610615, + 0.006446363870054483, + 0.011956926435232162 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 1 + ], + "selected_reward": -0.014207640662789345 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a J-20 stealth fighter aircraft?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the aircraft performing a takeoff roll on a rain-slicked runway?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera use a low-angle approach tracking smoothly alongside the accelerating aircraft?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are visible visual effects including water streams on the fuselage, a shimmering heat haze from the engines, and billowing white steam?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene dominated by a monochromatic palette of grays and blacks under an overcast sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01608443818986416 + ], + "opposite_logprob_tokens": [ + -4.141084671020508 + ], + "logprob_reward": -0.01608443818986416, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.01608443818986416, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16797170042991638, + 0.11150025576353073, + 0.185973659157753, + 0.10692014545202255, + 0.0727386549115181, + 0.04024985805153847, + 0.06718584895133972, + 0.08661426603794098, + 0.07838078588247299, + 0.08246488124132156 + ], + "credit_map": [ + 0.028864508494734764, + 0.02032710611820221, + 0.019009778276085854, + 0.012363397516310215, + 0.028061434626579285, + 0.021415159106254578, + 0.011393241584300995, + 0.026537062600255013, + 0.016445716843008995, + 0.017624016851186752, + 0.012379968538880348, + 0.013374772854149342, + 0.012871514074504375, + 0.014730008319020271, + 0.011974968016147614, + 0.012099280953407288, + 0.03111923113465309, + 0.025102995336055756, + 0.01823549158871174, + 0.013579217717051506, + 0.02289407141506672, + 0.031877174973487854, + 0.012165026739239693, + 0.031000424176454544, + 0.01728680729866028, + 0.01803658902645111, + 0.014248955994844437, + 0.008525016717612743, + 0.012292412109673023, + 0.017488135024905205, + 0.006450008135288954, + 0.012592204846441746, + 0.008035900071263313, + 0.017793908715248108, + 0.007489820010960102, + 0.009467514231801033, + 0.00638932827860117, + 0.012598673813045025, + 0.0041992501355707645, + 0.006764252670109272, + 0.004875972867012024, + 0.009491617791354656, + 0.0030031423084437847, + 0.003262648358941078, + 0.004772573243826628, + 0.006953049451112747, + 0.00425209803506732, + 0.0036387525033205748, + 0.012228787876665592, + 0.013825538568198681, + 0.005039950367063284, + 0.005509028676897287, + 0.006902762688696384, + 0.011564033105969429, + 0.007045738399028778, + 0.005070006009191275, + 0.01953584887087345, + 0.012103798799216747, + 0.006557957269251347, + 0.008419369347393513, + 0.014130420051515102, + 0.010661923326551914, + 0.008132027462124825, + 0.007072913460433483, + 0.009398472495377064, + 0.00758630083873868, + 0.011172208935022354, + 0.016703100875020027, + 0.005808647256344557, + 0.011099672876298428, + 0.007385591976344585, + 0.009226790629327297, + 0.00766770401969552, + 0.006841178983449936, + 0.01199120283126831, + 0.01435726135969162, + 0.005897916853427887, + 0.010663795284926891, + 0.014404253102838993, + 0.010641558095812798 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 1 + ], + "selected_reward": -0.01608443818986416 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a J-20 stealth fighter aircraft?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the aircraft performing a takeoff roll on a rain-slicked runway?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera use a low-angle approach tracking smoothly alongside the accelerating aircraft?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are visible visual effects including water streams on the fuselage, a shimmering heat haze from the engines, and billowing white steam?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene dominated by a monochromatic palette of grays and blacks under an overcast sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0110772168263793 + ], + "opposite_logprob_tokens": [ + -4.511077404022217 + ], + "logprob_reward": -0.0110772168263793, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.0110772168263793, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22857291996479034, + 0.08953241258859634, + 0.056860484182834625, + 0.10783465206623077, + 0.14260227978229523, + 0.025299154222011566, + 0.08129788190126419, + 0.06590548902750015, + 0.07352937757968903, + 0.12856534123420715 + ], + "credit_map": [ + 0.033843912184238434, + 0.03436017408967018, + 0.025129329413175583, + 0.020233219489455223, + 0.035468656569719315, + 0.03218137100338936, + 0.014870508573949337, + 0.032485757023096085, + 0.014175486750900745, + 0.011600716039538383, + 0.010909944772720337, + 0.005316086113452911, + 0.015320828184485435, + 0.013796094805002213, + 0.007447198964655399, + 0.010966052301228046, + 0.009867957793176174, + 0.009807102382183075, + 0.0047372174449265, + 0.0057717422023415565, + 0.008263863623142242, + 0.008719957433640957, + 0.004163653589785099, + 0.005528988316655159, + 0.01174150686711073, + 0.026493731886148453, + 0.013067906722426414, + 0.015140779316425323, + 0.009895166382193565, + 0.01764603517949581, + 0.006159754004329443, + 0.007689768448472023, + 0.015343276783823967, + 0.03668802231550217, + 0.016211355105042458, + 0.01329062320291996, + 0.013572251424193382, + 0.024491749703884125, + 0.008055384270846844, + 0.014949616976082325, + 0.003313800785690546, + 0.0036198552697896957, + 0.0021854930091649294, + 0.0028018115554004908, + 0.0028598878998309374, + 0.0035095515195280313, + 0.0029789286199957132, + 0.004029824864119291, + 0.009539532475173473, + 0.02244153805077076, + 0.005908885505050421, + 0.005769347306340933, + 0.00919119268655777, + 0.014545137993991375, + 0.009040473960340023, + 0.004861772991716862, + 0.01470083650201559, + 0.011254481039941311, + 0.007331325206905603, + 0.007647916208952665, + 0.007830658927559853, + 0.007397746201604605, + 0.005037520546466112, + 0.004705002065747976, + 0.012208906002342701, + 0.00687909172847867, + 0.008752540685236454, + 0.014343279413878918, + 0.005217373371124268, + 0.011000629514455795, + 0.007128296885639429, + 0.007999258115887642, + 0.01582382433116436, + 0.011206001974642277, + 0.02431555837392807, + 0.03617224097251892, + 0.013054374605417252, + 0.010879984125494957, + 0.009979026392102242, + 0.007134328596293926 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 9 + ], + "selected_reward": -0.0110772168263793 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "surreal low-angle approaching shot tracks smoothly alongside a J-20 stealth fighter as it roars down a rain-slicked runway with a powerful takeoff roll. The camera maintains a constant speed with the accelerating aircraft, emphasizing its immense velocity. The fighter's dark, angular fuselage, coated in a matte, radar-absorbing material, gleams in the diffuse light of an overcast sky, every rivet, seam, and panel line rendered in crisp, lifelike detail. Water streams across its geometric surfaces and sleek, monolithic canopy. As the aircraft surges forward, its powerful twin engines generate a shimmering, translucent heat haze, violently vaporizing the layer of water on the dark asphalt, creating billowing clouds of white steam that blend with the plumes of spray kicked up by its landing gear. The entire scene, dominated by a sober, monochromatic palette of grays and blacks, creates a tense and breathtaking atmosphere of raw technological power unleashed.", + "seed": 19025, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000056/group_0/index_8869_seed_19025_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": -0.03661809703335166, + "vlm_logprob": -0.03661809703335166, + "vlm_yes_no_margin": 4.225, + "vlm_token_credit": -0.03661809703335166, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a J-20 stealth fighter aircraft?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the aircraft performing a takeoff roll on a rain-slicked runway?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera use a low-angle approach tracking smoothly alongside the accelerating aircraft?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are visible visual effects including water streams on the fuselage, a shimmering heat haze from the engines, and billowing white steam?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene dominated by a monochromatic palette of grays and blacks under an overcast sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.13873858749866486 + ], + "opposite_logprob_tokens": [ + -2.1387386322021484 + ], + "logprob_reward": -0.13873858749866486, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.13873858749866486, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2405426949262619, + 0.1018456220626831, + 0.08487040549516678, + 0.08048171550035477, + 0.09526904672384262, + 0.09962902963161469, + 0.054996080696582794, + 0.08685039728879929, + 0.12125883251428604, + 0.034256163984537125 + ], + "credit_map": [ + 0.02462717704474926, + 0.030815381556749344, + 0.040788907557725906, + 0.03588493913412094, + 0.03134395182132721, + 0.02912086993455887, + 0.03086121752858162, + 0.017100242897868156, + 0.009313984774053097, + 0.019165512174367905, + 0.008851969614624977, + 0.015597577206790447, + 0.010760458186268806, + 0.012289855629205704, + 0.015485413372516632, + 0.01038084551692009, + 0.009566220454871655, + 0.017577357590198517, + 0.008836980909109116, + 0.009529699571430683, + 0.013440366834402084, + 0.010815329849720001, + 0.0052745407447218895, + 0.009829910472035408, + 0.016076695173978806, + 0.009304427541792393, + 0.01087311003357172, + 0.009167240932583809, + 0.008163206279277802, + 0.008786365389823914, + 0.01020003855228424, + 0.007910630665719509, + 0.012149450369179249, + 0.006243513431400061, + 0.01017194427549839, + 0.033192019909620285, + 0.005367674399167299, + 0.0051904586143791676, + 0.012384526431560516, + 0.01056945975869894, + 0.015074362978339195, + 0.011260186322033405, + 0.017991920933127403, + 0.021066587418317795, + 0.008127043023705482, + 0.008712477050721645, + 0.00956379808485508, + 0.007832652889192104, + 0.006744415033608675, + 0.01016317680478096, + 0.006386102177202702, + 0.00512987794354558, + 0.009722008369863033, + 0.008275029249489307, + 0.005277309101074934, + 0.0032981655094772577, + 0.013015855103731155, + 0.009826370514929295, + 0.01364170853048563, + 0.007476477883756161, + 0.01693645864725113, + 0.012629120610654354, + 0.0077173225581645966, + 0.005607083905488253, + 0.011478000320494175, + 0.02086499147117138, + 0.009840716607868671, + 0.008773965761065483, + 0.033558424562215805, + 0.02149955928325653, + 0.008356752805411816, + 0.00688642542809248, + 0.0056184642016887665, + 0.0040056840516626835, + 0.0025490224361419678, + 0.0026623308658599854, + 0.003527869936078787, + 0.0077936905436217785, + 0.0036623631604015827, + 0.004436739720404148 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 1 + ], + "selected_reward": -0.13873858749866486 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a J-20 stealth fighter aircraft?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the aircraft performing a takeoff roll on a rain-slicked runway?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera use a low-angle approach tracking smoothly alongside the accelerating aircraft?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are visible visual effects including water streams on the fuselage, a shimmering heat haze from the engines, and billowing white steam?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene dominated by a monochromatic palette of grays and blacks under an overcast sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02636047638952732 + ], + "opposite_logprob_tokens": [ + -3.651360511779785 + ], + "logprob_reward": -0.02636047638952732, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.02636047638952732, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23999521136283875, + 0.0971931591629982, + 0.14634118974208832, + 0.08677622675895691, + 0.0387575700879097, + 0.08321749418973923, + 0.08056323230266571, + 0.13279618322849274, + 0.07526827603578568, + 0.019091423600912094 + ], + "credit_map": [ + 0.02940523624420166, + 0.03325464949011803, + 0.03769595921039581, + 0.03156086802482605, + 0.03152533620595932, + 0.029831843450665474, + 0.02256048284471035, + 0.0241608377546072, + 0.007761947810649872, + 0.010986777022480965, + 0.011660746298730373, + 0.014232348650693893, + 0.00997872743755579, + 0.011520850472152233, + 0.015688197687268257, + 0.015363560058176517, + 0.017774172127246857, + 0.02666730061173439, + 0.01633714698255062, + 0.01470136921852827, + 0.02252470888197422, + 0.019541120156645775, + 0.010209299623966217, + 0.018586069345474243, + 0.013651465997099876, + 0.012978825718164444, + 0.01569289155304432, + 0.01063399575650692, + 0.008036921732127666, + 0.010881610214710236, + 0.006846118252724409, + 0.008054401725530624, + 0.00451201805844903, + 0.0038624743465334177, + 0.00603350019082427, + 0.008892172947525978, + 0.002296141814440489, + 0.00257508410140872, + 0.0066970037296414375, + 0.0038891760632395744, + 0.011567031033337116, + 0.007040061056613922, + 0.01582382619380951, + 0.010689438320696354, + 0.007868196815252304, + 0.009496032260358334, + 0.010331392288208008, + 0.01040151622146368, + 0.008272112347185612, + 0.014695716090500355, + 0.009703644551336765, + 0.007591248955577612, + 0.014524595811963081, + 0.00985732115805149, + 0.01063415315002203, + 0.005284443031996489, + 0.015113238245248795, + 0.013272665441036224, + 0.023572450503706932, + 0.010925175622105598, + 0.027489017695188522, + 0.019160008057951927, + 0.012721029110252857, + 0.01054259855300188, + 0.015603553503751755, + 0.010974432341754436, + 0.006007358897477388, + 0.005034775473177433, + 0.013818973675370216, + 0.011416585184633732, + 0.0066049061715602875, + 0.005807684268802404, + 0.0027909334748983383, + 0.002154262736439705, + 0.001779737533070147, + 0.0016365662449970841, + 0.0018773458432406187, + 0.0039100563153624535, + 0.0021198480390012264, + 0.002822672249749303 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 7 + ], + "selected_reward": -0.02636047638952732 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a J-20 stealth fighter aircraft?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the aircraft performing a takeoff roll on a rain-slicked runway?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera use a low-angle approach tracking smoothly alongside the accelerating aircraft?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are visible visual effects including water streams on the fuselage, a shimmering heat haze from the engines, and billowing white steam?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene dominated by a monochromatic palette of grays and blacks under an overcast sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007632143795490265 + ], + "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": [ + 0.30954545736312866, + 0.2801443934440613, + 0.05656275153160095, + 0.03890962153673172, + 0.05914433300495148, + 0.05187522992491722, + 0.05280936509370804, + 0.046615567058324814, + 0.05704031139612198, + 0.04735303670167923 + ], + "credit_map": [ + 0.029290413483977318, + 0.03279067948460579, + 0.05799437686800957, + 0.04898359254002571, + 0.040378786623477936, + 0.032430846244096756, + 0.039854686707258224, + 0.027822017669677734, + 0.02782696671783924, + 0.038318175822496414, + 0.039402469992637634, + 0.02896256186068058, + 0.028254909440875053, + 0.026210827752947807, + 0.05432083457708359, + 0.036847617477178574, + 0.006549453362822533, + 0.011254662647843361, + 0.007526275236159563, + 0.008559172041714191, + 0.006510385777801275, + 0.006495486479252577, + 0.00384192424826324, + 0.005825386382639408, + 0.005632263608276844, + 0.005835120100528002, + 0.005704602692276239, + 0.004886765033006668, + 0.006688648834824562, + 0.0041570356115698814, + 0.0028721163980662823, + 0.0031330655328929424, + 0.007503300905227661, + 0.0059400820173323154, + 0.008909293450415134, + 0.016551626846194267, + 0.0036033811047673225, + 0.0051638162694871426, + 0.005897440481930971, + 0.005575386807322502, + 0.008029995486140251, + 0.0070208655670285225, + 0.006104531232267618, + 0.005059582646936178, + 0.006878087762743235, + 0.0067826928570866585, + 0.005919323768466711, + 0.006080145016312599, + 0.0061231739819049835, + 0.007824323140084743, + 0.006391465198248625, + 0.0048235259018838406, + 0.007966074161231518, + 0.00632620370015502, + 0.008755040355026722, + 0.004599550738930702, + 0.004959947429597378, + 0.0049261730164289474, + 0.009096410125494003, + 0.003862862940877676, + 0.00716125313192606, + 0.006712098605930805, + 0.004603649489581585, + 0.0052931662648916245, + 0.013355177827179432, + 0.007932400330901146, + 0.003789731999859214, + 0.0042425538413226604, + 0.010641172528266907, + 0.008021637797355652, + 0.004255826119333506, + 0.0048018041998147964, + 0.006576312240213156, + 0.006160279270261526, + 0.0036302884109318256, + 0.0039025305304676294, + 0.005187247879803181, + 0.012486375868320465, + 0.004529713653028011, + 0.004880281165242195 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 4 + ], + "selected_reward": -0.007632143795490265 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a J-20 stealth fighter aircraft?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the aircraft performing a takeoff roll on a rain-slicked runway?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera use a low-angle approach tracking smoothly alongside the accelerating aircraft?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are visible visual effects including water streams on the fuselage, a shimmering heat haze from the engines, and billowing white steam?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene dominated by a monochromatic palette of grays and blacks under an overcast sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006744122598320246 + ], + "opposite_logprob_tokens": [ + -5.006743907928467 + ], + "logprob_reward": -0.006744122598320246, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006744122598320246, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.286723256111145, + 0.2064063400030136, + 0.05083591490983963, + 0.05074040964245796, + 0.037773389369249344, + 0.08453692495822906, + 0.0789845883846283, + 0.05735809728503227, + 0.10949855297803879, + 0.03714250773191452 + ], + "credit_map": [ + 0.025711365044116974, + 0.03215298429131508, + 0.04714912921190262, + 0.04801561310887337, + 0.04118800908327103, + 0.02898229844868183, + 0.039787814021110535, + 0.023736052215099335, + 0.023795519024133682, + 0.026809433475136757, + 0.031912628561258316, + 0.021470094099640846, + 0.02302197553217411, + 0.019693473353981972, + 0.04183677211403847, + 0.017866458743810654, + 0.0057564773596823215, + 0.011420533061027527, + 0.00471031479537487, + 0.007004878483712673, + 0.006859642919152975, + 0.005680735222995281, + 0.004206690471619368, + 0.005196647252887487, + 0.00876539945602417, + 0.00715244933962822, + 0.008353051729500294, + 0.004873867146670818, + 0.007805194240063429, + 0.006629227660596371, + 0.003399394918233156, + 0.0037618286442011595, + 0.0036587021313607693, + 0.006017706822603941, + 0.005917989183217287, + 0.003918083384633064, + 0.0030673672445118427, + 0.003786460030823946, + 0.007686929777264595, + 0.003720153821632266, + 0.013399574905633926, + 0.008798672817647457, + 0.014884237200021744, + 0.015231030993163586, + 0.0075531816110014915, + 0.00840013474225998, + 0.00718337669968605, + 0.009086721576750278, + 0.009026533924043179, + 0.012124961242079735, + 0.014764534309506416, + 0.007073157466948032, + 0.014009903185069561, + 0.010435947217047215, + 0.00697398092597723, + 0.0045755826868116856, + 0.009207292459905148, + 0.008060695603489876, + 0.006159311160445213, + 0.0045417677611112595, + 0.012645707465708256, + 0.009294742718338966, + 0.004462779499590397, + 0.0029858031775802374, + 0.01919614151120186, + 0.017210647463798523, + 0.007138334214687347, + 0.006746190134435892, + 0.024754168465733528, + 0.022109709680080414, + 0.00701073557138443, + 0.005332641769200563, + 0.004797501489520073, + 0.004249788820743561, + 0.003196526551619172, + 0.003302051220089197, + 0.0047078668139874935, + 0.008339589461684227, + 0.0031368855852633715, + 0.0054122996516525745 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 8 + ], + "selected_reward": -0.006744122598320246 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a J-20 stealth fighter aircraft?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the aircraft performing a takeoff roll on a rain-slicked runway?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera use a low-angle approach tracking smoothly alongside the accelerating aircraft?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are visible visual effects including water streams on the fuselage, a shimmering heat haze from the engines, and billowing white steam?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene dominated by a monochromatic palette of grays and blacks under an overcast sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0036151548847556114 + ], + "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": [ + 0.26154932379722595, + 0.12531569600105286, + 0.04615512117743492, + 0.04506648704409599, + 0.08554109185934067, + 0.04393701255321503, + 0.0450018011033535, + 0.16255991160869598, + 0.08194085955619812, + 0.10293275117874146 + ], + "credit_map": [ + 0.02385730855166912, + 0.030726337805390358, + 0.04859275370836258, + 0.03755128011107445, + 0.03679100051522255, + 0.03165002167224884, + 0.034311696887016296, + 0.01806890033185482, + 0.009166311472654343, + 0.020644165575504303, + 0.013425655663013458, + 0.01827489212155342, + 0.012130404822528362, + 0.0106609296053648, + 0.027454877272248268, + 0.01355845108628273, + 0.004720969125628471, + 0.009134550578892231, + 0.005473518744111061, + 0.005772329866886139, + 0.0057825022377073765, + 0.0063657136633992195, + 0.0037392722442746162, + 0.005166261922568083, + 0.006838564295321703, + 0.007638289127498865, + 0.005113473627716303, + 0.004235617350786924, + 0.009922251105308533, + 0.003747600130736828, + 0.0043783774599432945, + 0.0031923060305416584, + 0.011795337311923504, + 0.006395840551704168, + 0.009381681680679321, + 0.02950354665517807, + 0.0038321176543831825, + 0.0055159288458526134, + 0.008666020818054676, + 0.01045061182230711, + 0.005806996487081051, + 0.0051335059106349945, + 0.008453759364783764, + 0.007737495005130768, + 0.003578450530767441, + 0.004562949761748314, + 0.004483435768634081, + 0.004180419258773327, + 0.006119907833635807, + 0.008146986365318298, + 0.008619578555226326, + 0.0044334144331514835, + 0.003930755890905857, + 0.005621060263365507, + 0.004451044835150242, + 0.0036790489684790373, + 0.03787919133901596, + 0.016955509781837463, + 0.01753118634223938, + 0.012078196741640568, + 0.025804318487644196, + 0.025013545528054237, + 0.011542906053364277, + 0.01575505919754505, + 0.010143304243683815, + 0.013502233661711216, + 0.005278394557535648, + 0.0026797261089086533, + 0.022114230319857597, + 0.02123645879328251, + 0.003677185857668519, + 0.0033093199599534273, + 0.013465036638081074, + 0.010522781871259212, + 0.009399414993822575, + 0.007034610956907272, + 0.008276628330349922, + 0.01793968863785267, + 0.01369547937065363, + 0.022599097341299057 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 1 + ], + "selected_reward": -0.0036151548847556114 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "surreal low-angle approaching shot tracks smoothly alongside a J-20 stealth fighter as it roars down a rain-slicked runway with a powerful takeoff roll. The camera maintains a constant speed with the accelerating aircraft, emphasizing its immense velocity. The fighter's dark, angular fuselage, coated in a matte, radar-absorbing material, gleams in the diffuse light of an overcast sky, every rivet, seam, and panel line rendered in crisp, lifelike detail. Water streams across its geometric surfaces and sleek, monolithic canopy. As the aircraft surges forward, its powerful twin engines generate a shimmering, translucent heat haze, violently vaporizing the layer of water on the dark asphalt, creating billowing clouds of white steam that blend with the plumes of spray kicked up by its landing gear. The entire scene, dominated by a sober, monochromatic palette of grays and blacks, creates a tense and breathtaking atmosphere of raw technological power unleashed.", + "seed": 19026, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000056/group_0/index_8869_seed_19026_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": -0.03612936083227396, + "vlm_logprob": -0.03612936083227396, + "vlm_yes_no_margin": 4.075, + "vlm_token_credit": -0.03612936083227396, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a J-20 stealth fighter aircraft?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the aircraft performing a takeoff roll on a rain-slicked runway?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera use a low-angle approach tracking smoothly alongside the accelerating aircraft?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are visible visual effects including water streams on the fuselage, a shimmering heat haze from the engines, and billowing white steam?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene dominated by a monochromatic palette of grays and blacks under an overcast sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.09973616898059845 + ], + "opposite_logprob_tokens": [ + -2.474736213684082 + ], + "logprob_reward": -0.09973616898059845, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.09973616898059845, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2344432771205902, + 0.082657091319561, + 0.23025833070278168, + 0.12726885080337524, + 0.031388767063617706, + 0.13065144419670105, + 0.07452400028705597, + 0.036450378596782684, + 0.017170051112771034, + 0.035187769681215286 + ], + "credit_map": [ + 0.017737576737999916, + 0.02152923308312893, + 0.01650916412472725, + 0.017039036378264427, + 0.023224882781505585, + 0.04613157734274864, + 0.020747531205415726, + 0.07152428478002548, + 0.007397692650556564, + 0.010014740750193596, + 0.006899064872413874, + 0.011194498278200626, + 0.007490676362067461, + 0.012331428937613964, + 0.010427925735712051, + 0.016901059076189995, + 0.03673090413212776, + 0.03423662856221199, + 0.017702287063002586, + 0.02344672568142414, + 0.05619765818119049, + 0.028968404978513718, + 0.01627931371331215, + 0.01669641211628914, + 0.02047738991677761, + 0.012130992487072945, + 0.008731449954211712, + 0.009754087775945663, + 0.019171273335814476, + 0.015709351748228073, + 0.012357768602669239, + 0.028936538845300674, + 0.002860513748601079, + 0.002998304320499301, + 0.002860644832253456, + 0.0021286518312990665, + 0.004732547327876091, + 0.0066072652116417885, + 0.0020521157421171665, + 0.007148723118007183, + 0.011945664882659912, + 0.04548023268580437, + 0.030654579401016235, + 0.009607820771634579, + 0.007334917318075895, + 0.012104220688343048, + 0.007619359064847231, + 0.0059046498499810696, + 0.008730043657124043, + 0.0106557197868824, + 0.007621503435075283, + 0.007871950976550579, + 0.013733543455600739, + 0.011682932265102863, + 0.006462892517447472, + 0.007765408139675856, + 0.004938341677188873, + 0.00444132462143898, + 0.0035479944199323654, + 0.0040278653614223, + 0.003878043731674552, + 0.0031605891417711973, + 0.004985600244253874, + 0.007470620330423117, + 0.0028845469933003187, + 0.0017635347321629524, + 0.002737380564212799, + 0.002252407604828477, + 0.0018852206412702799, + 0.0014327153330668807, + 0.0024715790059417486, + 0.001742666237987578, + 0.004414704162627459, + 0.005731001030653715, + 0.003495200537145138, + 0.003508687252178788, + 0.005389851052314043, + 0.004633259028196335, + 0.00442893011495471, + 0.003586134174838662 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 5 + ], + "selected_reward": -0.09973616898059845 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a J-20 stealth fighter aircraft?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the aircraft performing a takeoff roll on a rain-slicked runway?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera use a low-angle approach tracking smoothly alongside the accelerating aircraft?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are visible visual effects including water streams on the fuselage, a shimmering heat haze from the engines, and billowing white steam?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene dominated by a monochromatic palette of grays and blacks under an overcast sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06207102909684181 + ], + "opposite_logprob_tokens": [ + -2.8120710849761963 + ], + "logprob_reward": -0.06207102909684181, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06207102909684181, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09480994939804077, + 0.10367707908153534, + 0.3917903006076813, + 0.0563976988196373, + 0.1299048513174057, + 0.08278439939022064, + 0.032578207552433014, + 0.034922122955322266, + 0.024308107793331146, + 0.048827264457941055 + ], + "credit_map": [ + 0.00856887549161911, + 0.010920912027359009, + 0.010192690417170525, + 0.00605744868516922, + 0.016359446570277214, + 0.016355356201529503, + 0.006949672009795904, + 0.019405551254749298, + 0.0070613776333630085, + 0.010590317659080029, + 0.008489025756716728, + 0.022804779931902885, + 0.014688133262097836, + 0.013783684000372887, + 0.014216050505638123, + 0.012043706141412258, + 0.06118498370051384, + 0.059962499886751175, + 0.024851549416780472, + 0.032859522849321365, + 0.10629299283027649, + 0.04772835224866867, + 0.024537313729524612, + 0.034373097121715546, + 0.007885508239269257, + 0.0064126853831112385, + 0.0045325662940740585, + 0.003256592433899641, + 0.017980843782424927, + 0.006269855424761772, + 0.004742092918604612, + 0.005317557603120804, + 0.013524757698178291, + 0.016060177236795425, + 0.01098592858761549, + 0.009250566363334656, + 0.01937018893659115, + 0.01970706693828106, + 0.011011053808033466, + 0.02999512106180191, + 0.006693932227790356, + 0.03313843160867691, + 0.005547714419662952, + 0.0052263010293245316, + 0.006758484523743391, + 0.01195122767239809, + 0.004293251316994429, + 0.009175049141049385, + 0.005119276233017445, + 0.005816732067614794, + 0.003783295163884759, + 0.003228123765438795, + 0.004618078004568815, + 0.003670173930004239, + 0.0029152671340852976, + 0.003427262417972088, + 0.004388573579490185, + 0.0054795709438622, + 0.00322520500048995, + 0.004136015195399523, + 0.0048974561505019665, + 0.0049764132127165794, + 0.0038354923017323017, + 0.003983395639806986, + 0.0038346198853105307, + 0.002887033624574542, + 0.0031373617239296436, + 0.0030993365217000246, + 0.003275912720710039, + 0.0019125408725813031, + 0.004001905210316181, + 0.0021593987476080656, + 0.0031337328255176544, + 0.014380215667188168, + 0.0025447234511375427, + 0.0031338436529040337, + 0.014479035511612892, + 0.0037641250528395176, + 0.003911561332643032, + 0.003480028361082077 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 4, + 1 + ], + "selected_reward": -0.06207102909684181 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a J-20 stealth fighter aircraft?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the aircraft performing a takeoff roll on a rain-slicked runway?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera use a low-angle approach tracking smoothly alongside the accelerating aircraft?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are visible visual effects including water streams on the fuselage, a shimmering heat haze from the engines, and billowing white steam?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene dominated by a monochromatic palette of grays and blacks under an overcast sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007634628098458052 + ], + "opposite_logprob_tokens": [ + -4.88263463973999 + ], + "logprob_reward": -0.007634628098458052, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007634628098458052, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2624664008617401, + 0.028429443016648293, + 0.15453115105628967, + 0.07697698473930359, + 0.059304554015398026, + 0.17949894070625305, + 0.07336762547492981, + 0.025114918127655983, + 0.1002165824174881, + 0.040093328803777695 + ], + "credit_map": [ + 0.026428762823343277, + 0.05306286737322807, + 0.024173656478524208, + 0.015721356496214867, + 0.04704170674085617, + 0.04682668298482895, + 0.025373831391334534, + 0.023837508633732796, + 0.0023117430973798037, + 0.0023435503244400024, + 0.002662101062014699, + 0.005272631533443928, + 0.0034769156482070684, + 0.004309894051402807, + 0.002640731865540147, + 0.00541187496855855, + 0.023239275440573692, + 0.024514896795153618, + 0.01120659988373518, + 0.014556389302015305, + 0.03960566222667694, + 0.01852339319884777, + 0.011695672757923603, + 0.011189267039299011, + 0.010989644564688206, + 0.0073375399224460125, + 0.005458219908177853, + 0.003991756122559309, + 0.02367684245109558, + 0.008335966616868973, + 0.0064462353475391865, + 0.01074078306555748, + 0.005157284438610077, + 0.008682656101882458, + 0.004828975535929203, + 0.005377659108489752, + 0.0160512812435627, + 0.00808277539908886, + 0.004661264829337597, + 0.006462662480771542, + 0.015122371725738049, + 0.06567839533090591, + 0.03393852338194847, + 0.010861613787710667, + 0.011403456330299377, + 0.019730519503355026, + 0.01021635252982378, + 0.01254771463572979, + 0.007555743679404259, + 0.01700577326118946, + 0.005777310114353895, + 0.006312662269920111, + 0.013586226850748062, + 0.008887149393558502, + 0.00655191158875823, + 0.007690852973610163, + 0.003594702109694481, + 0.004631782881915569, + 0.002490233164280653, + 0.002610884141176939, + 0.003440418979153037, + 0.002674676477909088, + 0.0025589873548597097, + 0.0031132311560213566, + 0.013567210175096989, + 0.010753188282251358, + 0.013285244815051556, + 0.01970515586435795, + 0.012320357374846935, + 0.008920454420149326, + 0.014276938512921333, + 0.007388032972812653, + 0.004583247005939484, + 0.01070876233279705, + 0.002380805555731058, + 0.0025199425872415304, + 0.010858090594410896, + 0.00493417214602232, + 0.0024283051025122404, + 0.0016800032462924719 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 2 + ], + "selected_reward": -0.007634628098458052 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a J-20 stealth fighter aircraft?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the aircraft performing a takeoff roll on a rain-slicked runway?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera use a low-angle approach tracking smoothly alongside the accelerating aircraft?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are visible visual effects including water streams on the fuselage, a shimmering heat haze from the engines, and billowing white steam?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene dominated by a monochromatic palette of grays and blacks under an overcast sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005953317973762751 + ], + "opposite_logprob_tokens": [ + -5.130953311920166 + ], + "logprob_reward": -0.005953317973762751, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005953317973762751, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.376740962266922, + 0.08312239497900009, + 0.06083425506949425, + 0.08612769842147827, + 0.06474798172712326, + 0.1485825479030609, + 0.045582905411720276, + 0.04804239049553871, + 0.026942379772663116, + 0.059276510030031204 + ], + "credit_map": [ + 0.03498067334294319, + 0.0841379463672638, + 0.026765137910842896, + 0.019107988104224205, + 0.06648235023021698, + 0.07225894927978516, + 0.03829168900847435, + 0.034716248512268066, + 0.005699377506971359, + 0.007945271208882332, + 0.007037927396595478, + 0.014447790570557117, + 0.011283809319138527, + 0.013000610284507275, + 0.006415479816496372, + 0.01729213073849678, + 0.009716547094285488, + 0.007620447780936956, + 0.005054691806435585, + 0.005360268522053957, + 0.010254133492708206, + 0.00946035049855709, + 0.005889388266950846, + 0.0074784280732274055, + 0.011143208481371403, + 0.007557498756796122, + 0.005818080622702837, + 0.005376686807721853, + 0.020538439974188805, + 0.01057936530560255, + 0.0072609917260706425, + 0.017853423953056335, + 0.0068540372885763645, + 0.007207264192402363, + 0.005276376847177744, + 0.005588632542639971, + 0.014466542750597, + 0.007814399898052216, + 0.0048308963887393475, + 0.012709833681583405, + 0.014338832348585129, + 0.022098291665315628, + 0.033958740532398224, + 0.011066251434385777, + 0.011848709546029568, + 0.018288586288690567, + 0.014198715798556805, + 0.022784432396292686, + 0.005807484500110149, + 0.007470513228327036, + 0.004968839697539806, + 0.005550520960241556, + 0.005759755149483681, + 0.0062791830860078335, + 0.004662705585360527, + 0.005083902273327112, + 0.006349785253405571, + 0.007028668653219938, + 0.004898494109511375, + 0.004910585004836321, + 0.005414249841123819, + 0.005539441481232643, + 0.0056395260617136955, + 0.008261638693511486, + 0.002884490182623267, + 0.0025595631450414658, + 0.003935092128813267, + 0.0061948965303599834, + 0.0031035614665597677, + 0.002673044800758362, + 0.003962260205298662, + 0.001629468984901905, + 0.0057550882920622826, + 0.01491598691791296, + 0.0038111077155917883, + 0.0044305468909442425, + 0.014085493981838226, + 0.006388641893863678, + 0.004517866764217615, + 0.005371781997382641 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 3 + ], + "selected_reward": -0.005953317973762751 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a J-20 stealth fighter aircraft?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the aircraft performing a takeoff roll on a rain-slicked runway?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera use a low-angle approach tracking smoothly alongside the accelerating aircraft?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are visible visual effects including water streams on the fuselage, a shimmering heat haze from the engines, and billowing white steam?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene dominated by a monochromatic palette of grays and blacks under an overcast sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005251660011708736 + ], + "opposite_logprob_tokens": [ + -5.255251884460449 + ], + "logprob_reward": -0.005251660011708736, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005251660011708736, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2666798233985901, + 0.14927493035793304, + 0.22684483230113983, + 0.048837870359420776, + 0.07752734422683716, + 0.06787931174039841, + 0.06037179380655289, + 0.038728225976228714, + 0.022138603031635284, + 0.04171723872423172 + ], + "credit_map": [ + 0.025808364152908325, + 0.05073808878660202, + 0.022092672064900398, + 0.015984194353222847, + 0.044039707630872726, + 0.048048924654722214, + 0.02572823502123356, + 0.034239619970321655, + 0.010717730969190598, + 0.012961314059793949, + 0.0115889897570014, + 0.02366754598915577, + 0.015005296096205711, + 0.024221720173954964, + 0.01309189759194851, + 0.03802042081952095, + 0.03351084887981415, + 0.0354643352329731, + 0.014030947349965572, + 0.017634764313697815, + 0.06354334950447083, + 0.026815136894583702, + 0.014058507047593594, + 0.021786902099847794, + 0.005759182386100292, + 0.005835669580847025, + 0.0036584523040801287, + 0.003354489803314209, + 0.0142374811694026, + 0.006433167494833469, + 0.004329686518758535, + 0.005229739472270012, + 0.007939686998724937, + 0.009997150860726833, + 0.004501177929341793, + 0.004798620939254761, + 0.0209492277354002, + 0.010020353831350803, + 0.005380834452807903, + 0.01394028589129448, + 0.006409651134163141, + 0.012548999860882759, + 0.01890716701745987, + 0.007390896324068308, + 0.006280192174017429, + 0.005210369359701872, + 0.00674363411962986, + 0.004388391971588135, + 0.005937321111559868, + 0.01001695729792118, + 0.006839904468506575, + 0.008187012746930122, + 0.011711642146110535, + 0.009413391351699829, + 0.0037233445327728987, + 0.00454222084954381, + 0.005116703920066357, + 0.00763469934463501, + 0.0030385043937712908, + 0.003319776151329279, + 0.006626208312809467, + 0.004105738364160061, + 0.004444706719368696, + 0.004441883414983749, + 0.002442816970869899, + 0.0025809481739997864, + 0.002731172600761056, + 0.0031708306632936, + 0.0036953615490347147, + 0.002021667081862688, + 0.0028405096381902695, + 0.002655292861163616, + 0.006771421525627375, + 0.004711422603577375, + 0.0040095229633152485, + 0.004732181783765554, + 0.0043928371742367744, + 0.006182209122925997, + 0.006456359755247831, + 0.004461281467229128 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 1 + ], + "selected_reward": -0.005251660011708736 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "surreal low-angle approaching shot tracks smoothly alongside a J-20 stealth fighter as it roars down a rain-slicked runway with a powerful takeoff roll. The camera maintains a constant speed with the accelerating aircraft, emphasizing its immense velocity. The fighter's dark, angular fuselage, coated in a matte, radar-absorbing material, gleams in the diffuse light of an overcast sky, every rivet, seam, and panel line rendered in crisp, lifelike detail. Water streams across its geometric surfaces and sleek, monolithic canopy. As the aircraft surges forward, its powerful twin engines generate a shimmering, translucent heat haze, violently vaporizing the layer of water on the dark asphalt, creating billowing clouds of white steam that blend with the plumes of spray kicked up by its landing gear. The entire scene, dominated by a sober, monochromatic palette of grays and blacks, creates a tense and breathtaking atmosphere of raw technological power unleashed.", + "seed": 19027, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000056/group_0/index_8869_seed_19027_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": -0.03905917149968445, + "vlm_logprob": -0.03905917149968445, + "vlm_yes_no_margin": 4.125, + "vlm_token_credit": -0.03905917149968445, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a J-20 stealth fighter aircraft?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the aircraft performing a takeoff roll on a rain-slicked runway?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera use a low-angle approach tracking smoothly alongside the accelerating aircraft?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are visible visual effects including water streams on the fuselage, a shimmering heat haze from the engines, and billowing white steam?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene dominated by a monochromatic palette of grays and blacks under an overcast sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.13423360884189606 + ], + "opposite_logprob_tokens": [ + -2.1342337131500244 + ], + "logprob_reward": -0.13423360884189606, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.13423360884189606, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1523199826478958, + 0.14270661771297455, + 0.11445170640945435, + 0.04513935372233391, + 0.1336953341960907, + 0.22294636070728302, + 0.056203488260507584, + 0.03162267431616783, + 0.05213773995637894, + 0.04877679795026779 + ], + "credit_map": [ + 0.01614045724272728, + 0.023268572986125946, + 0.01925831288099289, + 0.017264103516936302, + 0.022898264229297638, + 0.02561817690730095, + 0.012509740889072418, + 0.015362347476184368, + 0.017503898590803146, + 0.03211477771401405, + 0.01683490350842476, + 0.017087876796722412, + 0.018633702769875526, + 0.01779199205338955, + 0.011602000333368778, + 0.011137464083731174, + 0.011116700246930122, + 0.013886582106351852, + 0.008408395573496819, + 0.011896541342139244, + 0.012264152988791466, + 0.011899858713150024, + 0.028487935662269592, + 0.016491536051034927, + 0.00780832301825285, + 0.007231903728097677, + 0.006581680383533239, + 0.002963944571092725, + 0.009779470041394234, + 0.00379470968618989, + 0.0035696770064532757, + 0.0034096436575055122, + 0.020329412072896957, + 0.027419673278927803, + 0.008556848391890526, + 0.02326660044491291, + 0.014612749218940735, + 0.011874518357217312, + 0.014127704314887524, + 0.013507827185094357, + 0.036668360233306885, + 0.05127200856804848, + 0.012879764661192894, + 0.024230729788541794, + 0.02072330005466938, + 0.024420905858278275, + 0.02483997493982315, + 0.027911314740777016, + 0.00995706021785736, + 0.009444668889045715, + 0.005532726179808378, + 0.007955645211040974, + 0.005728958640247583, + 0.004085829481482506, + 0.007952946238219738, + 0.005545652471482754, + 0.004570671357214451, + 0.003960064146667719, + 0.0030212767887860537, + 0.003130012657493353, + 0.003798103192821145, + 0.003612611209973693, + 0.004030799958854914, + 0.0054991357028484344, + 0.005487381014972925, + 0.006275922991335392, + 0.00719092832878232, + 0.006484551355242729, + 0.007191593758761883, + 0.005442367400974035, + 0.006564851850271225, + 0.007500142324715853, + 0.0054808082059025764, + 0.007272268179804087, + 0.005927762016654015, + 0.004903585184365511, + 0.005156983155757189, + 0.007413842715322971, + 0.0075660464353859425, + 0.005055501591414213 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 1 + ], + "selected_reward": -0.13423360884189606 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a J-20 stealth fighter aircraft?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the aircraft performing a takeoff roll on a rain-slicked runway?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera use a low-angle approach tracking smoothly alongside the accelerating aircraft?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are visible visual effects including water streams on the fuselage, a shimmering heat haze from the engines, and billowing white steam?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene dominated by a monochromatic palette of grays and blacks under an overcast sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.043074361979961395 + ], + "opposite_logprob_tokens": [ + -3.168074369430542 + ], + "logprob_reward": -0.043074361979961395, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.043074361979961395, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08715324103832245, + 0.4261547923088074, + 0.0923619493842125, + 0.08504755049943924, + 0.06519944965839386, + 0.047201987355947495, + 0.10673028230667114, + 0.03138819709420204, + 0.04132330045104027, + 0.017439255490899086 + ], + "credit_map": [ + 0.00833925511687994, + 0.016512194648385048, + 0.01066578272730112, + 0.012530100531876087, + 0.007255976554006338, + 0.011595167219638824, + 0.00974519643932581, + 0.010509563609957695, + 0.04707099497318268, + 0.10278119891881943, + 0.05022265389561653, + 0.0409238375723362, + 0.0562271885573864, + 0.05220159515738487, + 0.05262032896280289, + 0.02410699799656868, + 0.009553736075758934, + 0.010317175649106503, + 0.008853791281580925, + 0.010966427624225616, + 0.011809415183961391, + 0.009936174377799034, + 0.02460896223783493, + 0.0063162632286548615, + 0.014349986799061298, + 0.01285118144005537, + 0.011483908630907536, + 0.0062136719934642315, + 0.01992734707891941, + 0.008026078343391418, + 0.005790822673588991, + 0.006404551677405834, + 0.011259193532168865, + 0.014377846382558346, + 0.004257206339389086, + 0.004835961386561394, + 0.008099338039755821, + 0.005526969209313393, + 0.012541258707642555, + 0.0043016765266656876, + 0.007122730370610952, + 0.008540361188352108, + 0.006233834661543369, + 0.005800031125545502, + 0.004510146100074053, + 0.004546786658465862, + 0.004866933450102806, + 0.005581161007285118, + 0.016226164996623993, + 0.022167880088090897, + 0.011345128528773785, + 0.007033670786768198, + 0.018142985180020332, + 0.010069821029901505, + 0.013454077765345573, + 0.008290553465485573, + 0.0031009595841169357, + 0.002739820396527648, + 0.005068827886134386, + 0.003781678155064583, + 0.0037956160958856344, + 0.004271084908396006, + 0.0030540009029209614, + 0.0055762059055268764, + 0.004605898633599281, + 0.0046656434424221516, + 0.006583776790648699, + 0.006720823235809803, + 0.004598659463226795, + 0.004058578982949257, + 0.004798286594450474, + 0.005291634239256382, + 0.002244784263893962, + 0.003011794527992606, + 0.0021654891315847635, + 0.0016788857756182551, + 0.002198928501456976, + 0.002438501687720418, + 0.0021385399159044027, + 0.0015623304061591625 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 6, + 2 + ], + "selected_reward": -0.043074361979961395 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a J-20 stealth fighter aircraft?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the aircraft performing a takeoff roll on a rain-slicked runway?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera use a low-angle approach tracking smoothly alongside the accelerating aircraft?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are visible visual effects including water streams on the fuselage, a shimmering heat haze from the engines, and billowing white steam?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene dominated by a monochromatic palette of grays and blacks under an overcast sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0036190745886415243 + ], + "opposite_logprob_tokens": [ + -5.628619194030762 + ], + "logprob_reward": -0.0036190745886415243, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.0036190745886415243, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17746159434318542, + 0.08734296262264252, + 0.12077482789754868, + 0.07217114418745041, + 0.14837853610515594, + 0.06656461954116821, + 0.0725085586309433, + 0.05084916576743126, + 0.14096671342849731, + 0.06298191845417023 + ], + "credit_map": [ + 0.01743953302502632, + 0.024313615635037422, + 0.014205245301127434, + 0.017411647364497185, + 0.025211045518517494, + 0.02655804716050625, + 0.01709124632179737, + 0.03523121029138565, + 0.010015376843512058, + 0.01655985601246357, + 0.011204706504940987, + 0.009628525003790855, + 0.011381646618247032, + 0.00985474232584238, + 0.007983372546732426, + 0.010714733973145485, + 0.01626726984977722, + 0.015334445051848888, + 0.014956789091229439, + 0.013931374996900558, + 0.017425527796149254, + 0.012317085638642311, + 0.019049126654863358, + 0.011493206955492496, + 0.012394324876368046, + 0.00958882924169302, + 0.009511967189610004, + 0.005456596612930298, + 0.011000722646713257, + 0.010945877060294151, + 0.005795217584818602, + 0.007477602455765009, + 0.02185726910829544, + 0.015186882577836514, + 0.008468921296298504, + 0.023207176476716995, + 0.017026560381054878, + 0.010358685627579689, + 0.038590386509895325, + 0.013682653196156025, + 0.011532958596944809, + 0.012638294138014317, + 0.008961907587945461, + 0.008311695419251919, + 0.005975893698632717, + 0.005492125637829304, + 0.006886598654091358, + 0.006765149999409914, + 0.009972505271434784, + 0.01270940899848938, + 0.0077039762400090694, + 0.005901265423744917, + 0.01216396875679493, + 0.007754598744213581, + 0.0101290512830019, + 0.006173781119287014, + 0.010731372982263565, + 0.007872024551033974, + 0.00471745990216732, + 0.004629326518625021, + 0.005700820125639439, + 0.005982713773846626, + 0.004613501485437155, + 0.006601948756724596, + 0.016632825136184692, + 0.016468120738863945, + 0.023760709911584854, + 0.017012011259794235, + 0.015847984701395035, + 0.015166866593062878, + 0.015212236903607845, + 0.02086595445871353, + 0.006071899551898241, + 0.009471707046031952, + 0.00633231783285737, + 0.00673217186704278, + 0.006297856569290161, + 0.008157770149409771, + 0.013084786012768745, + 0.0068334066309034824 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 8 + ], + "selected_reward": -0.0036190745886415243 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a J-20 stealth fighter aircraft?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the aircraft performing a takeoff roll on a rain-slicked runway?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera use a low-angle approach tracking smoothly alongside the accelerating aircraft?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are visible visual effects including water streams on the fuselage, a shimmering heat haze from the engines, and billowing white steam?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene dominated by a monochromatic palette of grays and blacks under an overcast sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006742938421666622 + ], + "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": [ + 0.28830620646476746, + 0.08505847305059433, + 0.0872775986790657, + 0.15618950128555298, + 0.07896632701158524, + 0.09363451600074768, + 0.0534309484064579, + 0.06596385687589645, + 0.0544583760201931, + 0.036714185029268265 + ], + "credit_map": [ + 0.023781396448612213, + 0.03808678686618805, + 0.029358621686697006, + 0.03511707857251167, + 0.04123423993587494, + 0.04171592742204666, + 0.03658156096935272, + 0.042430583387613297, + 0.009385817684233189, + 0.01569059118628502, + 0.011756635271012783, + 0.010018308646976948, + 0.00931988749653101, + 0.009780515916645527, + 0.011013639159500599, + 0.008093079552054405, + 0.012731818482279778, + 0.014702741988003254, + 0.011100292205810547, + 0.008268777281045914, + 0.01185450330376625, + 0.009274430572986603, + 0.010398547165095806, + 0.008946487680077553, + 0.031173793599009514, + 0.031993672251701355, + 0.01911807991564274, + 0.009346106089651585, + 0.029354272410273552, + 0.01869572140276432, + 0.007306021638214588, + 0.009201832115650177, + 0.012680422514677048, + 0.014448353089392185, + 0.004033312201499939, + 0.007359156385064125, + 0.01125691644847393, + 0.009103000164031982, + 0.012851246632635593, + 0.007233921438455582, + 0.015803463757038116, + 0.018911637365818024, + 0.010117044672369957, + 0.009394383989274502, + 0.01085361372679472, + 0.0110396733507514, + 0.007814256474375725, + 0.00970043707638979, + 0.009397398680448532, + 0.009076268412172794, + 0.00582340219989419, + 0.004703539423644543, + 0.008560056798160076, + 0.005354590248316526, + 0.004898081533610821, + 0.005617610644549131, + 0.011761651374399662, + 0.009426013566553593, + 0.006490921601653099, + 0.007463520858436823, + 0.0056923758238554, + 0.006022756919264793, + 0.00811329297721386, + 0.010993321426212788, + 0.006190514657646418, + 0.00594271719455719, + 0.0077014705166220665, + 0.0067347572185099125, + 0.008502986282110214, + 0.006375915836542845, + 0.005815449170768261, + 0.007194566540420055, + 0.003949865698814392, + 0.006168147549033165, + 0.0037590316496789455, + 0.004084229469299316, + 0.0045911530032753944, + 0.005245033651590347, + 0.005357585847377777, + 0.0035591369960457087 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 5 + ], + "selected_reward": -0.006742938421666622 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a J-20 stealth fighter aircraft?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the aircraft performing a takeoff roll on a rain-slicked runway?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera use a low-angle approach tracking smoothly alongside the accelerating aircraft?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are visible visual effects including water streams on the fuselage, a shimmering heat haze from the engines, and billowing white steam?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene dominated by a monochromatic palette of grays and blacks under an overcast sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007625873666256666 + ], + "opposite_logprob_tokens": [ + -4.882626056671143 + ], + "logprob_reward": -0.007625873666256666, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007625873666256666, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13706772029399872, + 0.13481304049491882, + 0.05671070143580437, + 0.02578636258840561, + 0.3905508816242218, + 0.08134433627128601, + 0.08062952011823654, + 0.034923046827316284, + 0.021498609334230423, + 0.036675747483968735 + ], + "credit_map": [ + 0.013570629991590977, + 0.027530740946531296, + 0.011972651816904545, + 0.01913200132548809, + 0.011589224450290203, + 0.015589077956974506, + 0.015046280808746815, + 0.022637123242020607, + 0.016791492700576782, + 0.02732902579009533, + 0.017977293580770493, + 0.01305350661277771, + 0.015587669797241688, + 0.016110606491565704, + 0.01907758228480816, + 0.008885866962373257, + 0.0071134972386062145, + 0.00873815268278122, + 0.00760607747361064, + 0.0053732795640826225, + 0.007954137399792671, + 0.005811120383441448, + 0.007177678402513266, + 0.0069367541000247, + 0.0041979593224823475, + 0.004855609964579344, + 0.003034195862710476, + 0.0024455622769892216, + 0.0039698476903140545, + 0.002537688473239541, + 0.002211853163316846, + 0.002533646998926997, + 0.05708731710910797, + 0.0677008181810379, + 0.02806202881038189, + 0.09025580435991287, + 0.03819070756435394, + 0.03156968206167221, + 0.03776724264025688, + 0.039917267858982086, + 0.013158915564417839, + 0.015924330800771713, + 0.0059785000048577785, + 0.007015890907496214, + 0.01148010604083538, + 0.011071115732192993, + 0.007599315606057644, + 0.009116161614656448, + 0.010340798646211624, + 0.01519701536744833, + 0.009264077991247177, + 0.0070743015967309475, + 0.013332909904420376, + 0.007925176993012428, + 0.011058615520596504, + 0.006436623632907867, + 0.006561203394085169, + 0.0048569669015705585, + 0.002425354439765215, + 0.003294575260952115, + 0.0038406657986342907, + 0.0038098131772130728, + 0.004233971703797579, + 0.005900495685636997, + 0.0026315529830753803, + 0.0027043786831200123, + 0.0024091405794024467, + 0.00309185404330492, + 0.0032308241352438927, + 0.0021491723600775003, + 0.0022947241086512804, + 0.0029869626741856337, + 0.003247418673709035, + 0.003918683156371117, + 0.0036022381391376257, + 0.004589495714753866, + 0.0033993700053542852, + 0.004902040120214224, + 0.00896819494664669, + 0.0040483055636286736 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 1 + ], + "selected_reward": -0.007625873666256666 + } + ], + "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/0000056/group_1/index_tensor([3297, 3297])_group_1_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000056/group_1/index_tensor([3297, 3297])_group_1_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..b1840bb667dc2c4f516cbf31c35cc146447fcc55 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000056/group_1/index_tensor([3297, 3297])_group_1_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "static, tight close-up focuses on a pair of manicured hands making the final, precise adjustments to the knot of a tie. The tie, a deep charcoal silk with a subtle satin sheen, is tied in a perfect Windsor knot atop the solid white collar of a crisp, high-line shirt. The thumb and index finger of the right hand gently pinch the top of the knot, while the left hand deliberately slides it upward, securing it firmly in place. With practiced, unhurried confidence, the fingers meticulously create and perfect a single, deep dimple in the silk beneath the knot. The rich texture of a dark, fine wool suit jacket, its sharp lapels clearly visible, fills the frame. The scene is bathed in warm, directional light from a single, off-camera source, casting soft highlights on the joints and uneven fabric while deepening the shadows, creating an intimate atmosphere of focused focus and quiet, formal preparation.", + "seed": 25005, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000056/group_1/index_3297_seed_25005_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": -0.15703381709754466, + "vlm_logprob": -0.15703381709754466, + "vlm_yes_no_margin": 3.025, + "vlm_token_credit": -0.15703381709754466, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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, tight close-up focusing on a pair of hands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tie made of deep charcoal silk with a satin sheen and tied in a Windsor knot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the hands performing the specific actions of pinching the knot with the right hand and sliding it upward with the left hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a deep dimple created in the silk beneath the knot by the fingers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by warm, directional light from a single off-camera source casting soft highlights and deep shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02131553925573826 + ], + "opposite_logprob_tokens": [ + -5.271315574645996 + ], + "logprob_reward": -0.02131553925573826, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.02131553925573826, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22059297561645508, + 0.09315139800310135, + 0.08591555058956146, + 0.10725823044776917, + 0.043602921068668365, + 0.07601020485162735, + 0.10193969309329987, + 0.15032094717025757, + 0.04442540183663368, + 0.0767826959490776 + ], + "credit_map": [ + 0.022508244961500168, + 0.021299049258232117, + 0.03584310784935951, + 0.020938346162438393, + 0.041559867560863495, + 0.01621108688414097, + 0.026042785495519638, + 0.036190468817949295, + 0.007441973779350519, + 0.008063209243118763, + 0.01693536899983883, + 0.014605240896344185, + 0.013375578448176384, + 0.007312432862818241, + 0.012012169696390629, + 0.013405420817434788, + 0.012680302374064922, + 0.014708481729030609, + 0.013834672048687935, + 0.009693932719528675, + 0.011953383684158325, + 0.004991667810827494, + 0.008350507356226444, + 0.009702591225504875, + 0.013587082736194134, + 0.01526732463389635, + 0.018071288242936134, + 0.013680631294846535, + 0.0166463665664196, + 0.006689031608402729, + 0.010551651008427143, + 0.01276483479887247, + 0.007813362404704094, + 0.004789792001247406, + 0.005026980768889189, + 0.003968632314354181, + 0.00867566466331482, + 0.003168330993503332, + 0.0047829290851950645, + 0.005377223715186119, + 0.00650746887549758, + 0.007174832746386528, + 0.011652467772364616, + 0.012377951294183731, + 0.00845326203852892, + 0.005705077201128006, + 0.014184454455971718, + 0.009954680688679218, + 0.011839606799185276, + 0.014481976628303528, + 0.009580465033650398, + 0.0073487162590026855, + 0.02256632409989834, + 0.009794372133910656, + 0.012891296297311783, + 0.013436932116746902, + 0.02734912373125553, + 0.01116883009672165, + 0.04836238920688629, + 0.017325684428215027, + 0.013654373586177826, + 0.005653558298945427, + 0.012641898356378078, + 0.01416507177054882, + 0.00629805913195014, + 0.0039519877173006535, + 0.006455633789300919, + 0.0064492118544876575, + 0.007206408306956291, + 0.0025773567613214254, + 0.005765154026448727, + 0.005721587222069502, + 0.010637452825903893, + 0.015343936160206795, + 0.008945140056312084, + 0.010053456760942936, + 0.007976861670613289, + 0.0065264166332781315, + 0.008132935501635075, + 0.009166485629975796 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 3 + ], + "selected_reward": -0.02131553925573826 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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, tight close-up focusing on a pair of hands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tie made of deep charcoal silk with a satin sheen and tied in a Windsor knot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the hands performing the specific actions of pinching the knot with the right hand and sliding it upward with the left hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a deep dimple created in the silk beneath the knot by the fingers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by warm, directional light from a single off-camera source casting soft highlights and deep shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.577013373374939 + ], + "opposite_logprob_tokens": [ + -0.827013373374939 + ], + "logprob_reward": -0.577013373374939, + "yes_no_margin_reward": 0.25, + "token_credit_reward": -0.577013373374939, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.29930439591407776, + 0.15120291709899902, + 0.09424861520528793, + 0.12079666554927826, + 0.023413153365254402, + 0.028880663216114044, + 0.11959241330623627, + 0.06445135921239853, + 0.04455671086907387, + 0.05355307087302208 + ], + "credit_map": [ + 0.02915925160050392, + 0.027441661804914474, + 0.04745667427778244, + 0.025096090510487556, + 0.0653156116604805, + 0.022401509806513786, + 0.03354606032371521, + 0.04888753592967987, + 0.011379913426935673, + 0.01645912602543831, + 0.021455522626638412, + 0.014807292260229588, + 0.029680103063583374, + 0.014793206006288528, + 0.0183760616928339, + 0.024251695722341537, + 0.013696066103875637, + 0.008767849765717983, + 0.0135369086638093, + 0.008108659647405148, + 0.015478874556720257, + 0.00831814669072628, + 0.013274666853249073, + 0.013067435473203659, + 0.017489193007349968, + 0.012468090280890465, + 0.02817169763147831, + 0.0104034673422575, + 0.02199537493288517, + 0.0059113884344697, + 0.011087454855442047, + 0.013270006515085697, + 0.0029373210854828358, + 0.003228128422051668, + 0.003439203603193164, + 0.002887256909161806, + 0.003478093072772026, + 0.001970315584912896, + 0.002399996155872941, + 0.0030728373676538467, + 0.0030799233354628086, + 0.002376057906076312, + 0.003580146236345172, + 0.005188691429793835, + 0.0028672553598880768, + 0.0022284353617578745, + 0.006081948522478342, + 0.003478206694126129, + 0.012661932967603207, + 0.012761359103024006, + 0.011723256669938564, + 0.010962085798382759, + 0.029096269980072975, + 0.00891218613833189, + 0.019748415797948837, + 0.013726911507546902, + 0.012419980950653553, + 0.004955675452947617, + 0.020731300115585327, + 0.006940276362001896, + 0.006191721651703119, + 0.00268815946765244, + 0.005650151986628771, + 0.004874091595411301, + 0.006990992464125156, + 0.005207487381994724, + 0.004681802354753017, + 0.003373365616425872, + 0.01020012702792883, + 0.0031191264279186726, + 0.006709784734994173, + 0.0042740292847156525, + 0.00731267873197794, + 0.008872617967426777, + 0.007342389784753323, + 0.0054347156547009945, + 0.005597397685050964, + 0.004877694416791201, + 0.005692524369806051, + 0.008423049002885818 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 3 + ], + "selected_reward": -0.577013373374939 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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, tight close-up focusing on a pair of hands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tie made of deep charcoal silk with a satin sheen and tied in a Windsor knot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the hands performing the specific actions of pinching the knot with the right hand and sliding it upward with the left hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a deep dimple created in the silk beneath the knot by the fingers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by warm, directional light from a single off-camera source casting soft highlights and deep shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029807114973664284 + ], + "opposite_logprob_tokens": [ + -3.5298070907592773 + ], + "logprob_reward": -0.029807114973664284, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029807114973664284, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.30082929134368896, + 0.15809236466884613, + 0.11710059642791748, + 0.06184515357017517, + 0.045096877962350845, + 0.05503392592072487, + 0.08016715198755264, + 0.07869384437799454, + 0.06129883974790573, + 0.04184191673994064 + ], + "credit_map": [ + 0.033587053418159485, + 0.02879038080573082, + 0.05757014825940132, + 0.0262734517455101, + 0.05179442837834358, + 0.02301567606627941, + 0.030917897820472717, + 0.048880260437726974, + 0.01222932431846857, + 0.019862407818436623, + 0.025910133495926857, + 0.012336558662354946, + 0.0316561795771122, + 0.012792973779141903, + 0.020349029451608658, + 0.0229557603597641, + 0.017104709520936012, + 0.011147982440888882, + 0.016896797344088554, + 0.010828482918441296, + 0.014740822836756706, + 0.010626550763845444, + 0.012617383152246475, + 0.02313787303864956, + 0.007991219870746136, + 0.008612524718046188, + 0.008910378441214561, + 0.006954994052648544, + 0.01207411102950573, + 0.0038885660469532013, + 0.006165155675262213, + 0.007248201873153448, + 0.006548407021909952, + 0.004983618855476379, + 0.0061210026033222675, + 0.004610970616340637, + 0.0067051611840724945, + 0.0035573432687669992, + 0.004919544793665409, + 0.007650825660675764, + 0.004261549096554518, + 0.008604020811617374, + 0.008997486904263496, + 0.006086185574531555, + 0.0071059903129935265, + 0.0039057221729308367, + 0.00813432689756155, + 0.007938644848763943, + 0.00835388619452715, + 0.011776657775044441, + 0.006857431493699551, + 0.00610870448872447, + 0.018905501812696457, + 0.007061248179525137, + 0.011293317191302776, + 0.009810403920710087, + 0.01448510680347681, + 0.0069425986148417, + 0.022629057988524437, + 0.011123355478048325, + 0.00528419204056263, + 0.0026252882089465857, + 0.008368109352886677, + 0.007236135192215443, + 0.00845412164926529, + 0.007535516284406185, + 0.006485200487077236, + 0.006518694572150707, + 0.015897968783974648, + 0.0032789648976176977, + 0.00762570695951581, + 0.005502671003341675, + 0.006105844862759113, + 0.00499518821015954, + 0.006719940807670355, + 0.004887117072939873, + 0.0037365411408245564, + 0.004323891364037991, + 0.005376644432544708, + 0.005696750245988369 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.029807114973664284 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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, tight close-up focusing on a pair of hands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tie made of deep charcoal silk with a satin sheen and tied in a Windsor knot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the hands performing the specific actions of pinching the knot with the right hand and sliding it upward with the left hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a deep dimple created in the silk beneath the knot by the fingers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by warm, directional light from a single off-camera source casting soft highlights and deep shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.14283011853694916 + ], + "opposite_logprob_tokens": [ + -2.0178301334381104 + ], + "logprob_reward": -0.14283011853694916, + "yes_no_margin_reward": 1.875, + "token_credit_reward": -0.14283011853694916, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.28267282247543335, + 0.20176801085472107, + 0.1037120521068573, + 0.08196187019348145, + 0.02328849770128727, + 0.022401636466383934, + 0.12129975110292435, + 0.055177923291921616, + 0.07004369795322418, + 0.037673771381378174 + ], + "credit_map": [ + 0.027245383709669113, + 0.025418788194656372, + 0.04016429930925369, + 0.022941231727600098, + 0.05936434492468834, + 0.022579187527298927, + 0.034837573766708374, + 0.05012201517820358, + 0.015251955948770046, + 0.020537935197353363, + 0.03146963566541672, + 0.01836402341723442, + 0.037053026258945465, + 0.019316524267196655, + 0.025658346712589264, + 0.034116558730602264, + 0.016504811123013496, + 0.009993457235395908, + 0.0159841850399971, + 0.008747413754463196, + 0.013914459384977818, + 0.008037940599024296, + 0.012467596679925919, + 0.01806218922138214, + 0.011437802575528622, + 0.009294806979596615, + 0.009154660627245903, + 0.007512669079005718, + 0.021087566390633583, + 0.005407234653830528, + 0.00991279911249876, + 0.008154327981173992, + 0.002527636708691716, + 0.003665440483018756, + 0.0031181194353848696, + 0.0023671274539083242, + 0.004909627605229616, + 0.0018961941823363304, + 0.0018816373776644468, + 0.0029227135237306356, + 0.0030051579233258963, + 0.003022859338670969, + 0.0029338046442717314, + 0.0024012774229049683, + 0.0026068976148962975, + 0.001675706240348518, + 0.0033572642132639885, + 0.003398669185116887, + 0.013167144730687141, + 0.014659746550023556, + 0.011032680049538612, + 0.010223031044006348, + 0.03065885417163372, + 0.010138731449842453, + 0.018488766625523567, + 0.012930789962410927, + 0.012657595798373222, + 0.004917595535516739, + 0.01737072877585888, + 0.004723553545773029, + 0.005661545321345329, + 0.0020152840297669172, + 0.00469578942283988, + 0.0031358301639556885, + 0.010541108436882496, + 0.007118256762623787, + 0.007278993260115385, + 0.007817700505256653, + 0.017569513991475105, + 0.003891425207257271, + 0.009829285554587841, + 0.005997415166348219, + 0.005862036254256964, + 0.006407762877643108, + 0.004925795365124941, + 0.0036438670940697193, + 0.004245220683515072, + 0.0034887127112597227, + 0.005014636553823948, + 0.00408574054017663 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 6 + ], + "selected_reward": -0.14283011853694916 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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, tight close-up focusing on a pair of hands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tie made of deep charcoal silk with a satin sheen and tied in a Windsor knot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the hands performing the specific actions of pinching the knot with the right hand and sliding it upward with the left hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a deep dimple created in the silk beneath the knot by the fingers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by warm, directional light from a single off-camera source casting soft highlights and deep shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014202939346432686 + ], + "opposite_logprob_tokens": [ + -4.264203071594238 + ], + "logprob_reward": -0.014202939346432686, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014202939346432686, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17459191381931305, + 0.07327742874622345, + 0.05877165123820305, + 0.288279265165329, + 0.08609673380851746, + 0.022050516679883003, + 0.04969991371035576, + 0.033865928649902344, + 0.02936684712767601, + 0.18399980664253235 + ], + "credit_map": [ + 0.018868155777454376, + 0.012263058684766293, + 0.040990594774484634, + 0.015893783420324326, + 0.024594945833086967, + 0.013180801644921303, + 0.018662773072719574, + 0.030137816444039345, + 0.0058676148764789104, + 0.006116051226854324, + 0.012369737029075623, + 0.010773038491606712, + 0.011148972436785698, + 0.0065660481341183186, + 0.008787943050265312, + 0.01164802722632885, + 0.009190027602016926, + 0.0067655364982783794, + 0.007136335130780935, + 0.007299675606191158, + 0.00932403840124607, + 0.004220352042466402, + 0.006024648901075125, + 0.0088110426440835, + 0.04787954315543175, + 0.03225313127040863, + 0.07152868807315826, + 0.03325981646776199, + 0.03637493774294853, + 0.014387675561010838, + 0.021128512918949127, + 0.03146695718169212, + 0.011505089700222015, + 0.00982804223895073, + 0.020830055698752403, + 0.010333206504583359, + 0.006079233717173338, + 0.007500851526856422, + 0.007017649244517088, + 0.013002600520849228, + 0.004070281982421875, + 0.0030777561478316784, + 0.003143284237012267, + 0.0023825655225664377, + 0.001763134729117155, + 0.0014797470066696405, + 0.0027179555036127567, + 0.003415794111788273, + 0.005132066085934639, + 0.005941214505583048, + 0.00705397455021739, + 0.004541180096566677, + 0.010277067311108112, + 0.004317660350352526, + 0.005859874654561281, + 0.006576879415661097, + 0.005775979720056057, + 0.002084574196487665, + 0.012612514197826385, + 0.001955744344741106, + 0.0028254676144570112, + 0.0010216482914984226, + 0.003274180693551898, + 0.004315821919590235, + 0.004965443164110184, + 0.0029037653002887964, + 0.003335246816277504, + 0.0027428928297013044, + 0.007280067075043917, + 0.001715756836347282, + 0.0035971456672996283, + 0.0028265330474823713, + 0.023791030049324036, + 0.03778180107474327, + 0.018864763900637627, + 0.022165773436427116, + 0.017910877242684364, + 0.017021944746375084, + 0.021655719727277756, + 0.02480790577828884 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 9, + 0 + ], + "selected_reward": -0.014202939346432686 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, tight close-up focuses on a pair of manicured hands making the final, precise adjustments to the knot of a tie. The tie, a deep charcoal silk with a subtle satin sheen, is tied in a perfect Windsor knot atop the solid white collar of a crisp, high-line shirt. The thumb and index finger of the right hand gently pinch the top of the knot, while the left hand deliberately slides it upward, securing it firmly in place. With practiced, unhurried confidence, the fingers meticulously create and perfect a single, deep dimple in the silk beneath the knot. The rich texture of a dark, fine wool suit jacket, its sharp lapels clearly visible, fills the frame. The scene is bathed in warm, directional light from a single, off-camera source, casting soft highlights on the joints and uneven fabric while deepening the shadows, creating an intimate atmosphere of focused focus and quiet, formal preparation.", + "seed": 25006, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000056/group_1/index_3297_seed_25006_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": -0.10167061612010002, + "vlm_logprob": -0.10167061612010002, + "vlm_yes_no_margin": 3.375, + "vlm_token_credit": -0.10167061612010002, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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, tight close-up focusing on a pair of hands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tie made of deep charcoal silk with a satin sheen and tied in a Windsor knot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the hands performing the specific actions of pinching the knot with the right hand and sliding it upward with the left hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a deep dimple created in the silk beneath the knot by the fingers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by warm, directional light from a single off-camera source casting soft highlights and deep shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01616314984858036 + ], + "opposite_logprob_tokens": [ + -5.641163349151611 + ], + "logprob_reward": -0.01616314984858036, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.01616314984858036, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06625047326087952, + 0.3176591694355011, + 0.01223740354180336, + 0.014330350793898106, + 0.012793456204235554, + 0.5061842799186707, + 0.007211681455373764, + 0.0102127930149436, + 0.01026259083300829, + 0.0428578220307827 + ], + "credit_map": [ + 0.011580569669604301, + 0.0070077539421617985, + 0.010590826161205769, + 0.007842861115932465, + 0.008876973763108253, + 0.00624843081459403, + 0.005528372246772051, + 0.00857468880712986, + 0.00932164303958416, + 0.009805472567677498, + 0.019002974033355713, + 0.015346519649028778, + 0.009198848158121109, + 0.11680257320404053, + 0.007860573939979076, + 0.13032056391239166, + 0.0014220785815268755, + 0.0012427206384018064, + 0.002337678102776408, + 0.0011347507825121284, + 0.0013548436108976603, + 0.0011326521635055542, + 0.0021318804938346148, + 0.0014807989355176687, + 0.0014887675642967224, + 0.0016351781086996198, + 0.0015530979726463556, + 0.0011837767669931054, + 0.0019641132093966007, + 0.0016018578317016363, + 0.003009252017363906, + 0.0018943084869533777, + 0.000687768217176199, + 0.0008465602877549827, + 0.0009835358941927552, + 0.0013044964289292693, + 0.0012957919389009476, + 0.003176948521286249, + 0.0007918588235042989, + 0.003706495277583599, + 0.022353926673531532, + 0.010507332161068916, + 0.029687484726309776, + 0.00825245026499033, + 0.010517257265746593, + 0.19839170575141907, + 0.01209153514355421, + 0.2143825739622116, + 0.0006531146937049925, + 0.0009472166420891881, + 0.001267762971110642, + 0.000915879150852561, + 0.0007749205688014627, + 0.0008566626347601414, + 0.0009930302621796727, + 0.000803094357252121, + 0.0008418567595072091, + 0.0009748081211000681, + 0.0015513940015807748, + 0.0012255535693839192, + 0.0016073222504928708, + 0.001680551329627633, + 0.0012983669294044375, + 0.0010329410433769226, + 0.0009697572095319629, + 0.0014225505292415619, + 0.0012140420731157064, + 0.0016836883733049035, + 0.0012242995435371995, + 0.0013241558335721493, + 0.0011948638129979372, + 0.0012292328756302595, + 0.0036313871387392282, + 0.00406757602468133, + 0.004256119951605797, + 0.008387506939470768, + 0.006735570263117552, + 0.003379873465746641, + 0.007059311959892511, + 0.005340478848665953 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 1, + 0 + ], + "selected_reward": -0.01616314984858036 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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, tight close-up focusing on a pair of hands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tie made of deep charcoal silk with a satin sheen and tied in a Windsor knot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the hands performing the specific actions of pinching the knot with the right hand and sliding it upward with the left hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a deep dimple created in the silk beneath the knot by the fingers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by warm, directional light from a single off-camera source casting soft highlights and deep shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.3877088129520416 + ], + "opposite_logprob_tokens": [ + -1.1377087831497192 + ], + "logprob_reward": -0.3877088129520416, + "yes_no_margin_reward": 0.75, + "token_credit_reward": -0.3877088129520416, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.027752086520195007, + 0.4490036964416504, + 0.012633212842047215, + 0.007952030748128891, + 0.014233391731977463, + 0.406613290309906, + 0.0053780595771968365, + 0.009460684843361378, + 0.006735874805599451, + 0.06023764610290527 + ], + "credit_map": [ + 0.0033791344612836838, + 0.002613076940178871, + 0.002989518688991666, + 0.003826214000582695, + 0.0049162497743964195, + 0.0029827351681888103, + 0.0035579903051257133, + 0.003487164154648781, + 0.011908975429832935, + 0.013482109643518925, + 0.026752976700663567, + 0.021812593564391136, + 0.014610100537538528, + 0.16493543982505798, + 0.01113407127559185, + 0.18436740338802338, + 0.0010763920145109296, + 0.0013708735350519419, + 0.0012697952333837748, + 0.0018943845061585307, + 0.001844488549977541, + 0.001977406907826662, + 0.001230144058354199, + 0.00196972768753767, + 0.00048779137432575226, + 0.0008079675026237965, + 0.0009488484356552362, + 0.0007744082249701023, + 0.0014191805385053158, + 0.0012394716031849384, + 0.001331329345703125, + 0.0009430327918380499, + 0.0009094833512790501, + 0.0014232700923457742, + 0.0012413235381245613, + 0.0014664465561509132, + 0.001709572970867157, + 0.002837856998667121, + 0.001313514425419271, + 0.003331922460347414, + 0.01797940954566002, + 0.008362545631825924, + 0.023687714710831642, + 0.006579614244401455, + 0.008444772101938725, + 0.15967854857444763, + 0.009723931550979614, + 0.1721567064523697, + 0.0003206255496479571, + 0.0007440162589773536, + 0.0005378745845519006, + 0.000709637941326946, + 0.0005135645042173564, + 0.0008407821878790855, + 0.0007774876430630684, + 0.000934070092625916, + 0.0006947588408365846, + 0.0007535478216595948, + 0.0013907832326367497, + 0.0015821382403373718, + 0.0012462179875001311, + 0.0017264079069718719, + 0.0011562173021957278, + 0.0009106126381084323, + 0.001043238677084446, + 0.0008624211768619716, + 0.0009735539788380265, + 0.0006660323706455529, + 0.0007468897383660078, + 0.0009428639314137399, + 0.0006233312888070941, + 0.0008775431197136641, + 0.005137964151799679, + 0.00666309567168355, + 0.006274833343923092, + 0.010957736521959305, + 0.00930213276296854, + 0.006079886574298143, + 0.007635501213371754, + 0.008186494931578636 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 5, + 9 + ], + "selected_reward": -0.3877088129520416 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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, tight close-up focusing on a pair of hands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tie made of deep charcoal silk with a satin sheen and tied in a Windsor knot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the hands performing the specific actions of pinching the knot with the right hand and sliding it upward with the left hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a deep dimple created in the silk beneath the knot by the fingers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by warm, directional light from a single off-camera source casting soft highlights and deep shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026353277266025543 + ], + "opposite_logprob_tokens": [ + -3.651353359222412 + ], + "logprob_reward": -0.026353277266025543, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026353277266025543, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07924237847328186, + 0.2186819165945053, + 0.02357589825987816, + 0.017361946403980255, + 0.02775595895946026, + 0.4438400864601135, + 0.012003360316157341, + 0.014643397182226181, + 0.07802456617355347, + 0.08487047255039215 + ], + "credit_map": [ + 0.015760725364089012, + 0.006673253607004881, + 0.011625012382864952, + 0.00954459048807621, + 0.006921656895428896, + 0.010384440422058105, + 0.008455468341708183, + 0.009877228178083897, + 0.007063126191496849, + 0.006813403218984604, + 0.011692077852785587, + 0.011533508077263832, + 0.006958106532692909, + 0.08128094673156738, + 0.006199097726494074, + 0.08714165538549423, + 0.00258844718337059, + 0.001964333700016141, + 0.003260358702391386, + 0.0035491175949573517, + 0.0026034233160316944, + 0.0030603609047830105, + 0.003574593458324671, + 0.002975263400003314, + 0.0015536792343482375, + 0.0017143174773082137, + 0.0032753103878349066, + 0.002066145185381174, + 0.002103673294186592, + 0.002051703864708543, + 0.002462084637954831, + 0.002135033020749688, + 0.0020538249518722296, + 0.0018242171499878168, + 0.0029383511282503605, + 0.0027539250440895557, + 0.0021273395977914333, + 0.0056876386515796185, + 0.003651449689641595, + 0.006719212047755718, + 0.019715869799256325, + 0.009170825593173504, + 0.026244288310408592, + 0.007017972879111767, + 0.009327786043286324, + 0.17340241372585297, + 0.010540075600147247, + 0.1884208470582962, + 0.0011070577893406153, + 0.0017788138939067721, + 0.0017880029045045376, + 0.00186622841283679, + 0.001617797533981502, + 0.0012220790376886725, + 0.0015000682324171066, + 0.001123312977142632, + 0.0013013072311878204, + 0.0017285292269662023, + 0.0025414249394088984, + 0.001597836147993803, + 0.001924827229231596, + 0.002357825404033065, + 0.0016222483245655894, + 0.0015693983295932412, + 0.007511378265917301, + 0.012388481758534908, + 0.009780346415936947, + 0.012077090330421925, + 0.00869661569595337, + 0.010177058167755604, + 0.007396128494292498, + 0.009997467510402203, + 0.008442790247499943, + 0.011175894178450108, + 0.0081942742690444, + 0.01426713727414608, + 0.01662781462073326, + 0.008890101686120033, + 0.010612349957227707, + 0.006660110782831907 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 1, + 9 + ], + "selected_reward": -0.026353277266025543 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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, tight close-up focusing on a pair of hands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tie made of deep charcoal silk with a satin sheen and tied in a Windsor knot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the hands performing the specific actions of pinching the knot with the right hand and sliding it upward with the left hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a deep dimple created in the silk beneath the knot by the fingers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by warm, directional light from a single off-camera source casting soft highlights and deep shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06205231696367264 + ], + "opposite_logprob_tokens": [ + -2.8120522499084473 + ], + "logprob_reward": -0.06205231696367264, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06205231696367264, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11588778346776962, + 0.149346724152565, + 0.04562222212553024, + 0.02436305209994316, + 0.048871058970689774, + 0.4331226646900177, + 0.044661395251750946, + 0.02668343111872673, + 0.05693913251161575, + 0.05450260266661644 + ], + "credit_map": [ + 0.01584247313439846, + 0.009428023360669613, + 0.012222115881741047, + 0.014268332161009312, + 0.014899319969117641, + 0.014774433337152004, + 0.01759282872080803, + 0.016860246658325195, + 0.0087303901091218, + 0.012751944363117218, + 0.013321097008883953, + 0.024507272988557816, + 0.021455978974699974, + 0.023351602256298065, + 0.012446223758161068, + 0.03278220072388649, + 0.003997204825282097, + 0.00531494477763772, + 0.004341889638453722, + 0.007139214780181646, + 0.006160159595310688, + 0.007095274981111288, + 0.004495820961892605, + 0.00707771023735404, + 0.002110856818035245, + 0.0023617844562977552, + 0.0031088010873645544, + 0.0028054751455783844, + 0.003915851004421711, + 0.0034602859523147345, + 0.003786381334066391, + 0.002813612576574087, + 0.003642626339569688, + 0.004845849238336086, + 0.004595140460878611, + 0.005296891555190086, + 0.006274210289120674, + 0.00962857436388731, + 0.0036041266284883022, + 0.010983636602759361, + 0.019010214135050774, + 0.009211207740008831, + 0.025533730164170265, + 0.007301884237676859, + 0.009541047737002373, + 0.1679757982492447, + 0.01101441029459238, + 0.18353432416915894, + 0.003073078813031316, + 0.006585912313312292, + 0.005599652417004108, + 0.00479159876704216, + 0.005624300334602594, + 0.006064447108656168, + 0.005172040779143572, + 0.007750358432531357, + 0.0021700705401599407, + 0.002514560241252184, + 0.004002530127763748, + 0.004197831265628338, + 0.004189185332506895, + 0.004059788305312395, + 0.0028560764621943235, + 0.002693386049941182, + 0.005588270723819733, + 0.011200332082808018, + 0.005682299844920635, + 0.007942364551126957, + 0.00601923419162631, + 0.007848070934414864, + 0.004716765601187944, + 0.007941788993775845, + 0.005189208313822746, + 0.005530407652258873, + 0.005192199721932411, + 0.010449716821312904, + 0.009006435051560402, + 0.004925592336803675, + 0.007868095301091671, + 0.006340940948575735 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 1, + 0 + ], + "selected_reward": -0.06205231696367264 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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, tight close-up focusing on a pair of hands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tie made of deep charcoal silk with a satin sheen and tied in a Windsor knot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the hands performing the specific actions of pinching the knot with the right hand and sliding it upward with the left hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a deep dimple created in the silk beneath the knot by the fingers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by warm, directional light from a single off-camera source casting soft highlights and deep shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01607552357017994 + ], + "opposite_logprob_tokens": [ + -4.141075611114502 + ], + "logprob_reward": -0.01607552357017994, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.01607552357017994, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16451489925384521, + 0.15578006207942963, + 0.03568965941667557, + 0.036383431404829025, + 0.03819780424237251, + 0.3750506341457367, + 0.02864728681743145, + 0.019772417843341827, + 0.015689276158809662, + 0.13027454912662506 + ], + "credit_map": [ + 0.03401617705821991, + 0.019122451543807983, + 0.023835793137550354, + 0.018663424998521805, + 0.01583593524992466, + 0.01832771673798561, + 0.014184045605361462, + 0.020529355853796005, + 0.007926443591713905, + 0.010725867003202438, + 0.011717651039361954, + 0.022487083449959755, + 0.018461724743247032, + 0.03203011676669121, + 0.011406120844185352, + 0.041025061160326004, + 0.003986770752817392, + 0.003895907197147608, + 0.004651418887078762, + 0.0054915901273489, + 0.005204560700803995, + 0.003841477446258068, + 0.005077199079096317, + 0.0035407361574470997, + 0.003262696322053671, + 0.003916479647159576, + 0.0051566786132752895, + 0.003195533761754632, + 0.004050066694617271, + 0.0038993346970528364, + 0.00680367648601532, + 0.0060989647172391415, + 0.00335471099242568, + 0.004406715743243694, + 0.0057419706135988235, + 0.0038529711309820414, + 0.005922464653849602, + 0.004060583189129829, + 0.006765526253730059, + 0.004092863295227289, + 0.01685664989054203, + 0.014368132688105106, + 0.023024560883641243, + 0.009333111345767975, + 0.01342017948627472, + 0.14032196998596191, + 0.013893337920308113, + 0.14383269846439362, + 0.0027269062120467424, + 0.004407667089253664, + 0.004413606598973274, + 0.0022762257140129805, + 0.003794863121584058, + 0.004554822575300932, + 0.0032499651424586773, + 0.003223229432478547, + 0.0022711383644491434, + 0.0018993803532794118, + 0.002665779087692499, + 0.001382353249937296, + 0.002331712283194065, + 0.005286569241434336, + 0.0019617860671132803, + 0.0019737007096409798, + 0.0023888263385742903, + 0.0019170516170561314, + 0.0019593352917581797, + 0.0020546128507703543, + 0.001460813800804317, + 0.0020901598036289215, + 0.0016677891835570335, + 0.0021506871562451124, + 0.019921453669667244, + 0.018438618630170822, + 0.010685368441045284, + 0.021114828065037727, + 0.022554725408554077, + 0.013071633875370026, + 0.010259753093123436, + 0.014228173531591892 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 1 + ], + "selected_reward": -0.01607552357017994 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, tight close-up focuses on a pair of manicured hands making the final, precise adjustments to the knot of a tie. The tie, a deep charcoal silk with a subtle satin sheen, is tied in a perfect Windsor knot atop the solid white collar of a crisp, high-line shirt. The thumb and index finger of the right hand gently pinch the top of the knot, while the left hand deliberately slides it upward, securing it firmly in place. With practiced, unhurried confidence, the fingers meticulously create and perfect a single, deep dimple in the silk beneath the knot. The rich texture of a dark, fine wool suit jacket, its sharp lapels clearly visible, fills the frame. The scene is bathed in warm, directional light from a single, off-camera source, casting soft highlights on the joints and uneven fabric while deepening the shadows, creating an intimate atmosphere of focused focus and quiet, formal preparation.", + "seed": 25007, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000056/group_1/index_3297_seed_25007_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": -0.06350122150033713, + "vlm_logprob": -0.06350122150033713, + "vlm_yes_no_margin": 3.725, + "vlm_token_credit": -0.06350122150033713, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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, tight close-up focusing on a pair of hands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tie made of deep charcoal silk with a satin sheen and tied in a Windsor knot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the hands performing the specific actions of pinching the knot with the right hand and sliding it upward with the left hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a deep dimple created in the silk beneath the knot by the fingers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by warm, directional light from a single off-camera source casting soft highlights and deep shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016150949522852898 + ], + "opposite_logprob_tokens": [ + -6.266150951385498 + ], + "logprob_reward": -0.016150949522852898, + "yes_no_margin_reward": 6.25, + "token_credit_reward": -0.016150949522852898, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1437983214855194, + 0.1174994632601738, + 0.0858241394162178, + 0.06806642562150955, + 0.07282587140798569, + 0.11870609223842621, + 0.15552765130996704, + 0.1207251250743866, + 0.05902169272303581, + 0.05800526216626167 + ], + "credit_map": [ + 0.016541242599487305, + 0.014883734285831451, + 0.02499375119805336, + 0.011819256469607353, + 0.014434708282351494, + 0.023192906752228737, + 0.01801472343504429, + 0.01991797424852848, + 0.018620552495121956, + 0.016264131292700768, + 0.026530081406235695, + 0.015593807213008404, + 0.0064261313527822495, + 0.007592146750539541, + 0.009355315007269382, + 0.017117293551564217, + 0.008324355818331242, + 0.008779674768447876, + 0.02874021604657173, + 0.004534638952463865, + 0.0028404430486261845, + 0.0038780851755291224, + 0.004296031314879656, + 0.024430686607956886, + 0.00820133276283741, + 0.005298849660903215, + 0.013156517408788204, + 0.00563539145514369, + 0.004500275012105703, + 0.010214947164058685, + 0.005778880789875984, + 0.01528022438287735, + 0.010094494558870792, + 0.014354770071804523, + 0.012571348808705807, + 0.006233313586562872, + 0.005168766248971224, + 0.006722773890942335, + 0.005963726434856653, + 0.011716674081981182, + 0.012883408926427364, + 0.011540167964994907, + 0.02343769744038582, + 0.008561369962990284, + 0.013134520500898361, + 0.019283432513475418, + 0.008400700986385345, + 0.021464787423610687, + 0.014778225682675838, + 0.02524283342063427, + 0.022090090438723564, + 0.013641281984746456, + 0.022294873371720314, + 0.018423331901431084, + 0.01611032709479332, + 0.022946679964661598, + 0.01490061916410923, + 0.01345417182892561, + 0.024033186957240105, + 0.017280278727412224, + 0.010269690304994583, + 0.010316796600818634, + 0.008484503254294395, + 0.021985869854688644, + 0.008035863749682903, + 0.006100899539887905, + 0.014119579456746578, + 0.005252923350781202, + 0.004971578251570463, + 0.005605468992143869, + 0.004413984250277281, + 0.010521390475332737, + 0.009748267941176891, + 0.00686128344386816, + 0.011020546779036522, + 0.004550598561763763, + 0.004294533748179674, + 0.006380119826644659, + 0.004203958436846733, + 0.010945950634777546 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 7 + ], + "selected_reward": -0.016150949522852898 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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, tight close-up focusing on a pair of hands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tie made of deep charcoal silk with a satin sheen and tied in a Windsor knot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the hands performing the specific actions of pinching the knot with the right hand and sliding it upward with the left hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a deep dimple created in the silk beneath the knot by the fingers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by warm, directional light from a single off-camera source casting soft highlights and deep shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.20189127326011658 + ], + "opposite_logprob_tokens": [ + -1.701891303062439 + ], + "logprob_reward": -0.20189127326011658, + "yes_no_margin_reward": 1.5, + "token_credit_reward": -0.20189127326011658, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2095881849527359, + 0.0913548693060875, + 0.07164597511291504, + 0.09827525168657303, + 0.16498009860515594, + 0.09410324692726135, + 0.06736727803945541, + 0.11619119346141815, + 0.04050315544009209, + 0.0459907092154026 + ], + "credit_map": [ + 0.033462341874837875, + 0.027200058102607727, + 0.022627048194408417, + 0.02137104794383049, + 0.024434830993413925, + 0.02753416635096073, + 0.021482892334461212, + 0.031475819647312164, + 0.016178783029317856, + 0.009938369505107403, + 0.019076963886618614, + 0.009332744404673576, + 0.007213486824184656, + 0.008368232287466526, + 0.007500938139855862, + 0.013745368458330631, + 0.008250386454164982, + 0.00872382428497076, + 0.010009617544710636, + 0.0074080815538764, + 0.006810672581195831, + 0.006982987746596336, + 0.009070356376469135, + 0.014390059746801853, + 0.010625027120113373, + 0.01115416269749403, + 0.01594390720129013, + 0.007993833161890507, + 0.011074851267039776, + 0.01411530002951622, + 0.011229992844164371, + 0.016138194128870964, + 0.020773623138666153, + 0.024819539859890938, + 0.022938784211874008, + 0.015541699714958668, + 0.019893575459718704, + 0.018764948472380638, + 0.014669612981379032, + 0.02757834456861019, + 0.01163035724312067, + 0.010437638498842716, + 0.019664062187075615, + 0.007380146533250809, + 0.009486178867518902, + 0.010419552214443684, + 0.007067762315273285, + 0.018017560243606567, + 0.00540921138599515, + 0.007134077604860067, + 0.010801658034324646, + 0.005449972581118345, + 0.011917747557163239, + 0.009036914445459843, + 0.009164697490632534, + 0.008453012444078922, + 0.014528396539390087, + 0.013461691327393055, + 0.024095598608255386, + 0.010669952258467674, + 0.007868506945669651, + 0.015695692971348763, + 0.011096537113189697, + 0.018774839118123055, + 0.0038583888672292233, + 0.004118912387639284, + 0.008527507074177265, + 0.003321831114590168, + 0.005585134960711002, + 0.00403432035818696, + 0.004933343734592199, + 0.0061237262561917305, + 0.0064931041561067104, + 0.005065298639237881, + 0.006217727903276682, + 0.0026362102944403887, + 0.005195771809667349, + 0.008361971005797386, + 0.004293456673622131, + 0.007727175951004028 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 7 + ], + "selected_reward": -0.20189127326011658 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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, tight close-up focusing on a pair of hands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tie made of deep charcoal silk with a satin sheen and tied in a Windsor knot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the hands performing the specific actions of pinching the knot with the right hand and sliding it upward with the left hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a deep dimple created in the silk beneath the knot by the fingers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by warm, directional light from a single off-camera source casting soft highlights and deep shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02634631097316742 + ], + "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": [ + 0.09297306090593338, + 0.12220049649477005, + 0.1269454061985016, + 0.16207119822502136, + 0.04249934107065201, + 0.09477946162223816, + 0.17852282524108887, + 0.0607752799987793, + 0.08838029205799103, + 0.030852608382701874 + ], + "credit_map": [ + 0.009069130755960941, + 0.011166287586092949, + 0.01562635600566864, + 0.009187700226902962, + 0.01371836569160223, + 0.01053234376013279, + 0.01106982957571745, + 0.01260304730385542, + 0.018498919904232025, + 0.01600654236972332, + 0.029653232544660568, + 0.012604950927197933, + 0.007065530866384506, + 0.008492964319884777, + 0.011027800850570202, + 0.018850555643439293, + 0.014111844822764397, + 0.009629282169044018, + 0.04146658256649971, + 0.008627649396657944, + 0.005199712701141834, + 0.010827848687767982, + 0.005861611105501652, + 0.031220881268382072, + 0.0221488568931818, + 0.022076645866036415, + 0.023868141695857048, + 0.012899939902126789, + 0.014724938198924065, + 0.02891293168067932, + 0.010916046798229218, + 0.026523683220148087, + 0.006215750239789486, + 0.004747702274471521, + 0.007960020564496517, + 0.004321495536714792, + 0.0042990208603441715, + 0.004550370387732983, + 0.003467216622084379, + 0.006937767378985882, + 0.009155110456049442, + 0.010881795547902584, + 0.014033048413693905, + 0.007116593886166811, + 0.014235388487577438, + 0.012060985900461674, + 0.009587795473635197, + 0.017708750441670418, + 0.0242739487439394, + 0.027063382789492607, + 0.03449103981256485, + 0.020669426769018173, + 0.012990105897188187, + 0.015721552073955536, + 0.014835753478109837, + 0.028477631509304047, + 0.008835352025926113, + 0.008957972750067711, + 0.012159740552306175, + 0.0057579530403018, + 0.005240368191152811, + 0.0053947181440889835, + 0.004650426097214222, + 0.009778747335076332, + 0.009623758494853973, + 0.007986051961779594, + 0.021937170997262, + 0.008655029349029064, + 0.007986152544617653, + 0.010482131503522396, + 0.006900792475789785, + 0.014809207059442997, + 0.0043645077385008335, + 0.0043223099783062935, + 0.004864584654569626, + 0.002223007380962372, + 0.0030612575355917215, + 0.004217011854052544, + 0.0025910602416843176, + 0.005208869930356741 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 3, + 2 + ], + "selected_reward": -0.02634631097316742 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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, tight close-up focusing on a pair of hands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tie made of deep charcoal silk with a satin sheen and tied in a Windsor knot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the hands performing the specific actions of pinching the knot with the right hand and sliding it upward with the left hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a deep dimple created in the silk beneath the knot by the fingers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by warm, directional light from a single off-camera source casting soft highlights and deep shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.062042124569416046 + ], + "opposite_logprob_tokens": [ + -2.812042236328125 + ], + "logprob_reward": -0.062042124569416046, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.062042124569416046, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12874330580234528, + 0.11405709385871887, + 0.10359594970941544, + 0.04098450392484665, + 0.05979179963469505, + 0.11245707422494888, + 0.10582958906888962, + 0.19339993596076965, + 0.09530860185623169, + 0.04583219811320305 + ], + "credit_map": [ + 0.022288331761956215, + 0.015265348367393017, + 0.016008740290999413, + 0.012111682444810867, + 0.013982665725052357, + 0.015006857924163342, + 0.013734466396272182, + 0.020345209166407585, + 0.01945711486041546, + 0.01318596675992012, + 0.026948152109980583, + 0.012126588262617588, + 0.007869375869631767, + 0.00916188396513462, + 0.00815971102565527, + 0.017148297280073166, + 0.014166110195219517, + 0.012212496250867844, + 0.02409445308148861, + 0.011799141764640808, + 0.007621333468705416, + 0.006953494157642126, + 0.009522899985313416, + 0.01722601428627968, + 0.006765200290828943, + 0.00764814717695117, + 0.0071562775410711765, + 0.003582343691959977, + 0.0040629091672599316, + 0.0033243512734770775, + 0.002710076980292797, + 0.005735195241868496, + 0.007668312173336744, + 0.009492023847997189, + 0.008985255844891071, + 0.004328317940235138, + 0.006773978006094694, + 0.006374017801135778, + 0.005608750507235527, + 0.010561144910752773, + 0.015029549598693848, + 0.014181924052536488, + 0.024580832570791245, + 0.010530787520110607, + 0.009400885552167892, + 0.011123140342533588, + 0.008169390261173248, + 0.01944057084619999, + 0.007305807899683714, + 0.00968075729906559, + 0.014208801090717316, + 0.0074996668845415115, + 0.020898453891277313, + 0.016994651407003403, + 0.013474684208631516, + 0.01576676405966282, + 0.0297860074788332, + 0.027034888043999672, + 0.04164678230881691, + 0.011856446973979473, + 0.015322361141443253, + 0.017355481162667274, + 0.015639547258615494, + 0.03475840762257576, + 0.010626931674778461, + 0.012845966033637524, + 0.021034950390458107, + 0.008680122904479504, + 0.008854152634739876, + 0.00928198080509901, + 0.008062317967414856, + 0.015922175720334053, + 0.006614089477807283, + 0.005398355890065432, + 0.006398620083928108, + 0.003461624262854457, + 0.004963058512657881, + 0.0075239925645291805, + 0.00413055345416069, + 0.007341901771724224 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 1 + ], + "selected_reward": -0.062042124569416046 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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, tight close-up focusing on a pair of hands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tie made of deep charcoal silk with a satin sheen and tied in a Windsor knot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the hands performing the specific actions of pinching the knot with the right hand and sliding it upward with the left hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a deep dimple created in the silk beneath the knot by the fingers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by warm, directional light from a single off-camera source casting soft highlights and deep shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011075449176132679 + ], + "opposite_logprob_tokens": [ + -4.511075496673584 + ], + "logprob_reward": -0.011075449176132679, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011075449176132679, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1749579757452011, + 0.057085856795310974, + 0.34377554059028625, + 0.03230345621705055, + 0.04772984981536865, + 0.13510514795780182, + 0.04568009078502655, + 0.10160641372203827, + 0.04139144718647003, + 0.020364174619317055 + ], + "credit_map": [ + 0.023436924442648888, + 0.015226582065224648, + 0.027359144762158394, + 0.023729050531983376, + 0.01189730316400528, + 0.02816481702029705, + 0.013632826507091522, + 0.031511325389146805, + 0.008055726066231728, + 0.007404378615319729, + 0.012214316986501217, + 0.00513392873108387, + 0.004718452226370573, + 0.006552965380251408, + 0.004675532691180706, + 0.008330554701387882, + 0.024626700207591057, + 0.018449626863002777, + 0.12381008267402649, + 0.018992090597748756, + 0.008983106352388859, + 0.01809234358370304, + 0.009700139984488487, + 0.12112144380807877, + 0.00440860353410244, + 0.004534659907221794, + 0.0068426006473600864, + 0.002595688682049513, + 0.0016275043599307537, + 0.0035056639462709427, + 0.0021752926986664534, + 0.006613444071263075, + 0.0068871513940393925, + 0.00730902049690485, + 0.010730062611401081, + 0.003484909189864993, + 0.0032261726446449757, + 0.004818703047931194, + 0.003089996287599206, + 0.008183833211660385, + 0.018828028813004494, + 0.016752710565924644, + 0.03031199611723423, + 0.010683765634894371, + 0.00969791505485773, + 0.017246292904019356, + 0.009089095517992973, + 0.02249533124268055, + 0.0063067954033613205, + 0.004332721699029207, + 0.00872022658586502, + 0.0046620965003967285, + 0.0029886115808039904, + 0.0067893643863499165, + 0.003269011154770851, + 0.00861126184463501, + 0.01586538553237915, + 0.015468082390725613, + 0.02158724144101143, + 0.011029212735593319, + 0.005693044513463974, + 0.008339973166584969, + 0.006214265711605549, + 0.017409205436706543, + 0.0037581517826765776, + 0.004634801298379898, + 0.00639578839763999, + 0.004584097769111395, + 0.005343256983906031, + 0.006196148227900267, + 0.003980410750955343, + 0.006498794537037611, + 0.0025671329349279404, + 0.002638960722833872, + 0.003289084881544113, + 0.0015785321593284607, + 0.002171268919482827, + 0.0028147685807198286, + 0.0022625995334237814, + 0.0030418250244110823 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 5 + ], + "selected_reward": -0.011075449176132679 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, tight close-up focuses on a pair of manicured hands making the final, precise adjustments to the knot of a tie. The tie, a deep charcoal silk with a subtle satin sheen, is tied in a perfect Windsor knot atop the solid white collar of a crisp, high-line shirt. The thumb and index finger of the right hand gently pinch the top of the knot, while the left hand deliberately slides it upward, securing it firmly in place. With practiced, unhurried confidence, the fingers meticulously create and perfect a single, deep dimple in the silk beneath the knot. The rich texture of a dark, fine wool suit jacket, its sharp lapels clearly visible, fills the frame. The scene is bathed in warm, directional light from a single, off-camera source, casting soft highlights on the joints and uneven fabric while deepening the shadows, creating an intimate atmosphere of focused focus and quiet, formal preparation.", + "seed": 25008, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000056/group_1/index_3297_seed_25008_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": -0.061972457356750964, + "vlm_logprob": -0.061972457356750964, + "vlm_yes_no_margin": 3.575, + "vlm_token_credit": -0.061972457356750964, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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, tight close-up focusing on a pair of hands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tie made of deep charcoal silk with a satin sheen and tied in a Windsor knot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the hands performing the specific actions of pinching the knot with the right hand and sliding it upward with the left hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a deep dimple created in the silk beneath the knot by the fingers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by warm, directional light from a single off-camera source casting soft highlights and deep shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01720573566854 + ], + "opposite_logprob_tokens": [ + -5.392205715179443 + ], + "logprob_reward": -0.01720573566854, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.01720573566854, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17579594254493713, + 0.0872502326965332, + 0.21671658754348755, + 0.05916439741849899, + 0.06478705257177353, + 0.050408605486154556, + 0.06642867624759674, + 0.0644308552145958, + 0.10402320325374603, + 0.11099443584680557 + ], + "credit_map": [ + 0.029423363506793976, + 0.009759715758264065, + 0.06931084394454956, + 0.025563787668943405, + 0.00584857352077961, + 0.017835242673754692, + 0.006540169008076191, + 0.01151425950229168, + 0.006386280991137028, + 0.0048855748027563095, + 0.028219250962138176, + 0.015949226915836334, + 0.006462119985371828, + 0.007932630367577076, + 0.0071607013233006, + 0.010254444554448128, + 0.018085381016135216, + 0.015412041917443275, + 0.07916413992643356, + 0.042784661054611206, + 0.009853294119238853, + 0.017757566645741463, + 0.01098956074565649, + 0.022669950500130653, + 0.0066505419090390205, + 0.006104767322540283, + 0.011986329220235348, + 0.009854961186647415, + 0.004828331526368856, + 0.009827718138694763, + 0.004287111107259989, + 0.005624635145068169, + 0.004840577952563763, + 0.008315620943903923, + 0.012246135622262955, + 0.01103682816028595, + 0.006012791767716408, + 0.00982668623328209, + 0.00625860458239913, + 0.006249810568988323, + 0.0037473139818757772, + 0.008210595697164536, + 0.008833307772874832, + 0.005183475557714701, + 0.004939178004860878, + 0.007711269427090883, + 0.007152747828513384, + 0.004630719777196646, + 0.004591529257595539, + 0.013204910792410374, + 0.01037587784230709, + 0.016001051291823387, + 0.004103674553334713, + 0.010347031988203526, + 0.003523603780195117, + 0.0042809983715415, + 0.0052906833589077, + 0.009616819210350513, + 0.00860605388879776, + 0.013592596165835857, + 0.0043968698009848595, + 0.008727148175239563, + 0.007433523423969746, + 0.006767160724848509, + 0.006509287282824516, + 0.01098576933145523, + 0.02669169194996357, + 0.014231734909117222, + 0.007186830043792725, + 0.017287729308009148, + 0.008298644796013832, + 0.012831512838602066, + 0.007707276847213507, + 0.00782170332968235, + 0.011999706737697124, + 0.009300445206463337, + 0.0069352868013083935, + 0.03200611099600792, + 0.006276485975831747, + 0.02894742041826248 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 9 + ], + "selected_reward": -0.01720573566854 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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, tight close-up focusing on a pair of hands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tie made of deep charcoal silk with a satin sheen and tied in a Windsor knot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the hands performing the specific actions of pinching the knot with the right hand and sliding it upward with the left hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a deep dimple created in the silk beneath the knot by the fingers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by warm, directional light from a single off-camera source casting soft highlights and deep shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.20186544954776764 + ], + "opposite_logprob_tokens": [ + -1.7018654346466064 + ], + "logprob_reward": -0.20186544954776764, + "yes_no_margin_reward": 1.5, + "token_credit_reward": -0.20186544954776764, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5896925926208496, + 0.158797487616539, + 0.04415034502744675, + 0.01981595903635025, + 0.054664961993694305, + 0.03610726073384285, + 0.017606692388653755, + 0.02064441703259945, + 0.01743110828101635, + 0.041089169681072235 + ], + "credit_map": [ + 0.1227821633219719, + 0.031556855887174606, + 0.21725483238697052, + 0.07957788556814194, + 0.029715847223997116, + 0.05052619427442551, + 0.026456793770194054, + 0.031822044402360916, + 0.01009573694318533, + 0.0037296598311513662, + 0.062054041773080826, + 0.03204621747136116, + 0.005104538518935442, + 0.008655172772705555, + 0.006548722740262747, + 0.03056340292096138, + 0.002943395171314478, + 0.006613672245293856, + 0.006140754558146, + 0.004865841008722782, + 0.009423494338989258, + 0.004800054244697094, + 0.004646345041692257, + 0.004716792143881321, + 0.002470507984980941, + 0.001706991926766932, + 0.004460588563233614, + 0.00226151617243886, + 0.0020689498633146286, + 0.0033356808125972748, + 0.0018880544230341911, + 0.0016236704541370273, + 0.0038740888703614473, + 0.006874819286167622, + 0.01029984001070261, + 0.00704118050634861, + 0.004046754911541939, + 0.009355951100587845, + 0.008361773565411568, + 0.004810560029000044, + 0.004962735343724489, + 0.008584153838455677, + 0.005244921892881393, + 0.00440590875223279, + 0.0033534541726112366, + 0.005091167986392975, + 0.0019332434749230742, + 0.0025316800456494093, + 0.0010859081521630287, + 0.003224137704819441, + 0.0025505826342850924, + 0.0033326679840683937, + 0.001595226931385696, + 0.0027926259208470583, + 0.0012614934239536524, + 0.001764051616191864, + 0.0014080856926739216, + 0.0037947618402540684, + 0.002249669749289751, + 0.004044319037348032, + 0.0019097001058980823, + 0.002846131334081292, + 0.002188778016716242, + 0.0022029734682291746, + 0.0020154037047177553, + 0.0022566956467926502, + 0.0024929193314164877, + 0.00393160292878747, + 0.0012031327933073044, + 0.002404570346698165, + 0.001645147567614913, + 0.0014816370094195008, + 0.0042729731649160385, + 0.003177827689796686, + 0.005893738940358162, + 0.0032154079526662827, + 0.002351405331864953, + 0.009171959012746811, + 0.002750614657998085, + 0.01025525014847517 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 4 + ], + "selected_reward": -0.20186544954776764 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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, tight close-up focusing on a pair of hands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tie made of deep charcoal silk with a satin sheen and tied in a Windsor knot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the hands performing the specific actions of pinching the knot with the right hand and sliding it upward with the left hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a deep dimple created in the silk beneath the knot by the fingers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by warm, directional light from a single off-camera source casting soft highlights and deep shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023292383179068565 + ], + "opposite_logprob_tokens": [ + -3.773292303085327 + ], + "logprob_reward": -0.023292383179068565, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023292383179068565, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.25774675607681274, + 0.05191529169678688, + 0.10962969809770584, + 0.0940045639872551, + 0.12173274904489517, + 0.08815830200910568, + 0.06186170130968094, + 0.093808114528656, + 0.07507215440273285, + 0.04607061669230461 + ], + "credit_map": [ + 0.047803670167922974, + 0.010435645468533039, + 0.10255417972803116, + 0.03754451870918274, + 0.009401235729455948, + 0.022943370044231415, + 0.01065530814230442, + 0.01640884205698967, + 0.0034139424096792936, + 0.004758843686431646, + 0.013284789398312569, + 0.007243189495056868, + 0.004524853080511093, + 0.004909207113087177, + 0.005474167410284281, + 0.008306301198899746, + 0.008380918763577938, + 0.011260122060775757, + 0.03190368413925171, + 0.018090924248099327, + 0.012931153178215027, + 0.00811243336647749, + 0.007771189324557781, + 0.011179272085428238, + 0.009959661401808262, + 0.011609968729317188, + 0.019537903368473053, + 0.012574897147715092, + 0.009438018314540386, + 0.015060950070619583, + 0.008415521122515202, + 0.007407643832266331, + 0.007822059094905853, + 0.013172658160328865, + 0.026695616543293, + 0.01653425395488739, + 0.00726708909496665, + 0.02274255082011223, + 0.017453279346227646, + 0.010045243427157402, + 0.011328750289976597, + 0.021142959594726562, + 0.014901760965585709, + 0.008873696438968182, + 0.00803652685135603, + 0.011541456915438175, + 0.005823349580168724, + 0.006509801838546991, + 0.005304612684994936, + 0.012823039665818214, + 0.007673898246139288, + 0.010998119600117207, + 0.003941470291465521, + 0.009210800752043724, + 0.0060365875251591206, + 0.00587317394092679, + 0.00531592033803463, + 0.011744719929993153, + 0.01344249490648508, + 0.020227231085300446, + 0.004510647151619196, + 0.014922954142093658, + 0.012575722299516201, + 0.011068432591855526, + 0.005688470788300037, + 0.006560931913554668, + 0.0133580407127738, + 0.009758013300597668, + 0.007375472225248814, + 0.011729844845831394, + 0.009702369570732117, + 0.010899013839662075, + 0.004031284246593714, + 0.00500914454460144, + 0.008093764074146748, + 0.005543735809624195, + 0.0025873947888612747, + 0.008355277590453625, + 0.0039257025346159935, + 0.008524308912456036 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 2 + ], + "selected_reward": -0.023292383179068565 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, tight close-up focusing on a pair of hands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tie made of deep charcoal silk with a satin sheen and tied in a Windsor knot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the hands performing the specific actions of pinching the knot with the right hand and sliding it upward with the left hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a deep dimple created in the silk beneath the knot by the fingers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by warm, directional light from a single off-camera source casting soft highlights and deep shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05495418235659599 + ], + "opposite_logprob_tokens": [ + -2.9299542903900146 + ], + "logprob_reward": -0.05495418235659599, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05495418235659599, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21474750339984894, + 0.13831205666065216, + 0.13149526715278625, + 0.05080818012356758, + 0.13471387326717377, + 0.08638709783554077, + 0.04657287523150444, + 0.034658242017030716, + 0.07263436168432236, + 0.08967052400112152 + ], + "credit_map": [ + 0.041195645928382874, + 0.01261149812489748, + 0.07674629241228104, + 0.02912461757659912, + 0.011744623072445393, + 0.020938141271471977, + 0.010111802257597446, + 0.012274895794689655, + 0.010293352417647839, + 0.004816982429474592, + 0.052741099148988724, + 0.028011653572320938, + 0.005471101962029934, + 0.009294318035244942, + 0.006018966902047396, + 0.021664585918188095, + 0.010260223411023617, + 0.008604060858488083, + 0.04108225181698799, + 0.023247217759490013, + 0.014410835690796375, + 0.009910155087709427, + 0.008233909495174885, + 0.015746628865599632, + 0.005071532912552357, + 0.0057486374862492085, + 0.00921971257776022, + 0.005152239929884672, + 0.005119577515870333, + 0.008177381940186024, + 0.004965821281075478, + 0.0073532769456505775, + 0.009908683598041534, + 0.014585164375603199, + 0.0277395099401474, + 0.01667443849146366, + 0.008584795519709587, + 0.02340853586792946, + 0.019837945699691772, + 0.013974800705909729, + 0.00925997830927372, + 0.018005574122071266, + 0.012648549862205982, + 0.009110791608691216, + 0.006580348592251539, + 0.012649358250200748, + 0.010744928382337093, + 0.00738756638020277, + 0.006342535372823477, + 0.00773330545052886, + 0.005960484035313129, + 0.005277012474834919, + 0.003449187148362398, + 0.006592053920030594, + 0.005664879456162453, + 0.0055534145794808865, + 0.002370459260419011, + 0.004999898374080658, + 0.004636820871382952, + 0.00600911071524024, + 0.003414882579818368, + 0.005546569358557463, + 0.003605091478675604, + 0.0040754093788564205, + 0.005618292838335037, + 0.010553033091127872, + 0.016732528805732727, + 0.012418599799275398, + 0.004296695347875357, + 0.011517815291881561, + 0.005017699208110571, + 0.006479693576693535, + 0.005714690778404474, + 0.005561051424592733, + 0.01031377725303173, + 0.01007152907550335, + 0.005448885262012482, + 0.02439439482986927, + 0.006502680946141481, + 0.021663518622517586 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 4 + ], + "selected_reward": -0.05495418235659599 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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, tight close-up focusing on a pair of hands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tie made of deep charcoal silk with a satin sheen and tied in a Windsor knot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the hands performing the specific actions of pinching the knot with the right hand and sliding it upward with the left hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a deep dimple created in the silk beneath the knot by the fingers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by warm, directional light from a single off-camera source casting soft highlights and deep shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012544536031782627 + ], + "opposite_logprob_tokens": [ + -4.387544631958008 + ], + "logprob_reward": -0.012544536031782627, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012544536031782627, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3287884294986725, + 0.051083117723464966, + 0.06059341877698898, + 0.10390716791152954, + 0.030957140028476715, + 0.06473788619041443, + 0.02476051263511181, + 0.07545420527458191, + 0.1724153608083725, + 0.08730274438858032 + ], + "credit_map": [ + 0.04705921560525894, + 0.020796995609998703, + 0.11366730183362961, + 0.04446084797382355, + 0.023653322830796242, + 0.03617720305919647, + 0.015176303684711456, + 0.027797244489192963, + 0.004266411531716585, + 0.0036894960794597864, + 0.015612955205142498, + 0.008307944983243942, + 0.003485008142888546, + 0.0035174302756786346, + 0.0034781305585056543, + 0.008725741878151894, + 0.0055678775534033775, + 0.011893854476511478, + 0.0071685342118144035, + 0.007165934890508652, + 0.011015226133167744, + 0.005829074885696173, + 0.0055351355113089085, + 0.006417776923626661, + 0.00933054555207491, + 0.015588750131428242, + 0.017102394253015518, + 0.01232639979571104, + 0.013778517954051495, + 0.015398599207401276, + 0.010543940588831902, + 0.009838019497692585, + 0.002294103614985943, + 0.0032817134633660316, + 0.004900282248854637, + 0.004032233264297247, + 0.0027847664896398783, + 0.00491756247356534, + 0.004951612092554569, + 0.0037948654498904943, + 0.007673315238207579, + 0.011004622094333172, + 0.01180704589933157, + 0.007232058327645063, + 0.006965033244341612, + 0.011000523343682289, + 0.0044623794965445995, + 0.0045929038897156715, + 0.0017145239980891347, + 0.004289919510483742, + 0.003133755410090089, + 0.004142997320741415, + 0.0016839156160131097, + 0.0037761135026812553, + 0.002815887099131942, + 0.003203399246558547, + 0.003477296559140086, + 0.007054597605019808, + 0.010238568298518658, + 0.012640035711228848, + 0.006582371890544891, + 0.013272261247038841, + 0.00943115446716547, + 0.012757924385368824, + 0.01171598769724369, + 0.011494467966258526, + 0.03286202996969223, + 0.02511291392147541, + 0.016449743881821632, + 0.027180463075637817, + 0.022518804296851158, + 0.025080952793359756, + 0.007338265422731638, + 0.00815825629979372, + 0.016794657334685326, + 0.009493133053183556, + 0.007451038341969252, + 0.01630326919257641, + 0.0062974183820188046, + 0.01546670775860548 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 3 + ], + "selected_reward": -0.012544536031782627 + } + ], + "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/0000056/group_10/index_tensor([3911, 3911])_group_10_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000056/group_10/index_tensor([3911, 3911])_group_10_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..6d2f8ea9bc1782639d33ea39943c8717b1d0f998 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000056/group_10/index_tensor([3911, 3911])_group_10_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "dramatic, low-angle tracking shot follows a man in a suit as he sprints down an empty, asphalt city street shrouded in torrential rain. The scene is shot in high-contrast black-and-white cinematic style, with the light from the street lamps casting long, distorted shadows on the wet pavement, further accentuating the darkness. The man wears a well-tailored, dark, single-breasted suit, now completely drenched in rain. The damp fabric clings to his leaning body, outlining the tense muscles of his running form. His short, black hair clings to his forehead, his features are sharply defined, his lips pursed, his expression firm and urgent. Each step he takes lands heavily on the puddled pavement, his arms swinging vigorously for balance. The huge splashes of water beneath his feet freeze into crystalline white forms in the strong light, leaving behind him a trail of gradually fading white, striking in the darkness. The light accurately captures the raindrops, splashing water and the wet sheen of the suit, making them shimmer silver in the darkness, creating a tense, resolute and fatalistic visual atmosphere.", + "seed": 16463, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000056/group_10/index_3911_seed_16463_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": -0.0156829203478992, + "vlm_logprob": -0.0156829203478992, + "vlm_yes_no_margin": 4.5, + "vlm_token_credit": -0.0156829203478992, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 in a high-contrast black-and-white cinematic style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera executing a low-angle tracking shot following a man in a suit sprinting down an empty asphalt city street?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment shrouded in torrential rain with puddled pavement and long, distorted shadows from street lamps?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the man's dark single-breasted suit appear completely drenched, clinging to his body to outline his running form?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the water splashes beneath the man's feet appear as crystalline white forms that shimmer silver in the darkness?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01228409819304943 + ], + "opposite_logprob_tokens": [ + -5.387284278869629 + ], + "logprob_reward": -0.01228409819304943, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.01228409819304943, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22012467682361603, + 0.048507772386074066, + 0.11202637851238251, + 0.2297273874282837, + 0.061644766479730606, + 0.058327119797468185, + 0.07420795410871506, + 0.105628602206707, + 0.03666815534234047, + 0.053137194365262985 + ], + "credit_map": [ + 0.018219824880361557, + 0.025219861418008804, + 0.011105571873486042, + 0.01775032840669155, + 0.03384050726890564, + 0.056091103702783585, + 0.028713639825582504, + 0.029183831065893173, + 0.003293790854513645, + 0.007817069068551064, + 0.0024160125758498907, + 0.0032895212061703205, + 0.013385939411818981, + 0.00991377979516983, + 0.003915771376341581, + 0.004475888796150684, + 0.005109891761094332, + 0.005245016887784004, + 0.006619845516979694, + 0.0031356276012957096, + 0.03700658306479454, + 0.041430551558732986, + 0.004906029906123877, + 0.008572827093303204, + 0.011862578801810741, + 0.010719988495111465, + 0.007873456925153732, + 0.00716579332947731, + 0.06403885781764984, + 0.08478106558322906, + 0.009420782327651978, + 0.03386486694216728, + 0.004799585323780775, + 0.0053033484145998955, + 0.003944133408367634, + 0.009142187424004078, + 0.012201067060232162, + 0.010334259830415249, + 0.007001339923590422, + 0.00891884882003069, + 0.003865350503474474, + 0.006022908724844456, + 0.003407471114769578, + 0.005158381536602974, + 0.012009277939796448, + 0.01359902136027813, + 0.007933015935122967, + 0.006331689655780792, + 0.006273900158703327, + 0.00754582742229104, + 0.008050227537751198, + 0.006651556584984064, + 0.016995146870613098, + 0.00782939326018095, + 0.006524663418531418, + 0.01433724258095026, + 0.012394716031849384, + 0.010150204412639141, + 0.010689117945730686, + 0.013424611650407314, + 0.010643220506608486, + 0.016834821552038193, + 0.013734564185142517, + 0.017757341265678406, + 0.0032461760565638542, + 0.00316346762701869, + 0.00849591102451086, + 0.003744449932128191, + 0.005191209260374308, + 0.005226327572017908, + 0.003848123364150524, + 0.0037524912040680647, + 0.0050398507155478, + 0.005119900684803724, + 0.0077716088853776455, + 0.0061221555806696415, + 0.00829071830958128, + 0.008153890259563923, + 0.0064370641484856606, + 0.006202007178217173 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 2 + ], + "selected_reward": -0.01228409819304943 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 in a high-contrast black-and-white cinematic style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera executing a low-angle tracking shot following a man in a suit sprinting down an empty asphalt city street?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment shrouded in torrential rain with puddled pavement and long, distorted shadows from street lamps?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the man's dark single-breasted suit appear completely drenched, clinging to his body to outline his running form?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the water splashes beneath the man's feet appear as crystalline white forms that shimmer silver in the darkness?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029821693897247314 + ], + "opposite_logprob_tokens": [ + -3.5298216342926025 + ], + "logprob_reward": -0.029821693897247314, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029821693897247314, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09148815274238586, + 0.21239762008190155, + 0.05772339180111885, + 0.1842949241399765, + 0.166954904794693, + 0.0342259481549263, + 0.12892182171344757, + 0.034512292593717575, + 0.043325792998075485, + 0.04615519195795059 + ], + "credit_map": [ + 0.009390577673912048, + 0.012803168036043644, + 0.005424316972494125, + 0.0077156974002718925, + 0.014267363585531712, + 0.016582058742642403, + 0.012683268636465073, + 0.01262169424444437, + 0.021282494068145752, + 0.03645336627960205, + 0.012876419350504875, + 0.015251834876835346, + 0.039988983422517776, + 0.03194396197795868, + 0.018620256334543228, + 0.03598027676343918, + 0.003800675505772233, + 0.006078270263969898, + 0.0033325706608593464, + 0.0033791745081543922, + 0.01284937746822834, + 0.015478321351110935, + 0.005788803566247225, + 0.00701619079336524, + 0.008439926430583, + 0.008591783232986927, + 0.006204539909958839, + 0.006353604607284069, + 0.053703878074884415, + 0.06887809932231903, + 0.009543041698634624, + 0.022580044344067574, + 0.016443321481347084, + 0.013972776010632515, + 0.010466476902365685, + 0.012924003414809704, + 0.021959125995635986, + 0.027363741770386696, + 0.01811663620173931, + 0.04570881277322769, + 0.0027988252695649862, + 0.004326656460762024, + 0.002704188460484147, + 0.004216123837977648, + 0.005056014750152826, + 0.005726922303438187, + 0.004844421986490488, + 0.004552791826426983, + 0.0102155776694417, + 0.0230315700173378, + 0.01498444750905037, + 0.018483279272913933, + 0.017426757141947746, + 0.015423106960952282, + 0.014924183487892151, + 0.014432894997298717, + 0.0027377537917345762, + 0.003486732253804803, + 0.0040930151008069515, + 0.004608266521245241, + 0.004418184980750084, + 0.005534951109439135, + 0.004678189288824797, + 0.004955196753144264, + 0.00343507737852633, + 0.005319847725331783, + 0.005718201864510775, + 0.004812414757907391, + 0.009180325083434582, + 0.005574896931648254, + 0.005311950575560331, + 0.0039730737917125225, + 0.004549733363091946, + 0.0048754336312413216, + 0.007784570567309856, + 0.006280788220465183, + 0.005025684833526611, + 0.006917363032698631, + 0.004803827032446861, + 0.005917785223573446 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 3, + 4 + ], + "selected_reward": -0.029821693897247314 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 in a high-contrast black-and-white cinematic style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera executing a low-angle tracking shot following a man in a suit sprinting down an empty asphalt city street?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment shrouded in torrential rain with puddled pavement and long, distorted shadows from street lamps?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the man's dark single-breasted suit appear completely drenched, clinging to his body to outline his running form?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the water splashes beneath the man's feet appear as crystalline white forms that shimmer silver in the darkness?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00594940735027194 + ], + "opposite_logprob_tokens": [ + -5.1309494972229 + ], + "logprob_reward": -0.00594940735027194, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.00594940735027194, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3329853117465973, + 0.060287389904260635, + 0.19555535912513733, + 0.09345299750566483, + 0.0514429435133934, + 0.03206649050116539, + 0.15000414848327637, + 0.02135254442691803, + 0.03867737203836441, + 0.02417544461786747 + ], + "credit_map": [ + 0.02660430409014225, + 0.037731412798166275, + 0.017419325187802315, + 0.027947457507252693, + 0.05106765404343605, + 0.08403536677360535, + 0.04032190516591072, + 0.04785788431763649, + 0.00479872664436698, + 0.00680071534588933, + 0.0033487591426819563, + 0.004372900817543268, + 0.01406452339142561, + 0.010042771697044373, + 0.006789071951061487, + 0.01006991881877184, + 0.008990184403955936, + 0.006869015749543905, + 0.012392934411764145, + 0.005430353805422783, + 0.06655567139387131, + 0.07562858611345291, + 0.008188304491341114, + 0.01150032039731741, + 0.008677857927978039, + 0.01327622402459383, + 0.003277112729847431, + 0.005040220450609922, + 0.02325090579688549, + 0.021338744089007378, + 0.007418495137244463, + 0.011173436418175697, + 0.00492289150133729, + 0.005092186853289604, + 0.002816451247781515, + 0.005907571874558926, + 0.008159098215401173, + 0.007997934706509113, + 0.006428350694477558, + 0.0101184556260705, + 0.0029147029854357243, + 0.002848560456186533, + 0.002751084277406335, + 0.003012781497091055, + 0.005295843351632357, + 0.006900231819599867, + 0.003604060737416148, + 0.0047392225824296474, + 0.015481844544410706, + 0.01955980435013771, + 0.0164972972124815, + 0.013232432305812836, + 0.026838820427656174, + 0.020165031775832176, + 0.015096878632903099, + 0.023132024332880974, + 0.0025829586666077375, + 0.0017939561512321234, + 0.0025527470279484987, + 0.002439632546156645, + 0.0025413245894014835, + 0.002916679484769702, + 0.00241750362329185, + 0.004107741639018059, + 0.003566133324056864, + 0.003890282940119505, + 0.007862313650548458, + 0.004593779798597097, + 0.005940844304859638, + 0.005134117789566517, + 0.004202843643724918, + 0.003487053792923689, + 0.0029961629770696163, + 0.002309606410562992, + 0.004170574247837067, + 0.002893876051530242, + 0.0026082892436534166, + 0.0035348478704690933, + 0.0020037724170833826, + 0.003658315632492304 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 6 + ], + "selected_reward": -0.00594940735027194 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 in a high-contrast black-and-white cinematic style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera executing a low-angle tracking shot following a man in a suit sprinting down an empty asphalt city street?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment shrouded in torrential rain with puddled pavement and long, distorted shadows from street lamps?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the man's dark single-breasted suit appear completely drenched, clinging to his body to outline his running form?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the water splashes beneath the man's feet appear as crystalline white forms that shimmer silver in the darkness?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009782234206795692 + ], + "opposite_logprob_tokens": [ + -4.634782314300537 + ], + "logprob_reward": -0.009782234206795692, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009782234206795692, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24472583830356598, + 0.06530312448740005, + 0.145649254322052, + 0.1052912175655365, + 0.05445314198732376, + 0.04579000920057297, + 0.19213663041591644, + 0.07174460589885712, + 0.030180737376213074, + 0.04472547024488449 + ], + "credit_map": [ + 0.01912374421954155, + 0.027528107166290283, + 0.013387092389166355, + 0.020678726956248283, + 0.04070812091231346, + 0.06218118593096733, + 0.031061936169862747, + 0.030056914314627647, + 0.004588329698890448, + 0.0058031040243804455, + 0.0023072895128279924, + 0.003078942187130451, + 0.02146855928003788, + 0.016107618808746338, + 0.004362450446933508, + 0.007586822845041752, + 0.007745552808046341, + 0.006492182612419128, + 0.008912314660847187, + 0.005184405017644167, + 0.046638038009405136, + 0.05230150744318962, + 0.008268250152468681, + 0.010107003152370453, + 0.005153906065970659, + 0.0093608433380723, + 0.0035599488765001297, + 0.007205670233815908, + 0.02358756586909294, + 0.032313477247953415, + 0.0101527264341712, + 0.013957069255411625, + 0.0031485268846154213, + 0.006061318796128035, + 0.0027649221010506153, + 0.007830650545656681, + 0.011657575145363808, + 0.009535817429423332, + 0.00793482456356287, + 0.005519501399248838, + 0.0034282929264009, + 0.004924330394715071, + 0.003164295107126236, + 0.004254745785146952, + 0.007994246669113636, + 0.009201204404234886, + 0.006192581728100777, + 0.006630311720073223, + 0.017155086621642113, + 0.02697792276740074, + 0.020166290923953056, + 0.020308706909418106, + 0.03234734386205673, + 0.027602026239037514, + 0.02167394943535328, + 0.025905296206474304, + 0.00473567470908165, + 0.007836788892745972, + 0.005435817409306765, + 0.011389444582164288, + 0.010672620497643948, + 0.010157614015042782, + 0.012259070761501789, + 0.009257573634386063, + 0.0020738220773637295, + 0.0027081093285232782, + 0.002998266601935029, + 0.0029340621549636126, + 0.007135334424674511, + 0.004909282550215721, + 0.003982409834861755, + 0.0034394480753690004, + 0.005234656389802694, + 0.004268014803528786, + 0.008368069306015968, + 0.0048980494029819965, + 0.005884060636162758, + 0.005746867042034864, + 0.004745711572468281, + 0.005580039229243994 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 2 + ], + "selected_reward": -0.009782234206795692 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 in a high-contrast black-and-white cinematic style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera executing a low-angle tracking shot following a man in a suit sprinting down an empty asphalt city street?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment shrouded in torrential rain with puddled pavement and long, distorted shadows from street lamps?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the man's dark single-breasted suit appear completely drenched, clinging to his body to outline his running form?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the water splashes beneath the man's feet appear as crystalline white forms that shimmer silver in the darkness?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020577168092131615 + ], + "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": [ + 0.25050029158592224, + 0.029626598581671715, + 0.35430222749710083, + 0.11829261481761932, + 0.021184727549552917, + 0.021822843700647354, + 0.1415834128856659, + 0.026685109362006187, + 0.012698307633399963, + 0.023303832858800888 + ], + "credit_map": [ + 0.01889747567474842, + 0.029666338115930557, + 0.014777479693293571, + 0.01936589926481247, + 0.04030732437968254, + 0.061674147844314575, + 0.03243045508861542, + 0.03338117152452469, + 0.0025749830529093742, + 0.004488956183195114, + 0.0014702773187309504, + 0.0020853402093052864, + 0.00481374142691493, + 0.004383797757327557, + 0.004548538941890001, + 0.005260961130261421, + 0.022411860525608063, + 0.012990321964025497, + 0.028269954025745392, + 0.010723840445280075, + 0.11527059972286224, + 0.13038502633571625, + 0.014443060383200645, + 0.019807584583759308, + 0.008635750971734524, + 0.014295846223831177, + 0.004217916168272495, + 0.00537450797855854, + 0.032769203186035156, + 0.0360533744096756, + 0.006657052785158157, + 0.01028895378112793, + 0.0024581653997302055, + 0.001738357706926763, + 0.0012411343632265925, + 0.002611459232866764, + 0.003578422125428915, + 0.003573023248463869, + 0.002600386505946517, + 0.003383779665455222, + 0.0016122182132676244, + 0.0027311700396239758, + 0.001366212498396635, + 0.002576868049800396, + 0.0033138112630695105, + 0.003534819232299924, + 0.003178904764354229, + 0.0035088404547423124, + 0.012931939214468002, + 0.020063556730747223, + 0.013693789951503277, + 0.0142679987475276, + 0.024588407948613167, + 0.02061464451253414, + 0.01628721132874489, + 0.019135857000947, + 0.002708980580791831, + 0.0022164920810610056, + 0.0035636750981211662, + 0.0024945628829300404, + 0.0030051881913095713, + 0.004647329915314913, + 0.0033354093320667744, + 0.0047134715132415295, + 0.001016492722555995, + 0.0011342361103743315, + 0.0021406360901892185, + 0.0010488616535440087, + 0.002025651978328824, + 0.0023090646136552095, + 0.0012807128950953484, + 0.001742652035318315, + 0.002218940993770957, + 0.0024085871409624815, + 0.0040062591433525085, + 0.002433015964925289, + 0.0022286432795226574, + 0.00360116851516068, + 0.002720411866903305, + 0.003686806419864297 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 6 + ], + "selected_reward": -0.020577168092131615 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "dramatic, low-angle tracking shot follows a man in a suit as he sprints down an empty, asphalt city street shrouded in torrential rain. The scene is shot in high-contrast black-and-white cinematic style, with the light from the street lamps casting long, distorted shadows on the wet pavement, further accentuating the darkness. The man wears a well-tailored, dark, single-breasted suit, now completely drenched in rain. The damp fabric clings to his leaning body, outlining the tense muscles of his running form. His short, black hair clings to his forehead, his features are sharply defined, his lips pursed, his expression firm and urgent. Each step he takes lands heavily on the puddled pavement, his arms swinging vigorously for balance. The huge splashes of water beneath his feet freeze into crystalline white forms in the strong light, leaving behind him a trail of gradually fading white, striking in the darkness. The light accurately captures the raindrops, splashing water and the wet sheen of the suit, making them shimmer silver in the darkness, creating a tense, resolute and fatalistic visual atmosphere.", + "seed": 16464, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000056/group_10/index_3911_seed_16464_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": -0.016751788184046747, + "vlm_logprob": -0.016751788184046747, + "vlm_yes_no_margin": 4.4, + "vlm_token_credit": -0.016751788184046747, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 in a high-contrast black-and-white cinematic style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera executing a low-angle tracking shot following a man in a suit sprinting down an empty asphalt city street?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment shrouded in torrential rain with puddled pavement and long, distorted shadows from street lamps?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the man's dark single-breasted suit appear completely drenched, clinging to his body to outline his running form?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the water splashes beneath the man's feet appear as crystalline white forms that shimmer silver in the darkness?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.030036769807338715 + ], + "opposite_logprob_tokens": [ + -3.780036687850952 + ], + "logprob_reward": -0.030036769807338715, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.030036769807338715, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17233526706695557, + 0.04851902276277542, + 0.05451032519340515, + 0.20031480491161346, + 0.042775388807058334, + 0.08021262288093567, + 0.1383119374513626, + 0.10581117123365402, + 0.09441402554512024, + 0.06279541552066803 + ], + "credit_map": [ + 0.019180383533239365, + 0.025822235271334648, + 0.020403875038027763, + 0.021467672660946846, + 0.02763698808848858, + 0.028842797502875328, + 0.013607780449092388, + 0.015373536385595798, + 0.0062182811088860035, + 0.004257489461451769, + 0.006093764677643776, + 0.005884326528757811, + 0.00911496952176094, + 0.006432989612221718, + 0.003935139626264572, + 0.0065820603631436825, + 0.007701305206865072, + 0.006052220240235329, + 0.005803077947348356, + 0.005507448222488165, + 0.00925632007420063, + 0.008605311624705791, + 0.005538717843592167, + 0.006045927293598652, + 0.02293771132826805, + 0.02485021948814392, + 0.012536423280835152, + 0.014720424078404903, + 0.015252661891281605, + 0.06305506825447083, + 0.024189157411456108, + 0.02277311123907566, + 0.005980340298265219, + 0.005409555044025183, + 0.0054328604601323605, + 0.004349823575466871, + 0.006010720506310463, + 0.004590242635458708, + 0.005515567027032375, + 0.005486277863383293, + 0.008597454987466335, + 0.00612220773473382, + 0.005580844823271036, + 0.009447867050766945, + 0.012416915036737919, + 0.020047180354595184, + 0.008703278377652168, + 0.009296871721744537, + 0.017205439507961273, + 0.017881222069263458, + 0.012320420704782009, + 0.023821430280804634, + 0.01639566756784916, + 0.01983925700187683, + 0.012944212183356285, + 0.01790429651737213, + 0.011861835606396198, + 0.016203366219997406, + 0.013756093569099903, + 0.012317812070250511, + 0.01978646218776703, + 0.009358693845570087, + 0.01178586296737194, + 0.01074104756116867, + 0.012329142540693283, + 0.009874913841485977, + 0.007150003220885992, + 0.0068651242181658745, + 0.01578674651682377, + 0.011488491669297218, + 0.024163546040654182, + 0.0067560565657913685, + 0.008877243846654892, + 0.00806337408721447, + 0.006524133961647749, + 0.006171163637191057, + 0.010695121251046658, + 0.006909850984811783, + 0.00864969752728939, + 0.006904830224812031 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 6 + ], + "selected_reward": -0.030036769807338715 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 in a high-contrast black-and-white cinematic style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera executing a low-angle tracking shot following a man in a suit sprinting down an empty asphalt city street?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment shrouded in torrential rain with puddled pavement and long, distorted shadows from street lamps?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the man's dark single-breasted suit appear completely drenched, clinging to his body to outline his running form?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the water splashes beneath the man's feet appear as crystalline white forms that shimmer silver in the darkness?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012557132169604301 + ], + "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": [ + 0.16686572134494781, + 0.08495009690523148, + 0.1134081557393074, + 0.13332593441009521, + 0.05442865565419197, + 0.08056453615427017, + 0.10205450654029846, + 0.07039845734834671, + 0.14481905102729797, + 0.0491848923265934 + ], + "credit_map": [ + 0.0178135447204113, + 0.030392155051231384, + 0.01690070703625679, + 0.0179439727216959, + 0.023514345288276672, + 0.026018723845481873, + 0.012293476611375809, + 0.021988794207572937, + 0.008694215677678585, + 0.00842150580137968, + 0.012629623524844646, + 0.0108938654884696, + 0.008978116326034069, + 0.010983281768858433, + 0.006077500991523266, + 0.01827198639512062, + 0.018232854083180428, + 0.011216063052415848, + 0.015622969716787338, + 0.009959883987903595, + 0.014407556504011154, + 0.015582585707306862, + 0.006652852054685354, + 0.021733395755290985, + 0.016766203567385674, + 0.019513266161084175, + 0.008099515922367573, + 0.00915359053760767, + 0.009000569581985474, + 0.036827970296144485, + 0.015787798911333084, + 0.018177010118961334, + 0.007389299105852842, + 0.012666965834796429, + 0.005375961773097515, + 0.0037971469573676586, + 0.0060006920248270035, + 0.009150189347565174, + 0.004372671712189913, + 0.0056757270358502865, + 0.011009841226041317, + 0.005852550733834505, + 0.008955102413892746, + 0.0068656750954687595, + 0.01060329470783472, + 0.01849362440407276, + 0.007224530912935734, + 0.011559919454157352, + 0.012195378541946411, + 0.011248844675719738, + 0.007337071932852268, + 0.011713044717907906, + 0.010819714516401291, + 0.022263996303081512, + 0.008484253659844398, + 0.017992215231060982, + 0.00726518128067255, + 0.0071764010936021805, + 0.010103766806423664, + 0.010435669682919979, + 0.009043197147548199, + 0.008936535567045212, + 0.010202172212302685, + 0.007235527504235506, + 0.015508351847529411, + 0.012562818825244904, + 0.0219182837754488, + 0.016245055943727493, + 0.013659617863595486, + 0.014487046748399734, + 0.038520947098731995, + 0.011916929855942726, + 0.0057616508565843105, + 0.007158210966736078, + 0.0057852501049637794, + 0.005807156674563885, + 0.006660172250121832, + 0.005450427066534758, + 0.006811758037656546, + 0.005750270094722509 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 3 + ], + "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 scene shot in a high-contrast black-and-white cinematic style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera executing a low-angle tracking shot following a man in a suit sprinting down an empty asphalt city street?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment shrouded in torrential rain with puddled pavement and long, distorted shadows from street lamps?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the man's dark single-breasted suit appear completely drenched, clinging to his body to outline his running form?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the water splashes beneath the man's feet appear as crystalline white forms that shimmer silver in the darkness?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004637320525944233 + ], + "opposite_logprob_tokens": [ + -5.379637241363525 + ], + "logprob_reward": -0.004637320525944233, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004637320525944233, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12236261367797852, + 0.05289595201611519, + 0.06545718014240265, + 0.07528802752494812, + 0.1343081295490265, + 0.060966216027736664, + 0.07041028887033463, + 0.20318108797073364, + 0.11732479184865952, + 0.0978056937456131 + ], + "credit_map": [ + 0.015307567082345486, + 0.016957340762019157, + 0.01511583011597395, + 0.009601537138223648, + 0.014965965412557125, + 0.026805398985743523, + 0.007223744876682758, + 0.016385231167078018, + 0.006077524274587631, + 0.004822235554456711, + 0.008144871331751347, + 0.005872973240911961, + 0.007252507843077183, + 0.007347302511334419, + 0.003439078340306878, + 0.009939461946487427, + 0.007823909632861614, + 0.005075846333056688, + 0.01012462843209505, + 0.009522510692477226, + 0.007932782173156738, + 0.009152219630777836, + 0.004864609334617853, + 0.01096067763864994, + 0.010219743475317955, + 0.009609573520720005, + 0.0048083290457725525, + 0.005574303679168224, + 0.006507202051579952, + 0.019627967849373817, + 0.00923069752752781, + 0.009710208512842655, + 0.019433829933404922, + 0.03840571269392967, + 0.012772955931723118, + 0.006305431015789509, + 0.009222475811839104, + 0.02456747554242611, + 0.008230768144130707, + 0.015369468368589878, + 0.007723247166723013, + 0.004354042001068592, + 0.0033399511594325304, + 0.004891187883913517, + 0.008494729176163673, + 0.01776261255145073, + 0.005878168623894453, + 0.00852227583527565, + 0.011803991161286831, + 0.008053196594119072, + 0.005006264429539442, + 0.007437935099005699, + 0.00895551685243845, + 0.012965262867510319, + 0.005885569844394922, + 0.010302552953362465, + 0.02299758419394493, + 0.02293897606432438, + 0.0286153145134449, + 0.029873548075556755, + 0.024629322811961174, + 0.02793678641319275, + 0.029931407421827316, + 0.01625814102590084, + 0.02028740383684635, + 0.01271815039217472, + 0.013522066175937653, + 0.010882935486733913, + 0.017973776906728745, + 0.011289820075035095, + 0.021751049906015396, + 0.008899588137865067, + 0.009371492080390453, + 0.015201287344098091, + 0.009139925241470337, + 0.008312307298183441, + 0.015460527502000332, + 0.010649141855537891, + 0.014641901478171349, + 0.015029117465019226 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 4, + 0 + ], + "selected_reward": -0.004637320525944233 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 in a high-contrast black-and-white cinematic style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera executing a low-angle tracking shot following a man in a suit sprinting down an empty asphalt city street?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment shrouded in torrential rain with puddled pavement and long, distorted shadows from street lamps?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the man's dark single-breasted suit appear completely drenched, clinging to his body to outline his running form?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the water splashes beneath the man's feet appear as crystalline white forms that shimmer silver in the darkness?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006737728603184223 + ], + "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": [ + 0.09973405301570892, + 0.08281483501195908, + 0.0445672944188118, + 0.051487937569618225, + 0.11840099096298218, + 0.272756963968277, + 0.06789865344762802, + 0.11100923269987106, + 0.075973741710186, + 0.07535623013973236 + ], + "credit_map": [ + 0.010073070414364338, + 0.014147242531180382, + 0.01158423162996769, + 0.011391588486731052, + 0.01296453271061182, + 0.01496327668428421, + 0.011096588335931301, + 0.013513541780412197, + 0.010386908426880836, + 0.008406615816056728, + 0.010535536333918571, + 0.01038987748324871, + 0.016433967277407646, + 0.010431778617203236, + 0.007525891996920109, + 0.00870426557958126, + 0.00426473980769515, + 0.004304032307118177, + 0.004872298799455166, + 0.005889919586479664, + 0.008195744827389717, + 0.0068356613628566265, + 0.003893765155225992, + 0.006311135832220316, + 0.006651409901678562, + 0.0064344885759055614, + 0.002999482210725546, + 0.003419445361942053, + 0.004820862784981728, + 0.01637260615825653, + 0.006514395587146282, + 0.004275250714272261, + 0.023508120328187943, + 0.02842102013528347, + 0.00833043735474348, + 0.008553040213882923, + 0.012346130795776844, + 0.018800247460603714, + 0.0075803110376000404, + 0.010861692018806934, + 0.03510650619864464, + 0.05402425676584244, + 0.02349398285150528, + 0.03166211023926735, + 0.03577255457639694, + 0.03140208497643471, + 0.026910755783319473, + 0.03438474237918854, + 0.007927394472062588, + 0.009520039893686771, + 0.006697044242173433, + 0.009132889099419117, + 0.009649191983044147, + 0.009461923502385616, + 0.0072351074777543545, + 0.00827507022768259, + 0.010812109336256981, + 0.015831220895051956, + 0.019045185297727585, + 0.012059672735631466, + 0.018874747678637505, + 0.010140059515833855, + 0.013438982889056206, + 0.010807263664901257, + 0.008962931111454964, + 0.0077844541519880295, + 0.006626259535551071, + 0.005415389779955149, + 0.005885208025574684, + 0.010322651825845242, + 0.02587437443435192, + 0.00510248402133584, + 0.008098674938082695, + 0.012378105893731117, + 0.008531266823410988, + 0.008785002864897251, + 0.01036074012517929, + 0.007049973122775555, + 0.012944070622324944, + 0.00720840273424983 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 4, + 7 + ], + "selected_reward": -0.006737728603184223 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 in a high-contrast black-and-white cinematic style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera executing a low-angle tracking shot following a man in a suit sprinting down an empty asphalt city street?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment shrouded in torrential rain with puddled pavement and long, distorted shadows from street lamps?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the man's dark single-breasted suit appear completely drenched, clinging to his body to outline his running form?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the water splashes beneath the man's feet appear as crystalline white forms that shimmer silver in the darkness?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029789989814162254 + ], + "opposite_logprob_tokens": [ + -3.529789924621582 + ], + "logprob_reward": -0.029789989814162254, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029789989814162254, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1100868210196495, + 0.17052258551120758, + 0.07763699442148209, + 0.03433791175484657, + 0.04303254559636116, + 0.07381342351436615, + 0.04125044122338295, + 0.17463237047195435, + 0.20168083906173706, + 0.07300613075494766 + ], + "credit_map": [ + 0.013317626900970936, + 0.01677635870873928, + 0.011392851360142231, + 0.009682903066277504, + 0.01665356568992138, + 0.021334722638130188, + 0.0075175645761191845, + 0.013411234132945538, + 0.02283622696995735, + 0.013729609549045563, + 0.0340069979429245, + 0.01803688146173954, + 0.017712051048874855, + 0.02063986472785473, + 0.011103727854788303, + 0.03245721384882927, + 0.013566168025135994, + 0.008380410261452198, + 0.008557964116334915, + 0.008074604906141758, + 0.00967367459088564, + 0.01073954813182354, + 0.006518988404422998, + 0.012125637382268906, + 0.006266766227781773, + 0.005226029083132744, + 0.0027564435731619596, + 0.0022924495860934258, + 0.003753466298803687, + 0.006939596030861139, + 0.003679953282698989, + 0.003423209534958005, + 0.010720868594944477, + 0.005620919167995453, + 0.004663995001465082, + 0.003494391916319728, + 0.003115670057013631, + 0.005868751555681229, + 0.005353683605790138, + 0.004194264765828848, + 0.008910231292247772, + 0.005802076309919357, + 0.004597686231136322, + 0.005759188439697027, + 0.012800606898963451, + 0.019357196986675262, + 0.007824108004570007, + 0.00876233447343111, + 0.004531738348305225, + 0.006357943639159203, + 0.004630116280168295, + 0.004549901932477951, + 0.005991258192807436, + 0.005907249171286821, + 0.004498250316828489, + 0.004783982411026955, + 0.018955647945404053, + 0.02449478767812252, + 0.02937815897166729, + 0.024151107296347618, + 0.022042645141482353, + 0.02086447738111019, + 0.018785789608955383, + 0.015959756448864937, + 0.019469615072011948, + 0.022153615951538086, + 0.020462725311517715, + 0.019580580294132233, + 0.012386517599225044, + 0.028998272493481636, + 0.06389035284519196, + 0.014739157631993294, + 0.011682799085974693, + 0.009889586828649044, + 0.008483397774398327, + 0.009204166010022163, + 0.011098598130047321, + 0.007010630331933498, + 0.007942521013319492, + 0.007694424130022526 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 1 + ], + "selected_reward": -0.029789989814162254 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "dramatic, low-angle tracking shot follows a man in a suit as he sprints down an empty, asphalt city street shrouded in torrential rain. The scene is shot in high-contrast black-and-white cinematic style, with the light from the street lamps casting long, distorted shadows on the wet pavement, further accentuating the darkness. The man wears a well-tailored, dark, single-breasted suit, now completely drenched in rain. The damp fabric clings to his leaning body, outlining the tense muscles of his running form. His short, black hair clings to his forehead, his features are sharply defined, his lips pursed, his expression firm and urgent. Each step he takes lands heavily on the puddled pavement, his arms swinging vigorously for balance. The huge splashes of water beneath his feet freeze into crystalline white forms in the strong light, leaving behind him a trail of gradually fading white, striking in the darkness. The light accurately captures the raindrops, splashing water and the wet sheen of the suit, making them shimmer silver in the darkness, creating a tense, resolute and fatalistic visual atmosphere.", + "seed": 16465, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000056/group_10/index_3911_seed_16465_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": -0.0147630020044744, + "vlm_logprob": -0.0147630020044744, + "vlm_yes_no_margin": 4.425, + "vlm_token_credit": -0.0147630020044744, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 in a high-contrast black-and-white cinematic style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera executing a low-angle tracking shot following a man in a suit sprinting down an empty asphalt city street?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment shrouded in torrential rain with puddled pavement and long, distorted shadows from street lamps?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the man's dark single-breasted suit appear completely drenched, clinging to his body to outline his running form?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the water splashes beneath the man's feet appear as crystalline white forms that shimmer silver in the darkness?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01728939637541771 + ], + "opposite_logprob_tokens": [ + -4.767289161682129 + ], + "logprob_reward": -0.01728939637541771, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.01728939637541771, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2952059507369995, + 0.06562307476997375, + 0.0523938424885273, + 0.10130005329847336, + 0.14563289284706116, + 0.04307126998901367, + 0.08594530820846558, + 0.11765611171722412, + 0.05355687439441681, + 0.039614662528038025 + ], + "credit_map": [ + 0.06839016824960709, + 0.047571130096912384, + 0.021787181496620178, + 0.0208567064255476, + 0.025804009288549423, + 0.04438319802284241, + 0.024144986644387245, + 0.0422685481607914, + 0.006389888934791088, + 0.009525388479232788, + 0.005876368377357721, + 0.009017812088131905, + 0.009582532569766045, + 0.008234703913331032, + 0.00849891733378172, + 0.008497464470565319, + 0.008490920066833496, + 0.0062667313031852245, + 0.005029695108532906, + 0.005833586677908897, + 0.005861358717083931, + 0.006288849748671055, + 0.0058201574720442295, + 0.00880253966897726, + 0.014877238310873508, + 0.01436540950089693, + 0.008853008970618248, + 0.008954965509474277, + 0.011894099414348602, + 0.013119078241288662, + 0.008669410832226276, + 0.02056683413684368, + 0.021350882947444916, + 0.028444483876228333, + 0.010151091031730175, + 0.015621056780219078, + 0.015920666977763176, + 0.0217356588691473, + 0.017975738272070885, + 0.01443331129848957, + 0.006829115096479654, + 0.005722414702177048, + 0.003640620270743966, + 0.002740125637501478, + 0.0039056139066815376, + 0.005970403086394072, + 0.0031042525079101324, + 0.011158723384141922, + 0.01045545656234026, + 0.014274557121098042, + 0.007996940053999424, + 0.008371110074222088, + 0.012836973182857037, + 0.012923679314553738, + 0.010462911799550056, + 0.008623678237199783, + 0.013722477480769157, + 0.01574556529521942, + 0.010634304024279118, + 0.01154777780175209, + 0.01910616271197796, + 0.01505239587277174, + 0.014690432697534561, + 0.017156988382339478, + 0.005161983426660299, + 0.010718842037022114, + 0.003801095299422741, + 0.00473926542326808, + 0.007177930325269699, + 0.007822506129741669, + 0.0051695965230464935, + 0.008965656161308289, + 0.004753895103931427, + 0.0040988377295434475, + 0.0063322060741484165, + 0.003934276755899191, + 0.005320414435118437, + 0.004675466567277908, + 0.004606184083968401, + 0.005893383640795946 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 7 + ], + "selected_reward": -0.01728939637541771 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 in a high-contrast black-and-white cinematic style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera executing a low-angle tracking shot following a man in a suit sprinting down an empty asphalt city street?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment shrouded in torrential rain with puddled pavement and long, distorted shadows from street lamps?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the man's dark single-breasted suit appear completely drenched, clinging to his body to outline his running form?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the water splashes beneath the man's feet appear as crystalline white forms that shimmer silver in the darkness?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018206853419542313 + ], + "opposite_logprob_tokens": [ + -4.01820707321167 + ], + "logprob_reward": -0.018206853419542313, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018206853419542313, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15200263261795044, + 0.07597675174474716, + 0.06370442360639572, + 0.10834911465644836, + 0.29678380489349365, + 0.05064799264073372, + 0.07996170222759247, + 0.0776817575097084, + 0.03410503268241882, + 0.060786787420511246 + ], + "credit_map": [ + 0.021465126425027847, + 0.022578585892915726, + 0.015244526788592339, + 0.011606918647885323, + 0.03281746804714203, + 0.017658086493611336, + 0.014797995798289776, + 0.01583392173051834, + 0.008015421219170094, + 0.014207866974174976, + 0.008220887742936611, + 0.008933275938034058, + 0.00914186891168356, + 0.009315616451203823, + 0.008017376065254211, + 0.010124442167580128, + 0.008782594464719296, + 0.008078046143054962, + 0.008129340596497059, + 0.00793527252972126, + 0.005181659013032913, + 0.007630036678165197, + 0.00971523579210043, + 0.008252246305346489, + 0.014049798250198364, + 0.01133593637496233, + 0.008091877214610577, + 0.015125993639230728, + 0.017238259315490723, + 0.014059866778552532, + 0.011796429753303528, + 0.016650961712002754, + 0.04721415415406227, + 0.0584334097802639, + 0.020097119733691216, + 0.024004528298974037, + 0.037029463797807693, + 0.04168004170060158, + 0.03410876914858818, + 0.03421632945537567, + 0.0060562980361282825, + 0.007984177209436893, + 0.005003728438168764, + 0.0054697440937161446, + 0.007141904905438423, + 0.0059351553209125996, + 0.006081358063966036, + 0.006975631695240736, + 0.008182976394891739, + 0.009840314276516438, + 0.005898216273635626, + 0.006769025232642889, + 0.016076048836112022, + 0.012997704558074474, + 0.010577737353742123, + 0.00961968582123518, + 0.005817190743982792, + 0.008751918561756611, + 0.0037015732377767563, + 0.009260797873139381, + 0.015372918918728828, + 0.010968807153403759, + 0.010211236774921417, + 0.013597317039966583, + 0.0025756065733730793, + 0.0072890338487923145, + 0.0030071912333369255, + 0.002325378591194749, + 0.006178195588290691, + 0.005020135082304478, + 0.003242583479732275, + 0.004466912243515253, + 0.005076655652374029, + 0.00546913780272007, + 0.0037267094012349844, + 0.008308710530400276, + 0.009817427955567837, + 0.008478831499814987, + 0.010354087688028812, + 0.009555233642458916 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 3 + ], + "selected_reward": -0.018206853419542313 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 in a high-contrast black-and-white cinematic style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera executing a low-angle tracking shot following a man in a suit sprinting down an empty asphalt city street?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment shrouded in torrential rain with puddled pavement and long, distorted shadows from street lamps?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the man's dark single-breasted suit appear completely drenched, clinging to his body to outline his running form?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the water splashes beneath the man's feet appear as crystalline white forms that shimmer silver in the darkness?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0059483409859240055 + ], + "opposite_logprob_tokens": [ + -5.130948543548584 + ], + "logprob_reward": -0.0059483409859240055, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.0059483409859240055, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10137628018856049, + 0.22258667647838593, + 0.045056767761707306, + 0.039336908608675, + 0.37015897035598755, + 0.046543654054403305, + 0.07839685678482056, + 0.0318305566906929, + 0.022061046212911606, + 0.0426521971821785 + ], + "credit_map": [ + 0.012717229314148426, + 0.015176018700003624, + 0.011033265851438046, + 0.015028120949864388, + 0.011183331720530987, + 0.011275402270257473, + 0.010538795031607151, + 0.014424133114516735, + 0.03523029386997223, + 0.06943690031766891, + 0.018072394654154778, + 0.017173875123262405, + 0.014281091280281544, + 0.038099829107522964, + 0.010947714559733868, + 0.01934458315372467, + 0.006842466536909342, + 0.005710837431252003, + 0.004170773550868034, + 0.005064206197857857, + 0.004845521412789822, + 0.00584412319585681, + 0.0056938547641038895, + 0.0068849907256662846, + 0.006039282772690058, + 0.004778570495545864, + 0.003521634265780449, + 0.0042364466935396194, + 0.0038184088189154863, + 0.0050615062937140465, + 0.004001332446932793, + 0.007879729382693768, + 0.07049334794282913, + 0.08401719480752945, + 0.023862574249505997, + 0.025716232135891914, + 0.03914114460349083, + 0.051741521805524826, + 0.03172074630856514, + 0.0434662401676178, + 0.005453412886708975, + 0.006175285205245018, + 0.007928011007606983, + 0.004767595790326595, + 0.004467124585062265, + 0.005516854580491781, + 0.0038989426102489233, + 0.00833643414080143, + 0.011466816067695618, + 0.01101944874972105, + 0.006630835123360157, + 0.005350781138986349, + 0.00795885268598795, + 0.013557184487581253, + 0.009857776574790478, + 0.012555169872939587, + 0.004335326608270407, + 0.00559525890275836, + 0.003420941298827529, + 0.0034755845554172993, + 0.0036562916357070208, + 0.0041373237036168575, + 0.0026222998276352882, + 0.004587531089782715, + 0.001737862010486424, + 0.003457306418567896, + 0.0015772448386996984, + 0.0022302919533103704, + 0.004550071433186531, + 0.002865962916985154, + 0.002797337481752038, + 0.0028449702076613903, + 0.004593800753355026, + 0.003392923856154084, + 0.007111282553523779, + 0.0040966784581542015, + 0.006273819133639336, + 0.005542283412069082, + 0.004174721892923117, + 0.007466690149158239 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 1, + 0 + ], + "selected_reward": -0.0059483409859240055 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 in a high-contrast black-and-white cinematic style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera executing a low-angle tracking shot following a man in a suit sprinting down an empty asphalt city street?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment shrouded in torrential rain with puddled pavement and long, distorted shadows from street lamps?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the man's dark single-breasted suit appear completely drenched, clinging to his body to outline his running form?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the water splashes beneath the man's feet appear as crystalline white forms that shimmer silver in the darkness?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014192831702530384 + ], + "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": [ + 0.22445671260356903, + 0.1294701248407364, + 0.06646563112735748, + 0.06398110091686249, + 0.11493546515703201, + 0.0577736534178257, + 0.08154378086328506, + 0.0762345939874649, + 0.14423485100269318, + 0.040904153138399124 + ], + "credit_map": [ + 0.037039898335933685, + 0.030161410570144653, + 0.02414974384009838, + 0.021074393764138222, + 0.02651819959282875, + 0.032311875373125076, + 0.027382396161556244, + 0.025818802416324615, + 0.017507607117295265, + 0.03820570185780525, + 0.009002131409943104, + 0.009809974581003189, + 0.01414122898131609, + 0.019782982766628265, + 0.00842293817549944, + 0.012597562745213509, + 0.010188218206167221, + 0.008899700827896595, + 0.006548215635120869, + 0.008067319169640541, + 0.004981122445315123, + 0.00925108976662159, + 0.00915409717708826, + 0.00937587022781372, + 0.007243370171636343, + 0.006350407842546701, + 0.007212117314338684, + 0.006883428432047367, + 0.009555087424814701, + 0.008415838703513145, + 0.006902392487972975, + 0.011418459005653858, + 0.022911544889211655, + 0.02519548498094082, + 0.006739291828125715, + 0.009153402410447598, + 0.01345977932214737, + 0.014547827653586864, + 0.012544379569590092, + 0.010383743792772293, + 0.00562296574935317, + 0.006452250760048628, + 0.003640171140432358, + 0.006213259883224964, + 0.011869611218571663, + 0.007646837737411261, + 0.00832277163863182, + 0.008005782030522823, + 0.009758196771144867, + 0.011916917748749256, + 0.007446610368788242, + 0.005638387519866228, + 0.01080147735774517, + 0.012554873712360859, + 0.012512034736573696, + 0.010915281251072884, + 0.007706670090556145, + 0.012235553935170174, + 0.003853798843920231, + 0.01288711279630661, + 0.012767448090016842, + 0.010110356844961643, + 0.007319975644350052, + 0.009353676810860634, + 0.009266261011362076, + 0.011412808671593666, + 0.008210279047489166, + 0.01721014641225338, + 0.0339571088552475, + 0.018823644146323204, + 0.02321338653564453, + 0.02214120142161846, + 0.003270229324698448, + 0.0038397065363824368, + 0.0032669443171471357, + 0.004507214762270451, + 0.009533941745758057, + 0.005962796974927187, + 0.005560838617384434, + 0.004962481092661619 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 1 + ], + "selected_reward": -0.014192831702530384 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 in a high-contrast black-and-white cinematic style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera executing a low-angle tracking shot following a man in a suit sprinting down an empty asphalt city street?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment shrouded in torrential rain with puddled pavement and long, distorted shadows from street lamps?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the man's dark single-breasted suit appear completely drenched, clinging to his body to outline his running form?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the water splashes beneath the man's feet appear as crystalline white forms that shimmer silver in the darkness?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018177587538957596 + ], + "opposite_logprob_tokens": [ + -4.018177509307861 + ], + "logprob_reward": -0.018177587538957596, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018177587538957596, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23423950374126434, + 0.27333176136016846, + 0.0482499934732914, + 0.056920625269412994, + 0.05115462839603424, + 0.0891735628247261, + 0.08279233425855637, + 0.05656914785504341, + 0.06347869336605072, + 0.044089701026678085 + ], + "credit_map": [ + 0.039927203208208084, + 0.037893570959568024, + 0.017867853865027428, + 0.026436200365424156, + 0.024453463032841682, + 0.028954410925507545, + 0.02038038708269596, + 0.03832640498876572, + 0.04424762725830078, + 0.061179373413324356, + 0.023716570809483528, + 0.028693895787000656, + 0.022336430847644806, + 0.03708407282829285, + 0.02480211667716503, + 0.03127168491482735, + 0.008477278985083103, + 0.0069878618232905865, + 0.0032101599499583244, + 0.005819991230964661, + 0.004982836544513702, + 0.006698453798890114, + 0.005247773136943579, + 0.006825638003647327, + 0.007694749627262354, + 0.006030723452568054, + 0.007600252982228994, + 0.005271612666547298, + 0.006303529720753431, + 0.0082605155184865, + 0.005970955826342106, + 0.009788284078240395, + 0.006455052178353071, + 0.007760170381516218, + 0.004725432023406029, + 0.005238085053861141, + 0.007215173915028572, + 0.006430742796510458, + 0.006776296067982912, + 0.006553675513714552, + 0.014118720777332783, + 0.012106996960937977, + 0.01008053869009018, + 0.006201194599270821, + 0.008620028384029865, + 0.01164721418172121, + 0.008194097317755222, + 0.018204770982265472, + 0.012184500694274902, + 0.01300980243831873, + 0.006605400703847408, + 0.006904826033860445, + 0.011071487329900265, + 0.01283499225974083, + 0.010125245898962021, + 0.010056083090603352, + 0.005618805065751076, + 0.007834038697183132, + 0.0045640477910637856, + 0.006662504281848669, + 0.010686557739973068, + 0.00811947975307703, + 0.006223894655704498, + 0.006859824061393738, + 0.005858430173248053, + 0.010668625123798847, + 0.005240513011813164, + 0.006935675162822008, + 0.01018836535513401, + 0.00814133882522583, + 0.006689866539090872, + 0.009755884297192097, + 0.004581639543175697, + 0.005132816266268492, + 0.007266216445714235, + 0.006127489730715752, + 0.005272787529975176, + 0.005586425308138132, + 0.003916068933904171, + 0.0062062558718025684 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 5 + ], + "selected_reward": -0.018177587538957596 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "dramatic, low-angle tracking shot follows a man in a suit as he sprints down an empty, asphalt city street shrouded in torrential rain. The scene is shot in high-contrast black-and-white cinematic style, with the light from the street lamps casting long, distorted shadows on the wet pavement, further accentuating the darkness. The man wears a well-tailored, dark, single-breasted suit, now completely drenched in rain. The damp fabric clings to his leaning body, outlining the tense muscles of his running form. His short, black hair clings to his forehead, his features are sharply defined, his lips pursed, his expression firm and urgent. Each step he takes lands heavily on the puddled pavement, his arms swinging vigorously for balance. The huge splashes of water beneath his feet freeze into crystalline white forms in the strong light, leaving behind him a trail of gradually fading white, striking in the darkness. The light accurately captures the raindrops, splashing water and the wet sheen of the suit, making them shimmer silver in the darkness, creating a tense, resolute and fatalistic visual atmosphere.", + "seed": 16466, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000056/group_10/index_3911_seed_16466_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": -0.018104939628392457, + "vlm_logprob": -0.018104939628392457, + "vlm_yes_no_margin": 4.4, + "vlm_token_credit": -0.018104939628392457, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 in a high-contrast black-and-white cinematic style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera executing a low-angle tracking shot following a man in a suit sprinting down an empty asphalt city street?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment shrouded in torrential rain with puddled pavement and long, distorted shadows from street lamps?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the man's dark single-breasted suit appear completely drenched, clinging to his body to outline his running form?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the water splashes beneath the man's feet appear as crystalline white forms that shimmer silver in the darkness?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01740538328886032 + ], + "opposite_logprob_tokens": [ + -4.6424055099487305 + ], + "logprob_reward": -0.01740538328886032, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.01740538328886032, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13025981187820435, + 0.19732043147087097, + 0.06332643330097198, + 0.038533758372068405, + 0.07688778638839722, + 0.3261292576789856, + 0.033281970769166946, + 0.03292493149638176, + 0.06110899895429611, + 0.040226612240076065 + ], + "credit_map": [ + 0.018120359629392624, + 0.016207735985517502, + 0.016857799142599106, + 0.014242848381400108, + 0.018625373020768166, + 0.020324843004345894, + 0.01165985967963934, + 0.01422099769115448, + 0.028935467824339867, + 0.038990166038274765, + 0.032598502933979034, + 0.0282390546053648, + 0.01588909514248371, + 0.020575493574142456, + 0.013033771887421608, + 0.019058875739574432, + 0.007855355739593506, + 0.009250405244529247, + 0.016278116032481194, + 0.006510848179459572, + 0.004988348577171564, + 0.005404675379395485, + 0.005838805343955755, + 0.007199881132692099, + 0.0037210837472230196, + 0.007754719816148281, + 0.004347439389675856, + 0.004907527472823858, + 0.0040473369881510735, + 0.00543492566794157, + 0.005205138120800257, + 0.0031155897304415703, + 0.009121118113398552, + 0.005561067722737789, + 0.02223849855363369, + 0.007815361954271793, + 0.004357511643320322, + 0.007685994729399681, + 0.006897160783410072, + 0.013211076147854328, + 0.09515145421028137, + 0.01481679268181324, + 0.12394021451473236, + 0.013717691414058208, + 0.01152548473328352, + 0.0216666329652071, + 0.01733265444636345, + 0.02797831781208515, + 0.004126374144107103, + 0.003251797752454877, + 0.006805306766182184, + 0.0038746448699384928, + 0.0033853300847113132, + 0.0030309364665299654, + 0.004646973684430122, + 0.0041606067679822445, + 0.006683948915451765, + 0.0037179840728640556, + 0.006745403166860342, + 0.0031339721754193306, + 0.0036387909203767776, + 0.0029384621884673834, + 0.003156195627525449, + 0.0029101758264005184, + 0.01115262508392334, + 0.005018949508666992, + 0.017173146829009056, + 0.007867644540965557, + 0.006527286022901535, + 0.006682324688881636, + 0.003117890562862158, + 0.0035691310185939074, + 0.002533895894885063, + 0.003419332206249237, + 0.007360030431300402, + 0.005652065854519606, + 0.0061644804663956165, + 0.005187718663364649, + 0.004438941832631826, + 0.0054701450280845165 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 1, + 0 + ], + "selected_reward": -0.01740538328886032 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 in a high-contrast black-and-white cinematic style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera executing a low-angle tracking shot following a man in a suit sprinting down an empty asphalt city street?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment shrouded in torrential rain with puddled pavement and long, distorted shadows from street lamps?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the man's dark single-breasted suit appear completely drenched, clinging to his body to outline his running form?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the water splashes beneath the man's feet appear as crystalline white forms that shimmer silver in the darkness?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.038115888833999634 + ], + "opposite_logprob_tokens": [ + -3.288115978240967 + ], + "logprob_reward": -0.038115888833999634, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.038115888833999634, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09445194154977798, + 0.0922192707657814, + 0.35056743025779724, + 0.13955473899841309, + 0.02540316991508007, + 0.16780678927898407, + 0.01704052835702896, + 0.02905149571597576, + 0.034887392073869705, + 0.04901726543903351 + ], + "credit_map": [ + 0.012771520763635635, + 0.010984132997691631, + 0.019052349030971527, + 0.0099837901070714, + 0.012372730299830437, + 0.010729202069342136, + 0.010543441399931908, + 0.008014778606593609, + 0.011967840604484081, + 0.020830053836107254, + 0.01524331048130989, + 0.014498181641101837, + 0.008036521263420582, + 0.008670318871736526, + 0.006414574105292559, + 0.006558464840054512, + 0.043620720505714417, + 0.048201292753219604, + 0.09432020783424377, + 0.038191135972738266, + 0.030394526198506355, + 0.028754787519574165, + 0.03413065895438194, + 0.03295412287116051, + 0.01536177471280098, + 0.030330142006278038, + 0.018354341387748718, + 0.013225731439888477, + 0.014799377880990505, + 0.024047544226050377, + 0.01079980656504631, + 0.012636013329029083, + 0.002648681867867708, + 0.002211714629083872, + 0.005653666332364082, + 0.003058597445487976, + 0.0024496051482856274, + 0.0027698727790266275, + 0.002214370761066675, + 0.004396661650389433, + 0.032737672328948975, + 0.017972147092223167, + 0.03798237070441246, + 0.014371386729180813, + 0.011212633922696114, + 0.019160106778144836, + 0.014725408516824245, + 0.01964505761861801, + 0.002112689893692732, + 0.0015139700844883919, + 0.004112839233130217, + 0.0013978551141917706, + 0.0024024886079132557, + 0.0015632702270522714, + 0.002145063364878297, + 0.001792351366020739, + 0.0051737078465521336, + 0.0026907941792160273, + 0.006260276306420565, + 0.0037503372877836227, + 0.0022730042692273855, + 0.003080429509282112, + 0.002843028400093317, + 0.0029799186158925295, + 0.0065711745992302895, + 0.003239769022911787, + 0.0074360426515340805, + 0.003946586512029171, + 0.0047240727581083775, + 0.002864815527573228, + 0.002786983037367463, + 0.003317947033792734, + 0.002145201899111271, + 0.002899123588576913, + 0.007034044247120619, + 0.006536559201776981, + 0.014607882127165794, + 0.010207241401076317, + 0.003252452239394188, + 0.0023347612004727125 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 5, + 3 + ], + "selected_reward": -0.038115888833999634 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 in a high-contrast black-and-white cinematic style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera executing a low-angle tracking shot following a man in a suit sprinting down an empty asphalt city street?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment shrouded in torrential rain with puddled pavement and long, distorted shadows from street lamps?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the man's dark single-breasted suit appear completely drenched, clinging to his body to outline his running form?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the water splashes beneath the man's feet appear as crystalline white forms that shimmer silver in the darkness?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004094432573765516 + ], + "opposite_logprob_tokens": [ + -5.50409460067749 + ], + "logprob_reward": -0.004094432573765516, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004094432573765516, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10408537834882736, + 0.14162097871303558, + 0.07536721229553223, + 0.1293456107378006, + 0.10080981999635696, + 0.22430483996868134, + 0.018465643748641014, + 0.02048518694937229, + 0.034080423414707184, + 0.15143486857414246 + ], + "credit_map": [ + 0.008861653506755829, + 0.00997925829142332, + 0.011888758279383183, + 0.012241987511515617, + 0.018585456535220146, + 0.015432562679052353, + 0.013921483419835567, + 0.013174225576221943, + 0.01783565990626812, + 0.03205898031592369, + 0.0260806605219841, + 0.020428717136383057, + 0.013249050825834274, + 0.013806496746838093, + 0.01000421866774559, + 0.008157198317348957, + 0.009119825437664986, + 0.010228592902421951, + 0.01578347757458687, + 0.009119940921664238, + 0.008515654131770134, + 0.005114381201565266, + 0.009147449396550655, + 0.008337891660630703, + 0.01330638024955988, + 0.02737417444586754, + 0.0174845140427351, + 0.013847341760993004, + 0.012050896883010864, + 0.023630861192941666, + 0.01028047688305378, + 0.011370981112122536, + 0.010560167953372002, + 0.005914126057177782, + 0.03646863251924515, + 0.008706831373274326, + 0.00486207427456975, + 0.01129012368619442, + 0.005857398267835379, + 0.01715046726167202, + 0.06770197302103043, + 0.012170148082077503, + 0.08475830405950546, + 0.00939563475549221, + 0.007126601878553629, + 0.013459005393087864, + 0.00879841297864914, + 0.020894769579172134, + 0.0021048467606306076, + 0.0018534036353230476, + 0.002204104093834758, + 0.002071499591693282, + 0.002776889828965068, + 0.0018625064985826612, + 0.00347368186339736, + 0.0021187139209359884, + 0.003015021560713649, + 0.0019058403559029102, + 0.004626305773854256, + 0.0022463228087872267, + 0.0020036513451486826, + 0.002163124270737171, + 0.0023680829908698797, + 0.002156841102987528, + 0.005283188074827194, + 0.003533804789185524, + 0.006616809871047735, + 0.0058264220133423805, + 0.00351573433727026, + 0.0031946073286235332, + 0.002871360396966338, + 0.0032385012600570917, + 0.008132251910865307, + 0.00931883230805397, + 0.03265291824936867, + 0.02303355000913143, + 0.03132186084985733, + 0.02484027110040188, + 0.011525099165737629, + 0.010610097087919712 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 9, + 1 + ], + "selected_reward": -0.004094432573765516 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 in a high-contrast black-and-white cinematic style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera executing a low-angle tracking shot following a man in a suit sprinting down an empty asphalt city street?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment shrouded in torrential rain with puddled pavement and long, distorted shadows from street lamps?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the man's dark single-breasted suit appear completely drenched, clinging to his body to outline his running form?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the water splashes beneath the man's feet appear as crystalline white forms that shimmer silver in the darkness?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007629304192960262 + ], + "opposite_logprob_tokens": [ + -4.88262939453125 + ], + "logprob_reward": -0.007629304192960262, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007629304192960262, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08454466611146927, + 0.04064572602510452, + 0.4690307378768921, + 0.036259569227695465, + 0.04092792794108391, + 0.23209641873836517, + 0.015865245833992958, + 0.015006745234131813, + 0.011495810002088547, + 0.05412721261382103 + ], + "credit_map": [ + 0.011402644217014313, + 0.010557021014392376, + 0.010711271315813065, + 0.008079592138528824, + 0.012899879366159439, + 0.011619343422353268, + 0.008477097377181053, + 0.010797811672091484, + 0.005605091340839863, + 0.005853711627423763, + 0.005925772711634636, + 0.00499557564035058, + 0.0036329461727291346, + 0.006039369385689497, + 0.003440889297053218, + 0.005152370780706406, + 0.05571446567773819, + 0.06509103626012802, + 0.13113951683044434, + 0.044928163290023804, + 0.039809904992580414, + 0.03843916952610016, + 0.04857306927442551, + 0.045335400849580765, + 0.0034524775110185146, + 0.007170261815190315, + 0.004167849197983742, + 0.004330627154558897, + 0.004342402331531048, + 0.005689909681677818, + 0.00467223534360528, + 0.002433805726468563, + 0.004051589872688055, + 0.0034259979147464037, + 0.012167935259640217, + 0.003833763999864459, + 0.0025646379217505455, + 0.0037182308733463287, + 0.004588515497744083, + 0.006577257066965103, + 0.04921848326921463, + 0.023191383108496666, + 0.058269355446100235, + 0.019143596291542053, + 0.014375810511410236, + 0.025040926411747932, + 0.016625160351395607, + 0.02623169869184494, + 0.001401898916810751, + 0.0012607285752892494, + 0.0017401898512616754, + 0.002049488015472889, + 0.002244815230369568, + 0.0015777584630995989, + 0.0031370846554636955, + 0.002453280845656991, + 0.0024883945006877184, + 0.0012634099693968892, + 0.0035190547350794077, + 0.0013874401338398457, + 0.001464864588342607, + 0.0014962641289457679, + 0.0016691595083102584, + 0.0017181577859446406, + 0.0017627928173169494, + 0.0011251990217715502, + 0.0014450086746364832, + 0.0014245671918615699, + 0.001675403444096446, + 0.0015321591636165977, + 0.0013457626337185502, + 0.001184917287901044, + 0.002265502931550145, + 0.0025536641478538513, + 0.006082805804908276, + 0.00609429320320487, + 0.015359499491751194, + 0.011956761591136456, + 0.005017055198550224, + 0.0047976309433579445 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 5, + 0 + ], + "selected_reward": -0.007629304192960262 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 in a high-contrast black-and-white cinematic style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera executing a low-angle tracking shot following a man in a suit sprinting down an empty asphalt city street?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment shrouded in torrential rain with puddled pavement and long, distorted shadows from street lamps?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the man's dark single-breasted suit appear completely drenched, clinging to his body to outline his running form?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the water splashes beneath the man's feet appear as crystalline white forms that shimmer silver in the darkness?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023279689252376556 + ], + "opposite_logprob_tokens": [ + -3.7732796669006348 + ], + "logprob_reward": -0.023279689252376556, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023279689252376556, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08117277920246124, + 0.05643749237060547, + 0.20217522978782654, + 0.04665094614028931, + 0.15948785841464996, + 0.24208873510360718, + 0.02885311283171177, + 0.03507617861032486, + 0.06406065821647644, + 0.08399710059165955 + ], + "credit_map": [ + 0.011131536215543747, + 0.009955204091966152, + 0.01148710586130619, + 0.008803765289485455, + 0.01222727820277214, + 0.011407124809920788, + 0.007308219559490681, + 0.00885253120213747, + 0.007240307051688433, + 0.009380683302879333, + 0.011545234359800816, + 0.006128482986241579, + 0.005500809755176306, + 0.007057284936308861, + 0.004285795614123344, + 0.005298888310790062, + 0.02437863126397133, + 0.028493648394942284, + 0.05765164643526077, + 0.01902524009346962, + 0.016355756670236588, + 0.017718542367219925, + 0.019733428955078125, + 0.018818311393260956, + 0.0046892971731722355, + 0.008490731939673424, + 0.005572973750531673, + 0.004849004093557596, + 0.005367179401218891, + 0.007317451294511557, + 0.005313854198902845, + 0.0050504496321082115, + 0.016623156145215034, + 0.012759624049067497, + 0.05332721024751663, + 0.01123902015388012, + 0.007085626479238272, + 0.015763889998197556, + 0.007730097509920597, + 0.03495921567082405, + 0.06417886912822723, + 0.01743072271347046, + 0.07833141833543777, + 0.014818859286606312, + 0.010965455323457718, + 0.018875794485211372, + 0.01297708135098219, + 0.024510491639375687, + 0.002818529726937413, + 0.002797900466248393, + 0.004633513279259205, + 0.0032308546360582113, + 0.003678438253700733, + 0.003418891690671444, + 0.004513655789196491, + 0.003761327126994729, + 0.006096682045608759, + 0.0029926039278507233, + 0.008830494247376919, + 0.004026831593364477, + 0.00361604243516922, + 0.0034530304837971926, + 0.0026571485213935375, + 0.003403344890102744, + 0.0072951246984303, + 0.006306419149041176, + 0.01813911460340023, + 0.00900939293205738, + 0.008686944842338562, + 0.006651111878454685, + 0.004279430024325848, + 0.0036931112408638, + 0.004283740185201168, + 0.005743179004639387, + 0.014451299794018269, + 0.008152935653924942, + 0.02534162439405918, + 0.017700305208563805, + 0.005012737587094307, + 0.0033112680539488792 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 2, + 4 + ], + "selected_reward": -0.023279689252376556 + } + ], + "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/0000056/group_11/index_tensor([7663, 7663])_group_11_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000056/group_11/index_tensor([7663, 7663])_group_11_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..fa206f1fd7bd131465099b24121d529cbd640a52 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000056/group_11/index_tensor([7663, 7663])_group_11_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "static, eye-level close-up focuses on a vintage desk fan with a polished chrome cage, its blades a translucent blur as they generate a steady stream of air. Placed on the rich, dark wood surface of a desk, the fan's target is a slender ribbon of brilliant emerald green silk. Tethered just out of frame, the ribbon is caught in the airflow and performs a hypnotic, silent dance. It gracefully undulates, coiling and unfurling in a continuous, fluid motion, its high-gloss surface catching the warm, soft light from a nearby window. As the ribbon twists and flutters, it casts fleeting, subtle shadows on the polished desk. In the soft-focus background, a glimpse of a glossy black keyboard and a stack of papers can be seen, situating the scene in an office environment. The entire image is imbued with a warm color temperature, creating a tranquil, contemplative, and almost meditative atmosphere, focused on the simple, beautiful physics of the present moment.", + "seed": 16226, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000056/group_11/index_7663_seed_16226_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": -0.012763404566794635, + "vlm_logprob": -0.012763404566794635, + "vlm_yes_no_margin": 4.7, + "vlm_token_credit": -0.012763404566794635, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a vintage desk fan with a polished chrome cage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the fan's blade motion depicted as a translucent blur generating a steady stream of air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a slender ribbon of brilliant emerald green silk visible, undulating and coiling in the airflow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set on a rich, dark wood desk surface with a soft-focus background showing a glossy black keyboard and stack of papers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting warm and soft, casting subtle shadows from the ribbon onto the polished desk?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03227825462818146 + ], + "opposite_logprob_tokens": [ + -3.532278299331665 + ], + "logprob_reward": -0.03227825462818146, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.03227825462818146, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2785813510417938, + 0.1192554235458374, + 0.054028645157814026, + 0.24632331728935242, + 0.04164402186870575, + 0.07707799971103668, + 0.021870780736207962, + 0.09607335180044174, + 0.03503197059035301, + 0.030113162472844124 + ], + "credit_map": [ + 0.047897543758153915, + 0.03776318579912186, + 0.023978013545274734, + 0.03360242024064064, + 0.033029187470674515, + 0.04001397639513016, + 0.027164507657289505, + 0.03513246774673462, + 0.01868806965649128, + 0.017180277034640312, + 0.007157084532082081, + 0.010626154020428658, + 0.010919854044914246, + 0.02206510677933693, + 0.016367856413125992, + 0.016251012682914734, + 0.009965645149350166, + 0.00797864980995655, + 0.00422183983027935, + 0.00417868560180068, + 0.005955986212939024, + 0.008797123096883297, + 0.006569951307028532, + 0.006360754836350679, + 0.03350119665265083, + 0.031069327145814896, + 0.018795769661664963, + 0.021205956116318703, + 0.02984505146741867, + 0.05916300788521767, + 0.027865998446941376, + 0.024876991286873817, + 0.006506094243377447, + 0.005580936558544636, + 0.004816015250980854, + 0.004183363635092974, + 0.005207538604736328, + 0.00605433015152812, + 0.00441831024363637, + 0.004877431318163872, + 0.010069223120808601, + 0.010688957758247852, + 0.009367289021611214, + 0.00904015451669693, + 0.008130067028105259, + 0.011165914125740528, + 0.007900793105363846, + 0.010715587064623833, + 0.00390397640876472, + 0.002432592213153839, + 0.002047393238171935, + 0.0020939912647008896, + 0.0029454610776156187, + 0.0028352513909339905, + 0.0020906038116663694, + 0.0035215069074183702, + 0.014566360041499138, + 0.014683220535516739, + 0.0056938608177006245, + 0.0071918242610991, + 0.011921914294362068, + 0.019883722066879272, + 0.012664761394262314, + 0.009467674419283867, + 0.00529811205342412, + 0.00404030317440629, + 0.0027934478130191565, + 0.00305137038230896, + 0.003876953385770321, + 0.009396299719810486, + 0.0025268024764955044, + 0.004048679023981094, + 0.004297574050724506, + 0.005024286452680826, + 0.00404768530279398, + 0.003105700248852372, + 0.002662650076672435, + 0.0043821875005960464, + 0.0024573300033807755, + 0.0041357469744980335 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 1 + ], + "selected_reward": -0.03227825462818146 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a vintage desk fan with a polished chrome cage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the fan's blade motion depicted as a translucent blur generating a steady stream of air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a slender ribbon of brilliant emerald green silk visible, undulating and coiling in the airflow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set on a rich, dark wood desk surface with a soft-focus background showing a glossy black keyboard and stack of papers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting warm and soft, casting subtle shadows from the ribbon onto the polished desk?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01608221046626568 + ], + "opposite_logprob_tokens": [ + -4.141082286834717 + ], + "logprob_reward": -0.01608221046626568, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.01608221046626568, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3300875723361969, + 0.18155868351459503, + 0.036030132323503494, + 0.1091371402144432, + 0.038022298365831375, + 0.10259981453418732, + 0.04713740572333336, + 0.05714035406708717, + 0.05520613491535187, + 0.04308049753308296 + ], + "credit_map": [ + 0.0588035061955452, + 0.039776287972927094, + 0.034709200263023376, + 0.033948905766010284, + 0.03675190731883049, + 0.04998832195997238, + 0.03451675921678543, + 0.041592687368392944, + 0.03264098986983299, + 0.020763594657182693, + 0.023253317922353745, + 0.015910107642412186, + 0.016381073743104935, + 0.03275061026215553, + 0.014840487390756607, + 0.025018511340022087, + 0.005106324329972267, + 0.005407084245234728, + 0.003658552886918187, + 0.0038530719466507435, + 0.004507622215896845, + 0.00629149004817009, + 0.003547915956005454, + 0.0036580709274858236, + 0.016198862344026566, + 0.015019292943179607, + 0.0103134885430336, + 0.010664262808859348, + 0.010436163283884525, + 0.020752927288413048, + 0.010787423700094223, + 0.014964716508984566, + 0.00630812207236886, + 0.004945194814354181, + 0.005154822953045368, + 0.0043398234993219376, + 0.004464463330805302, + 0.00457137543708086, + 0.0031872799154371023, + 0.005051217507570982, + 0.017491400241851807, + 0.012258225120604038, + 0.014149188064038754, + 0.010058379732072353, + 0.01065263245254755, + 0.017162462696433067, + 0.008362258784472942, + 0.012465264648199081, + 0.006233423948287964, + 0.007196035701781511, + 0.004751567728817463, + 0.004514132160693407, + 0.0058305044658482075, + 0.009031175635755062, + 0.004288410302251577, + 0.005292158108204603, + 0.010084659792482853, + 0.008593893609941006, + 0.006045197602361441, + 0.005181583575904369, + 0.005323055665940046, + 0.008861685171723366, + 0.0057924906723201275, + 0.0072577884420752525, + 0.00959235429763794, + 0.004907522350549698, + 0.0072897737845778465, + 0.0062756165862083435, + 0.004226932302117348, + 0.008842266164720058, + 0.005551363807171583, + 0.008520307019352913, + 0.007212607190012932, + 0.005029590334743261, + 0.005767445545643568, + 0.005505944602191448, + 0.00410100445151329, + 0.006766777019947767, + 0.0036427797749638557, + 0.005054347217082977 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 3 + ], + "selected_reward": -0.01608221046626568 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a vintage desk fan with a polished chrome cage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the fan's blade motion depicted as a translucent blur generating a steady stream of air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a slender ribbon of brilliant emerald green silk visible, undulating and coiling in the airflow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set on a rich, dark wood desk surface with a soft-focus background showing a glossy black keyboard and stack of papers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting warm and soft, casting subtle shadows from the ribbon onto the polished desk?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004635065793991089 + ], + "opposite_logprob_tokens": [ + -5.379634857177734 + ], + "logprob_reward": -0.004635065793991089, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004635065793991089, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3804376423358917, + 0.08356139063835144, + 0.10392238199710846, + 0.1057547777891159, + 0.08462844789028168, + 0.06917073577642441, + 0.0596814826130867, + 0.05306695029139519, + 0.03390088677406311, + 0.025875259190797806 + ], + "credit_map": [ + 0.05622225999832153, + 0.04666830226778984, + 0.030961066484451294, + 0.035488031804561615, + 0.04711805656552315, + 0.0644845962524414, + 0.052878379821777344, + 0.04661700502038002, + 0.014854715205729008, + 0.009590832516551018, + 0.01062853168696165, + 0.00738429045304656, + 0.0095071317628026, + 0.01326019037514925, + 0.008077110163867474, + 0.010258601978421211, + 0.014065131545066833, + 0.013591586612164974, + 0.008073532022535801, + 0.009765276685357094, + 0.011177935637533665, + 0.018662376329302788, + 0.012100089341402054, + 0.01648646593093872, + 0.01488992478698492, + 0.01268193032592535, + 0.008404052816331387, + 0.00821551401168108, + 0.012012219056487083, + 0.02591043710708618, + 0.01175681035965681, + 0.011883899569511414, + 0.011431049555540085, + 0.010202791541814804, + 0.0073411124758422375, + 0.008307607844471931, + 0.010881494730710983, + 0.014488316141068935, + 0.00897476077079773, + 0.013001322746276855, + 0.011767868883907795, + 0.009144727140665054, + 0.006989858578890562, + 0.006281637586653233, + 0.007026450242847204, + 0.010738522745668888, + 0.00766952196136117, + 0.00955215748399496, + 0.007983813993632793, + 0.008732962422072887, + 0.006703081075102091, + 0.00640473235398531, + 0.006872075609862804, + 0.01170867495238781, + 0.0048528872430324554, + 0.006423255894333124, + 0.010661697946488857, + 0.007092562038451433, + 0.005423077382147312, + 0.00638045696541667, + 0.005176883656531572, + 0.007037958595901728, + 0.004937220364809036, + 0.006357098463922739, + 0.004889839328825474, + 0.0034495985601097345, + 0.0031154886819422245, + 0.0030976401176303625, + 0.0046532778069376945, + 0.007158150430768728, + 0.003409491153433919, + 0.0041274032555520535, + 0.004606976173818111, + 0.003084429306909442, + 0.002857496263459325, + 0.002962269354611635, + 0.002399258315563202, + 0.004322305787354708, + 0.002439952688291669, + 0.0032025736290961504 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 2 + ], + "selected_reward": -0.004635065793991089 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a vintage desk fan with a polished chrome cage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the fan's blade motion depicted as a translucent blur generating a steady stream of air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a slender ribbon of brilliant emerald green silk visible, undulating and coiling in the airflow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set on a rich, dark wood desk surface with a soft-focus background showing a glossy black keyboard and stack of papers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting warm and soft, casting subtle shadows from the ribbon onto the polished desk?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006731689441949129 + ], + "opposite_logprob_tokens": [ + -5.0067315101623535 + ], + "logprob_reward": -0.006731689441949129, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006731689441949129, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2142598181962967, + 0.1372140347957611, + 0.058594029396772385, + 0.048551369458436966, + 0.08045991510152817, + 0.07816189527511597, + 0.1923978477716446, + 0.1304849535226822, + 0.034111302345991135, + 0.025764845311641693 + ], + "credit_map": [ + 0.03840227425098419, + 0.02328021265566349, + 0.019171960651874542, + 0.02356252446770668, + 0.026472263038158417, + 0.031133681535720825, + 0.025953426957130432, + 0.02628345601260662, + 0.02565370872616768, + 0.018000725656747818, + 0.018376953899860382, + 0.016722889617085457, + 0.013017549179494381, + 0.017584435641765594, + 0.010744619183242321, + 0.01711314357817173, + 0.011742163449525833, + 0.00551856542006135, + 0.009659972041845322, + 0.006202422082424164, + 0.005099097266793251, + 0.007586005609482527, + 0.005916113033890724, + 0.0068696835078299046, + 0.007873325608670712, + 0.005582503974437714, + 0.006430207751691341, + 0.005080885253846645, + 0.004746431019157171, + 0.007428017444908619, + 0.004599785432219505, + 0.006810210645198822, + 0.013228733092546463, + 0.006933534052222967, + 0.0130630386993289, + 0.011800143867731094, + 0.006867361254990101, + 0.012201022356748581, + 0.007302064914256334, + 0.009064010344445705, + 0.011097743175923824, + 0.011138033121824265, + 0.010950721800327301, + 0.008595362305641174, + 0.007532962132245302, + 0.011693386361002922, + 0.007793105207383633, + 0.00936057697981596, + 0.02926887571811676, + 0.02591993473470211, + 0.019913723692297935, + 0.019292166456580162, + 0.02180686965584755, + 0.03862311318516731, + 0.014426103793084621, + 0.02314705401659012, + 0.02057957649230957, + 0.013604137115180492, + 0.01978997327387333, + 0.01154728140681982, + 0.010310391895473003, + 0.023876626044511795, + 0.01122152991592884, + 0.019555430859327316, + 0.005543878301978111, + 0.003617840586230159, + 0.0032183434814214706, + 0.003504863241687417, + 0.0033239813055843115, + 0.007609602063894272, + 0.003395740408450365, + 0.00389704923145473, + 0.004581121727824211, + 0.003563104895874858, + 0.0028780086431652308, + 0.0025315526872873306, + 0.002363587962463498, + 0.00421210378408432, + 0.002208361867815256, + 0.003426999319344759 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 1 + ], + "selected_reward": -0.006731689441949129 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a vintage desk fan with a polished chrome cage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the fan's blade motion depicted as a translucent blur generating a steady stream of air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a slender ribbon of brilliant emerald green silk visible, undulating and coiling in the airflow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set on a rich, dark wood desk surface with a soft-focus background showing a glossy black keyboard and stack of papers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting warm and soft, casting subtle shadows from the ribbon onto the polished desk?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004089802503585815 + ], + "opposite_logprob_tokens": [ + -5.504089832305908 + ], + "logprob_reward": -0.004089802503585815, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004089802503585815, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21914298832416534, + 0.08609890192747116, + 0.03573347628116608, + 0.045330993831157684, + 0.05395113676786423, + 0.08476151525974274, + 0.23128797113895416, + 0.1591760814189911, + 0.04376380890607834, + 0.040753137320280075 + ], + "credit_map": [ + 0.029025783762335777, + 0.02836250513792038, + 0.026468560099601746, + 0.0270056314766407, + 0.022542782127857208, + 0.03705572336912155, + 0.01962769404053688, + 0.02905430644750595, + 0.013214629143476486, + 0.01193061750382185, + 0.013512399047613144, + 0.01072879321873188, + 0.007368454243987799, + 0.010345463640987873, + 0.007210924755781889, + 0.01178761851042509, + 0.007008643355220556, + 0.004180684685707092, + 0.004520622082054615, + 0.003236127784475684, + 0.0037568460684269667, + 0.004297245759516954, + 0.004048365633934736, + 0.004684941377490759, + 0.007597457151859999, + 0.004367480985820293, + 0.006416648160666227, + 0.007519387640058994, + 0.004578637424856424, + 0.00537788076326251, + 0.0035307020880281925, + 0.005942795891314745, + 0.007812813855707645, + 0.006962187588214874, + 0.008490360341966152, + 0.005692900158464909, + 0.004589879885315895, + 0.00815882720053196, + 0.004979156889021397, + 0.0072650122456252575, + 0.013002386316657066, + 0.011110706254839897, + 0.009700111113488674, + 0.009276543743908405, + 0.00850253738462925, + 0.013891340233385563, + 0.00914517417550087, + 0.010132710449397564, + 0.0348392054438591, + 0.03114495240151882, + 0.025701625272631645, + 0.023216279223561287, + 0.02563544362783432, + 0.04545172303915024, + 0.017094740644097328, + 0.028204012662172318, + 0.024577785283327103, + 0.019723089411854744, + 0.02186467871069908, + 0.019394146278500557, + 0.014533761888742447, + 0.025693578645586967, + 0.012153279967606068, + 0.0212357547134161, + 0.0073420703411102295, + 0.004843342117965221, + 0.004823833238333464, + 0.004066356923431158, + 0.004004394169896841, + 0.008643326349556446, + 0.004666617140173912, + 0.005373871419578791, + 0.007236808072775602, + 0.004894312936812639, + 0.00449482724070549, + 0.005329430103302002, + 0.004342429805546999, + 0.005541290622204542, + 0.0038368189707398415, + 0.005077215377241373 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 7 + ], + "selected_reward": -0.004089802503585815 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, eye-level close-up focuses on a vintage desk fan with a polished chrome cage, its blades a translucent blur as they generate a steady stream of air. Placed on the rich, dark wood surface of a desk, the fan's target is a slender ribbon of brilliant emerald green silk. Tethered just out of frame, the ribbon is caught in the airflow and performs a hypnotic, silent dance. It gracefully undulates, coiling and unfurling in a continuous, fluid motion, its high-gloss surface catching the warm, soft light from a nearby window. As the ribbon twists and flutters, it casts fleeting, subtle shadows on the polished desk. In the soft-focus background, a glimpse of a glossy black keyboard and a stack of papers can be seen, situating the scene in an office environment. The entire image is imbued with a warm color temperature, creating a tranquil, contemplative, and almost meditative atmosphere, focused on the simple, beautiful physics of the present moment.", + "seed": 16227, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000056/group_11/index_7663_seed_16227_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": -0.050444012694060805, + "vlm_logprob": -0.050444012694060805, + "vlm_yes_no_margin": 3.55, + "vlm_token_credit": -0.050444012694060805, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a vintage desk fan with a polished chrome cage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the fan's blade motion depicted as a translucent blur generating a steady stream of air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a slender ribbon of brilliant emerald green silk visible, undulating and coiling in the airflow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set on a rich, dark wood desk surface with a soft-focus background showing a glossy black keyboard and stack of papers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting warm and soft, casting subtle shadows from the ribbon onto the polished desk?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1684889942407608 + ], + "opposite_logprob_tokens": [ + -1.9184889793395996 + ], + "logprob_reward": -0.1684889942407608, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.1684889942407608, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16287913918495178, + 0.14578670263290405, + 0.04842417687177658, + 0.1816423088312149, + 0.2226419597864151, + 0.056940604001283646, + 0.055161453783512115, + 0.05426153913140297, + 0.020015669986605644, + 0.05224645137786865 + ], + "credit_map": [ + 0.013128184713423252, + 0.018018806353211403, + 0.020221784710884094, + 0.01788475178182125, + 0.025501534342765808, + 0.01590763032436371, + 0.03702705726027489, + 0.015189391560852528, + 0.009650759398937225, + 0.012743109837174416, + 0.016750141978263855, + 0.01714414730668068, + 0.019077302888035774, + 0.03329336270689964, + 0.02135758101940155, + 0.01577029563486576, + 0.0031085575465112925, + 0.005358516238629818, + 0.0058168889954686165, + 0.005008669104427099, + 0.005755353253334761, + 0.010472305119037628, + 0.008027042262256145, + 0.004876843653619289, + 0.01300546433776617, + 0.026973625645041466, + 0.014572561718523502, + 0.016572872176766396, + 0.017322981730103493, + 0.02675955928862095, + 0.04675519838929176, + 0.019680054858326912, + 0.018002675846219063, + 0.025868862867355347, + 0.023828426375985146, + 0.023600652813911438, + 0.017513154074549675, + 0.019034378230571747, + 0.06422032415866852, + 0.030573483556509018, + 0.004052834119647741, + 0.004385944455862045, + 0.006355322897434235, + 0.0066582574509084225, + 0.008833209984004498, + 0.009270156733691692, + 0.011134220287203789, + 0.006250658072531223, + 0.003434299724176526, + 0.0062480503693223, + 0.006562442053109407, + 0.00485847145318985, + 0.004624097142368555, + 0.004493067041039467, + 0.01744633913040161, + 0.007494690362364054, + 0.0037439968436956406, + 0.0075175403617322445, + 0.005031467881053686, + 0.005794372409582138, + 0.006853088270872831, + 0.008273370563983917, + 0.010314001701772213, + 0.006733699236065149, + 0.0014107887400314212, + 0.0016705713933333755, + 0.0023478169459849596, + 0.003120848909020424, + 0.0025620802771300077, + 0.003518429584801197, + 0.0032370840199291706, + 0.0021480496507138014, + 0.004770250990986824, + 0.004539224784821272, + 0.0051603433676064014, + 0.007432636339217424, + 0.008423165418207645, + 0.006567548029124737, + 0.010044044815003872, + 0.00530923530459404 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 3, + 0 + ], + "selected_reward": -0.1684889942407608 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a vintage desk fan with a polished chrome cage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the fan's blade motion depicted as a translucent blur generating a steady stream of air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a slender ribbon of brilliant emerald green silk visible, undulating and coiling in the airflow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set on a rich, dark wood desk surface with a soft-focus background showing a glossy black keyboard and stack of papers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting warm and soft, casting subtle shadows from the ribbon onto the polished desk?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023314394056797028 + ], + "opposite_logprob_tokens": [ + -3.7733144760131836 + ], + "logprob_reward": -0.023314394056797028, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023314394056797028, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23892229795455933, + 0.15184509754180908, + 0.07031779736280441, + 0.06567277759313583, + 0.03642028942704201, + 0.03299126774072647, + 0.0968499481678009, + 0.12328863143920898, + 0.14220625162124634, + 0.041485678404569626 + ], + "credit_map": [ + 0.022037731483578682, + 0.039056457579135895, + 0.032149024307727814, + 0.029516220092773438, + 0.018314315006136894, + 0.017109563574194908, + 0.050251226872205734, + 0.030487755313515663, + 0.011317289434373379, + 0.012981844134628773, + 0.014474164694547653, + 0.02142767608165741, + 0.01713167317211628, + 0.01981934905052185, + 0.03570185974240303, + 0.01899123378098011, + 0.007037362549453974, + 0.004506487865000963, + 0.01568002812564373, + 0.010102984495460987, + 0.0053358725272119045, + 0.009899942204356194, + 0.009649486280977726, + 0.008105621673166752, + 0.005227131303399801, + 0.009535443969070911, + 0.006477897055447102, + 0.006990232039242983, + 0.005622298456728458, + 0.00589103763923049, + 0.018502196297049522, + 0.0074265338480472565, + 0.0023792756255716085, + 0.0033560486044734716, + 0.004847213625907898, + 0.004613123368471861, + 0.0032449818681925535, + 0.004749610088765621, + 0.008034826256334782, + 0.005195206962525845, + 0.0023281059693545103, + 0.003357855137437582, + 0.003891485743224621, + 0.003297241870313883, + 0.0032719429582357407, + 0.0053113410249352455, + 0.007597821764647961, + 0.003935468383133411, + 0.006025176960974932, + 0.007658375892788172, + 0.009062611497938633, + 0.006889190059155226, + 0.00658158166334033, + 0.007512006442993879, + 0.038144659250974655, + 0.014976343140006065, + 0.009112396277487278, + 0.01109554897993803, + 0.012451034039258957, + 0.01104714348912239, + 0.01410314068198204, + 0.01884797401726246, + 0.03139762580394745, + 0.015233751386404037, + 0.010274575091898441, + 0.01246732659637928, + 0.01517324335873127, + 0.02518649213016033, + 0.018219923600554466, + 0.018940022215247154, + 0.02658000774681568, + 0.01536465622484684, + 0.003179411171004176, + 0.003130149794742465, + 0.005653144791722298, + 0.005412809085100889, + 0.004802229814231396, + 0.008225468918681145, + 0.005088134203106165, + 0.005994325969368219 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 8 + ], + "selected_reward": -0.023314394056797028 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a vintage desk fan with a polished chrome cage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the fan's blade motion depicted as a translucent blur generating a steady stream of air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a slender ribbon of brilliant emerald green silk visible, undulating and coiling in the airflow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set on a rich, dark wood desk surface with a soft-focus background showing a glossy black keyboard and stack of papers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting warm and soft, casting subtle shadows from the ribbon onto the polished desk?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00978435855358839 + ], + "opposite_logprob_tokens": [ + -4.63478422164917 + ], + "logprob_reward": -0.00978435855358839, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.00978435855358839, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.25187525153160095, + 0.12527728080749512, + 0.1105770394206047, + 0.10304614156484604, + 0.05754685774445534, + 0.027586786076426506, + 0.061088185757398605, + 0.07657302170991898, + 0.1265566200017929, + 0.059872839599847794 + ], + "credit_map": [ + 0.02301196940243244, + 0.024657748639583588, + 0.038783103227615356, + 0.03286058083176613, + 0.03868784382939339, + 0.024837985634803772, + 0.03652648627758026, + 0.03250950574874878, + 0.011139395646750927, + 0.019083717837929726, + 0.017498861998319626, + 0.01310228556394577, + 0.009310133755207062, + 0.013241052627563477, + 0.028728678822517395, + 0.013173157349228859, + 0.008138841018080711, + 0.00998397171497345, + 0.015489947982132435, + 0.008352159522473812, + 0.011223006062209606, + 0.01988285407423973, + 0.026445364579558372, + 0.011060894466936588, + 0.007854505442082882, + 0.011970199644565582, + 0.010348206385970116, + 0.010810690000653267, + 0.008425320498645306, + 0.010983855463564396, + 0.030437612906098366, + 0.012215760536491871, + 0.003902187803760171, + 0.004686807282269001, + 0.006350966170430183, + 0.005436015781015158, + 0.004558582790195942, + 0.006318308413028717, + 0.017221907153725624, + 0.009072085842490196, + 0.0014919587410986423, + 0.002038300968706608, + 0.00373259000480175, + 0.0024698504712432623, + 0.0033747204579412937, + 0.006231287494301796, + 0.0048256320878863335, + 0.003422444686293602, + 0.005184484180063009, + 0.007201948203146458, + 0.009552795439958572, + 0.004770267754793167, + 0.00401210505515337, + 0.004884847905486822, + 0.017198458313941956, + 0.008283279836177826, + 0.005278116092085838, + 0.004860924556851387, + 0.007946844212710857, + 0.006475063506513834, + 0.0073852622881531715, + 0.009211516007781029, + 0.025854388251900673, + 0.009560907259583473, + 0.009118340909481049, + 0.01117892749607563, + 0.012844291515648365, + 0.02174762450158596, + 0.018283342942595482, + 0.015191115438938141, + 0.025991255417466164, + 0.01220172643661499, + 0.004583968315273523, + 0.00505652092397213, + 0.007082934956997633, + 0.010880670510232449, + 0.008123265579342842, + 0.008983220905065536, + 0.00839472096413374, + 0.006767536513507366 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 1 + ], + "selected_reward": -0.00978435855358839 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a vintage desk fan with a polished chrome cage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the fan's blade motion depicted as a translucent blur generating a steady stream of air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a slender ribbon of brilliant emerald green silk visible, undulating and coiling in the airflow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set on a rich, dark wood desk surface with a soft-focus background showing a glossy black keyboard and stack of papers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting warm and soft, casting subtle shadows from the ribbon onto the polished desk?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"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": [ + 0.16916145384311676, + 0.10705577582120895, + 0.07375821471214294, + 0.03695711866021156, + 0.11045382916927338, + 0.037833794951438904, + 0.031549062579870224, + 0.06753610819578171, + 0.12724216282367706, + 0.2384524941444397 + ], + "credit_map": [ + 0.014007046818733215, + 0.029683779925107956, + 0.022004077211022377, + 0.021096860989928246, + 0.015235058031976223, + 0.00978616252541542, + 0.04135401174426079, + 0.015994448214769363, + 0.009339803829789162, + 0.017615601420402527, + 0.01321351807564497, + 0.012681704945862293, + 0.009523607790470123, + 0.011134397238492966, + 0.02280350960791111, + 0.010743632912635803, + 0.005622385535389185, + 0.009966186247766018, + 0.009088225662708282, + 0.008136867545545101, + 0.006360085681080818, + 0.0125731797888875, + 0.015201989561319351, + 0.006809293292462826, + 0.00244427053257823, + 0.0032086067367345095, + 0.004372997209429741, + 0.003578315721824765, + 0.005473769269883633, + 0.006596570368856192, + 0.006756101734936237, + 0.0045264894142746925, + 0.00828926544636488, + 0.010833745822310448, + 0.010713941417634487, + 0.011259426362812519, + 0.008190248161554337, + 0.010494711808860302, + 0.03407689929008484, + 0.016595585271716118, + 0.00237465207464993, + 0.0035058530047535896, + 0.005067234858870506, + 0.004338140599429607, + 0.00358151295222342, + 0.005915450863540173, + 0.008060422725975513, + 0.004990528803318739, + 0.002476002089679241, + 0.0038988706655800343, + 0.005286138970404863, + 0.003102158661931753, + 0.0022495626471936703, + 0.0031275933142751455, + 0.0072930725291371346, + 0.004115664400160313, + 0.004994148388504982, + 0.004991187714040279, + 0.008705763146281242, + 0.0065162102691829205, + 0.004416387528181076, + 0.006883115507662296, + 0.022689519450068474, + 0.0083397775888443, + 0.01065551396459341, + 0.010928780771791935, + 0.01330622099339962, + 0.014635284431278706, + 0.011256047524511814, + 0.011303946375846863, + 0.04247257113456726, + 0.01268378272652626, + 0.018925132229924202, + 0.022265715524554253, + 0.028367789462208748, + 0.04508345574140549, + 0.02754012867808342, + 0.03037305176258087, + 0.03982866182923317, + 0.026068564504384995 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 8 + ], + "selected_reward": -0.03809661045670509 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a vintage desk fan with a polished chrome cage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the fan's blade motion depicted as a translucent blur generating a steady stream of air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a slender ribbon of brilliant emerald green silk visible, undulating and coiling in the airflow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set on a rich, dark wood desk surface with a soft-focus background showing a glossy black keyboard and stack of papers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting warm and soft, casting subtle shadows from the ribbon onto the polished desk?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012535706162452698 + ], + "opposite_logprob_tokens": [ + -4.387535572052002 + ], + "logprob_reward": -0.012535706162452698, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012535706162452698, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24388134479522705, + 0.11303409934043884, + 0.0764254704117775, + 0.05813198164105415, + 0.08035397529602051, + 0.038395922631025314, + 0.04242892563343048, + 0.1310911923646927, + 0.07316974550485611, + 0.14308728277683258 + ], + "credit_map": [ + 0.02031814120709896, + 0.041992466896772385, + 0.029098492115736008, + 0.027726171538233757, + 0.02721031941473484, + 0.01759491115808487, + 0.056055858731269836, + 0.023885000497102737, + 0.00986891146749258, + 0.018193570896983147, + 0.01633167825639248, + 0.015118592418730259, + 0.011586911045014858, + 0.013573315925896168, + 0.017739949747920036, + 0.010621168650686741, + 0.005206528585404158, + 0.007020206190645695, + 0.008333345875144005, + 0.005545033607631922, + 0.008129540830850601, + 0.015041234903037548, + 0.019798519089818, + 0.0073510645888745785, + 0.004482992459088564, + 0.006123111583292484, + 0.008394945412874222, + 0.007049912586808205, + 0.0063210478983819485, + 0.008656916208565235, + 0.010665631853044033, + 0.006437423173338175, + 0.00598174100741744, + 0.008066182024776936, + 0.008134977892041206, + 0.007778409402817488, + 0.006867539137601852, + 0.008756909519433975, + 0.0228925459086895, + 0.01187567226588726, + 0.0030336640775203705, + 0.004514453466981649, + 0.005267795640975237, + 0.005294782109558582, + 0.005063212942332029, + 0.005210443399846554, + 0.005737741477787495, + 0.004273830913007259, + 0.003029338549822569, + 0.0031392450910061598, + 0.004343156237155199, + 0.004314373712986708, + 0.0036910860799252987, + 0.003898272290825844, + 0.01465519331395626, + 0.005358259659260511, + 0.009623566642403603, + 0.010484181344509125, + 0.013749903999269009, + 0.01197698526084423, + 0.010162130929529667, + 0.015468922443687916, + 0.04311637952923775, + 0.016509123146533966, + 0.006292575038969517, + 0.00532766617834568, + 0.007029481697827578, + 0.008731069043278694, + 0.008425010368227959, + 0.006689974572509527, + 0.02261788584291935, + 0.00805608369410038, + 0.010865788906812668, + 0.012319016270339489, + 0.01806195080280304, + 0.02575157955288887, + 0.017540596425533295, + 0.019081905484199524, + 0.02285628393292427, + 0.016610153019428253 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 7 + ], + "selected_reward": -0.012535706162452698 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, eye-level close-up focuses on a vintage desk fan with a polished chrome cage, its blades a translucent blur as they generate a steady stream of air. Placed on the rich, dark wood surface of a desk, the fan's target is a slender ribbon of brilliant emerald green silk. Tethered just out of frame, the ribbon is caught in the airflow and performs a hypnotic, silent dance. It gracefully undulates, coiling and unfurling in a continuous, fluid motion, its high-gloss surface catching the warm, soft light from a nearby window. As the ribbon twists and flutters, it casts fleeting, subtle shadows on the polished desk. In the soft-focus background, a glimpse of a glossy black keyboard and a stack of papers can be seen, situating the scene in an office environment. The entire image is imbued with a warm color temperature, creating a tranquil, contemplative, and almost meditative atmosphere, focused on the simple, beautiful physics of the present moment.", + "seed": 16228, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000056/group_11/index_7663_seed_16228_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": -0.016421728022396564, + "vlm_logprob": -0.016421728022396564, + "vlm_yes_no_margin": 4.325, + "vlm_token_credit": -0.016421728022396564, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a vintage desk fan with a polished chrome cage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the fan's blade motion depicted as a translucent blur generating a steady stream of air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a slender ribbon of brilliant emerald green silk visible, undulating and coiling in the airflow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set on a rich, dark wood desk surface with a soft-focus background showing a glossy black keyboard and stack of papers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting warm and soft, casting subtle shadows from the ribbon onto the polished desk?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.035405244678258896 + ], + "opposite_logprob_tokens": [ + -3.410405158996582 + ], + "logprob_reward": -0.035405244678258896, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.035405244678258896, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16581448912620544, + 0.1287667453289032, + 0.16180419921875, + 0.15872728824615479, + 0.11645226925611496, + 0.05555308237671852, + 0.043368108570575714, + 0.039898473769426346, + 0.033643659204244614, + 0.09597170352935791 + ], + "credit_map": [ + 0.02655908279120922, + 0.012723876163363457, + 0.018841277807950974, + 0.017818612977862358, + 0.009160158224403858, + 0.009545502252876759, + 0.046756330877542496, + 0.024409644305706024, + 0.013765968382358551, + 0.01665625162422657, + 0.024620190262794495, + 0.018871966749429703, + 0.010011114180088043, + 0.014710000716149807, + 0.016500424593687057, + 0.013630833476781845, + 0.011122329160571098, + 0.0067216865718364716, + 0.06626003235578537, + 0.010101240128278732, + 0.0056254868395626545, + 0.00720260152593255, + 0.0455065593123436, + 0.009264258667826653, + 0.04635341838002205, + 0.00790280383080244, + 0.027673816308379173, + 0.017077626660466194, + 0.006174450274556875, + 0.012243270874023438, + 0.026465659961104393, + 0.014836237765848637, + 0.007597580086439848, + 0.005118133034557104, + 0.007042560260742903, + 0.044031355530023575, + 0.003092140192165971, + 0.003407539799809456, + 0.03886903449892998, + 0.007293918170034885, + 0.004538584966212511, + 0.002465418539941311, + 0.018260415643453598, + 0.017283303663134575, + 0.001993036363273859, + 0.0031855034176260233, + 0.004176450427621603, + 0.0036503742448985577, + 0.00503131840378046, + 0.00375421904027462, + 0.002814623061567545, + 0.011498083360493183, + 0.0027200912591069937, + 0.003829431952908635, + 0.009140973910689354, + 0.004579363856464624, + 0.004371657967567444, + 0.0048910765908658504, + 0.006787494290620089, + 0.005733427591621876, + 0.0032432822044938803, + 0.005415458232164383, + 0.004287579096853733, + 0.005168494302779436, + 0.0059360298328101635, + 0.0025267968885600567, + 0.0039872704073786736, + 0.005166169721633196, + 0.0023395526222884655, + 0.0038063840474933386, + 0.006625982001423836, + 0.0032554755453020334, + 0.007026101462543011, + 0.004335579928010702, + 0.005015602335333824, + 0.03776303678750992, + 0.003210194408893585, + 0.004033650737255812, + 0.028873223811388016, + 0.005714312195777893 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 3 + ], + "selected_reward": -0.035405244678258896 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a vintage desk fan with a polished chrome cage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the fan's blade motion depicted as a translucent blur generating a steady stream of air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a slender ribbon of brilliant emerald green silk visible, undulating and coiling in the airflow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set on a rich, dark wood desk surface with a soft-focus background showing a glossy black keyboard and stack of papers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting warm and soft, casting subtle shadows from the ribbon onto the polished desk?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02059772238135338 + ], + "opposite_logprob_tokens": [ + -3.8955976963043213 + ], + "logprob_reward": -0.02059772238135338, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.02059772238135338, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1855209916830063, + 0.10877810418605804, + 0.18137513101100922, + 0.05660469830036163, + 0.2107737809419632, + 0.06928253173828125, + 0.04963046312332153, + 0.03852654621005058, + 0.06960512697696686, + 0.029902582988142967 + ], + "credit_map": [ + 0.028341175988316536, + 0.025525255128741264, + 0.022928936406970024, + 0.026589522138237953, + 0.01375647634267807, + 0.018113592639565468, + 0.030644357204437256, + 0.019621672108769417, + 0.016429634764790535, + 0.013766898773610592, + 0.014531704597175121, + 0.011153250001370907, + 0.01207053940743208, + 0.016777677461504936, + 0.009960899129509926, + 0.014087509363889694, + 0.012794953770935535, + 0.009059698320925236, + 0.07390287518501282, + 0.012962644919753075, + 0.0075729014351964, + 0.007208323571830988, + 0.04804574325680733, + 0.009827992878854275, + 0.010855509899556637, + 0.005939049180597067, + 0.009941092692315578, + 0.006326489616185427, + 0.003913423512130976, + 0.007393769919872284, + 0.0063174134120345116, + 0.005917954724282026, + 0.01402209047228098, + 0.007864038459956646, + 0.00816372036933899, + 0.08179634064435959, + 0.005612242501229048, + 0.005438548978418112, + 0.07485663890838623, + 0.013020156882703304, + 0.0044433725997805595, + 0.004373094532638788, + 0.020533304661512375, + 0.024360312148928642, + 0.0023288980592042208, + 0.003493372816592455, + 0.004778696224093437, + 0.004971478134393692, + 0.004031293094158173, + 0.0029804417863488197, + 0.002909563248977065, + 0.01444720197468996, + 0.0023259276058524847, + 0.003970245365053415, + 0.0116624366492033, + 0.007303355261683464, + 0.003606021637097001, + 0.003373251296579838, + 0.0071695726364851, + 0.004429874010384083, + 0.002871930366382003, + 0.003973489161580801, + 0.009251048788428307, + 0.0038513571489602327, + 0.0070092943497002125, + 0.00488269655033946, + 0.011307425796985626, + 0.0084030507132411, + 0.00362419243901968, + 0.003507935907691717, + 0.02452278323471546, + 0.006347751710563898, + 0.005972464568912983, + 0.0034643651451915503, + 0.0032241998706012964, + 0.0050497218035161495, + 0.0016366863856092095, + 0.0024988821242004633, + 0.0026218818966299295, + 0.005434381775557995 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 2 + ], + "selected_reward": -0.02059772238135338 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a vintage desk fan with a polished chrome cage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the fan's blade motion depicted as a translucent blur generating a steady stream of air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a slender ribbon of brilliant emerald green silk visible, undulating and coiling in the airflow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set on a rich, dark wood desk surface with a soft-focus background showing a glossy black keyboard and stack of papers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting warm and soft, casting subtle shadows from the ribbon onto the polished desk?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005948578007519245 + ], + "opposite_logprob_tokens": [ + -5.130948543548584 + ], + "logprob_reward": -0.005948578007519245, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005948578007519245, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21487654745578766, + 0.15054285526275635, + 0.09981155395507812, + 0.09307079762220383, + 0.06431534886360168, + 0.17372705042362213, + 0.057369131594896317, + 0.06267654150724411, + 0.03711730241775513, + 0.046492863446474075 + ], + "credit_map": [ + 0.03125559911131859, + 0.01355073507875204, + 0.019253209233283997, + 0.03314948454499245, + 0.010499955154955387, + 0.01351465005427599, + 0.06883038580417633, + 0.024822525680065155, + 0.017690572887659073, + 0.023357639089226723, + 0.018816061317920685, + 0.030127283185720444, + 0.013474172912538052, + 0.01940682716667652, + 0.014512547291815281, + 0.013157756999135017, + 0.010469475761055946, + 0.00911655928939581, + 0.028441453352570534, + 0.011346443556249142, + 0.00877225212752819, + 0.007795644924044609, + 0.017352990806102753, + 0.006516729947179556, + 0.02857300266623497, + 0.0069245933555066586, + 0.016627537086606026, + 0.008555509150028229, + 0.0039977990090847015, + 0.007269605994224548, + 0.012040620669722557, + 0.009082125499844551, + 0.00436005461961031, + 0.0036950190551579, + 0.0042956178076565266, + 0.023474516347050667, + 0.002703967737033963, + 0.003090908518061042, + 0.01916077546775341, + 0.0035344939678907394, + 0.008387528359889984, + 0.007147621363401413, + 0.06478677690029144, + 0.05788373202085495, + 0.005607894621789455, + 0.006251717917621136, + 0.015228346921503544, + 0.008433429524302483, + 0.005976537242531776, + 0.00594164477661252, + 0.003937684930860996, + 0.015330594033002853, + 0.003673289902508259, + 0.005104048643261194, + 0.012566929683089256, + 0.004838399589061737, + 0.006918750237673521, + 0.0057662599720060825, + 0.015022207982838154, + 0.008974246680736542, + 0.005331999622285366, + 0.00657234899699688, + 0.007823312655091286, + 0.006267414428293705, + 0.00505585502833128, + 0.0028525497764348984, + 0.006350724026560783, + 0.0038127820007503033, + 0.0019077174365520477, + 0.002505520125851035, + 0.010825865902006626, + 0.0038062899839133024, + 0.00562503095716238, + 0.0033940556459128857, + 0.005293430294841528, + 0.016222137957811356, + 0.0025493400171399117, + 0.0028579721692949533, + 0.006371143274009228, + 0.004179756157100201 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 1 + ], + "selected_reward": -0.005948578007519245 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a vintage desk fan with a polished chrome cage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the fan's blade motion depicted as a translucent blur generating a steady stream of air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a slender ribbon of brilliant emerald green silk visible, undulating and coiling in the airflow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set on a rich, dark wood desk surface with a soft-focus background showing a glossy black keyboard and stack of papers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting warm and soft, casting subtle shadows from the ribbon onto the polished desk?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0076235076412558556 + ], + "opposite_logprob_tokens": [ + -4.882623672485352 + ], + "logprob_reward": -0.0076235076412558556, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.0076235076412558556, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12250741571187973, + 0.13308003544807434, + 0.2963784337043762, + 0.07101647555828094, + 0.05434415489435196, + 0.11422204226255417, + 0.07506756484508514, + 0.09569951891899109, + 0.010339203290641308, + 0.027345124632120132 + ], + "credit_map": [ + 0.015672650188207626, + 0.00956228282302618, + 0.011103015393018723, + 0.011253791861236095, + 0.006749169901013374, + 0.009045558981597424, + 0.037640634924173355, + 0.021480320021510124, + 0.014247583225369453, + 0.017397234216332436, + 0.02634703926742077, + 0.021730436012148857, + 0.010962308384478092, + 0.012405411340296268, + 0.01863955892622471, + 0.011350453831255436, + 0.021642835810780525, + 0.011746864765882492, + 0.12306862324476242, + 0.017513606697320938, + 0.010760861448943615, + 0.011172491125762463, + 0.08273142576217651, + 0.017741721123456955, + 0.017521237954497337, + 0.005237098317593336, + 0.005875331349670887, + 0.006845060735940933, + 0.004729938227683306, + 0.005087618716061115, + 0.01769605465233326, + 0.008024141192436218, + 0.0057348390109837055, + 0.002945292042568326, + 0.003722731489688158, + 0.016161693260073662, + 0.0020349512342363596, + 0.0018823172431439161, + 0.018046123906970024, + 0.003816207405179739, + 0.009233378805220127, + 0.0067452057264745235, + 0.037500690668821335, + 0.022066129371523857, + 0.00625125877559185, + 0.004420163109898567, + 0.019972970709204674, + 0.008032247424125671, + 0.00920011568814516, + 0.007461621426045895, + 0.006480230484157801, + 0.018232662230730057, + 0.00568418251350522, + 0.004447900224477053, + 0.015753094106912613, + 0.0078077614307403564, + 0.009639463387429714, + 0.00682848272845149, + 0.025071511045098305, + 0.010054352693259716, + 0.005549750290811062, + 0.00646858150139451, + 0.02324395813047886, + 0.008843403309583664, + 0.0021113555412739515, + 0.0009287630091421306, + 0.0011788217816501856, + 0.001610713661648333, + 0.0007422392955049872, + 0.0007398321758955717, + 0.001797072240151465, + 0.0012304052943363786, + 0.003551701782271266, + 0.0015977364964783192, + 0.001884974422864616, + 0.0114651620388031, + 0.0011021109530702233, + 0.0011899475939571857, + 0.004918953403830528, + 0.0016345377080142498 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 1, + 0 + ], + "selected_reward": -0.0076235076412558556 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a vintage desk fan with a polished chrome cage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the fan's blade motion depicted as a translucent blur generating a steady stream of air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a slender ribbon of brilliant emerald green silk visible, undulating and coiling in the airflow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set on a rich, dark wood desk surface with a soft-focus background showing a glossy black keyboard and stack of papers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting warm and soft, casting subtle shadows from the ribbon onto the polished desk?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"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": [ + 0.2815377116203308, + 0.08392224460840225, + 0.3512820899486542, + 0.06294691562652588, + 0.046961210668087006, + 0.04718594253063202, + 0.044578708708286285, + 0.027043577283620834, + 0.02964339405298233, + 0.02489824779331684 + ], + "credit_map": [ + 0.04071230813860893, + 0.017930973321199417, + 0.021059101447463036, + 0.02502734586596489, + 0.013859209604561329, + 0.0183882899582386, + 0.09613437205553055, + 0.04842609912157059, + 0.009125657379627228, + 0.011456483043730259, + 0.015705866739153862, + 0.012973291799426079, + 0.007169070187956095, + 0.005970668513327837, + 0.013773748651146889, + 0.007747463416308165, + 0.024073420092463493, + 0.013536294922232628, + 0.15122321248054504, + 0.020837174728512764, + 0.012513858266174793, + 0.012616160325706005, + 0.09648158401250839, + 0.02000040002167225, + 0.014649549499154091, + 0.004193732514977455, + 0.01600361056625843, + 0.005436339415609837, + 0.003927689511328936, + 0.0062185171991586685, + 0.007222612854093313, + 0.005294866859912872, + 0.0046730670146644115, + 0.003082080977037549, + 0.005924288183450699, + 0.018383799120783806, + 0.0019046782981604338, + 0.0017963590798899531, + 0.008043345995247364, + 0.0031535914167761803, + 0.002961411839351058, + 0.0029956072103232145, + 0.01474691927433014, + 0.016368458047509193, + 0.0020245425403118134, + 0.001833472982980311, + 0.003313112771138549, + 0.0029424161184579134, + 0.005231770221143961, + 0.0050025503151118755, + 0.00405719643458724, + 0.010803865268826485, + 0.003736920887604356, + 0.0026070834137499332, + 0.008813311345875263, + 0.004326010588556528, + 0.002203278010711074, + 0.0024240072816610336, + 0.005953719839453697, + 0.0037704494316130877, + 0.0017344520892947912, + 0.002220424823462963, + 0.005285793449729681, + 0.0034514525905251503, + 0.004377246368676424, + 0.0023503140546381474, + 0.005088355857878923, + 0.003064562100917101, + 0.0017597441328689456, + 0.001985495910048485, + 0.007448472082614899, + 0.0035692027304321527, + 0.00439996924251318, + 0.0014607735211029649, + 0.0017907590372487903, + 0.00300781661644578, + 0.0014357888139784336, + 0.0012357884552329779, + 0.008821993134915829, + 0.0027453608345240355 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 1 + ], + "selected_reward": -0.012533587403595448 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, eye-level close-up focuses on a vintage desk fan with a polished chrome cage, its blades a translucent blur as they generate a steady stream of air. Placed on the rich, dark wood surface of a desk, the fan's target is a slender ribbon of brilliant emerald green silk. Tethered just out of frame, the ribbon is caught in the airflow and performs a hypnotic, silent dance. It gracefully undulates, coiling and unfurling in a continuous, fluid motion, its high-gloss surface catching the warm, soft light from a nearby window. As the ribbon twists and flutters, it casts fleeting, subtle shadows on the polished desk. In the soft-focus background, a glimpse of a glossy black keyboard and a stack of papers can be seen, situating the scene in an office environment. The entire image is imbued with a warm color temperature, creating a tranquil, contemplative, and almost meditative atmosphere, focused on the simple, beautiful physics of the present moment.", + "seed": 16229, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000056/group_11/index_7663_seed_16229_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": -0.012658426654525101, + "vlm_logprob": -0.012658426654525101, + "vlm_yes_no_margin": 5.175, + "vlm_token_credit": -0.012658426654525101, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a vintage desk fan with a polished chrome cage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the fan's blade motion depicted as a translucent blur generating a steady stream of air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a slender ribbon of brilliant emerald green silk visible, undulating and coiling in the airflow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set on a rich, dark wood desk surface with a soft-focus background showing a glossy black keyboard and stack of papers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting warm and soft, casting subtle shadows from the ribbon onto the polished desk?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04553950950503349 + ], + "opposite_logprob_tokens": [ + -3.170539617538452 + ], + "logprob_reward": -0.04553950950503349, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04553950950503349, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4064333736896515, + 0.09123621881008148, + 0.04526761919260025, + 0.07891494035720825, + 0.04475977644324303, + 0.13102839887142181, + 0.062422629445791245, + 0.07220083475112915, + 0.044750384986400604, + 0.022985773161053658 + ], + "credit_map": [ + 0.025963908061385155, + 0.12727153301239014, + 0.020242774859070778, + 0.019981127232313156, + 0.03377395495772362, + 0.03132286295294762, + 0.125286266207695, + 0.02259097807109356, + 0.014055009000003338, + 0.01082400232553482, + 0.007603962440043688, + 0.010032055899500847, + 0.010875619947910309, + 0.010647977702319622, + 0.016258060932159424, + 0.010939544066786766, + 0.004725021310150623, + 0.0038459207862615585, + 0.0053608715534210205, + 0.007358168251812458, + 0.0056057036854326725, + 0.004735526628792286, + 0.005793999880552292, + 0.007842411287128925, + 0.018692197278141975, + 0.008450469002127647, + 0.006647076457738876, + 0.007071343716233969, + 0.00955056119710207, + 0.009290569461882114, + 0.009896744973957539, + 0.00931598525494337, + 0.006224983371794224, + 0.003759231884032488, + 0.006248049903661013, + 0.005199764389544725, + 0.007926755584776402, + 0.0069239879958331585, + 0.005232295487076044, + 0.0032447075936943293, + 0.01253590453416109, + 0.01276592630892992, + 0.017930030822753906, + 0.01923845149576664, + 0.0183675829321146, + 0.02607092820107937, + 0.014566030353307724, + 0.009553554467856884, + 0.010329649783670902, + 0.0060097225941717625, + 0.011169731616973877, + 0.005840555764734745, + 0.007643889170140028, + 0.008912515826523304, + 0.006559692323207855, + 0.005956877022981644, + 0.009477131068706512, + 0.014035100117325783, + 0.009461387991905212, + 0.008251940831542015, + 0.008901015855371952, + 0.008241504430770874, + 0.008109236136078835, + 0.005723525770008564, + 0.008140393532812595, + 0.007971282117068768, + 0.0038946839049458504, + 0.004972582217305899, + 0.006286723539233208, + 0.004034806974232197, + 0.004767262376844883, + 0.004682651720941067, + 0.002475888468325138, + 0.003033439861610532, + 0.0025952672585844994, + 0.0022460687905550003, + 0.002915536519140005, + 0.0040311566554009914, + 0.003671998390927911, + 0.002016422338783741 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 1 + ], + "selected_reward": -0.04553950950503349 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a vintage desk fan with a polished chrome cage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the fan's blade motion depicted as a translucent blur generating a steady stream of air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a slender ribbon of brilliant emerald green silk visible, undulating and coiling in the airflow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set on a rich, dark wood desk surface with a soft-focus background showing a glossy black keyboard and stack of papers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting warm and soft, casting subtle shadows from the ribbon onto the polished desk?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"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": [ + 0.2429603785276413, + 0.15093861520290375, + 0.254275381565094, + 0.03799130395054817, + 0.037183891981840134, + 0.05628187581896782, + 0.04825205355882645, + 0.05593437701463699, + 0.07495331019163132, + 0.0412287674844265 + ], + "credit_map": [ + 0.02853362262248993, + 0.0450165830552578, + 0.021572038531303406, + 0.02351040206849575, + 0.03234120458364487, + 0.027969103306531906, + 0.043160293251276016, + 0.02085714042186737, + 0.016601480543613434, + 0.028146911412477493, + 0.012113172560930252, + 0.017359627410769463, + 0.018135834485292435, + 0.014820328913629055, + 0.03178292140364647, + 0.011978348717093468, + 0.043458499014377594, + 0.02284301444888115, + 0.027232222259044647, + 0.043515097349882126, + 0.03457614406943321, + 0.022642647847533226, + 0.03813648596405983, + 0.021871298551559448, + 0.006396614946424961, + 0.003592029446735978, + 0.003272139234468341, + 0.003763469634577632, + 0.004891488701105118, + 0.004278014879673719, + 0.004420306999236345, + 0.007377239875495434, + 0.004944436717778444, + 0.003381720744073391, + 0.004335267469286919, + 0.00561510631814599, + 0.004952483344823122, + 0.00362923601642251, + 0.004113297909498215, + 0.0062123448587954044, + 0.005530857015401125, + 0.004786663688719273, + 0.007821986451745033, + 0.007332039065659046, + 0.008497730828821659, + 0.007402072194963694, + 0.009520584717392921, + 0.005389941856265068, + 0.0076358928345143795, + 0.004084673244506121, + 0.008385928347706795, + 0.005247642286121845, + 0.005549330729991198, + 0.006137754302471876, + 0.007307523861527443, + 0.00390331051312387, + 0.006203729659318924, + 0.01381857879459858, + 0.004508695565164089, + 0.005649522878229618, + 0.006425570696592331, + 0.007154925726354122, + 0.006059195846319199, + 0.00611415971070528, + 0.011301952414214611, + 0.016062606126070023, + 0.0065328762866556644, + 0.007681410294026136, + 0.007599410135298967, + 0.008325873874127865, + 0.010807981714606285, + 0.0066411979496479034, + 0.005019783973693848, + 0.004152929410338402, + 0.003798970952630043, + 0.004876154940575361, + 0.006253273691982031, + 0.0052828011102974415, + 0.007579668890684843, + 0.0042651863768696785 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 1 + ], + "selected_reward": -0.008649858646094799 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a vintage desk fan with a polished chrome cage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the fan's blade motion depicted as a translucent blur generating a steady stream of air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a slender ribbon of brilliant emerald green silk visible, undulating and coiling in the airflow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set on a rich, dark wood desk surface with a soft-focus background showing a glossy black keyboard and stack of papers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting warm and soft, casting subtle shadows from the ribbon onto the polished desk?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0010413468116894364 + ], + "opposite_logprob_tokens": [ + -6.876041412353516 + ], + "logprob_reward": -0.0010413468116894364, + "yes_no_margin_reward": 6.875, + "token_credit_reward": -0.0010413468116894364, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4194492697715759, + 0.08226026594638824, + 0.2010989785194397, + 0.028067899867892265, + 0.03773435205221176, + 0.04051562398672104, + 0.04999293386936188, + 0.04635393247008324, + 0.06598382443189621, + 0.028542889282107353 + ], + "credit_map": [ + 0.02940826117992401, + 0.13688130676746368, + 0.016197971999645233, + 0.020254967734217644, + 0.03285109996795654, + 0.032010503113269806, + 0.13529330492019653, + 0.01655186526477337, + 0.009379366412758827, + 0.007773514837026596, + 0.00686251325532794, + 0.011197945103049278, + 0.011963178403675556, + 0.011037284508347511, + 0.014874887652695179, + 0.009171576239168644, + 0.0262308306992054, + 0.01785183697938919, + 0.031820960342884064, + 0.040434468537569046, + 0.02958555333316326, + 0.015511990524828434, + 0.028468085452914238, + 0.011195245198905468, + 0.006268285680562258, + 0.0023164108861237764, + 0.0033477868419140577, + 0.0031454197596758604, + 0.003465123940259218, + 0.00310550257563591, + 0.003511114278808236, + 0.0029082554392516613, + 0.0038477119524031878, + 0.003135523060336709, + 0.005579410120844841, + 0.005717176478356123, + 0.00725168501958251, + 0.0037015436682850122, + 0.004318691790103912, + 0.00418261019513011, + 0.0038652115035802126, + 0.0037454268895089626, + 0.005134721286594868, + 0.00590461865067482, + 0.006826332770287991, + 0.0063901240937411785, + 0.006234955508261919, + 0.0024142323527485132, + 0.0056557380594313145, + 0.003971180412918329, + 0.0068677254021167755, + 0.005464344751089811, + 0.007426787633448839, + 0.005357042886316776, + 0.009332449175417423, + 0.005917663685977459, + 0.007331897504627705, + 0.004010559991002083, + 0.005373757798224688, + 0.005707571282982826, + 0.005851724650710821, + 0.006154763512313366, + 0.006911533419042826, + 0.005012126639485359, + 0.011809291318058968, + 0.008489863947033882, + 0.005317834671586752, + 0.006338868290185928, + 0.007694787345826626, + 0.005395219195634127, + 0.012140612117946148, + 0.008797353133559227, + 0.0028334937524050474, + 0.002918149344623089, + 0.002444789744913578, + 0.004206894896924496, + 0.0035845560487359762, + 0.0034496018197387457, + 0.006497275084257126, + 0.0026081290561705828 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 1 + ], + "selected_reward": -0.0010413468116894364 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a vintage desk fan with a polished chrome cage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the fan's blade motion depicted as a translucent blur generating a steady stream of air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a slender ribbon of brilliant emerald green silk visible, undulating and coiling in the airflow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set on a rich, dark wood desk surface with a soft-focus background showing a glossy black keyboard and stack of papers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting warm and soft, casting subtle shadows from the ribbon onto the polished desk?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0028147385455667973 + ], + "opposite_logprob_tokens": [ + -5.877814769744873 + ], + "logprob_reward": -0.0028147385455667973, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.0028147385455667973, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16599032282829285, + 0.199098140001297, + 0.12916299700737, + 0.09145595133304596, + 0.06762842833995819, + 0.042976025491952896, + 0.07493051141500473, + 0.07518503814935684, + 0.10832002758979797, + 0.045252569019794464 + ], + "credit_map": [ + 0.016378529369831085, + 0.023109450936317444, + 0.01696309633553028, + 0.03259454667568207, + 0.01984410360455513, + 0.016180824488401413, + 0.030467918142676353, + 0.01045185700058937, + 0.027215315029025078, + 0.030892804265022278, + 0.01382975373417139, + 0.017362503334879875, + 0.019819531589746475, + 0.021560244262218475, + 0.0384039506316185, + 0.030014049261808395, + 0.01927785761654377, + 0.01195685938000679, + 0.014247521758079529, + 0.017194826155900955, + 0.016672207042574883, + 0.013904153369367123, + 0.016231123358011246, + 0.019678447395563126, + 0.015570569783449173, + 0.007575818803161383, + 0.007650409359484911, + 0.007206512615084648, + 0.01022478099912405, + 0.011621474288403988, + 0.013414569199085236, + 0.018191814422607422, + 0.0076153697445988655, + 0.006454033777117729, + 0.00887712650001049, + 0.00921445433050394, + 0.009805756621062756, + 0.00857348833233118, + 0.009268967434763908, + 0.007819226011633873, + 0.005769775714725256, + 0.0032825723756104708, + 0.004316509235650301, + 0.00417099567130208, + 0.006625643000006676, + 0.006087270565330982, + 0.007889775559306145, + 0.004833482205867767, + 0.008048778399825096, + 0.005627296399325132, + 0.00955637451261282, + 0.008448207750916481, + 0.008923993445932865, + 0.0075914436019957066, + 0.013883359730243683, + 0.012851059436798096, + 0.008789712563157082, + 0.013251665979623795, + 0.007447501178830862, + 0.008693094365298748, + 0.0069183725863695145, + 0.008923498913645744, + 0.01019964087754488, + 0.010961556807160378, + 0.01528965588659048, + 0.01190294697880745, + 0.008965984918177128, + 0.011074641719460487, + 0.012059199623763561, + 0.011818831786513329, + 0.02386423386633396, + 0.013344536535441875, + 0.004947008099406958, + 0.005295896902680397, + 0.005025546532124281, + 0.006304656621068716, + 0.006815155036747456, + 0.006011262536048889, + 0.006923384964466095, + 0.003929661121219397 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 2 + ], + "selected_reward": -0.0028147385455667973 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a vintage desk fan with a polished chrome cage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the fan's blade motion depicted as a translucent blur generating a steady stream of air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a slender ribbon of brilliant emerald green silk visible, undulating and coiling in the airflow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set on a rich, dark wood desk surface with a soft-focus background showing a glossy black keyboard and stack of papers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting warm and soft, casting subtle shadows from the ribbon onto the polished desk?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00524667976424098 + ], + "opposite_logprob_tokens": [ + -5.255246639251709 + ], + "logprob_reward": -0.00524667976424098, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.00524667976424098, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.320830762386322, + 0.1325760930776596, + 0.09429173171520233, + 0.03399369493126869, + 0.025202028453350067, + 0.03468102961778641, + 0.04134829714894295, + 0.08123886585235596, + 0.1777581125497818, + 0.05807935819029808 + ], + "credit_map": [ + 0.02688831463456154, + 0.08759929984807968, + 0.018820209428668022, + 0.025959137827157974, + 0.03391842171549797, + 0.024474458768963814, + 0.08747480809688568, + 0.01569611392915249, + 0.017006246373057365, + 0.0205154437571764, + 0.010563203133642673, + 0.013507837429642677, + 0.01775246299803257, + 0.014676533639431, + 0.025087645277380943, + 0.0134667307138443, + 0.011048357002437115, + 0.008988807909190655, + 0.011925898492336273, + 0.014791183173656464, + 0.017713414505124092, + 0.008859057910740376, + 0.012604247778654099, + 0.008360767737030983, + 0.007265517488121986, + 0.0029521321412175894, + 0.00296696275472641, + 0.003214114811271429, + 0.0039060537237674, + 0.004071151837706566, + 0.004409319255501032, + 0.00520844291895628, + 0.0036504853051155806, + 0.0024714034516364336, + 0.002528427168726921, + 0.0030462672002613544, + 0.0038675167597830296, + 0.0034229864832013845, + 0.0036362342070788145, + 0.002578709740191698, + 0.0041495575569570065, + 0.0027972024399787188, + 0.004941735416650772, + 0.004095313139259815, + 0.005467210430651903, + 0.004218826536089182, + 0.005859450902789831, + 0.0031517362222075462, + 0.006980606820434332, + 0.0034501885529607534, + 0.007672259118407965, + 0.0048159267753362656, + 0.004806476179510355, + 0.004472843836992979, + 0.005928249564021826, + 0.0032217504922300577, + 0.01390707679092884, + 0.01642766408622265, + 0.00667984876781702, + 0.006826410070061684, + 0.008822732605040073, + 0.009553605690598488, + 0.01054629310965538, + 0.008475236594676971, + 0.026371970772743225, + 0.02002336084842682, + 0.016080046072602272, + 0.014703149907290936, + 0.016619428992271423, + 0.014927364885807037, + 0.04311378300189972, + 0.025919022038578987, + 0.005660130176693201, + 0.008162634447216988, + 0.005345397628843784, + 0.006923753768205643, + 0.008296222425997257, + 0.006452292203903198, + 0.010988752357661724, + 0.006250174716114998 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 1 + ], + "selected_reward": -0.00524667976424098 + } + ], + "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/0000056/group_12/index_tensor([5874, 5874])_group_12_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000056/group_12/index_tensor([5874, 5874])_group_12_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..de8a7f2529e315bbeb16550755730ff8e3e50a6c --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000056/group_12/index_tensor([5874, 5874])_group_12_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "In a fixed medium shot, the frame steadily captures a moment of magical tranquility. A gentle, moonlit overhead light, emanating from an unidentified source, illuminates the center of the scene with a fairytale-like softness. The light pierces the endless darkness above and focuses on a mirror-like surface of marble. The marble itself is a work of art, its deep base shimmering with dreamy veins of milky white and lavender, like a frozen nebula. From above, countless shimmering golden dust particles slowly fall with an elegant and soothing rhythm. Each particle, like a miniature star, catches and refracts light as it falls, leaving fleeting, wispy trails of light. These golden particles float and swirl like a silent, light snow, gently and effortlessly, finally touching the cool, smooth marble, whose mirror-like qualities perfectly reflect the golden rain. As the dust continues to adhere and accumulate, it gradually forms a thin layer of golden velvet carpet with a warm luster on the marble. The whole process is filled with a quiet, pure and fascinating dream atmosphere.", + "seed": 6818, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000056/group_12/index_5874_seed_6818_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": -0.05969088405836374, + "vlm_logprob": -0.05969088405836374, + "vlm_yes_no_margin": 3.7, + "vlm_token_credit": -0.05969088405836374, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot fixed and medium in framing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a moonlit overhead light source illuminating the center of the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the central surface a mirror-like marble with deep base colors and milky white and lavender veins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do countless shimmering golden dust particles fall slowly from above and reflect light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the falling golden particles accumulate to form a thin layer of golden velvet carpet on the marble?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07611610740423203 + ], + "opposite_logprob_tokens": [ + -3.2011160850524902 + ], + "logprob_reward": -0.07611610740423203, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.07611610740423203, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3269992172718048, + 0.07557417452335358, + 0.0569910891354084, + 0.03739861026406288, + 0.0396229587495327, + 0.047052692621946335, + 0.15130817890167236, + 0.11804192513227463, + 0.054762981832027435, + 0.09224816411733627 + ], + "credit_map": [ + 0.034438926726579666, + 0.05768502131104469, + 0.037651415914297104, + 0.03302725777029991, + 0.04480699077248573, + 0.04363729804754257, + 0.04128001257777214, + 0.034472305327653885, + 0.009846781380474567, + 0.015545902773737907, + 0.005850286688655615, + 0.006290442310273647, + 0.01539439707994461, + 0.01103955414146185, + 0.005801668856292963, + 0.005805143155157566, + 0.006892069708555937, + 0.011226732283830643, + 0.00413152389228344, + 0.005741773638874292, + 0.010378492996096611, + 0.008704700507223606, + 0.004049927927553654, + 0.005865868180990219, + 0.004920827690511942, + 0.006443434860557318, + 0.003983065951615572, + 0.003155661979690194, + 0.006074538454413414, + 0.005483605898916721, + 0.003258305136114359, + 0.004079172387719154, + 0.003642384195700288, + 0.005574085749685764, + 0.0062309266068041325, + 0.0033773297909647226, + 0.00632711173966527, + 0.005708019249141216, + 0.003444140078499913, + 0.005318962968885899, + 0.004722336772829294, + 0.010894037783145905, + 0.0056245774030685425, + 0.0030369421001523733, + 0.007973913103342056, + 0.00746577512472868, + 0.003033643588423729, + 0.00430146511644125, + 0.01626000367105007, + 0.04771331697702408, + 0.00872773490846157, + 0.007596840150654316, + 0.029037831351161003, + 0.027363155037164688, + 0.006945775821805, + 0.007663531694561243, + 0.012304001487791538, + 0.029336629435420036, + 0.010479740798473358, + 0.00806591659784317, + 0.0196742732077837, + 0.01902182400226593, + 0.0077753509394824505, + 0.011384185403585434, + 0.005557862110435963, + 0.014100603759288788, + 0.005334210582077503, + 0.004214730579406023, + 0.00876111164689064, + 0.00787146296352148, + 0.002766466001048684, + 0.006156533025205135, + 0.00818079523742199, + 0.022916095331311226, + 0.011345751583576202, + 0.005713461432605982, + 0.010611695237457752, + 0.013231136836111546, + 0.004505698103457689, + 0.015743529424071312 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 7 + ], + "selected_reward": -0.07611610740423203 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot fixed and medium in framing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a moonlit overhead light source illuminating the center of the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the central surface a mirror-like marble with deep base colors and milky white and lavender veins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do countless shimmering golden dust particles fall slowly from above and reflect light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the falling golden particles accumulate to form a thin layer of golden velvet carpet on the marble?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06208626553416252 + ], + "opposite_logprob_tokens": [ + -2.812086343765259 + ], + "logprob_reward": -0.06208626553416252, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06208626553416252, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09279026091098785, + 0.09380356222391129, + 0.22993332147598267, + 0.15080583095550537, + 0.06008248031139374, + 0.09698095917701721, + 0.049188971519470215, + 0.0885942354798317, + 0.06185678020119667, + 0.07596364617347717 + ], + "credit_map": [ + 0.011796113103628159, + 0.012486962601542473, + 0.010462216101586819, + 0.010272320359945297, + 0.01196040865033865, + 0.010692666284739971, + 0.013859898783266544, + 0.011259679682552814, + 0.010338966734707355, + 0.018973667174577713, + 0.008228635415434837, + 0.008319111540913582, + 0.015332335606217384, + 0.013879948295652866, + 0.008046549744904041, + 0.010684345848858356, + 0.034528207033872604, + 0.05676473304629326, + 0.015758104622364044, + 0.017015818506479263, + 0.042154185473918915, + 0.03953203186392784, + 0.01391618326306343, + 0.010264069773256779, + 0.01881709136068821, + 0.03976799175143242, + 0.011673185043036938, + 0.009705076925456524, + 0.027243508026003838, + 0.024502906948328018, + 0.008013153448700905, + 0.011082923039793968, + 0.006498180329799652, + 0.018206777051091194, + 0.0044570304453372955, + 0.0031037991866469383, + 0.010817956179380417, + 0.009842177852988243, + 0.0033828953746706247, + 0.0037736634258180857, + 0.010655881837010384, + 0.02852087840437889, + 0.006560876499861479, + 0.004796463530510664, + 0.018637241795659065, + 0.017085671424865723, + 0.004400461446493864, + 0.006323486100882292, + 0.004676987882703543, + 0.012953334487974644, + 0.0037666228599846363, + 0.003031413070857525, + 0.010283173061907291, + 0.007377123925834894, + 0.0031657570507377386, + 0.003934554755687714, + 0.008264366537332535, + 0.025571277365088463, + 0.007465542294085026, + 0.005446706432849169, + 0.013407430611550808, + 0.013981277123093605, + 0.006590085569769144, + 0.007867549546062946, + 0.005178236868232489, + 0.01662718690931797, + 0.00560876727104187, + 0.0041333651170134544, + 0.00977700762450695, + 0.009992271661758423, + 0.0045663719065487385, + 0.005973572842776775, + 0.005802605301141739, + 0.021795647218823433, + 0.005684577394276857, + 0.005957007873803377, + 0.01355308759957552, + 0.012686558067798615, + 0.0042081899009644985, + 0.0062759690918028355 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 3, + 5 + ], + "selected_reward": -0.06208626553416252 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot fixed and medium in framing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a moonlit overhead light source illuminating the center of the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the central surface a mirror-like marble with deep base colors and milky white and lavender veins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do countless shimmering golden dust particles fall slowly from above and reflect light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the falling golden particles accumulate to form a thin layer of golden velvet carpet on the marble?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"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": [ + 0.1640383005142212, + 0.11713331192731857, + 0.17776355147361755, + 0.11038035154342651, + 0.060756586492061615, + 0.06778571009635925, + 0.0684564858675003, + 0.08459420502185822, + 0.08727873116731644, + 0.06181277707219124 + ], + "credit_map": [ + 0.01806456409394741, + 0.015065072104334831, + 0.03026011772453785, + 0.015462798997759819, + 0.013756823725998402, + 0.018243150785565376, + 0.018381228670477867, + 0.03480452671647072, + 0.01236763782799244, + 0.015843527391552925, + 0.01863132044672966, + 0.007978410460054874, + 0.019072186201810837, + 0.01610492542386055, + 0.008439401164650917, + 0.01869589276611805, + 0.021768666803836823, + 0.036149151623249054, + 0.019409744068980217, + 0.010382609441876411, + 0.029901256784796715, + 0.027251381427049637, + 0.011287461966276169, + 0.021613260731101036, + 0.011088152416050434, + 0.027186773717403412, + 0.009873475879430771, + 0.007559521123766899, + 0.020105982199311256, + 0.016542911529541016, + 0.006922017317265272, + 0.01110150758177042, + 0.005627182777971029, + 0.012404311448335648, + 0.008586847223341465, + 0.0045599499717354774, + 0.010104544460773468, + 0.00873626209795475, + 0.005483813118189573, + 0.005253673531115055, + 0.007427657954394817, + 0.0151909738779068, + 0.005503594875335693, + 0.003981827292591333, + 0.013157851994037628, + 0.009776039980351925, + 0.004027969669550657, + 0.008719788864254951, + 0.007260207552462816, + 0.015986142680048943, + 0.006110835820436478, + 0.004380068276077509, + 0.011885619722306728, + 0.011010415852069855, + 0.004837269429117441, + 0.006985917221754789, + 0.008592630736529827, + 0.02431841567158699, + 0.008627395145595074, + 0.005521348677575588, + 0.011636891402304173, + 0.01201638299971819, + 0.00709605822339654, + 0.006785065401345491, + 0.007917512208223343, + 0.026367273181676865, + 0.006997251883149147, + 0.005465065594762564, + 0.014205994084477425, + 0.014356338419020176, + 0.005926649086177349, + 0.006042630411684513, + 0.006332503631711006, + 0.015057574026286602, + 0.007885982282459736, + 0.005665704142302275, + 0.0073937526904046535, + 0.008290442638099194, + 0.005195400677621365, + 0.005991407670080662 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 1 + ], + "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 shot fixed and medium in framing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a moonlit overhead light source illuminating the center of the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the central surface a mirror-like marble with deep base colors and milky white and lavender veins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do countless shimmering golden dust particles fall slowly from above and reflect light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the falling golden particles accumulate to form a thin layer of golden velvet carpet on the marble?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0013399679446592927 + ], + "opposite_logprob_tokens": [ + -6.626339912414551 + ], + "logprob_reward": -0.0013399679446592927, + "yes_no_margin_reward": 6.625, + "token_credit_reward": -0.0013399679446592927, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13873055577278137, + 0.058684129267930984, + 0.07347290217876434, + 0.06100156903266907, + 0.17775312066078186, + 0.16060900688171387, + 0.13206161558628082, + 0.08766576647758484, + 0.0683789774775505, + 0.04164232686161995 + ], + "credit_map": [ + 0.02013375237584114, + 0.029295790940523148, + 0.01314616296440363, + 0.014157448895275593, + 0.020262420177459717, + 0.019041404128074646, + 0.010798917151987553, + 0.0118946498259902, + 0.007389751262962818, + 0.010170175693929195, + 0.00427747517824173, + 0.004455127753317356, + 0.013784931972622871, + 0.007891482673585415, + 0.0054413313046097755, + 0.005273852497339249, + 0.010119537822902203, + 0.011403459124267101, + 0.007734264247119427, + 0.004420279059559107, + 0.011283127591013908, + 0.011647788807749748, + 0.004441269673407078, + 0.012423179112374783, + 0.006395314354449511, + 0.014737814664840698, + 0.00491653848439455, + 0.0030250914860516787, + 0.013278813101351261, + 0.009825112298130989, + 0.004626315087080002, + 0.004196572117507458, + 0.01648421213030815, + 0.04683394357562065, + 0.019776733592152596, + 0.010604382492601871, + 0.032184671610593796, + 0.033101897686719894, + 0.01009315811097622, + 0.00867412332445383, + 0.016383137553930283, + 0.045471422374248505, + 0.01241533923894167, + 0.0078058703802526, + 0.032477106899023056, + 0.029475385323166847, + 0.00849741417914629, + 0.008083319291472435, + 0.015408032573759556, + 0.04069065675139427, + 0.007677889429032803, + 0.004939568694680929, + 0.02510804310441017, + 0.025889666751027107, + 0.005915685556828976, + 0.006432077381759882, + 0.009323726408183575, + 0.025874663144350052, + 0.006791387684643269, + 0.004037796054035425, + 0.015846364200115204, + 0.015191183425486088, + 0.005974506493657827, + 0.0046261390671133995, + 0.006398958154022694, + 0.020373279228806496, + 0.005401590373367071, + 0.0035620550625026226, + 0.013860508799552917, + 0.011911069042980671, + 0.003787259804084897, + 0.003084265859797597, + 0.0035792726557701826, + 0.011651674285531044, + 0.003946494776755571, + 0.0025789993815124035, + 0.006793451961129904, + 0.006861406844109297, + 0.0022407863289117813, + 0.0039902362041175365 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 5, + 0 + ], + "selected_reward": -0.0013399679446592927 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot fixed and medium in framing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a moonlit overhead light source illuminating the center of the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the central surface a mirror-like marble with deep base colors and milky white and lavender veins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do countless shimmering golden dust particles fall slowly from above and reflect light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the falling golden particles accumulate to form a thin layer of golden velvet carpet on the marble?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.14283104240894318 + ], + "opposite_logprob_tokens": [ + -2.0178310871124268 + ], + "logprob_reward": -0.14283104240894318, + "yes_no_margin_reward": 1.875, + "token_credit_reward": -0.14283104240894318, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14693902432918549, + 0.04399732127785683, + 0.10554544627666473, + 0.07868684083223343, + 0.19675077497959137, + 0.06921612471342087, + 0.09836260229349136, + 0.10595172643661499, + 0.10386786609888077, + 0.05068231746554375 + ], + "credit_map": [ + 0.018364112824201584, + 0.016672199591994286, + 0.01877211034297943, + 0.018186237663030624, + 0.017455661669373512, + 0.016590574756264687, + 0.027627315372228622, + 0.013270808383822441, + 0.005171309690922499, + 0.007808533962816, + 0.004585199058055878, + 0.004467824939638376, + 0.00883513130247593, + 0.005804820917546749, + 0.003444491419941187, + 0.0038800097536295652, + 0.010317469015717506, + 0.016573062166571617, + 0.012222814373672009, + 0.007657254580408335, + 0.015181466937065125, + 0.014320033602416515, + 0.007723939139395952, + 0.021549412980675697, + 0.009930435568094254, + 0.014525029808282852, + 0.009231907315552235, + 0.003966685384511948, + 0.013643497601151466, + 0.013069847598671913, + 0.006686833221465349, + 0.007632608059793711, + 0.017947494983673096, + 0.04873121902346611, + 0.02100255899131298, + 0.011829077266156673, + 0.03740891069173813, + 0.03772678226232529, + 0.01232653297483921, + 0.009778204374015331, + 0.007434837520122528, + 0.014410370960831642, + 0.0074619511142373085, + 0.0047345287166535854, + 0.013276664540171623, + 0.009760851971805096, + 0.004418039228767157, + 0.007718883920460939, + 0.01076188962906599, + 0.02761596441268921, + 0.006096731871366501, + 0.004879541229456663, + 0.018564818426966667, + 0.01901814714074135, + 0.004517180845141411, + 0.006908322684466839, + 0.011184830218553543, + 0.037043482065200806, + 0.006349548697471619, + 0.006801696494221687, + 0.015010977163910866, + 0.015214096754789352, + 0.006908566690981388, + 0.00743853161111474, + 0.009031741879880428, + 0.02700216881930828, + 0.008418379351496696, + 0.007343880366533995, + 0.01752401888370514, + 0.016750024631619453, + 0.007741178851574659, + 0.01005647238343954, + 0.004529612138867378, + 0.012645157985389233, + 0.0057348813861608505, + 0.003621281124651432, + 0.006694153416901827, + 0.007113484665751457, + 0.004408716689795256, + 0.005935029126703739 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 7 + ], + "selected_reward": -0.14283104240894318 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In a fixed medium shot, the frame steadily captures a moment of magical tranquility. A gentle, moonlit overhead light, emanating from an unidentified source, illuminates the center of the scene with a fairytale-like softness. The light pierces the endless darkness above and focuses on a mirror-like surface of marble. The marble itself is a work of art, its deep base shimmering with dreamy veins of milky white and lavender, like a frozen nebula. From above, countless shimmering golden dust particles slowly fall with an elegant and soothing rhythm. Each particle, like a miniature star, catches and refracts light as it falls, leaving fleeting, wispy trails of light. These golden particles float and swirl like a silent, light snow, gently and effortlessly, finally touching the cool, smooth marble, whose mirror-like qualities perfectly reflect the golden rain. As the dust continues to adhere and accumulate, it gradually forms a thin layer of golden velvet carpet with a warm luster on the marble. The whole process is filled with a quiet, pure and fascinating dream atmosphere.", + "seed": 6819, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000056/group_12/index_5874_seed_6819_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": -0.14253121335059404, + "vlm_logprob": -0.14253121335059404, + "vlm_yes_no_margin": 2.824999988079071, + "vlm_token_credit": -0.14253121335059404, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot fixed and medium in framing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a moonlit overhead light source illuminating the center of the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the central surface a mirror-like marble with deep base colors and milky white and lavender veins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do countless shimmering golden dust particles fall slowly from above and reflect light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the falling golden particles accumulate to form a thin layer of golden velvet carpet on the marble?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.34984689950942993 + ], + "opposite_logprob_tokens": [ + -1.3498468399047852 + ], + "logprob_reward": -0.34984689950942993, + "yes_no_margin_reward": 0.9999999403953552, + "token_credit_reward": -0.34984689950942993, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2198363095521927, + 0.08320033550262451, + 0.030189022421836853, + 0.08110804110765457, + 0.06880566477775574, + 0.05645720660686493, + 0.18620069324970245, + 0.05008991062641144, + 0.16209068894386292, + 0.06202214956283569 + ], + "credit_map": [ + 0.016382716596126556, + 0.020878557115793228, + 0.04103139042854309, + 0.031616877764463425, + 0.02271786704659462, + 0.026515575125813484, + 0.03225855156779289, + 0.028434762731194496, + 0.005091995932161808, + 0.006605171132832766, + 0.018986569717526436, + 0.011141195893287659, + 0.008774195797741413, + 0.006722382269799709, + 0.0160947497934103, + 0.009784070774912834, + 0.0019824288319796324, + 0.003432132303714752, + 0.006720053963363171, + 0.0035756067372858524, + 0.0037360021378844976, + 0.0029903892427682877, + 0.004335866309702396, + 0.0034165421966463327, + 0.00409075478091836, + 0.017753006890416145, + 0.01228735875338316, + 0.006276885513216257, + 0.015619957819581032, + 0.006027107127010822, + 0.010270955972373486, + 0.00878201238811016, + 0.003289462998509407, + 0.009057367220520973, + 0.015891816467046738, + 0.007641674019396305, + 0.007992339320480824, + 0.0047105541452765465, + 0.012932301498949528, + 0.007290143519639969, + 0.00466521643102169, + 0.005103647243231535, + 0.007472688797861338, + 0.007708044722676277, + 0.006576206535100937, + 0.006571803707629442, + 0.010434619151055813, + 0.007924981415271759, + 0.009434076957404613, + 0.020765045657753944, + 0.0407572016119957, + 0.021795758977532387, + 0.019867556169629097, + 0.012628154829144478, + 0.04249479994177818, + 0.018458113074302673, + 0.003021571785211563, + 0.007911033928394318, + 0.01078084111213684, + 0.006307859439402819, + 0.006040408741682768, + 0.0036922076251357794, + 0.0072885300032794476, + 0.005047456827014685, + 0.009288400411605835, + 0.01684102602303028, + 0.027873393148183823, + 0.029962070286273956, + 0.016740018501877785, + 0.016193754971027374, + 0.022686954587697983, + 0.022505080327391624, + 0.0036837374791502953, + 0.005395137704908848, + 0.012868785299360752, + 0.012573369778692722, + 0.005998580250889063, + 0.006379876751452684, + 0.008689005859196186, + 0.006433655042201281 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 8 + ], + "selected_reward": -0.34984689950942993 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot fixed and medium in framing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a moonlit overhead light source illuminating the center of the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the central surface a mirror-like marble with deep base colors and milky white and lavender veins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do countless shimmering golden dust particles fall slowly from above and reflect light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the falling golden particles accumulate to form a thin layer of golden velvet carpet on the marble?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.14286234974861145 + ], + "opposite_logprob_tokens": [ + -2.017862319946289 + ], + "logprob_reward": -0.14286234974861145, + "yes_no_margin_reward": 1.875, + "token_credit_reward": -0.14286234974861145, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.28282997012138367, + 0.09435422718524933, + 0.07683386653661728, + 0.13145706057548523, + 0.0376286581158638, + 0.037135880440473557, + 0.057048652321100235, + 0.13226169347763062, + 0.027158737182617188, + 0.12329117208719254 + ], + "credit_map": [ + 0.016537489369511604, + 0.03696059435606003, + 0.05843053758144379, + 0.030615078285336494, + 0.045915793627500534, + 0.029218019917607307, + 0.039866212755441666, + 0.02528625912964344, + 0.005973959341645241, + 0.01468158233910799, + 0.013392027467489243, + 0.01131808664649725, + 0.015687977895140648, + 0.010223755612969398, + 0.013035445474088192, + 0.010041390545666218, + 0.0046440111473202705, + 0.011492195539176464, + 0.009350750595331192, + 0.008945265784859657, + 0.01787143386900425, + 0.00831394549459219, + 0.008869769051671028, + 0.007346495054662228, + 0.006256496533751488, + 0.016717514023184776, + 0.012537773698568344, + 0.014980399049818516, + 0.020334621891379356, + 0.011926202103495598, + 0.03236442804336548, + 0.016339626163244247, + 0.0020715887658298016, + 0.004945529159158468, + 0.0055465903133153915, + 0.004612633027136326, + 0.006120749283581972, + 0.0045933797955513, + 0.00500134564936161, + 0.004736841190606356, + 0.0019134830217808485, + 0.004629672039300203, + 0.00617131358012557, + 0.004885857924818993, + 0.004459738731384277, + 0.0036232343409210443, + 0.006874097511172295, + 0.0045784879475831985, + 0.003323903540149331, + 0.005605571437627077, + 0.010062213987112045, + 0.006859394256025553, + 0.007344372104853392, + 0.005856967531144619, + 0.010457196272909641, + 0.007539032958447933, + 0.007730316836386919, + 0.016294753178954124, + 0.02374259941279888, + 0.020400920882821083, + 0.017148081213235855, + 0.01270014327019453, + 0.02323753386735916, + 0.011007341556251049, + 0.0017801172798499465, + 0.003054976463317871, + 0.003778039012104273, + 0.005221029743552208, + 0.003273236798122525, + 0.00296446168795228, + 0.0038976604118943214, + 0.0031892149709165096, + 0.006693734787404537, + 0.014034394174814224, + 0.017524827271699905, + 0.019758012145757675, + 0.016274144873023033, + 0.012370756827294827, + 0.01733655482530594, + 0.019298752769827843 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 3 + ], + "selected_reward": -0.14286234974861145 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot fixed and medium in framing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a moonlit overhead light source illuminating the center of the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the central surface a mirror-like marble with deep base colors and milky white and lavender veins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do countless shimmering golden dust particles fall slowly from above and reflect light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the falling golden particles accumulate to form a thin layer of golden velvet carpet on the marble?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014209521003067493 + ], + "opposite_logprob_tokens": [ + -4.264209747314453 + ], + "logprob_reward": -0.014209521003067493, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014209521003067493, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17654284834861755, + 0.1580515056848526, + 0.11201819032430649, + 0.12505678832530975, + 0.06323353946208954, + 0.09074007719755173, + 0.03919783979654312, + 0.07783377170562744, + 0.08225061744451523, + 0.07507486641407013 + ], + "credit_map": [ + 0.01023445650935173, + 0.023844555020332336, + 0.026213202625513077, + 0.029972899705171585, + 0.020321015268564224, + 0.015438498929142952, + 0.032796379178762436, + 0.017721818760037422, + 0.008323037065565586, + 0.014890521764755249, + 0.025708001106977463, + 0.024121487513184547, + 0.015621103346347809, + 0.015030584298074245, + 0.03511740639805794, + 0.019239338114857674, + 0.006475965492427349, + 0.019701775163412094, + 0.018666014075279236, + 0.010680992156267166, + 0.01971546746790409, + 0.008305283263325691, + 0.017421161755919456, + 0.011051516979932785, + 0.00669697904959321, + 0.014249628409743309, + 0.02658863551914692, + 0.014165833592414856, + 0.017108533531427383, + 0.010790273547172546, + 0.019048569723963737, + 0.016408316791057587, + 0.0030975216068327427, + 0.00858418457210064, + 0.015971863642334938, + 0.009364892728626728, + 0.007047863211482763, + 0.00308103091083467, + 0.010310432873666286, + 0.005775745492428541, + 0.003909578081220388, + 0.008937899954617023, + 0.01415472012013197, + 0.013384565711021423, + 0.010594162158668041, + 0.005602066870778799, + 0.024316711351275444, + 0.009840362705290318, + 0.0020209525246173143, + 0.007831351831555367, + 0.00531762046739459, + 0.005879784002900124, + 0.005923552438616753, + 0.0030549117363989353, + 0.005536587443202734, + 0.0036330760922282934, + 0.004593510180711746, + 0.009807392954826355, + 0.013488094322383404, + 0.012486299499869347, + 0.01039399579167366, + 0.006729385815560818, + 0.013473516330122948, + 0.006861573085188866, + 0.006127516273409128, + 0.007854223251342773, + 0.010573784820735455, + 0.01678006909787655, + 0.009839427657425404, + 0.009418237023055553, + 0.01247330754995346, + 0.00918404757976532, + 0.003921184688806534, + 0.01058516837656498, + 0.010253294371068478, + 0.015052450820803642, + 0.009256369434297085, + 0.00629065278917551, + 0.009506131522357464, + 0.010209605097770691 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 3 + ], + "selected_reward": -0.014209521003067493 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot fixed and medium in framing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a moonlit overhead light source illuminating the center of the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the central surface a mirror-like marble with deep base colors and milky white and lavender veins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do countless shimmering golden dust particles fall slowly from above and reflect light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the falling golden particles accumulate to form a thin layer of golden velvet carpet on the marble?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0041033364832401276 + ], + "opposite_logprob_tokens": [ + -5.504103183746338 + ], + "logprob_reward": -0.0041033364832401276, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.0041033364832401276, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1507556289434433, + 0.15242211520671844, + 0.09016460180282593, + 0.14476534724235535, + 0.052624333649873734, + 0.10975366830825806, + 0.033249374479055405, + 0.06490267813205719, + 0.04655493050813675, + 0.15480738878250122 + ], + "credit_map": [ + 0.011050170287489891, + 0.017633425071835518, + 0.03425515070557594, + 0.021550873294472694, + 0.013878394849598408, + 0.014038868248462677, + 0.021978138014674187, + 0.016370585188269615, + 0.00802754145115614, + 0.017237139865756035, + 0.03873923420906067, + 0.018838046118617058, + 0.016154760494828224, + 0.00798137579113245, + 0.02948964759707451, + 0.015954354777932167, + 0.0046043600887060165, + 0.01446938794106245, + 0.019405260682106018, + 0.010423289611935616, + 0.011308207176625729, + 0.005999252200126648, + 0.014871214516460896, + 0.009083625860512257, + 0.0069120218977332115, + 0.016135431826114655, + 0.03387255594134331, + 0.017010753974318504, + 0.0183093398809433, + 0.010983645915985107, + 0.02609126828610897, + 0.015450307168066502, + 0.003364193020388484, + 0.005411691498011351, + 0.011368805542588234, + 0.00895780511200428, + 0.004855288192629814, + 0.0039520589634776115, + 0.009048131294548512, + 0.005666359327733517, + 0.005830219481140375, + 0.010604077950119972, + 0.026518531143665314, + 0.013208616524934769, + 0.010951612144708633, + 0.00687450310215354, + 0.02221287041902542, + 0.013553230091929436, + 0.0021670754067599773, + 0.004110758192837238, + 0.005450779106467962, + 0.004412462469190359, + 0.0042442637495696545, + 0.003475884208455682, + 0.005732402671128511, + 0.0036557461135089397, + 0.004421338904649019, + 0.00797218643128872, + 0.013446456752717495, + 0.009090343490242958, + 0.0077587622217834, + 0.006580262444913387, + 0.00900694727897644, + 0.00662637036293745, + 0.002940515289083123, + 0.007369921077042818, + 0.010500803589820862, + 0.007797637954354286, + 0.005230833310633898, + 0.0028805998153984547, + 0.004321017302572727, + 0.0055136000737547874, + 0.007424088194966316, + 0.016609758138656616, + 0.027328593656420708, + 0.03643316030502319, + 0.017874201759696007, + 0.011536854319274426, + 0.021599480882287025, + 0.01600123941898346 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 1, + 0 + ], + "selected_reward": -0.0041033364832401276 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot fixed and medium in framing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a moonlit overhead light source illuminating the center of the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the central surface a mirror-like marble with deep base colors and milky white and lavender veins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do countless shimmering golden dust particles fall slowly from above and reflect light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the falling golden particles accumulate to form a thin layer of golden velvet carpet on the marble?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.20163396000862122 + ], + "opposite_logprob_tokens": [ + -1.7016339302062988 + ], + "logprob_reward": -0.20163396000862122, + "yes_no_margin_reward": 1.5, + "token_credit_reward": -0.20163396000862122, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17685545980930328, + 0.07989640533924103, + 0.03724684938788414, + 0.09225349128246307, + 0.08713371306657791, + 0.13044962286949158, + 0.060104962438344955, + 0.051097795367240906, + 0.06539561599493027, + 0.21956609189510345 + ], + "credit_map": [ + 0.011235175654292107, + 0.026146020740270615, + 0.03161216899752617, + 0.026805665343999863, + 0.01973157748579979, + 0.01139911450445652, + 0.03086581639945507, + 0.01905994303524494, + 0.0037001995369791985, + 0.009052565321326256, + 0.015647048130631447, + 0.00920908898115158, + 0.00927599798887968, + 0.005645291414111853, + 0.018415287137031555, + 0.008950937539339066, + 0.0022986901458352804, + 0.005008447449654341, + 0.005594093818217516, + 0.003996267449110746, + 0.005914586596190929, + 0.003487410955131054, + 0.006475259084254503, + 0.004472094587981701, + 0.004303286783397198, + 0.015248123556375504, + 0.015670347958803177, + 0.009667347185313702, + 0.015525873750448227, + 0.005835042800754309, + 0.01716882735490799, + 0.008834641426801682, + 0.004153014160692692, + 0.00844667013734579, + 0.019208671525120735, + 0.011103867553174496, + 0.00849408470094204, + 0.006330410949885845, + 0.02013147436082363, + 0.009265528060495853, + 0.0062979403883218765, + 0.014010091312229633, + 0.028636574745178223, + 0.014783504419028759, + 0.013751634396612644, + 0.007650237064808607, + 0.029481276869773865, + 0.015838366001844406, + 0.0029826234094798565, + 0.005671667400747538, + 0.012395682744681835, + 0.0073945773765444756, + 0.0067648012191057205, + 0.005071205552667379, + 0.013090020045638084, + 0.006734386086463928, + 0.0031680758111178875, + 0.007160264067351818, + 0.008213374763727188, + 0.007525934372097254, + 0.006304506212472916, + 0.005153969395905733, + 0.008060358464717865, + 0.005511310417205095, + 0.003486916422843933, + 0.00774069270119071, + 0.009843823499977589, + 0.013703898526728153, + 0.007880747318267822, + 0.0058601307682693005, + 0.009005554020404816, + 0.007873854599893093, + 0.011545185931026936, + 0.023191384971141815, + 0.04512852802872658, + 0.048781849443912506, + 0.02413102053105831, + 0.01502505037933588, + 0.030670417472720146, + 0.021092664450407028 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 5 + ], + "selected_reward": -0.20163396000862122 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In a fixed medium shot, the frame steadily captures a moment of magical tranquility. A gentle, moonlit overhead light, emanating from an unidentified source, illuminates the center of the scene with a fairytale-like softness. The light pierces the endless darkness above and focuses on a mirror-like surface of marble. The marble itself is a work of art, its deep base shimmering with dreamy veins of milky white and lavender, like a frozen nebula. From above, countless shimmering golden dust particles slowly fall with an elegant and soothing rhythm. Each particle, like a miniature star, catches and refracts light as it falls, leaving fleeting, wispy trails of light. These golden particles float and swirl like a silent, light snow, gently and effortlessly, finally touching the cool, smooth marble, whose mirror-like qualities perfectly reflect the golden rain. As the dust continues to adhere and accumulate, it gradually forms a thin layer of golden velvet carpet with a warm luster on the marble. The whole process is filled with a quiet, pure and fascinating dream atmosphere.", + "seed": 6820, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000056/group_12/index_5874_seed_6820_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": -0.12947030300274492, + "vlm_logprob": -0.12947030300274492, + "vlm_yes_no_margin": 2.8, + "vlm_token_credit": -0.12947030300274492, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot fixed and medium in framing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a moonlit overhead light source illuminating the center of the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the central surface a mirror-like marble with deep base colors and milky white and lavender veins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do countless shimmering golden dust particles fall slowly from above and reflect light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the falling golden particles accumulate to form a thin layer of golden velvet carpet on the marble?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.39943429827690125 + ], + "opposite_logprob_tokens": [ + -1.2744343280792236 + ], + "logprob_reward": -0.39943429827690125, + "yes_no_margin_reward": 0.875, + "token_credit_reward": -0.39943429827690125, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23092307150363922, + 0.15955165028572083, + 0.08395645767450333, + 0.058632589876651764, + 0.08491133898496628, + 0.06542354077100754, + 0.04753606393933296, + 0.08271748572587967, + 0.05238449573516846, + 0.13396328687667847 + ], + "credit_map": [ + 0.02108309604227543, + 0.02078818902373314, + 0.04975467547774315, + 0.026406705379486084, + 0.016323501244187355, + 0.02149878442287445, + 0.026760058477520943, + 0.048308055847883224, + 0.01030091941356659, + 0.01289559155702591, + 0.030545219779014587, + 0.023411128669977188, + 0.008014002814888954, + 0.01718706078827381, + 0.029004046693444252, + 0.028193680569529533, + 0.00757954828441143, + 0.012815910391509533, + 0.011457366868853569, + 0.011683565564453602, + 0.009690622799098492, + 0.008122649043798447, + 0.009983867406845093, + 0.01262292917817831, + 0.005313587374985218, + 0.008793696761131287, + 0.00841311365365982, + 0.006595853250473738, + 0.007370494306087494, + 0.006463944911956787, + 0.006561423186212778, + 0.009120479226112366, + 0.007249732501804829, + 0.005651217419654131, + 0.014600430615246296, + 0.014024625532329082, + 0.005562720354646444, + 0.011621170677244663, + 0.011652285233139992, + 0.014549154788255692, + 0.005950292572379112, + 0.007486557587981224, + 0.00882981438189745, + 0.008435548283159733, + 0.00750304339453578, + 0.008660574443638325, + 0.006504496559500694, + 0.012053212150931358, + 0.002996182767674327, + 0.00523549597710371, + 0.008612878620624542, + 0.0067186541855335236, + 0.006620477885007858, + 0.007437983527779579, + 0.004799821879714727, + 0.005114572588354349, + 0.00406245281919837, + 0.003751369658857584, + 0.029318872839212418, + 0.005822630599141121, + 0.003907894715666771, + 0.0240537878125906, + 0.004832867998629808, + 0.006967606954276562, + 0.005302537232637405, + 0.005670165177434683, + 0.007016067858785391, + 0.007440636400133371, + 0.00527821434661746, + 0.008929941803216934, + 0.006092880852520466, + 0.00665405485779047, + 0.01124943420290947, + 0.008467328734695911, + 0.02389049343764782, + 0.01964515820145607, + 0.007783095818012953, + 0.029096996411681175, + 0.014389583840966225, + 0.019441185519099236 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 9 + ], + "selected_reward": -0.39943429827690125 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot fixed and medium in framing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a moonlit overhead light source illuminating the center of the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the central surface a mirror-like marble with deep base colors and milky white and lavender veins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do countless shimmering golden dust particles fall slowly from above and reflect light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the falling golden particles accumulate to form a thin layer of golden velvet carpet on the marble?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04310736060142517 + ], + "opposite_logprob_tokens": [ + -3.168107271194458 + ], + "logprob_reward": -0.04310736060142517, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04310736060142517, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06348248571157455, + 0.024114245548844337, + 0.05900127813220024, + 0.036972250789403915, + 0.01563885062932968, + 0.016131270676851273, + 0.018523676320910454, + 0.5350016355514526, + 0.04296104237437248, + 0.18817320466041565 + ], + "credit_map": [ + 0.004512535408139229, + 0.004688572604209185, + 0.010931316763162613, + 0.00744299590587616, + 0.0035841159988194704, + 0.006720537319779396, + 0.01114641036838293, + 0.0144560057669878, + 0.00245395814999938, + 0.0030103977769613266, + 0.0036118379794061184, + 0.0027988231740891933, + 0.002222494687885046, + 0.002602141350507736, + 0.002703065285459161, + 0.004711525049060583, + 0.00610122038051486, + 0.009594526141881943, + 0.010784942656755447, + 0.006334438920021057, + 0.007358410395681858, + 0.004997515119612217, + 0.007379158399999142, + 0.006451065186411142, + 0.0028481120243668556, + 0.004763399250805378, + 0.008226430974900723, + 0.0046423994936048985, + 0.004266036674380302, + 0.0030239620245993137, + 0.0052272421307861805, + 0.0039746686816215515, + 0.0013104514218866825, + 0.002012164331972599, + 0.002625787165015936, + 0.0016364868497475982, + 0.0016093060839921236, + 0.001310318592004478, + 0.003458787454292178, + 0.0016755490796640515, + 0.0013957482296973467, + 0.0016106711700558662, + 0.0021431564819067717, + 0.002325372537598014, + 0.0013089408166706562, + 0.0019200292881578207, + 0.002160650910809636, + 0.0032667024061083794, + 0.0014459509402513504, + 0.0016525075770914555, + 0.0029484624974429607, + 0.0030502781737595797, + 0.0014758172910660505, + 0.0018933819374069571, + 0.0030824707355350256, + 0.002974805887788534, + 0.025037264451384544, + 0.008935486897826195, + 0.21137569844722748, + 0.02536751702427864, + 0.015613901428878307, + 0.1837770938873291, + 0.021070707589387894, + 0.0438239723443985, + 0.004080486483871937, + 0.004017940256744623, + 0.006275932304561138, + 0.005302749574184418, + 0.004086796659976244, + 0.006469350773841143, + 0.005501975305378437, + 0.007225811015814543, + 0.014202623628079891, + 0.02667989954352379, + 0.024534091353416443, + 0.021563515067100525, + 0.029531728476285934, + 0.02941402606666088, + 0.01467360183596611, + 0.02757371962070465 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 9, + 0 + ], + "selected_reward": -0.04310736060142517 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot fixed and medium in framing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a moonlit overhead light source illuminating the center of the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the central surface a mirror-like marble with deep base colors and milky white and lavender veins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do countless shimmering golden dust particles fall slowly from above and reflect light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the falling golden particles accumulate to form a thin layer of golden velvet carpet on the marble?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10035992413759232 + ], + "opposite_logprob_tokens": [ + -2.3503599166870117 + ], + "logprob_reward": -0.10035992413759232, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10035992413759232, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1663869321346283, + 0.04679016023874283, + 0.039006251841783524, + 0.03144686296582222, + 0.03034270368516445, + 0.04100130870938301, + 0.03770458698272705, + 0.4074227809906006, + 0.08020959794521332, + 0.11968877166509628 + ], + "credit_map": [ + 0.010933205485343933, + 0.012052432633936405, + 0.04648476466536522, + 0.013030601665377617, + 0.009373053908348083, + 0.014205262996256351, + 0.034456077963113785, + 0.025851532816886902, + 0.0038632580544799566, + 0.006486553233116865, + 0.006653284654021263, + 0.005954574327915907, + 0.0035826456733047962, + 0.004390865098685026, + 0.0061301738023757935, + 0.009728805162012577, + 0.002825613133609295, + 0.005021398421376944, + 0.005020861513912678, + 0.004800456576049328, + 0.0062707699835300446, + 0.003363820491358638, + 0.005293354392051697, + 0.006409976631402969, + 0.0020663158502429724, + 0.003544761100783944, + 0.005971870850771666, + 0.004196938592940569, + 0.004208975937217474, + 0.0027249713893979788, + 0.004379796329885721, + 0.0043532331474125385, + 0.0024794642813503742, + 0.003049821127206087, + 0.004802288953214884, + 0.0036817274522036314, + 0.0036918409168720245, + 0.0023784709628671408, + 0.005227206274867058, + 0.005031882785260677, + 0.00307271396741271, + 0.003888295963406563, + 0.004843559581786394, + 0.005786709487438202, + 0.003887767205014825, + 0.005101875402033329, + 0.005954533815383911, + 0.008465851657092571, + 0.0028096719179302454, + 0.004672035109251738, + 0.006815663073211908, + 0.005321868695318699, + 0.004343823529779911, + 0.004506959114223719, + 0.00449630431830883, + 0.004738262854516506, + 0.019744526594877243, + 0.007984740659594536, + 0.1605934202671051, + 0.020756734535098076, + 0.013435782864689827, + 0.12467756122350693, + 0.024009620770812035, + 0.03622040897607803, + 0.007137659937143326, + 0.0043885670602321625, + 0.014166357927024364, + 0.011493824422359467, + 0.004665105137974024, + 0.015293039381504059, + 0.010438015684485435, + 0.012627030722796917, + 0.007166978903114796, + 0.017829054966568947, + 0.01467742957174778, + 0.016473405063152313, + 0.019210796803236008, + 0.015275324694812298, + 0.011578712612390518, + 0.01747707463800907 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 9 + ], + "selected_reward": -0.10035992413759232 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot fixed and medium in framing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a moonlit overhead light source illuminating the center of the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the central surface a mirror-like marble with deep base colors and milky white and lavender veins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do countless shimmering golden dust particles fall slowly from above and reflect light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the falling golden particles accumulate to form a thin layer of golden velvet carpet on the marble?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004101437050849199 + ], + "opposite_logprob_tokens": [ + -5.504101276397705 + ], + "logprob_reward": -0.004101437050849199, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004101437050849199, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09799082577228546, + 0.06673679500818253, + 0.0741431713104248, + 0.06403283029794693, + 0.0421028658747673, + 0.024518905207514763, + 0.0724743977189064, + 0.19616922736167908, + 0.09703373908996582, + 0.26479724049568176 + ], + "credit_map": [ + 0.0075324587523937225, + 0.00930633395910263, + 0.02163379266858101, + 0.01067010872066021, + 0.005659602582454681, + 0.010855204425752163, + 0.013504980131983757, + 0.01882834918797016, + 0.0068838102743029594, + 0.015420567244291306, + 0.008691866882145405, + 0.007984851486980915, + 0.006992128677666187, + 0.005691091064363718, + 0.007260967046022415, + 0.00781151233240962, + 0.007779587060213089, + 0.012533964589238167, + 0.009974555112421513, + 0.009583103470504284, + 0.01085985079407692, + 0.007045980077236891, + 0.008394631557166576, + 0.007971503771841526, + 0.0077179898507893085, + 0.011395454406738281, + 0.010221734642982483, + 0.008914713747799397, + 0.006410721689462662, + 0.0062902444042265415, + 0.005793318152427673, + 0.007288646884262562, + 0.004221735987812281, + 0.005326211452484131, + 0.005836467258632183, + 0.0057881963439285755, + 0.004279598128050566, + 0.00419269222766161, + 0.006965228356420994, + 0.005492733791470528, + 0.0020025211852043867, + 0.0019387747161090374, + 0.0035591265186667442, + 0.0030963062308728695, + 0.0017609215574339032, + 0.003471877658739686, + 0.0034603984095156193, + 0.005228979978710413, + 0.006001485977321863, + 0.010615144856274128, + 0.009497781284153461, + 0.010855153203010559, + 0.010885452851653099, + 0.00813476461917162, + 0.00801217183470726, + 0.008472446352243423, + 0.010745296254754066, + 0.008344604633748531, + 0.0739799290895462, + 0.012247187085449696, + 0.007894245907664299, + 0.05107232928276062, + 0.014349907636642456, + 0.017535723745822906, + 0.009013724513351917, + 0.007225280161947012, + 0.016254020854830742, + 0.013788900338113308, + 0.009356281720101833, + 0.012314126826822758, + 0.015551909804344177, + 0.013529494404792786, + 0.016817424446344376, + 0.03785938024520874, + 0.028070613741874695, + 0.03364507853984833, + 0.04178164526820183, + 0.04945187270641327, + 0.023024724796414375, + 0.034146521240472794 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 7, + 0 + ], + "selected_reward": -0.004101437050849199 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot fixed and medium in framing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a moonlit overhead light source illuminating the center of the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the central surface a mirror-like marble with deep base colors and milky white and lavender veins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do countless shimmering golden dust particles fall slowly from above and reflect light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the falling golden particles accumulate to form a thin layer of golden velvet carpet on the marble?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10034849494695663 + ], + "opposite_logprob_tokens": [ + -2.350348472595215 + ], + "logprob_reward": -0.10034849494695663, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10034849494695663, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06933262199163437, + 0.05073360726237297, + 0.07755683362483978, + 0.0748828575015068, + 0.05372827127575874, + 0.05125655233860016, + 0.0618291012942791, + 0.3519814610481262, + 0.07917561382055283, + 0.1295231133699417 + ], + "credit_map": [ + 0.004816185683012009, + 0.007624390069395304, + 0.007003428880125284, + 0.009293651208281517, + 0.007062854710966349, + 0.008995029143989086, + 0.010216157883405685, + 0.014320923015475273, + 0.004104815423488617, + 0.007647796534001827, + 0.008704292587935925, + 0.006100838538259268, + 0.00464332802221179, + 0.005316030699759722, + 0.0063680922612547874, + 0.007848413661122322, + 0.006747662555426359, + 0.009924488142132759, + 0.010258441790938377, + 0.009779558517038822, + 0.010189009830355644, + 0.008447064086794853, + 0.01050728838890791, + 0.011703326366841793, + 0.006204725708812475, + 0.009096718393266201, + 0.013794143684208393, + 0.01039972435683012, + 0.010216004215180874, + 0.006543644238263369, + 0.01017792895436287, + 0.008449970744550228, + 0.003969641402363777, + 0.004611952695995569, + 0.012590592727065086, + 0.005363910458981991, + 0.00413382425904274, + 0.005548766814172268, + 0.010071543976664543, + 0.00743804220110178, + 0.003680511610582471, + 0.004277278669178486, + 0.007687252480536699, + 0.007125661242753267, + 0.00474914675578475, + 0.006461647804826498, + 0.006219509989023209, + 0.011055542156100273, + 0.004977827426046133, + 0.008561423979699612, + 0.009437336586415768, + 0.008996888995170593, + 0.009072836488485336, + 0.006885171867907047, + 0.006934253964573145, + 0.006963360123336315, + 0.017084309831261635, + 0.009987016208469868, + 0.12105229496955872, + 0.018000952899456024, + 0.013434510678052902, + 0.12999752163887024, + 0.01382829062640667, + 0.02859657071530819, + 0.007908927276730537, + 0.008299232460558414, + 0.01203437615185976, + 0.011131241917610168, + 0.008068778552114964, + 0.009530717507004738, + 0.011723801493644714, + 0.01047853846102953, + 0.010080564767122269, + 0.017606394365429878, + 0.015335433185100555, + 0.018654340878129005, + 0.018064064905047417, + 0.016575684770941734, + 0.015545148402452469, + 0.01766149327158928 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 9, + 8 + ], + "selected_reward": -0.10034849494695663 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In a fixed medium shot, the frame steadily captures a moment of magical tranquility. A gentle, moonlit overhead light, emanating from an unidentified source, illuminates the center of the scene with a fairytale-like softness. The light pierces the endless darkness above and focuses on a mirror-like surface of marble. The marble itself is a work of art, its deep base shimmering with dreamy veins of milky white and lavender, like a frozen nebula. From above, countless shimmering golden dust particles slowly fall with an elegant and soothing rhythm. Each particle, like a miniature star, catches and refracts light as it falls, leaving fleeting, wispy trails of light. These golden particles float and swirl like a silent, light snow, gently and effortlessly, finally touching the cool, smooth marble, whose mirror-like qualities perfectly reflect the golden rain. As the dust continues to adhere and accumulate, it gradually forms a thin layer of golden velvet carpet with a warm luster on the marble. The whole process is filled with a quiet, pure and fascinating dream atmosphere.", + "seed": 6821, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000056/group_12/index_5874_seed_6821_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": -0.1015824228990823, + "vlm_logprob": -0.1015824228990823, + "vlm_yes_no_margin": 3.2, + "vlm_token_credit": -0.1015824228990823, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot fixed and medium in framing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a moonlit overhead light source illuminating the center of the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the central surface a mirror-like marble with deep base colors and milky white and lavender veins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do countless shimmering golden dust particles fall slowly from above and reflect light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the falling golden particles accumulate to form a thin layer of golden velvet carpet on the marble?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2804926633834839 + ], + "opposite_logprob_tokens": [ + -1.6554926633834839 + ], + "logprob_reward": -0.2804926633834839, + "yes_no_margin_reward": 1.375, + "token_credit_reward": -0.2804926633834839, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1571471393108368, + 0.020404761657118797, + 0.0552077479660511, + 0.03933842480182648, + 0.06919952481985092, + 0.5360351800918579, + 0.026603704318404198, + 0.041788939386606216, + 0.03660326078534126, + 0.017671389505267143 + ], + "credit_map": [ + 0.008180815726518631, + 0.018426991999149323, + 0.013435312546789646, + 0.06042977422475815, + 0.003706671530380845, + 0.007267522159963846, + 0.018802499398589134, + 0.026897547766566277, + 0.001560556935146451, + 0.001987675204873085, + 0.0025925065856426954, + 0.005258460063487291, + 0.0011476515792310238, + 0.0012774845818057656, + 0.002205662429332733, + 0.004374765325337648, + 0.0023799226619303226, + 0.005287604406476021, + 0.004990920890122652, + 0.021319331601262093, + 0.00225386512465775, + 0.0020208931528031826, + 0.0069145578891038895, + 0.010040649212896824, + 0.0019205197459086776, + 0.0035726979840546846, + 0.004500387702137232, + 0.010261211544275284, + 0.0019587776623666286, + 0.0030872842762619257, + 0.003671605372801423, + 0.010365946218371391, + 0.00406687892973423, + 0.0033711197320371866, + 0.012082246132194996, + 0.007379821967333555, + 0.004516771994531155, + 0.0043437425047159195, + 0.016948547214269638, + 0.016490397974848747, + 0.014195567928254604, + 0.01403784193098545, + 0.04573862999677658, + 0.11519755423069, + 0.026620348915457726, + 0.019480612128973007, + 0.19434651732444763, + 0.10641805827617645, + 0.0018954805564135313, + 0.0026672868989408016, + 0.006522704381495714, + 0.004208846483379602, + 0.0025527803227305412, + 0.0017689844826236367, + 0.0032387496903538704, + 0.0037488725502043962, + 0.005694039631634951, + 0.0032978670205920935, + 0.0049873231910169125, + 0.008433505892753601, + 0.0028381289448589087, + 0.0024972804822027683, + 0.004316100850701332, + 0.009724692441523075, + 0.001546255312860012, + 0.0027474937960505486, + 0.0033916328102350235, + 0.013327821157872677, + 0.0011530222836881876, + 0.0019121746299788356, + 0.003968631848692894, + 0.008556227199733257, + 0.00089140550699085, + 0.0012277258792892098, + 0.002843476366251707, + 0.006233359221369028, + 0.0008645639754831791, + 0.0007733174716122448, + 0.00309357070364058, + 0.001743970438838005 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 4 + ], + "selected_reward": -0.2804926633834839 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot fixed and medium in framing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a moonlit overhead light source illuminating the center of the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the central surface a mirror-like marble with deep base colors and milky white and lavender veins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do countless shimmering golden dust particles fall slowly from above and reflect light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the falling golden particles accumulate to form a thin layer of golden velvet carpet on the marble?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03814159333705902 + ], + "opposite_logprob_tokens": [ + -3.2881414890289307 + ], + "logprob_reward": -0.03814159333705902, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03814159333705902, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.42575904726982117, + 0.05305515602231026, + 0.06901445984840393, + 0.061380982398986816, + 0.09388834238052368, + 0.0545673593878746, + 0.024495933204889297, + 0.053177475929260254, + 0.13081951439380646, + 0.03384174406528473 + ], + "credit_map": [ + 0.015364370308816433, + 0.056751273572444916, + 0.027276089414954185, + 0.17006251215934753, + 0.009406950324773788, + 0.018089886754751205, + 0.05362127348780632, + 0.07518669962882996, + 0.002312879776582122, + 0.0055190217681229115, + 0.0040154303424060345, + 0.019494395703077316, + 0.0031652299221605062, + 0.002587450435385108, + 0.005628054030239582, + 0.010332692414522171, + 0.0038816758897155523, + 0.007031114771962166, + 0.006757368799299002, + 0.023768624290823936, + 0.004568145610392094, + 0.003285522572696209, + 0.008791428990662098, + 0.01093058381229639, + 0.002510793972760439, + 0.005634477827697992, + 0.005845073610544205, + 0.023517468944191933, + 0.0025384423788636923, + 0.0024941200390458107, + 0.01275478396564722, + 0.006085822358727455, + 0.004869254305958748, + 0.006881369277834892, + 0.008565027266740799, + 0.034935127943754196, + 0.0038993372581899166, + 0.0032023934181779623, + 0.018116282299160957, + 0.01341954804956913, + 0.002446176717057824, + 0.0022142883390188217, + 0.005204725544899702, + 0.00957164354622364, + 0.0029255624394863844, + 0.002653928939253092, + 0.019513849169015884, + 0.010037186555564404, + 0.0018455247627571225, + 0.0016286523314192891, + 0.0033345308620482683, + 0.006715763825923204, + 0.001620655762962997, + 0.0014209981309249997, + 0.004849917721003294, + 0.003079889575019479, + 0.0020092835184186697, + 0.0026022919919341803, + 0.0061159092001616955, + 0.011203384026885033, + 0.003473144257441163, + 0.0025194198824465275, + 0.01897740736603737, + 0.006276632659137249, + 0.00475924601778388, + 0.01007926557213068, + 0.009711435995995998, + 0.05209752172231674, + 0.0041315252892673016, + 0.006487807724624872, + 0.018587617203593254, + 0.024965081363916397, + 0.0014150537317618728, + 0.0028621689416468143, + 0.0032586408779025078, + 0.013679472729563713, + 0.0011305921943858266, + 0.0015321235405281186, + 0.006064117420464754, + 0.0038995707873255014 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 4 + ], + "selected_reward": -0.03814159333705902 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot fixed and medium in framing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a moonlit overhead light source illuminating the center of the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the central surface a mirror-like marble with deep base colors and milky white and lavender veins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do countless shimmering golden dust particles fall slowly from above and reflect light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the falling golden particles accumulate to form a thin layer of golden velvet carpet on the marble?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026369301602244377 + ], + "opposite_logprob_tokens": [ + -3.651369333267212 + ], + "logprob_reward": -0.026369301602244377, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026369301602244377, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20324409008026123, + 0.09188665449619293, + 0.1648760586977005, + 0.09591002762317657, + 0.07226277142763138, + 0.1072421595454216, + 0.02956542931497097, + 0.18053561449050903, + 0.021758757531642914, + 0.03271843120455742 + ], + "credit_map": [ + 0.009012291207909584, + 0.02389775589108467, + 0.017563899978995323, + 0.0778479352593422, + 0.005829507950693369, + 0.009203070774674416, + 0.029817640781402588, + 0.030071988701820374, + 0.004398004151880741, + 0.008314038626849651, + 0.007241369225084782, + 0.03369789198040962, + 0.003934179898351431, + 0.004552771337330341, + 0.008118350058794022, + 0.021630054339766502, + 0.007002539467066526, + 0.016366848722100258, + 0.013885264284908772, + 0.06511587649583817, + 0.00624556839466095, + 0.005854999180883169, + 0.02036179043352604, + 0.03004317730665207, + 0.003784785745665431, + 0.009454448707401752, + 0.007299415301531553, + 0.038333360105752945, + 0.0031725794542580843, + 0.004072947893291712, + 0.015764852985739708, + 0.014027634635567665, + 0.005259345751255751, + 0.006123319733887911, + 0.00979828741401434, + 0.01876426674425602, + 0.004980482626706362, + 0.0040834941901266575, + 0.006520009599626064, + 0.01673356629908085, + 0.0031237874645739794, + 0.002596257720142603, + 0.007961363531649113, + 0.02262832783162594, + 0.005374210420995951, + 0.003764066379517317, + 0.039441391825675964, + 0.022352756932377815, + 0.0014789848355576396, + 0.002980452962219715, + 0.004005293361842632, + 0.008620335720479488, + 0.002343327272683382, + 0.001737206825055182, + 0.002455479232594371, + 0.005944348871707916, + 0.005822793580591679, + 0.005754384677857161, + 0.01922529749572277, + 0.04157392308115959, + 0.009573806077241898, + 0.00707887951284647, + 0.05933486670255661, + 0.03217167779803276, + 0.001599423005245626, + 0.0020938673987984657, + 0.0037178185302764177, + 0.005326100159436464, + 0.0019397542346268892, + 0.001465601148083806, + 0.0023201252333819866, + 0.0032960674725472927, + 0.001545650651678443, + 0.0029178212862461805, + 0.005972324404865503, + 0.010606306605041027, + 0.001865952042862773, + 0.001426444505341351, + 0.005035922862589359, + 0.0033480064012110233 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 2 + ], + "selected_reward": -0.026369301602244377 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot fixed and medium in framing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a moonlit overhead light source illuminating the center of the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the central surface a mirror-like marble with deep base colors and milky white and lavender veins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do countless shimmering golden dust particles fall slowly from above and reflect light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the falling golden particles accumulate to form a thin layer of golden velvet carpet on the marble?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002493845531716943 + ], + "opposite_logprob_tokens": [ + -6.002493858337402 + ], + "logprob_reward": -0.002493845531716943, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.002493845531716943, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15940390527248383, + 0.05607044696807861, + 0.1222618892788887, + 0.030283300206065178, + 0.09667076170444489, + 0.2164410948753357, + 0.025030121207237244, + 0.1472926139831543, + 0.06781432777643204, + 0.07873158901929855 + ], + "credit_map": [ + 0.007777063641697168, + 0.021043632179498672, + 0.014940066263079643, + 0.05528996139764786, + 0.010983612388372421, + 0.008426574990153313, + 0.02397780865430832, + 0.01696518249809742, + 0.003849431173875928, + 0.010398454032838345, + 0.006927658338099718, + 0.005626159254461527, + 0.007178339641541243, + 0.004934447351843119, + 0.00654904730618, + 0.010606912896037102, + 0.006483125500380993, + 0.013782303780317307, + 0.010170516557991505, + 0.04389151185750961, + 0.007128125987946987, + 0.005710292607545853, + 0.015112776309251785, + 0.019983239471912384, + 0.00287827430292964, + 0.003754256060346961, + 0.004267233423888683, + 0.0050962381064891815, + 0.003659273264929652, + 0.0027046927716583014, + 0.004072339739650488, + 0.003850991604849696, + 0.007242455147206783, + 0.0074428366497159, + 0.011799643747508526, + 0.030981024727225304, + 0.005992475897073746, + 0.004828912205994129, + 0.015165569260716438, + 0.01321784034371376, + 0.006506149657070637, + 0.004990255460143089, + 0.012311077676713467, + 0.041300851851701736, + 0.011247520335018635, + 0.007868942804634571, + 0.08202030509710312, + 0.05019599199295044, + 0.0017183865420520306, + 0.002816135063767433, + 0.0033802122343331575, + 0.005570306908339262, + 0.0024774831254035234, + 0.0018066830234602094, + 0.0027302252128720284, + 0.004530688282102346, + 0.004157882649451494, + 0.003359058406203985, + 0.01290472038090229, + 0.03133844584226608, + 0.007238567806780338, + 0.005118310451507568, + 0.05206751078367233, + 0.031108126044273376, + 0.0036709881387650967, + 0.005109761841595173, + 0.00546359084546566, + 0.026573000475764275, + 0.0029113732744008303, + 0.002590940799564123, + 0.015135823749005795, + 0.006358850281685591, + 0.0031025384087115526, + 0.006651340518146753, + 0.0064060925506055355, + 0.03104565665125847, + 0.0034171398729085922, + 0.003057155990973115, + 0.014704611152410507, + 0.010347055271267891 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 7 + ], + "selected_reward": -0.002493845531716943 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot fixed and medium in framing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a moonlit overhead light source illuminating the center of the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the central surface a mirror-like marble with deep base colors and milky white and lavender veins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do countless shimmering golden dust particles fall slowly from above and reflect light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the falling golden particles accumulate to form a thin layer of golden velvet carpet on the marble?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1604147106409073 + ], + "opposite_logprob_tokens": [ + -1.910414695739746 + ], + "logprob_reward": -0.1604147106409073, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.1604147106409073, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13707740604877472, + 0.11307252943515778, + 0.07073629647493362, + 0.10556598752737045, + 0.1270633190870285, + 0.03132502734661102, + 0.03668660297989845, + 0.22125336527824402, + 0.11022473126649857, + 0.04699475318193436 + ], + "credit_map": [ + 0.010731196962296963, + 0.016677234321832657, + 0.013207140378654003, + 0.046686600893735886, + 0.0054588099010288715, + 0.006383090279996395, + 0.013841626234352589, + 0.02409171313047409, + 0.0050573390908539295, + 0.01209992729127407, + 0.008745145983994007, + 0.04293733462691307, + 0.004174035973846912, + 0.004724432248622179, + 0.012082141824066639, + 0.023252170532941818, + 0.003253673668950796, + 0.007226969581097364, + 0.005417416803538799, + 0.027498973533511162, + 0.003056152258068323, + 0.0026534933131188154, + 0.0079152537509799, + 0.013714363798499107, + 0.004161459393799305, + 0.010516222566366196, + 0.008129249326884747, + 0.040751710534095764, + 0.004775641486048698, + 0.005268332548439503, + 0.015124065801501274, + 0.016839299350976944, + 0.0059321364387869835, + 0.008426937274634838, + 0.01419633999466896, + 0.048481427133083344, + 0.007013639900833368, + 0.005762175656855106, + 0.024981239810585976, + 0.012269426137208939, + 0.0015667881816625595, + 0.002246081829071045, + 0.005423486698418856, + 0.003367470810189843, + 0.0027878861874341965, + 0.0019222144037485123, + 0.008028390817344189, + 0.005982708185911179, + 0.0028396614361554384, + 0.003681118367239833, + 0.006160958670079708, + 0.009148950688540936, + 0.0035587134771049023, + 0.0024702411610633135, + 0.004830043762922287, + 0.003996913321316242, + 0.009286322630941868, + 0.005482546519488096, + 0.013736998662352562, + 0.036394283175468445, + 0.009453046135604382, + 0.008205980062484741, + 0.08108873665332794, + 0.057605449110269547, + 0.004181708674877882, + 0.008494867011904716, + 0.00889250822365284, + 0.04403042793273926, + 0.0036194371059536934, + 0.005547656212002039, + 0.01595638133585453, + 0.019501736387610435, + 0.0027350543532520533, + 0.004570119548588991, + 0.00629595248028636, + 0.015866035595536232, + 0.0034759240224957466, + 0.002854762366041541, + 0.0066921291872859, + 0.004504775628447533 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 4 + ], + "selected_reward": -0.1604147106409073 + } + ], + "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/0000056/group_13/index_tensor([2363, 2363])_group_13_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000056/group_13/index_tensor([2363, 2363])_group_13_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..4d3ca5542753e8eaf2171e1aecb18f226bbe6441 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000056/group_13/index_tensor([2363, 2363])_group_13_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "An extreme close-up shot, captured from the perspective of the object itself, closely follows the steady and rapid descent of a round, matte-textured white tablet into clear water. The moment it is fully submerged, illuminated by bright, even overhead lighting, its surface begins to dissolve rapidly, and from its core, thousands of tiny, densely packed white bubbles erupt violently. These bubbles, captured in slow motion, form a turbulent, upward-moving stream of energetic white. As they rise, they collide, merge, and burst, stirring the surrounding still water in an uncontrolled microscopic storm. The tablet itself is stirred, tumbled, and gradually disintegrated in this violent chemical reaction, its shattered white particles merging with the surging stream of bubbles. The dominant tones of the image are the clear turquoise of the water and the pure white of the bubbles and tablet. Light filtering through the bubbles creates a shimmering, translucent sheen, transforming this microscopic chemical reaction into a visual feast of primal power and dynamic beauty, imbued with the ultimate explosive power and tension of competitive sports.", + "seed": 5830, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000056/group_13/index_2363_seed_5830_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": -0.2696846309117973, + "vlm_logprob": -0.2696846309117973, + "vlm_yes_no_margin": 4.05, + "vlm_token_credit": -0.2696846309117973, + "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 from the perspective of the object itself?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a round, matte-textured white tablet descending into clear water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the tablet dissolve rapidly upon submersion, emitting thousands of tiny white bubbles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the bubbles captured in slow motion forming a turbulent, upward-moving stream?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by bright, even overhead lighting with dominant turquoise water tones?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"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": [ + -1.3238760232925415 + ], + "opposite_logprob_tokens": [ + -0.3238759934902191 + ], + "logprob_reward": -1.3238760232925415, + "yes_no_margin_reward": -1.0, + "token_credit_reward": -1.3238760232925415, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.056316811591386795, + 0.050495028495788574, + 0.030317114666104317, + 0.06285495311021805, + 0.11580463498830795, + 0.243657186627388, + 0.1536303609609604, + 0.02060742862522602, + 0.0630558431148529, + 0.203260600566864 + ], + "credit_map": [ + 0.004306187387555838, + 0.009116188623011112, + 0.0052501000463962555, + 0.010457856580615044, + 0.006527354009449482, + 0.006883944850414991, + 0.006732074078172445, + 0.007043108344078064, + 0.003813880728557706, + 0.005820463411509991, + 0.0024735622573643923, + 0.005266392137855291, + 0.0027837941888719797, + 0.002915444318205118, + 0.014658862724900246, + 0.012762627564370632, + 0.0033634451683610678, + 0.00377236888743937, + 0.004272079095244408, + 0.004401028621941805, + 0.004220661241561174, + 0.003486385801807046, + 0.002864555222913623, + 0.003936590626835823, + 0.00859797839075327, + 0.00733366422355175, + 0.006840602960437536, + 0.007047960069030523, + 0.009800531901419163, + 0.00917824823409319, + 0.0070413886569440365, + 0.007014583796262741, + 0.01656213402748108, + 0.010969292372465134, + 0.00915975496172905, + 0.008884728886187077, + 0.011809209361672401, + 0.022166967391967773, + 0.010680553503334522, + 0.02557199075818062, + 0.00884782150387764, + 0.006266545969992876, + 0.006444354075938463, + 0.006437945645302534, + 0.014202041551470757, + 0.08834624290466309, + 0.016213634982705116, + 0.09689858555793762, + 0.006840886082500219, + 0.00878914911299944, + 0.00620238296687603, + 0.007929078303277493, + 0.011442696675658226, + 0.0470251739025116, + 0.01389035489410162, + 0.05151064693927765, + 0.0024029791820794344, + 0.0029491328168660402, + 0.0017862843815237284, + 0.0022623129189014435, + 0.002014495199546218, + 0.004011786077171564, + 0.0017845258116722107, + 0.0033959129359573126, + 0.002157893730327487, + 0.0038215848617255688, + 0.0021783809643238783, + 0.002853507874533534, + 0.006194116082042456, + 0.01937101036310196, + 0.0038553441409021616, + 0.022624002769589424, + 0.017065148800611496, + 0.01705831103026867, + 0.01680852472782135, + 0.020787909626960754, + 0.02057873271405697, + 0.04700174555182457, + 0.012451037764549255, + 0.051509201526641846 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 9, + 6 + ], + "selected_reward": -1.3238760232925415 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 from the perspective of the object itself?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a round, matte-textured white tablet descending into clear water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the tablet dissolve rapidly upon submersion, emitting thousands of tiny white bubbles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the bubbles captured in slow motion forming a turbulent, upward-moving stream?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by bright, even overhead lighting with dominant turquoise water tones?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012563135474920273 + ], + "opposite_logprob_tokens": [ + -4.387563228607178 + ], + "logprob_reward": -0.012563135474920273, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012563135474920273, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.045904677361249924, + 0.06903154402971268, + 0.031231122091412544, + 0.08709345757961273, + 0.07915084064006805, + 0.1719452142715454, + 0.25495225191116333, + 0.07515507936477661, + 0.12346244603395462, + 0.06207335740327835 + ], + "credit_map": [ + 0.0036027291789650917, + 0.0051017990335822105, + 0.0030695137102156878, + 0.006211866624653339, + 0.0047363052144646645, + 0.0044600265100598335, + 0.0090491883456707, + 0.009673249907791615, + 0.007680862210690975, + 0.014467190951108932, + 0.005221241619437933, + 0.011111218482255936, + 0.007590204942971468, + 0.006795549765229225, + 0.009084307588636875, + 0.007080969400703907, + 0.004320724401623011, + 0.004252534359693527, + 0.004131087101995945, + 0.0046821581199765205, + 0.0035746153444051743, + 0.0033203191123902798, + 0.002824835479259491, + 0.004124847706407309, + 0.012489108368754387, + 0.01195668987929821, + 0.010221322067081928, + 0.010441302321851254, + 0.011312909424304962, + 0.010751957073807716, + 0.010184423066675663, + 0.009735743515193462, + 0.01299026608467102, + 0.011442877352237701, + 0.010187825188040733, + 0.009575671516358852, + 0.008974522352218628, + 0.00747754005715251, + 0.008615915663540363, + 0.009886223822832108, + 0.008917920291423798, + 0.007459460292011499, + 0.006453065201640129, + 0.006172088906168938, + 0.009894009679555893, + 0.05755864083766937, + 0.012196588329970837, + 0.06329343467950821, + 0.010058025829494, + 0.01506403461098671, + 0.006220581941306591, + 0.012070409953594208, + 0.020677894353866577, + 0.08271143585443497, + 0.020185653120279312, + 0.08796420693397522, + 0.013141972944140434, + 0.011586138047277927, + 0.007478440646082163, + 0.008145169354975224, + 0.008662081323564053, + 0.009507478214800358, + 0.008250223472714424, + 0.008383573032915592, + 0.005828947760164738, + 0.009804390370845795, + 0.004540989175438881, + 0.007118229288607836, + 0.016858384013175964, + 0.03322929888963699, + 0.00850074551999569, + 0.037581462413072586, + 0.0058886040933430195, + 0.008727048523724079, + 0.008254649117588997, + 0.011020810343325138, + 0.00868247915059328, + 0.006677187047898769, + 0.006774592213332653, + 0.006047986447811127 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 5, + 8 + ], + "selected_reward": -0.012563135474920273 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 from the perspective of the object itself?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a round, matte-textured white tablet descending into clear water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the tablet dissolve rapidly upon submersion, emitting thousands of tiny white bubbles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the bubbles captured in slow motion forming a turbulent, upward-moving stream?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by bright, even overhead lighting with dominant turquoise water tones?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0059483409859240055 + ], + "opposite_logprob_tokens": [ + -5.130948543548584 + ], + "logprob_reward": -0.0059483409859240055, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.0059483409859240055, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06208796054124832, + 0.04841556027531624, + 0.056230444461107254, + 0.09794724732637405, + 0.03861525282263756, + 0.061203598976135254, + 0.34869384765625, + 0.09291817247867584, + 0.10166364908218384, + 0.09222428500652313 + ], + "credit_map": [ + 0.005933208856731653, + 0.00888009462505579, + 0.007299613207578659, + 0.014022179879248142, + 0.006495009642094374, + 0.005836581811308861, + 0.00768149271607399, + 0.005939776543527842, + 0.003435678081586957, + 0.0038160739932209253, + 0.0033444445580244064, + 0.0058419108390808105, + 0.0031876747962087393, + 0.0034121631179004908, + 0.01402329746633768, + 0.011354313232004642, + 0.008615154772996902, + 0.007358741480857134, + 0.007499547675251961, + 0.007092386484146118, + 0.007743400055915117, + 0.006055969279259443, + 0.0060819173231720924, + 0.005783319007605314, + 0.011603198945522308, + 0.01273530162870884, + 0.01416150201112032, + 0.01664494164288044, + 0.010262169875204563, + 0.010963892564177513, + 0.011314836330711842, + 0.010261391289532185, + 0.004722944926470518, + 0.0034227967262268066, + 0.002856384962797165, + 0.003850047942250967, + 0.0034038322046399117, + 0.0075661297887563705, + 0.003832727437838912, + 0.008960384875535965, + 0.004885976668447256, + 0.005415773019194603, + 0.004359893500804901, + 0.00558263948187232, + 0.0069316960871219635, + 0.013796389102935791, + 0.006080832332372665, + 0.014150397852063179, + 0.029907941818237305, + 0.05174653232097626, + 0.030739709734916687, + 0.04654465615749359, + 0.039079464972019196, + 0.04580006003379822, + 0.05549819767475128, + 0.04937727004289627, + 0.00901296827942133, + 0.014052116312086582, + 0.0104692243039608, + 0.016651568934321404, + 0.009644093923270702, + 0.014598566107451916, + 0.008104865439236164, + 0.010384763590991497, + 0.004408383741974831, + 0.008026622235774994, + 0.004722794983536005, + 0.005747051443904638, + 0.012584803625941277, + 0.028826864436268806, + 0.0061357272788882256, + 0.03121139481663704, + 0.006276859901845455, + 0.009402251802384853, + 0.00596868060529232, + 0.00833882112056017, + 0.007911537773907185, + 0.024144968017935753, + 0.005116737447679043, + 0.02506442554295063 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 8, + 3 + ], + "selected_reward": -0.0059483409859240055 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 from the perspective of the object itself?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a round, matte-textured white tablet descending into clear water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the tablet dissolve rapidly upon submersion, emitting thousands of tiny white bubbles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the bubbles captured in slow motion forming a turbulent, upward-moving stream?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by bright, even overhead lighting with dominant turquoise water tones?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004096094518899918 + ], + "opposite_logprob_tokens": [ + -5.504096031188965 + ], + "logprob_reward": -0.004096094518899918, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004096094518899918, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.037182487547397614, + 0.03335486724972725, + 0.02388322539627552, + 0.04530169069766998, + 0.08568188548088074, + 0.2702946066856384, + 0.2970331609249115, + 0.04696352779865265, + 0.07887260615825653, + 0.08143186569213867 + ], + "credit_map": [ + 0.003091941587626934, + 0.0034883725456893444, + 0.004148819949477911, + 0.007182232104241848, + 0.003585429396480322, + 0.0032277253922075033, + 0.006912319455295801, + 0.005545651074498892, + 0.002905072644352913, + 0.004728204570710659, + 0.00192291010171175, + 0.0038027188275009394, + 0.002409625332802534, + 0.0030403065029531717, + 0.007450514938682318, + 0.007095517590641975, + 0.0032614280935376883, + 0.0028698521200567484, + 0.0039566559717059135, + 0.003871825523674488, + 0.002793618245050311, + 0.0024696129839867353, + 0.0025092114228755236, + 0.0021510219667106867, + 0.00930081307888031, + 0.005413328763097525, + 0.004389118868857622, + 0.004561431240290403, + 0.006123567931354046, + 0.005384236574172974, + 0.004249787889420986, + 0.005879408214241266, + 0.009860025718808174, + 0.007474706042557955, + 0.006120013538748026, + 0.006036564242094755, + 0.007694828324019909, + 0.019344840198755264, + 0.007192418910562992, + 0.021958496421575546, + 0.007767833769321442, + 0.005751448683440685, + 0.005875216796994209, + 0.006180898752063513, + 0.015583833679556847, + 0.10037536174058914, + 0.017897164449095726, + 0.11086288839578629, + 0.012096748687326908, + 0.013423265889286995, + 0.008951105177402496, + 0.009061808697879314, + 0.020979847759008408, + 0.10439255088567734, + 0.018401499837636948, + 0.10972633957862854, + 0.007133061531931162, + 0.004663489758968353, + 0.003792050527408719, + 0.005475482903420925, + 0.004194498062133789, + 0.009179380722343922, + 0.004071001894772053, + 0.008454566821455956, + 0.0054778000339865685, + 0.009015969932079315, + 0.004380065482109785, + 0.006254889536648989, + 0.014512084424495697, + 0.015822840854525566, + 0.006780882366001606, + 0.01662808284163475, + 0.005424966104328632, + 0.009875241667032242, + 0.008866631425917149, + 0.01324604544788599, + 0.008956359699368477, + 0.013299508951604366, + 0.008020433597266674, + 0.013742687180638313 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 5, + 4 + ], + "selected_reward": -0.004096094518899918 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up from the perspective of the object itself?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a round, matte-textured white tablet descending into clear water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the tablet dissolve rapidly upon submersion, emitting thousands of tiny white bubbles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the bubbles captured in slow motion forming a turbulent, upward-moving stream?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by bright, even overhead lighting with dominant turquoise water tones?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0019395602867007256 + ], + "opposite_logprob_tokens": [ + -6.25193977355957 + ], + "logprob_reward": -0.0019395602867007256, + "yes_no_margin_reward": 6.25, + "token_credit_reward": -0.0019395602867007256, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1041024699807167, + 0.02765919268131256, + 0.022325851023197174, + 0.037510618567466736, + 0.10870704799890518, + 0.2145339995622635, + 0.3727063536643982, + 0.050301969051361084, + 0.019384395331144333, + 0.04276812821626663 + ], + "credit_map": [ + 0.009310782887041569, + 0.015452459454536438, + 0.010836604051291943, + 0.020273005589842796, + 0.011119555681943893, + 0.011771008372306824, + 0.012529321946203709, + 0.012809737585484982, + 0.0028813390526920557, + 0.004210491664707661, + 0.0019065005471929908, + 0.004769435152411461, + 0.0021263076923787594, + 0.0019369582878425717, + 0.005135405343025923, + 0.004692757036536932, + 0.0032671340741217136, + 0.0029345666989684105, + 0.002002821071073413, + 0.0033210262190550566, + 0.0024370327591896057, + 0.0022945082746446133, + 0.002502053277567029, + 0.00356670911423862, + 0.0040788245387375355, + 0.004709560889750719, + 0.00535453949123621, + 0.004491444211453199, + 0.004972073715180159, + 0.005137292668223381, + 0.005093513987958431, + 0.0036733730230480433, + 0.008310290984809399, + 0.00630365964025259, + 0.006624705623835325, + 0.0067596519365906715, + 0.007435766514390707, + 0.02983846515417099, + 0.00965651124715805, + 0.033778004348278046, + 0.005949508864432573, + 0.003828431712463498, + 0.004899105057120323, + 0.00430989870801568, + 0.012655220925807953, + 0.08047055453062057, + 0.01345384493470192, + 0.0889674499630928, + 0.01127589400857687, + 0.01606588624417782, + 0.007669898681342602, + 0.010526306927204132, + 0.029377123340964317, + 0.13215093314647675, + 0.023020517081022263, + 0.142619788646698, + 0.007545324973762035, + 0.005732255522161722, + 0.005570040550082922, + 0.006424987688660622, + 0.004718198906630278, + 0.007808503229171038, + 0.005215400364249945, + 0.007287259213626385, + 0.0030723176896572113, + 0.0023064485285431147, + 0.002097038784995675, + 0.0024845001753419638, + 0.0031139024067670107, + 0.001976943574845791, + 0.002241509733721614, + 0.0020917365327477455, + 0.00366110703907907, + 0.005327967926859856, + 0.0028329601045697927, + 0.009545392356812954, + 0.005153491161763668, + 0.004570962395519018, + 0.0045676217414438725, + 0.007108628284186125 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 5, + 4 + ], + "selected_reward": -0.0019395602867007256 + } + ], + "vlm_decoded_outputs": [ + "No", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An extreme close-up shot, captured from the perspective of the object itself, closely follows the steady and rapid descent of a round, matte-textured white tablet into clear water. The moment it is fully submerged, illuminated by bright, even overhead lighting, its surface begins to dissolve rapidly, and from its core, thousands of tiny, densely packed white bubbles erupt violently. These bubbles, captured in slow motion, form a turbulent, upward-moving stream of energetic white. As they rise, they collide, merge, and burst, stirring the surrounding still water in an uncontrolled microscopic storm. The tablet itself is stirred, tumbled, and gradually disintegrated in this violent chemical reaction, its shattered white particles merging with the surging stream of bubbles. The dominant tones of the image are the clear turquoise of the water and the pure white of the bubbles and tablet. Light filtering through the bubbles creates a shimmering, translucent sheen, transforming this microscopic chemical reaction into a visual feast of primal power and dynamic beauty, imbued with the ultimate explosive power and tension of competitive sports.", + "seed": 5831, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000056/group_13/index_2363_seed_5831_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": -0.008470721042249352, + "vlm_logprob": -0.008470721042249352, + "vlm_yes_no_margin": 6.0, + "vlm_token_credit": -0.008470721042249352, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 from the perspective of the object itself?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a round, matte-textured white tablet descending into clear water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the tablet dissolve rapidly upon submersion, emitting thousands of tiny white bubbles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the bubbles captured in slow motion forming a turbulent, upward-moving stream?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by bright, even overhead lighting with dominant turquoise water tones?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.036152295768260956 + ], + "opposite_logprob_tokens": [ + -3.6611523628234863 + ], + "logprob_reward": -0.036152295768260956, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.036152295768260956, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.27790674567222595, + 0.061842065304517746, + 0.0761154368519783, + 0.17126613855361938, + 0.04103042557835579, + 0.034637533128261566, + 0.06711667776107788, + 0.028151996433734894, + 0.1929188072681427, + 0.0490141436457634 + ], + "credit_map": [ + 0.03261743113398552, + 0.02039421908557415, + 0.026579374447464943, + 0.02822461538016796, + 0.03076908178627491, + 0.0656842589378357, + 0.025965141132473946, + 0.04767264425754547, + 0.005413503386080265, + 0.00832424871623516, + 0.005541852675378323, + 0.01084643229842186, + 0.00857748743146658, + 0.006979386787861586, + 0.0074020205065608025, + 0.008757130242884159, + 0.007095869164913893, + 0.010455360636115074, + 0.006873180158436298, + 0.015189437195658684, + 0.00903542060405016, + 0.008963220752775669, + 0.008938299492001534, + 0.00956464372575283, + 0.04821232333779335, + 0.010674425400793552, + 0.01700480654835701, + 0.010435407049953938, + 0.01005480159074068, + 0.02237735316157341, + 0.011740660294890404, + 0.04076635465025902, + 0.005083917174488306, + 0.003967018332332373, + 0.004391047637909651, + 0.005638668313622475, + 0.0049088080413639545, + 0.0069648995995521545, + 0.0066201770678162575, + 0.0034558901097625494, + 0.0032274876721203327, + 0.005785500165075064, + 0.0043424745090305805, + 0.007218278478831053, + 0.003750825533643365, + 0.0028663005214184523, + 0.004964757710695267, + 0.0024819052778184414, + 0.003993734251707792, + 0.009137380868196487, + 0.005899548064917326, + 0.015630876645445824, + 0.007090273778885603, + 0.003606423269957304, + 0.01384130772203207, + 0.007917128503322601, + 0.0021142938639968634, + 0.006117746699601412, + 0.0022425625938922167, + 0.005361365154385567, + 0.005336621776223183, + 0.0016070957062765956, + 0.003634120337665081, + 0.001738190185278654, + 0.01002516970038414, + 0.06729302555322647, + 0.06962288171052933, + 0.012609358876943588, + 0.012268037535250187, + 0.004272757098078728, + 0.011739962734282017, + 0.005087623838335276, + 0.0030152674298733473, + 0.007516145706176758, + 0.007731765508651733, + 0.008877950720489025, + 0.00361680262722075, + 0.006398576311767101, + 0.007024267688393593, + 0.0048333690501749516 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 3 + ], + "selected_reward": -0.036152295768260956 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 from the perspective of the object itself?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a round, matte-textured white tablet descending into clear water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the tablet dissolve rapidly upon submersion, emitting thousands of tiny white bubbles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the bubbles captured in slow motion forming a turbulent, upward-moving stream?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by bright, even overhead lighting with dominant turquoise water tones?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002840296132490039 + ], + "opposite_logprob_tokens": [ + -5.877840518951416 + ], + "logprob_reward": -0.002840296132490039, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.002840296132490039, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10310182720422745, + 0.1372307389974594, + 0.06461495906114578, + 0.07234619557857513, + 0.0586576946079731, + 0.08393354713916779, + 0.0216120146214962, + 0.06465496867895126, + 0.3687414526939392, + 0.02510654367506504 + ], + "credit_map": [ + 0.009811480529606342, + 0.009597968310117722, + 0.01302956510335207, + 0.01351979747414589, + 0.011685716919600964, + 0.019170349463820457, + 0.009915575385093689, + 0.01637137494981289, + 0.010787627659738064, + 0.016474811360239983, + 0.01279747299849987, + 0.02417760156095028, + 0.01429193839430809, + 0.021920647472143173, + 0.017321260645985603, + 0.019459377974271774, + 0.004971692338585854, + 0.008503417484462261, + 0.006832307670265436, + 0.008959694765508175, + 0.0077440752647817135, + 0.008701613172888756, + 0.007605460472404957, + 0.011296696029603481, + 0.006459223572164774, + 0.009595805779099464, + 0.009795406833291054, + 0.010020189918577671, + 0.008011233992874622, + 0.008774620480835438, + 0.009790577925741673, + 0.009899134747684002, + 0.009418929927051067, + 0.0060653844848275185, + 0.005192334298044443, + 0.008551209233701229, + 0.0069280159659683704, + 0.00867403857409954, + 0.009545299224555492, + 0.004282481502741575, + 0.005614690948277712, + 0.016109740361571312, + 0.009354698471724987, + 0.012215472757816315, + 0.00872371718287468, + 0.003841874422505498, + 0.022014763206243515, + 0.006058594677597284, + 0.0016765240579843521, + 0.00321337697096169, + 0.0034598165657371283, + 0.004160257056355476, + 0.00236170319840312, + 0.0020689840894192457, + 0.00315767340362072, + 0.0015136775327846408, + 0.0049022892490029335, + 0.009211750701069832, + 0.011003958061337471, + 0.012417901307344437, + 0.0093998359516263, + 0.0031170565634965897, + 0.010496031492948532, + 0.004106148146092892, + 0.014002298004925251, + 0.14044372737407684, + 0.08527836203575134, + 0.026536332443356514, + 0.0233094934374094, + 0.0074252355843782425, + 0.06425507366657257, + 0.007490960881114006, + 0.0021772608160972595, + 0.003682608949020505, + 0.0044007087126374245, + 0.0034675493370741606, + 0.001925322925671935, + 0.0045656440779566765, + 0.002481163013726473, + 0.0024062851443886757 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 1, + 0 + ], + "selected_reward": -0.002840296132490039 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 from the perspective of the object itself?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a round, matte-textured white tablet descending into clear water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the tablet dissolve rapidly upon submersion, emitting thousands of tiny white bubbles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the bubbles captured in slow motion forming a turbulent, upward-moving stream?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by bright, even overhead lighting with dominant turquoise water tones?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0009240409708581865 + ], + "opposite_logprob_tokens": [ + -7.000924110412598 + ], + "logprob_reward": -0.0009240409708581865, + "yes_no_margin_reward": 7.0, + "token_credit_reward": -0.0009240409708581865, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14003434777259827, + 0.03658100590109825, + 0.151924729347229, + 0.25147542357444763, + 0.04492051899433136, + 0.10155420750379562, + 0.020081812515854836, + 0.03390523046255112, + 0.20039962232112885, + 0.019123084843158722 + ], + "credit_map": [ + 0.012911347672343254, + 0.010367856360971928, + 0.019124485552310944, + 0.015505988150835037, + 0.015992160886526108, + 0.02583242580294609, + 0.0179063118994236, + 0.022393761202692986, + 0.0037603985983878374, + 0.003812747076153755, + 0.004096901509910822, + 0.005200664978474379, + 0.004405491054058075, + 0.0062014441937208176, + 0.004555328283458948, + 0.004548027645796537, + 0.012397038750350475, + 0.02219497039914131, + 0.016084318980574608, + 0.023261476308107376, + 0.022845325991511345, + 0.02061224728822708, + 0.02008216455578804, + 0.014447188936173916, + 0.06612187623977661, + 0.014880512841045856, + 0.023757265880703926, + 0.01910557597875595, + 0.013550757430493832, + 0.036248356103897095, + 0.013583718799054623, + 0.06422737240791321, + 0.008527415804564953, + 0.0048943995498120785, + 0.003964419011026621, + 0.007221234031021595, + 0.0037296751979738474, + 0.005041768774390221, + 0.007632886990904808, + 0.003908717539161444, + 0.00503607839345932, + 0.019088536500930786, + 0.01220544520765543, + 0.01684788055717945, + 0.009695837274193764, + 0.0060973926447331905, + 0.02474440261721611, + 0.007838642224669456, + 0.0010703008156269789, + 0.003218948608264327, + 0.002524441806599498, + 0.004749268293380737, + 0.002162009244784713, + 0.0011625098995864391, + 0.0035538277588784695, + 0.0016405053902417421, + 0.00251532974652946, + 0.0058849891647696495, + 0.006568278651684523, + 0.006649668328464031, + 0.00510956859216094, + 0.0015190760605037212, + 0.003497747238725424, + 0.0021605691872537136, + 0.00967467948794365, + 0.07460885494947433, + 0.06481184810400009, + 0.012553759850561619, + 0.012879635207355022, + 0.004712688270956278, + 0.015204104594886303, + 0.0059540485963225365, + 0.0018322294345125556, + 0.0022054158616811037, + 0.0028451718389987946, + 0.0026753824204206467, + 0.0016856376314535737, + 0.002891842508688569, + 0.002891765208914876, + 0.0020956401713192463 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 8, + 2 + ], + "selected_reward": -0.0009240409708581865 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 from the perspective of the object itself?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a round, matte-textured white tablet descending into clear water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the tablet dissolve rapidly upon submersion, emitting thousands of tiny white bubbles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the bubbles captured in slow motion forming a turbulent, upward-moving stream?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by bright, even overhead lighting with dominant turquoise water tones?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0009227308910340071 + ], + "opposite_logprob_tokens": [ + -7.000922679901123 + ], + "logprob_reward": -0.0009227308910340071, + "yes_no_margin_reward": 7.0, + "token_credit_reward": -0.0009227308910340071, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0812092199921608, + 0.043471090495586395, + 0.11198914796113968, + 0.4084121584892273, + 0.0349462665617466, + 0.11311503499746323, + 0.047071754932403564, + 0.08232928812503815, + 0.0442051962018013, + 0.03325086832046509 + ], + "credit_map": [ + 0.009765871800482273, + 0.004923615138977766, + 0.00957176461815834, + 0.006722337566316128, + 0.010821900330483913, + 0.018707454204559326, + 0.0076368991285562515, + 0.013059376738965511, + 0.0036212450359016657, + 0.0038753701373934746, + 0.007861065678298473, + 0.0037686212453991175, + 0.008388488553464413, + 0.006330475211143494, + 0.004379604943096638, + 0.005246218293905258, + 0.008975627832114697, + 0.01704651303589344, + 0.018931804224848747, + 0.01606615073978901, + 0.015510289929807186, + 0.011810293421149254, + 0.014058994129300117, + 0.00958947092294693, + 0.10998184978961945, + 0.025760654360055923, + 0.04387925937771797, + 0.03473352640867233, + 0.022311069071292877, + 0.048010293394327164, + 0.03004579246044159, + 0.0936897024512291, + 0.006704037543386221, + 0.003491380251944065, + 0.0027193170972168446, + 0.006272723898291588, + 0.0029201051220297813, + 0.004774757660925388, + 0.005581232253462076, + 0.002482715295627713, + 0.008439645171165466, + 0.023683900013566017, + 0.010049830190837383, + 0.017322013154625893, + 0.015513270162045956, + 0.00587106728926301, + 0.022992104291915894, + 0.00924320612102747, + 0.002544594928622246, + 0.007016266696155071, + 0.0073568024672567844, + 0.010065054520964622, + 0.004705230239778757, + 0.002485789591446519, + 0.007498913444578648, + 0.005399100948125124, + 0.005186926107853651, + 0.010770306922495365, + 0.016410131007432938, + 0.011585969477891922, + 0.012263042852282524, + 0.0038642296567559242, + 0.017711671069264412, + 0.0045370194129645824, + 0.0015607995446771383, + 0.010303162969648838, + 0.0050224377773702145, + 0.003551292000338435, + 0.0027942797169089317, + 0.001400136505253613, + 0.017526110634207726, + 0.0020469767041504383, + 0.002244455274194479, + 0.004750168416649103, + 0.005033130291849375, + 0.006155841983854771, + 0.002769581973552704, + 0.0046189408749341965, + 0.004526900593191385, + 0.0031518491450697184 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 5, + 2 + ], + "selected_reward": -0.0009227308910340071 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 from the perspective of the object itself?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a round, matte-textured white tablet descending into clear water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the tablet dissolve rapidly upon submersion, emitting thousands of tiny white bubbles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the bubbles captured in slow motion forming a turbulent, upward-moving stream?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by bright, even overhead lighting with dominant turquoise water tones?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0015142414486035705 + ], + "opposite_logprob_tokens": [ + -6.501514434814453 + ], + "logprob_reward": -0.0015142414486035705, + "yes_no_margin_reward": 6.5, + "token_credit_reward": -0.0015142414486035705, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17371241748332977, + 0.03196355700492859, + 0.05469704791903496, + 0.1883665770292282, + 0.04573799669742584, + 0.10470370203256607, + 0.06519947201013565, + 0.058865565806627274, + 0.2199765294790268, + 0.05677708610892296 + ], + "credit_map": [ + 0.023716231808066368, + 0.011965055949985981, + 0.0156339593231678, + 0.011052688583731651, + 0.023617707192897797, + 0.041492875665426254, + 0.01788507029414177, + 0.028348831459879875, + 0.002148842206224799, + 0.002923371270298958, + 0.002468397142365575, + 0.003702970687299967, + 0.004290697164833546, + 0.007260376121848822, + 0.0035987473092973232, + 0.005570152774453163, + 0.0054509337060153484, + 0.007299543358385563, + 0.013632310554385185, + 0.00741422176361084, + 0.005055597517639399, + 0.004492089152336121, + 0.004482949618250132, + 0.0068694003857672215, + 0.04011252894997597, + 0.019086966291069984, + 0.02281867153942585, + 0.023945072665810585, + 0.014520443975925446, + 0.018200309947133064, + 0.019839204847812653, + 0.029843369498848915, + 0.008437338285148144, + 0.005607690196484327, + 0.007401155773550272, + 0.005587119609117508, + 0.0038629944901913404, + 0.005700744688510895, + 0.005337109789252281, + 0.0038038468919694424, + 0.007911596447229385, + 0.019808517768979073, + 0.009474698454141617, + 0.013673284091055393, + 0.010889312252402306, + 0.009709130972623825, + 0.026095619425177574, + 0.007141544949263334, + 0.0042371624149382114, + 0.010555642656981945, + 0.00989365205168724, + 0.014596260152757168, + 0.008074818179011345, + 0.0037361644208431244, + 0.008195560425519943, + 0.005910216830670834, + 0.003734287340193987, + 0.01066483836621046, + 0.006226017139852047, + 0.010961193591356277, + 0.006621327251195908, + 0.0025174368638545275, + 0.013504866510629654, + 0.00463560177013278, + 0.006922377739101648, + 0.08151809871196747, + 0.025322727859020233, + 0.016089636832475662, + 0.015265021473169327, + 0.004651487339287996, + 0.06439859420061111, + 0.0058085862547159195, + 0.0047022514045238495, + 0.008448487147688866, + 0.007920297794044018, + 0.008680591359734535, + 0.005074306856840849, + 0.0076237027533352375, + 0.008255265653133392, + 0.006072185467928648 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 3, + 0 + ], + "selected_reward": -0.0015142414486035705 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An extreme close-up shot, captured from the perspective of the object itself, closely follows the steady and rapid descent of a round, matte-textured white tablet into clear water. The moment it is fully submerged, illuminated by bright, even overhead lighting, its surface begins to dissolve rapidly, and from its core, thousands of tiny, densely packed white bubbles erupt violently. These bubbles, captured in slow motion, form a turbulent, upward-moving stream of energetic white. As they rise, they collide, merge, and burst, stirring the surrounding still water in an uncontrolled microscopic storm. The tablet itself is stirred, tumbled, and gradually disintegrated in this violent chemical reaction, its shattered white particles merging with the surging stream of bubbles. The dominant tones of the image are the clear turquoise of the water and the pure white of the bubbles and tablet. Light filtering through the bubbles creates a shimmering, translucent sheen, transforming this microscopic chemical reaction into a visual feast of primal power and dynamic beauty, imbued with the ultimate explosive power and tension of competitive sports.", + "seed": 5832, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000056/group_13/index_2363_seed_5832_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": -0.13265451076440513, + "vlm_logprob": -0.13265451076440513, + "vlm_yes_no_margin": 4.65, + "vlm_token_credit": -0.13265451076440513, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 from the perspective of the object itself?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a round, matte-textured white tablet descending into clear water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the tablet dissolve rapidly upon submersion, emitting thousands of tiny white bubbles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the bubbles captured in slow motion forming a turbulent, upward-moving stream?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by bright, even overhead lighting with dominant turquoise water tones?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.6489024758338928 + ], + "opposite_logprob_tokens": [ + -0.7739024758338928 + ], + "logprob_reward": -0.6489024758338928, + "yes_no_margin_reward": 0.125, + "token_credit_reward": -0.6489024758338928, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2882658839225769, + 0.06641674041748047, + 0.11975856125354767, + 0.10072111338376999, + 0.036459121853113174, + 0.10219715535640717, + 0.06820018589496613, + 0.04343393072485924, + 0.05026693269610405, + 0.1242804229259491 + ], + "credit_map": [ + 0.041391126811504364, + 0.03239426389336586, + 0.0369114987552166, + 0.025475142523646355, + 0.056302260607481, + 0.0403716117143631, + 0.02676800824701786, + 0.028651950880885124, + 0.008029265329241753, + 0.00561096565797925, + 0.007890764623880386, + 0.006145901046693325, + 0.008136643096804619, + 0.009430112317204475, + 0.01376055832952261, + 0.007412525359541178, + 0.015914658084511757, + 0.015329700894653797, + 0.015157070942223072, + 0.01424461416900158, + 0.013507520779967308, + 0.00831860676407814, + 0.019533544778823853, + 0.017752831801772118, + 0.01832875795662403, + 0.01214878261089325, + 0.015476896427571774, + 0.010105740278959274, + 0.011159321293234825, + 0.009549804031848907, + 0.009368049912154675, + 0.014583754353225231, + 0.004640966188162565, + 0.003290391992777586, + 0.004066582769155502, + 0.004713424015790224, + 0.005916460417211056, + 0.004295863211154938, + 0.00485009653493762, + 0.004685333464294672, + 0.014232095330953598, + 0.01133943535387516, + 0.013770935125648975, + 0.011762990616261959, + 0.008840937167406082, + 0.010858667083084583, + 0.015459802933037281, + 0.01593228243291378, + 0.01145957875996828, + 0.00698729045689106, + 0.008244971744716167, + 0.008217849768698215, + 0.008138549514114857, + 0.006135615054517984, + 0.008491638116538525, + 0.010524689219892025, + 0.0038065360859036446, + 0.006948774214833975, + 0.004236300475895405, + 0.007358125410974026, + 0.005688098259270191, + 0.004398164805024862, + 0.006788552738726139, + 0.004209376405924559, + 0.0029743118211627007, + 0.0054472279734909534, + 0.004126064945012331, + 0.009607309475541115, + 0.00624214019626379, + 0.006211936939507723, + 0.008913708850741386, + 0.006744228303432465, + 0.013262378983199596, + 0.010019284673035145, + 0.010471988469362259, + 0.012863430194556713, + 0.013428058475255966, + 0.017052391543984413, + 0.01811159960925579, + 0.02907129004597664 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 2 + ], + "selected_reward": -0.6489024758338928 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 from the perspective of the object itself?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a round, matte-textured white tablet descending into clear water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the tablet dissolve rapidly upon submersion, emitting thousands of tiny white bubbles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the bubbles captured in slow motion forming a turbulent, upward-moving stream?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by bright, even overhead lighting with dominant turquoise water tones?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00528628658503294 + ], + "opposite_logprob_tokens": [ + -5.25528621673584 + ], + "logprob_reward": -0.00528628658503294, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.00528628658503294, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20077666640281677, + 0.08071444183588028, + 0.05859158933162689, + 0.20494374632835388, + 0.13704635202884674, + 0.0757221058011055, + 0.04524516686797142, + 0.12352628260850906, + 0.033076509833335876, + 0.04035714641213417 + ], + "credit_map": [ + 0.024665892124176025, + 0.017061080783605576, + 0.02394023723900318, + 0.016551878303289413, + 0.031798187643289566, + 0.0315520204603672, + 0.03495781496167183, + 0.02024954929947853, + 0.007385884877294302, + 0.00788151752203703, + 0.00935509242117405, + 0.007536567281931639, + 0.015649300068616867, + 0.008666343986988068, + 0.012796391732990742, + 0.011443349532783031, + 0.0074062226340174675, + 0.007691495586186647, + 0.007930420339107513, + 0.00818298663944006, + 0.007198244333267212, + 0.005875978618860245, + 0.007663066033273935, + 0.006643175147473812, + 0.024106236174702644, + 0.0300007201731205, + 0.022779516875743866, + 0.026647113263607025, + 0.0238035935908556, + 0.021568728610873222, + 0.02334846369922161, + 0.03268939629197121, + 0.012880505993962288, + 0.01947767101228237, + 0.015072617679834366, + 0.012050608173012733, + 0.021733282133936882, + 0.018672896549105644, + 0.020545803010463715, + 0.01661296933889389, + 0.010456289164721966, + 0.008847657591104507, + 0.009939813055098057, + 0.008420254103839397, + 0.007273458410054445, + 0.00790852215141058, + 0.011856901459395885, + 0.011019209399819374, + 0.005202724598348141, + 0.006370173301547766, + 0.00557282567024231, + 0.004544322844594717, + 0.007858030498027802, + 0.006137921940535307, + 0.005104533396661282, + 0.004454632755368948, + 0.030962124466896057, + 0.008598238229751587, + 0.03193508833646774, + 0.013394116424024105, + 0.01113856490701437, + 0.008569663390517235, + 0.009708977304399014, + 0.009219512343406677, + 0.002780090318992734, + 0.0033606954384595156, + 0.00263768108561635, + 0.005187331698834896, + 0.004783265758305788, + 0.004125495441257954, + 0.0045909699983894825, + 0.005610978696495295, + 0.0027629947289824486, + 0.0042554945684969425, + 0.002999552758410573, + 0.00832285825163126, + 0.005280445329844952, + 0.005921725183725357, + 0.004580062814056873, + 0.00623401440680027 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 4 + ], + "selected_reward": -0.00528628658503294 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 from the perspective of the object itself?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a round, matte-textured white tablet descending into clear water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the tablet dissolve rapidly upon submersion, emitting thousands of tiny white bubbles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the bubbles captured in slow motion forming a turbulent, upward-moving stream?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by bright, even overhead lighting with dominant turquoise water tones?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"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": [ + 0.08199498802423477, + 0.047277044504880905, + 0.03912634402513504, + 0.07066388428211212, + 0.0858912542462349, + 0.04993673413991928, + 0.5092675685882568, + 0.03911692276597023, + 0.037727270275354385, + 0.03899803385138512 + ], + "credit_map": [ + 0.01340409740805626, + 0.008174318820238113, + 0.010963993147015572, + 0.0054632811807096004, + 0.010850640013813972, + 0.014764776453375816, + 0.008947730995714664, + 0.009426153264939785, + 0.004798216745257378, + 0.004715071525424719, + 0.00521850073710084, + 0.004108724184334278, + 0.008716093376278877, + 0.005724383983761072, + 0.007966244593262672, + 0.006029804237186909, + 0.003602276323363185, + 0.004606933332979679, + 0.004632824566215277, + 0.004855726845562458, + 0.003544497536495328, + 0.0030572752002626657, + 0.007993935607373714, + 0.00683287438005209, + 0.009389190934598446, + 0.00942752044647932, + 0.008646732196211815, + 0.008882569149136543, + 0.0067390152253210545, + 0.005751802586019039, + 0.00958975963294506, + 0.012237290851771832, + 0.007357455790042877, + 0.009724934585392475, + 0.008650913834571838, + 0.0086150411516428, + 0.014634017832577229, + 0.012993060052394867, + 0.011868615634739399, + 0.012047210708260536, + 0.006067866459488869, + 0.005404253490269184, + 0.005993533879518509, + 0.0070467134937644005, + 0.004300217609852552, + 0.005420520901679993, + 0.00894993171095848, + 0.006753700785338879, + 0.07786574214696884, + 0.039224542677402496, + 0.10087809711694717, + 0.06885001808404922, + 0.10307420790195465, + 0.046311378479003906, + 0.03871658071875572, + 0.034346986562013626, + 0.005182464607059956, + 0.004285460337996483, + 0.005097173620015383, + 0.004961526487022638, + 0.0055878921411931515, + 0.004036432132124901, + 0.004030661191791296, + 0.0059353141114115715, + 0.002586313057690859, + 0.0037198136560618877, + 0.003207029541954398, + 0.010235665366053581, + 0.0041143037378787994, + 0.006007521413266659, + 0.004975021351128817, + 0.002881605178117752, + 0.0037859349977225065, + 0.0029914663173258305, + 0.002849605632945895, + 0.005952363833785057, + 0.005244721192866564, + 0.006006193812936544, + 0.004759365227073431, + 0.007408385630697012 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 4, + 0 + ], + "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 shot an extreme close-up from the perspective of the object itself?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a round, matte-textured white tablet descending into clear water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the tablet dissolve rapidly upon submersion, emitting thousands of tiny white bubbles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the bubbles captured in slow motion forming a turbulent, upward-moving stream?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by bright, even overhead lighting with dominant turquoise water tones?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002491348423063755 + ], + "opposite_logprob_tokens": [ + -6.002491474151611 + ], + "logprob_reward": -0.002491348423063755, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.002491348423063755, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11898173391819, + 0.0850299745798111, + 0.11125552654266357, + 0.11691246181726456, + 0.12253754585981369, + 0.07450002431869507, + 0.07395227253437042, + 0.11863589286804199, + 0.042180683463811874, + 0.13601389527320862 + ], + "credit_map": [ + 0.01754005067050457, + 0.010754955932497978, + 0.014377744868397713, + 0.008819571696221828, + 0.017511144280433655, + 0.01689315214753151, + 0.018060531467199326, + 0.015024579130113125, + 0.00992120336741209, + 0.008056534454226494, + 0.008840049616992474, + 0.008128189481794834, + 0.018122000619769096, + 0.007650268264114857, + 0.015240159817039967, + 0.009071560576558113, + 0.013560899533331394, + 0.012532556429505348, + 0.015125538222491741, + 0.012760014273226261, + 0.013918097130954266, + 0.007116798777133226, + 0.023465512320399284, + 0.012776109389960766, + 0.01206948608160019, + 0.020418591797351837, + 0.012836495414376259, + 0.013503768481314182, + 0.012110011652112007, + 0.014279305934906006, + 0.01367935910820961, + 0.018015440553426743, + 0.01304877270013094, + 0.013703959994018078, + 0.013095215894281864, + 0.01283893920481205, + 0.01440027542412281, + 0.011308933608233929, + 0.016437899321317673, + 0.027703547850251198, + 0.009249753318727016, + 0.009169850498437881, + 0.00848020613193512, + 0.008526523597538471, + 0.006218703463673592, + 0.008300244808197021, + 0.011550139635801315, + 0.013004593551158905, + 0.011878866702318192, + 0.005395243410021067, + 0.01569293439388275, + 0.0067141675390303135, + 0.015214234590530396, + 0.006386011838912964, + 0.006626818794757128, + 0.0060439929366111755, + 0.019118642434477806, + 0.011195939034223557, + 0.021215975284576416, + 0.01656317710876465, + 0.009384428150951862, + 0.010099627077579498, + 0.013674451038241386, + 0.01738365739583969, + 0.0031901062466204166, + 0.004885665141046047, + 0.0035205346066504717, + 0.00990584772080183, + 0.004768735263496637, + 0.0066560483537614346, + 0.0051390924490988255, + 0.004114653915166855, + 0.013042174279689789, + 0.00997262541204691, + 0.011725116521120071, + 0.011027912609279156, + 0.02218436636030674, + 0.019567104056477547, + 0.01926599256694317, + 0.029228610917925835 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 4, + 0 + ], + "selected_reward": -0.002491348423063755 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 from the perspective of the object itself?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a round, matte-textured white tablet descending into clear water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the tablet dissolve rapidly upon submersion, emitting thousands of tiny white bubbles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the bubbles captured in slow motion forming a turbulent, upward-moving stream?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by bright, even overhead lighting with dominant turquoise water tones?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0013360392767935991 + ], + "opposite_logprob_tokens": [ + -6.626336097717285 + ], + "logprob_reward": -0.0013360392767935991, + "yes_no_margin_reward": 6.625, + "token_credit_reward": -0.0013360392767935991, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21728909015655518, + 0.0653596892952919, + 0.030728016048669815, + 0.044194817543029785, + 0.08622237294912338, + 0.024142295122146606, + 0.3999549150466919, + 0.052966706454753876, + 0.033984243869781494, + 0.04515779763460159 + ], + "credit_map": [ + 0.04081924632191658, + 0.015895163640379906, + 0.03534584119915962, + 0.010710244998335838, + 0.0228310227394104, + 0.048051171004772186, + 0.019726140424609184, + 0.02391025796532631, + 0.007549808360636234, + 0.006586542818695307, + 0.007429429795593023, + 0.005250751506537199, + 0.012982659041881561, + 0.006513815838843584, + 0.011762062087655067, + 0.0072846319526433945, + 0.0036686293315142393, + 0.003907522186636925, + 0.0038038697093725204, + 0.0031936124432832003, + 0.004511560779064894, + 0.002044904511421919, + 0.005714979488402605, + 0.0038829424884170294, + 0.004311638884246349, + 0.008053564466536045, + 0.004490245599299669, + 0.003995575476437807, + 0.005189941730350256, + 0.00535061489790678, + 0.005271735601127148, + 0.007531499955803156, + 0.009290309622883797, + 0.010254697874188423, + 0.010116996243596077, + 0.007030114997178316, + 0.012952337972819805, + 0.01038024015724659, + 0.014204501174390316, + 0.011993177235126495, + 0.003152749501168728, + 0.002949966350570321, + 0.0032116028014570475, + 0.0032764191273599863, + 0.0020564261358231306, + 0.002358254510909319, + 0.003971558529883623, + 0.003165319561958313, + 0.06368060410022736, + 0.02926098182797432, + 0.08268178999423981, + 0.0518919862806797, + 0.08106233924627304, + 0.035515692085027695, + 0.02930178865790367, + 0.026559757068753242, + 0.012639600783586502, + 0.005203515291213989, + 0.0134159279987216, + 0.006345993839204311, + 0.003667635377496481, + 0.0033101870212703943, + 0.003900345880538225, + 0.004483502823859453, + 0.0030524854082614183, + 0.004371558781713247, + 0.00301445578224957, + 0.005637775640934706, + 0.004758069291710854, + 0.0039937240071594715, + 0.003912060987204313, + 0.005244117230176926, + 0.005452676676213741, + 0.0037795782554894686, + 0.0038296582642942667, + 0.00446282746270299, + 0.005805075168609619, + 0.006892346777021885, + 0.006219047587364912, + 0.008716591633856297 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 4 + ], + "selected_reward": -0.0013360392767935991 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An extreme close-up shot, captured from the perspective of the object itself, closely follows the steady and rapid descent of a round, matte-textured white tablet into clear water. The moment it is fully submerged, illuminated by bright, even overhead lighting, its surface begins to dissolve rapidly, and from its core, thousands of tiny, densely packed white bubbles erupt violently. These bubbles, captured in slow motion, form a turbulent, upward-moving stream of energetic white. As they rise, they collide, merge, and burst, stirring the surrounding still water in an uncontrolled microscopic storm. The tablet itself is stirred, tumbled, and gradually disintegrated in this violent chemical reaction, its shattered white particles merging with the surging stream of bubbles. The dominant tones of the image are the clear turquoise of the water and the pure white of the bubbles and tablet. Light filtering through the bubbles creates a shimmering, translucent sheen, transforming this microscopic chemical reaction into a visual feast of primal power and dynamic beauty, imbued with the ultimate explosive power and tension of competitive sports.", + "seed": 5833, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000056/group_13/index_2363_seed_5833_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": -0.06849179223645478, + "vlm_logprob": -0.06849179223645478, + "vlm_yes_no_margin": 4.95, + "vlm_token_credit": -0.06849179223645478, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 from the perspective of the object itself?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a round, matte-textured white tablet descending into clear water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the tablet dissolve rapidly upon submersion, emitting thousands of tiny white bubbles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the bubbles captured in slow motion forming a turbulent, upward-moving stream?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by bright, even overhead lighting with dominant turquoise water tones?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.3275047242641449 + ], + "opposite_logprob_tokens": [ + -1.3275047540664673 + ], + "logprob_reward": -0.3275047242641449, + "yes_no_margin_reward": 1.0, + "token_credit_reward": -0.3275047242641449, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.37756067514419556, + 0.07836052775382996, + 0.06510385125875473, + 0.08209563046693802, + 0.05999401956796646, + 0.07233516126871109, + 0.017752306535840034, + 0.03404930233955383, + 0.16288432478904724, + 0.04986424744129181 + ], + "credit_map": [ + 0.05705396085977554, + 0.03833452984690666, + 0.029182765632867813, + 0.025940103456377983, + 0.05394253134727478, + 0.13446439802646637, + 0.016447894275188446, + 0.022194525226950645, + 0.007693661842495203, + 0.00868805218487978, + 0.00449532363563776, + 0.01365840807557106, + 0.020377827808260918, + 0.010659812949597836, + 0.0050893439911305904, + 0.007698095869272947, + 0.008146354928612709, + 0.008551321923732758, + 0.004033826291561127, + 0.005265940446406603, + 0.019508471712470055, + 0.005560961551964283, + 0.0061868601478636265, + 0.007850105874240398, + 0.004055637866258621, + 0.01632629707455635, + 0.005493425764143467, + 0.011909163556993008, + 0.024040311574935913, + 0.008820713497698307, + 0.005404261406511068, + 0.006045816000550985, + 0.006447795778512955, + 0.006565080489963293, + 0.0065704951994121075, + 0.007625865750014782, + 0.005464815068989992, + 0.01117873378098011, + 0.007858985103666782, + 0.008282243274152279, + 0.0075542558915913105, + 0.005991305224597454, + 0.015506484545767307, + 0.008766951970756054, + 0.007006531115621328, + 0.008074167184531689, + 0.009058307856321335, + 0.010377157479524612, + 0.0013437130255624652, + 0.0019182205433025956, + 0.00264338799752295, + 0.001842749654315412, + 0.0020276163704693317, + 0.0023376273456960917, + 0.0022847314830869436, + 0.003354259068146348, + 0.00278034177608788, + 0.005780122708529234, + 0.0035629819612950087, + 0.0024728181306272745, + 0.00302623538300395, + 0.0034376068506389856, + 0.006232651881873608, + 0.006756541784852743, + 0.011835731565952301, + 0.020902609452605247, + 0.026304025202989578, + 0.022760488092899323, + 0.02128951996564865, + 0.014602147974073887, + 0.02803139016032219, + 0.01715840958058834, + 0.004128956235945225, + 0.007272137328982353, + 0.0070772734470665455, + 0.0067026037722826, + 0.004509201738983393, + 0.004671703092753887, + 0.007504371926188469, + 0.007997999899089336 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 3 + ], + "selected_reward": -0.3275047242641449 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 from the perspective of the object itself?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a round, matte-textured white tablet descending into clear water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the tablet dissolve rapidly upon submersion, emitting thousands of tiny white bubbles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the bubbles captured in slow motion forming a turbulent, upward-moving stream?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by bright, even overhead lighting with dominant turquoise water tones?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009806787595152855 + ], + "opposite_logprob_tokens": [ + -4.6348066329956055 + ], + "logprob_reward": -0.009806787595152855, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009806787595152855, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20020487904548645, + 0.12531521916389465, + 0.2952149510383606, + 0.07402951270341873, + 0.03821518272161484, + 0.09190027415752411, + 0.05635383725166321, + 0.04712635651230812, + 0.04766426980495453, + 0.02397560514509678 + ], + "credit_map": [ + 0.02985166572034359, + 0.02155405841767788, + 0.015507668256759644, + 0.009492278099060059, + 0.029451144859194756, + 0.07876106351613998, + 0.005454867146909237, + 0.010132123716175556, + 0.013579152524471283, + 0.01562565378844738, + 0.007248247973620892, + 0.026051022112369537, + 0.03702541068196297, + 0.007977548986673355, + 0.010372315533459187, + 0.007435846142470837, + 0.0362447127699852, + 0.030485156923532486, + 0.025050533935427666, + 0.045776836574077606, + 0.06118956208229065, + 0.02618495002388954, + 0.02849174290895462, + 0.0417914055287838, + 0.003325053956359625, + 0.01580049842596054, + 0.003805855754762888, + 0.00512184901162982, + 0.025970762595534325, + 0.010650458745658398, + 0.0042470907792449, + 0.0051079303957521915, + 0.004614804871380329, + 0.004922983702272177, + 0.0045277392491698265, + 0.004149433691054583, + 0.005296902731060982, + 0.004527772311121225, + 0.0055448454804718494, + 0.004630701150745153, + 0.007807999849319458, + 0.0077663385309278965, + 0.022812172770500183, + 0.009854194708168507, + 0.008607137948274612, + 0.011928658932447433, + 0.011929375119507313, + 0.011194389313459396, + 0.007432655896991491, + 0.007238802034407854, + 0.003924846649169922, + 0.008925375528633595, + 0.007027392275631428, + 0.00822023581713438, + 0.007953668013215065, + 0.0056308540515601635, + 0.003557230345904827, + 0.009657409973442554, + 0.005887292325496674, + 0.003604952711611986, + 0.003457685001194477, + 0.004505505785346031, + 0.009875099174678326, + 0.00658117700368166, + 0.0027977011632174253, + 0.005856785923242569, + 0.008260293863713741, + 0.006307473871856928, + 0.005010351538658142, + 0.003831874579191208, + 0.008597152307629585, + 0.007002633064985275, + 0.0015119106974452734, + 0.004757911432534456, + 0.0036208725068718195, + 0.0034313946962356567, + 0.0025394654367119074, + 0.003693957580253482, + 0.0023786674719303846, + 0.002041423227638006 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 1 + ], + "selected_reward": -0.009806787595152855 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 from the perspective of the object itself?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a round, matte-textured white tablet descending into clear water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the tablet dissolve rapidly upon submersion, emitting thousands of tiny white bubbles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the bubbles captured in slow motion forming a turbulent, upward-moving stream?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by bright, even overhead lighting with dominant turquoise water tones?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"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": [ + 0.18522915244102478, + 0.10174886137247086, + 0.07963509857654572, + 0.19320246577262878, + 0.056636836379766464, + 0.13673406839370728, + 0.14025434851646423, + 0.058832064270973206, + 0.022434093058109283, + 0.025293029844760895 + ], + "credit_map": [ + 0.02634279429912567, + 0.023651640862226486, + 0.01154190395027399, + 0.011497645638883114, + 0.03841453790664673, + 0.05717279016971588, + 0.0076761855743825436, + 0.008931663818657398, + 0.009406313300132751, + 0.013973969034850597, + 0.006199412047863007, + 0.01780419982969761, + 0.030434558168053627, + 0.008234073407948017, + 0.010332336649298668, + 0.005364001728594303, + 0.006638690363615751, + 0.00891770888119936, + 0.007768790703266859, + 0.015453712083399296, + 0.01779717393219471, + 0.006051627919077873, + 0.008103513158857822, + 0.008903875946998596, + 0.007586936000734568, + 0.04097661375999451, + 0.008132426999509335, + 0.01180567778646946, + 0.07496406883001328, + 0.02406124211847782, + 0.007883415557444096, + 0.017792081460356712, + 0.007019678130745888, + 0.005350316409021616, + 0.006194890011101961, + 0.009322118014097214, + 0.0065594869665801525, + 0.0065367859788239, + 0.00952840130776167, + 0.006125159095972776, + 0.010453981347382069, + 0.01178552582859993, + 0.02859780751168728, + 0.01707417331635952, + 0.014354303479194641, + 0.014644280076026917, + 0.017221206799149513, + 0.022602802142500877, + 0.015351683832705021, + 0.016032995656132698, + 0.014531970024108887, + 0.015134590677917004, + 0.01780501939356327, + 0.01823837123811245, + 0.01932237669825554, + 0.023837342858314514, + 0.0038461906369775534, + 0.00949781946837902, + 0.0066976603120565414, + 0.0060928999446332455, + 0.004255350679159164, + 0.004652232863008976, + 0.011056074872612953, + 0.012733838520944118, + 0.0017601719591766596, + 0.002745702862739563, + 0.004713962320238352, + 0.003088610479608178, + 0.0026076622307300568, + 0.0017865118570625782, + 0.002630081959068775, + 0.003101389855146408, + 0.0016268165782094002, + 0.00381658342666924, + 0.00290920352563262, + 0.003300436306744814, + 0.0022884442005306482, + 0.003679928369820118, + 0.004536798223853111, + 0.0031348213087767363 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 6 + ], + "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 an extreme close-up from the perspective of the object itself?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a round, matte-textured white tablet descending into clear water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the tablet dissolve rapidly upon submersion, emitting thousands of tiny white bubbles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the bubbles captured in slow motion forming a turbulent, upward-moving stream?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by bright, even overhead lighting with dominant turquoise water tones?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0017157370457425714 + ], + "opposite_logprob_tokens": [ + -6.376715660095215 + ], + "logprob_reward": -0.0017157370457425714, + "yes_no_margin_reward": 6.375, + "token_credit_reward": -0.0017157370457425714, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.28388330340385437, + 0.04109262675046921, + 0.03855038806796074, + 0.06733335554599762, + 0.06335749477148056, + 0.08374708890914917, + 0.23498457670211792, + 0.041795533150434494, + 0.10931916534900665, + 0.03593647852540016 + ], + "credit_map": [ + 0.04665859043598175, + 0.02908756583929062, + 0.019170960411429405, + 0.014468005858361721, + 0.039761848747730255, + 0.11115755140781403, + 0.008733212947845459, + 0.014845562167465687, + 0.003630097722634673, + 0.004592748824506998, + 0.002807295648381114, + 0.006000221706926823, + 0.009102120995521545, + 0.006532637868076563, + 0.003942970652133226, + 0.004484531469643116, + 0.004594677593559027, + 0.004997409414499998, + 0.0029889605939388275, + 0.0039249141700565815, + 0.010643931105732918, + 0.0034652631729841232, + 0.004159865435212851, + 0.0037753679789602757, + 0.0033094706013798714, + 0.013183742761611938, + 0.003815897973254323, + 0.0060235969722270966, + 0.02267940528690815, + 0.007751026656478643, + 0.003756053512915969, + 0.006814155727624893, + 0.00759325223043561, + 0.008576619438827038, + 0.008332556113600731, + 0.006705339066684246, + 0.006082039326429367, + 0.007556987460702658, + 0.010455258190631866, + 0.008055437356233597, + 0.006733706220984459, + 0.008302446454763412, + 0.014971068128943443, + 0.01071964018046856, + 0.00899248756468296, + 0.008444941602647305, + 0.011099270544946194, + 0.01448353286832571, + 0.031034883111715317, + 0.029503783211112022, + 0.015488664619624615, + 0.030764199793338776, + 0.030973564833402634, + 0.0409957617521286, + 0.03273575380444527, + 0.023487962782382965, + 0.0033622882328927517, + 0.0064414143562316895, + 0.00466881925240159, + 0.0060128103941679, + 0.0036063960287719965, + 0.004534459672868252, + 0.007605069316923618, + 0.005564278922975063, + 0.006967779714614153, + 0.013447538949549198, + 0.01958562061190605, + 0.015487262979149818, + 0.012988547794520855, + 0.008598849177360535, + 0.01734371855854988, + 0.014899850822985172, + 0.0031168246641755104, + 0.005579124204814434, + 0.005070685409009457, + 0.004871055018156767, + 0.0036406870931386948, + 0.004096576943993568, + 0.004967395216226578, + 0.004594129975885153 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 8 + ], + "selected_reward": -0.0017157370457425714 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 from the perspective of the object itself?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a round, matte-textured white tablet descending into clear water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the tablet dissolve rapidly upon submersion, emitting thousands of tiny white bubbles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the bubbles captured in slow motion forming a turbulent, upward-moving stream?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by bright, even overhead lighting with dominant turquoise water tones?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0017134760273620486 + ], + "opposite_logprob_tokens": [ + -6.376713275909424 + ], + "logprob_reward": -0.0017134760273620486, + "yes_no_margin_reward": 6.375, + "token_credit_reward": -0.0017134760273620486, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4131213128566742, + 0.018935436382889748, + 0.041344836354255676, + 0.0438879057765007, + 0.05994661897420883, + 0.04600435495376587, + 0.0963563323020935, + 0.05164606496691704, + 0.1593005657196045, + 0.06945659220218658 + ], + "credit_map": [ + 0.06585550308227539, + 0.0350567027926445, + 0.034278467297554016, + 0.027027826756238937, + 0.045008353888988495, + 0.1702912449836731, + 0.014363656751811504, + 0.021239571273326874, + 0.002744467696174979, + 0.001879072398878634, + 0.002364758402109146, + 0.0032377091702073812, + 0.0039837840013206005, + 0.0012789670145139098, + 0.0020598398987203836, + 0.0013868382666260004, + 0.004485408775508404, + 0.00616651913151145, + 0.00266768760047853, + 0.004168118815869093, + 0.01372677180916071, + 0.0030658284667879343, + 0.003449015785008669, + 0.0036154906265437603, + 0.003150632604956627, + 0.009078074246644974, + 0.0021606148220598698, + 0.002978026634082198, + 0.015113945119082928, + 0.0048161339946091175, + 0.002299567451700568, + 0.004290912300348282, + 0.0066276066936552525, + 0.0061926390044391155, + 0.007038658019155264, + 0.008079770021140575, + 0.008452107198536396, + 0.009578489698469639, + 0.007569784764200449, + 0.006407563108950853, + 0.0040161143988370895, + 0.004347828216850758, + 0.009544525295495987, + 0.005297569092363119, + 0.005304590333253145, + 0.004437214694917202, + 0.006142213940620422, + 0.006914299912750721, + 0.0136990612372756, + 0.010550013743340969, + 0.007337597664445639, + 0.01246852520853281, + 0.012154492549598217, + 0.016680818051099777, + 0.013588149100542068, + 0.009877672418951988, + 0.004332700278609991, + 0.008048426359891891, + 0.0063623483292758465, + 0.005097749177366495, + 0.003671217942610383, + 0.006285578943789005, + 0.009298519231379032, + 0.008549529127776623, + 0.009754737839102745, + 0.01978393830358982, + 0.026275699958205223, + 0.02041265368461609, + 0.019034041091799736, + 0.013744589872658253, + 0.02786390297114849, + 0.022431006655097008, + 0.006330494303256273, + 0.009170151315629482, + 0.01320202462375164, + 0.009581967256963253, + 0.007360012270510197, + 0.006799113471060991, + 0.008083922788500786, + 0.008928912691771984 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 6 + ], + "selected_reward": -0.0017134760273620486 + } + ], + "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/0000056/group_14/index_tensor([3036, 3036])_group_14_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000056/group_14/index_tensor([3036, 3036])_group_14_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..15c45cda7d647439c50b8429a58e5040cc53187a --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000056/group_14/index_tensor([3036, 3036])_group_14_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "long shot, starting in the distance and slowly moving forward, captured from a slightly overhead angle, captures a tranquil living room bathed in the golden afternoon sun. Warm light slants in through a large floor-to-ceiling window, casting long, soft shadows across the polished oak floor. Tiny particles of dust drifting in the air shimmer in the beams of light. At the center of the frame, a mother sits comfortably on a soft beige rug. Her long, softly wavy brown hair is casually tied back, and she wears a loose, off-white knit sweater. Before her lies a small pile of freshly washed, neatly folded baby clothes in soft shades of pale yellow, sky blue, and pure white. A gentle, contented smile graces her face, her eyes filled with maternal love. She reaches out her slender fingers and delicately strokes the soft cotton of a pale yellow baby bodysuit, her movements gentle and cherishing. Then, she picked up a small white T-shirt and, steadily and calmly, brought it to her nose. She closed her eyes, as if inhaling the comforting, fresh scent of sunlight and baby laundry soap. The entire scene, bathed in warm, soft tones, was shrouded in a tranquil, healing atmosphere, perfectly embodying the warmth and comfort of home.", + "seed": 2920, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000056/group_14/index_3036_seed_2920_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": -0.3980641119647771, + "vlm_logprob": -0.3980641119647771, + "vlm_yes_no_margin": 2.275, + "vlm_token_credit": -0.3980641119647771, + "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 living room with a large floor-to-ceiling window?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera shot a long shot starting in the distance and slowly moving forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera angle slightly overhead?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a mother sitting on a beige rug wearing an off-white knit sweater with long brown hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the mother pick up a white T-shirt and bring it to her nose while closing her eyes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.26983553171157837 + ], + "opposite_logprob_tokens": [ + -1.5198354721069336 + ], + "logprob_reward": -0.26983553171157837, + "yes_no_margin_reward": 1.25, + "token_credit_reward": -0.26983553171157837, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19581717252731323, + 0.03967725858092308, + 0.02625168301165104, + 0.02778874710202217, + 0.013118837028741837, + 0.014334863051772118, + 0.07092098146677017, + 0.019934002310037613, + 0.38182970881462097, + 0.2103266716003418 + ], + "credit_map": [ + 0.028802968561649323, + 0.023954646661877632, + 0.015277096070349216, + 0.03582097589969635, + 0.020201990380883217, + 0.01946566067636013, + 0.022222401574254036, + 0.0300714410841465, + 0.003922976087778807, + 0.005988357588648796, + 0.003894758876413107, + 0.007034142967313528, + 0.0058319359086453915, + 0.004645943641662598, + 0.0045716408640146255, + 0.003787505440413952, + 0.0030604011844843626, + 0.004269641358405352, + 0.002740902127698064, + 0.005659107584506273, + 0.0032169229816645384, + 0.0032251477241516113, + 0.0020513220224529505, + 0.002028242452070117, + 0.0035839753691107035, + 0.004489774815738201, + 0.0042946115136146545, + 0.003941722214221954, + 0.0029331680852919817, + 0.0032521639950573444, + 0.0024312459863722324, + 0.002862086519598961, + 0.001310555962845683, + 0.0016038898611441255, + 0.0015329961897805333, + 0.0018054209649562836, + 0.0018117970321327448, + 0.0012931074015796185, + 0.0012574968859553337, + 0.0025035736616700888, + 0.0018290738807991147, + 0.0022319727577269077, + 0.001364417257718742, + 0.001716301660053432, + 0.0021977927535772324, + 0.0023477445356547832, + 0.001329453894868493, + 0.001318107359111309, + 0.008037131279706955, + 0.0061515565030276775, + 0.007304940838366747, + 0.007764481008052826, + 0.006894581951200962, + 0.01478516310453415, + 0.009066286496818066, + 0.01091684028506279, + 0.0020441561937332153, + 0.002973877592012286, + 0.002623155014589429, + 0.0023874142207205296, + 0.002223316580057144, + 0.0019714150112122297, + 0.0034035893622785807, + 0.002307077869772911, + 0.04721809923648834, + 0.022316616028547287, + 0.16318003833293915, + 0.08351019769906998, + 0.013614336960017681, + 0.016320811584591866, + 0.02124195545911789, + 0.014427664689719677, + 0.02044268511235714, + 0.015655070543289185, + 0.016590498387813568, + 0.02719082683324814, + 0.020251259207725525, + 0.026355566456913948, + 0.052918750792741776, + 0.030922027304768562 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 0 + ], + "selected_reward": -0.26983553171157837 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 living room with a large floor-to-ceiling window?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera shot a long shot starting in the distance and slowly moving forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera angle slightly overhead?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a mother sitting on a beige rug wearing an off-white knit sweater with long brown hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the mother pick up a white T-shirt and bring it to her nose while closing her eyes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"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": [ + -1.6005505323410034 + ], + "opposite_logprob_tokens": [ + -0.22555048763751984 + ], + "logprob_reward": -1.6005505323410034, + "yes_no_margin_reward": -1.375, + "token_credit_reward": -1.6005505323410034, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.2020590901374817, + 0.0555766224861145, + 0.06768350303173065, + 0.030762020498514175, + 0.019411880522966385, + 0.03142835572361946, + 0.10204373300075531, + 0.03381342440843582, + 0.2983531355857849, + 0.15886813402175903 + ], + "credit_map": [ + 0.02093844674527645, + 0.02908242680132389, + 0.017335839569568634, + 0.03954271972179413, + 0.027889003977179527, + 0.027708474546670914, + 0.022364070639014244, + 0.017198126763105392, + 0.006689449772238731, + 0.007032951805740595, + 0.008498385548591614, + 0.010319439694285393, + 0.00815498735755682, + 0.007154019549489021, + 0.0037761747371405363, + 0.003951223101466894, + 0.007781401742249727, + 0.007756792940199375, + 0.007465712260454893, + 0.0171440951526165, + 0.007316987495869398, + 0.007495081052184105, + 0.00625018822029233, + 0.006473258603364229, + 0.003513432340696454, + 0.005372514482587576, + 0.0032702272292226553, + 0.005588812287896872, + 0.0031553299631923437, + 0.003502890933305025, + 0.0027186223305761814, + 0.003640197217464447, + 0.0026028170250356197, + 0.0024079305585473776, + 0.002362082013860345, + 0.0033623354975134134, + 0.002292423974722624, + 0.00230753468349576, + 0.001686274423263967, + 0.002390484558418393, + 0.004522626753896475, + 0.004114071372896433, + 0.004957256373018026, + 0.003911932464689016, + 0.0035510582383722067, + 0.004290309268981218, + 0.00283246161416173, + 0.00324864499270916, + 0.008888386189937592, + 0.011934333480894566, + 0.011537567712366581, + 0.012037099339067936, + 0.01376771554350853, + 0.018328815698623657, + 0.009749418124556541, + 0.01580040343105793, + 0.004019325599074364, + 0.005369109101593494, + 0.003077741712331772, + 0.0038709130603820086, + 0.004567947704344988, + 0.0030603203922510147, + 0.004575669765472412, + 0.005272404290735722, + 0.03858105093240738, + 0.01859641820192337, + 0.12039992958307266, + 0.055348288267850876, + 0.01394376065582037, + 0.014630420133471489, + 0.0225822851061821, + 0.014271027408540249, + 0.015236974693834782, + 0.01305395271629095, + 0.012972124852240086, + 0.022494696080684662, + 0.018047738820314407, + 0.01921173930168152, + 0.03843000903725624, + 0.0194209273904562 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 9 + ], + "selected_reward": -1.6005505323410034 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 living room with a large floor-to-ceiling window?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera shot a long shot starting in the distance and slowly moving forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera angle slightly overhead?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a mother sitting on a beige rug wearing an off-white knit sweater with long brown hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the mother pick up a white T-shirt and bring it to her nose while closing her eyes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05501861125230789 + ], + "opposite_logprob_tokens": [ + -2.930018663406372 + ], + "logprob_reward": -0.05501861125230789, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05501861125230789, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3086596429347992, + 0.0812661200761795, + 0.057792920619249344, + 0.06789536029100418, + 0.0287802591919899, + 0.022777941077947617, + 0.05711768940091133, + 0.040773071348667145, + 0.23728477954864502, + 0.09765219688415527 + ], + "credit_map": [ + 0.032269131392240524, + 0.04106435179710388, + 0.025318488478660583, + 0.05751567333936691, + 0.04476185515522957, + 0.03203656151890755, + 0.03515838086605072, + 0.040535230189561844, + 0.010634402744472027, + 0.009463026188313961, + 0.016505079343914986, + 0.016547823324799538, + 0.006673410069197416, + 0.008866773918271065, + 0.00537399435415864, + 0.007201614324003458, + 0.006220359820872545, + 0.009979181922972202, + 0.004818722140043974, + 0.013590891845524311, + 0.008089830167591572, + 0.006857342552393675, + 0.004690396599471569, + 0.00354619394056499, + 0.008525824174284935, + 0.01380511187016964, + 0.010918393731117249, + 0.012261626310646534, + 0.006056306418031454, + 0.007119521498680115, + 0.004405754618346691, + 0.0048028212040662766, + 0.0034118550829589367, + 0.004597667139023542, + 0.003902956610545516, + 0.004158494528383017, + 0.0026000337675213814, + 0.0030550078954547644, + 0.00309042283333838, + 0.003963821567595005, + 0.0027290410362184048, + 0.0030805275309830904, + 0.002538063796237111, + 0.0032705955673009157, + 0.003499341430142522, + 0.0036853139754384756, + 0.002007095841690898, + 0.0019679635297507048, + 0.005309185478836298, + 0.009987927041947842, + 0.00630387244746089, + 0.0075585064478218555, + 0.006306381896138191, + 0.005827015265822411, + 0.010775645263493061, + 0.005049156490713358, + 0.004359433893114328, + 0.005405294708907604, + 0.005752166733145714, + 0.005296944174915552, + 0.005665380973368883, + 0.006299718748778105, + 0.0037213952746242285, + 0.0042727370746433735, + 0.02811981737613678, + 0.013224869966506958, + 0.10132735222578049, + 0.046946123242378235, + 0.01128239743411541, + 0.011319493874907494, + 0.015482595190405846, + 0.009582124650478363, + 0.011027458123862743, + 0.01011044904589653, + 0.010319140739738941, + 0.01442713849246502, + 0.013302646577358246, + 0.010099846869707108, + 0.019773248583078384, + 0.008592270314693451 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 9 + ], + "selected_reward": -0.05501861125230789 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 living room with a large floor-to-ceiling window?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera shot a long shot starting in the distance and slowly moving forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera angle slightly overhead?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a mother sitting on a beige rug wearing an off-white knit sweater with long brown hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the mother pick up a white T-shirt and bring it to her nose while closing her eyes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0028299542609602213 + ], + "opposite_logprob_tokens": [ + -5.8778300285339355 + ], + "logprob_reward": -0.0028299542609602213, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.0028299542609602213, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17474614083766937, + 0.04904315993189812, + 0.026254380121827126, + 0.025052402168512344, + 0.024284332990646362, + 0.02046772465109825, + 0.04121338576078415, + 0.03345878794789314, + 0.5139231085777283, + 0.09155664592981339 + ], + "credit_map": [ + 0.023071760311722755, + 0.023608779534697533, + 0.012448160909116268, + 0.030573848634958267, + 0.026245413348078728, + 0.016913454979658127, + 0.01867353916168213, + 0.023211173713207245, + 0.005380233284085989, + 0.006628578063100576, + 0.0056549422442913055, + 0.008151503279805183, + 0.007822310552001, + 0.0066947126761078835, + 0.004875008016824722, + 0.0038358718156814575, + 0.003081165486946702, + 0.003496079705655575, + 0.0022002209443598986, + 0.00481818150728941, + 0.0037592072039842606, + 0.0030456408858299255, + 0.0026826143730431795, + 0.0031712702475488186, + 0.0029070626478642225, + 0.0026629921048879623, + 0.0022155155893415213, + 0.00679753115400672, + 0.0031623549293726683, + 0.002838817425072193, + 0.002294895239174366, + 0.002173233777284622, + 0.003092552302405238, + 0.0036724626552313566, + 0.0027352606412023306, + 0.004057678859680891, + 0.003877721494063735, + 0.0030142588075250387, + 0.0023090117610991, + 0.0015253874007612467, + 0.002582717686891556, + 0.003411330981180072, + 0.0029652253724634647, + 0.002549550961703062, + 0.002661116886883974, + 0.0027652725111693144, + 0.0017861712258309126, + 0.001746338908560574, + 0.0034467908553779125, + 0.009477951563894749, + 0.004840857349336147, + 0.005459219682961702, + 0.005314462818205357, + 0.004469715990126133, + 0.0048768664710223675, + 0.0033275210298597813, + 0.0026611709035933018, + 0.008181187324225903, + 0.0036061336286365986, + 0.004820276517421007, + 0.0040529221296310425, + 0.003826703177765012, + 0.003110384102910757, + 0.0032000101637095213, + 0.06257514655590057, + 0.02963694930076599, + 0.21819621324539185, + 0.11538359522819519, + 0.018197929486632347, + 0.022153640165925026, + 0.0290730781853199, + 0.01870652474462986, + 0.009079243056476116, + 0.013687623664736748, + 0.006554251536726952, + 0.01976473443210125, + 0.010924164205789566, + 0.008286719210445881, + 0.013348627835512161, + 0.009911279194056988 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 9 + ], + "selected_reward": -0.0028299542609602213 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 living room with a large floor-to-ceiling window?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera shot a long shot starting in the distance and slowly moving forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera angle slightly overhead?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a mother sitting on a beige rug wearing an off-white knit sweater with long brown hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the mother pick up a white T-shirt and bring it to her nose while closing her eyes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06208593025803566 + ], + "opposite_logprob_tokens": [ + -2.8120858669281006 + ], + "logprob_reward": -0.06208593025803566, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06208593025803566, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10637453198432922, + 0.020326299592852592, + 0.015310457907617092, + 0.014770393259823322, + 0.015253177843987942, + 0.028221243992447853, + 0.029613755643367767, + 0.028199108317494392, + 0.6818450093269348, + 0.0600859671831131 + ], + "credit_map": [ + 0.014629947952926159, + 0.012113755568861961, + 0.007737131789326668, + 0.015865281224250793, + 0.012597547844052315, + 0.009762541390955448, + 0.016230138018727303, + 0.01743820495903492, + 0.0019141698721796274, + 0.005233290139585733, + 0.0022788105998188257, + 0.002885768422856927, + 0.002102336846292019, + 0.0028782042209059, + 0.0015907606575638056, + 0.0014429610455408692, + 0.001675065839663148, + 0.002115948125720024, + 0.0019471660489216447, + 0.002723678946495056, + 0.0019535147584974766, + 0.0014327586395666003, + 0.0016855107387527823, + 0.0017768166726455092, + 0.0016164329135790467, + 0.0016911299899220467, + 0.0016304850578308105, + 0.002720800694078207, + 0.0018474820535629988, + 0.0014151382492855191, + 0.0016875447472557425, + 0.002161381533369422, + 0.0017979054246097803, + 0.0020647174678742886, + 0.0017936209915205836, + 0.0029190736822783947, + 0.0020206754561513662, + 0.001709552714601159, + 0.0015599122270941734, + 0.0013877220917493105, + 0.002873525023460388, + 0.004494667053222656, + 0.0027803818229585886, + 0.004854012746363878, + 0.003817666554823518, + 0.0037756403908133507, + 0.003625901648774743, + 0.0019994541071355343, + 0.002895328216254711, + 0.004051539581269026, + 0.002822163514792919, + 0.003468635492026806, + 0.004018653184175491, + 0.0032670265063643456, + 0.005404663737863302, + 0.0036857493687421083, + 0.0034048703964799643, + 0.005283549893647432, + 0.002907878952100873, + 0.00413223123177886, + 0.002842501737177372, + 0.0037527678068727255, + 0.0032295442651957273, + 0.0026457689236849546, + 0.08180389553308487, + 0.041824132204055786, + 0.28656771779060364, + 0.15292102098464966, + 0.02474951185286045, + 0.030422061681747437, + 0.03800159692764282, + 0.025555141270160675, + 0.008632832206785679, + 0.0049086688086390495, + 0.006424569990485907, + 0.009465433657169342, + 0.0068970019929111, + 0.005031954497098923, + 0.013924330472946167, + 0.0048011853359639645 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 9 + ], + "selected_reward": -0.06208593025803566 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "No", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "long shot, starting in the distance and slowly moving forward, captured from a slightly overhead angle, captures a tranquil living room bathed in the golden afternoon sun. Warm light slants in through a large floor-to-ceiling window, casting long, soft shadows across the polished oak floor. Tiny particles of dust drifting in the air shimmer in the beams of light. At the center of the frame, a mother sits comfortably on a soft beige rug. Her long, softly wavy brown hair is casually tied back, and she wears a loose, off-white knit sweater. Before her lies a small pile of freshly washed, neatly folded baby clothes in soft shades of pale yellow, sky blue, and pure white. A gentle, contented smile graces her face, her eyes filled with maternal love. She reaches out her slender fingers and delicately strokes the soft cotton of a pale yellow baby bodysuit, her movements gentle and cherishing. Then, she picked up a small white T-shirt and, steadily and calmly, brought it to her nose. She closed her eyes, as if inhaling the comforting, fresh scent of sunlight and baby laundry soap. The entire scene, bathed in warm, soft tones, was shrouded in a tranquil, healing atmosphere, perfectly embodying the warmth and comfort of home.", + "seed": 2921, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000056/group_14/index_3036_seed_2921_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": -0.3728902956470847, + "vlm_logprob": -0.3728902956470847, + "vlm_yes_no_margin": 2.575, + "vlm_token_credit": -0.3728902956470847, + "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 living room with a large floor-to-ceiling window?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera shot a long shot starting in the distance and slowly moving forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera angle slightly overhead?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a mother sitting on a beige rug wearing an off-white knit sweater with long brown hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the mother pick up a white T-shirt and bring it to her nose while closing her eyes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06091494485735893 + ], + "opposite_logprob_tokens": [ + -3.060914993286133 + ], + "logprob_reward": -0.06091494485735893, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.06091494485735893, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09217566251754761, + 0.03174266964197159, + 0.029222123324871063, + 0.03300514817237854, + 0.036641694605350494, + 0.17952735722064972, + 0.31896087527275085, + 0.0495365671813488, + 0.16808345913887024, + 0.061104413121938705 + ], + "credit_map": [ + 0.014491256326436996, + 0.015050877816975117, + 0.010097647085785866, + 0.01132702361792326, + 0.013209937140345573, + 0.00830129999667406, + 0.009881780482828617, + 0.009815841913223267, + 0.003877977142110467, + 0.003730621188879013, + 0.0029326691292226315, + 0.0028724928852170706, + 0.005567830055952072, + 0.0029973220080137253, + 0.006049942225217819, + 0.00371381314471364, + 0.004277682863175869, + 0.0030231496784836054, + 0.003455191617831588, + 0.0041696131229400635, + 0.004776039160788059, + 0.0026633627712726593, + 0.0035290031228214502, + 0.0033280823845416307, + 0.003108440199866891, + 0.0034095116425305605, + 0.004469109233468771, + 0.005887995008379221, + 0.004284759983420372, + 0.004124452825635672, + 0.003941898699849844, + 0.0037789822090417147, + 0.003998929169028997, + 0.0031805504113435745, + 0.003457068931311369, + 0.003836587304249406, + 0.0046875737607479095, + 0.0032431487925350666, + 0.007415242027491331, + 0.006822594907134771, + 0.009360109455883503, + 0.004317959304898977, + 0.05478692427277565, + 0.008219518698751926, + 0.007997370325028896, + 0.010359401814639568, + 0.010983267799019814, + 0.0735028088092804, + 0.025354918092489243, + 0.013331403955817223, + 0.02435530535876751, + 0.015892839059233665, + 0.013184520415961742, + 0.059248849749565125, + 0.028366897255182266, + 0.13922616839408875, + 0.004307103808969259, + 0.003950329031795263, + 0.013137327507138252, + 0.009128575213253498, + 0.0037895243149250746, + 0.006133997347205877, + 0.0034282428678125143, + 0.005661469418555498, + 0.017885703593492508, + 0.02268238179385662, + 0.018991056829690933, + 0.019711513072252274, + 0.03067884035408497, + 0.019709933549165726, + 0.019380800426006317, + 0.019043223932385445, + 0.006377503275871277, + 0.010433574207127094, + 0.010340099222958088, + 0.0132005475461483, + 0.005599924363195896, + 0.006001068279147148, + 0.00396574754267931, + 0.005185945890843868 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 5, + 8 + ], + "selected_reward": -0.06091494485735893 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 living room with a large floor-to-ceiling window?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera shot a long shot starting in the distance and slowly moving forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera angle slightly overhead?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a mother sitting on a beige rug wearing an off-white knit sweater with long brown hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the mother pick up a white T-shirt and bring it to her nose while closing her eyes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"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": [ + -1.701535701751709 + ], + "opposite_logprob_tokens": [ + -0.2015356421470642 + ], + "logprob_reward": -1.701535701751709, + "yes_no_margin_reward": -1.5, + "token_credit_reward": -1.701535701751709, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.11808261275291443, + 0.04743199795484543, + 0.06715788692235947, + 0.036021340638399124, + 0.04111659526824951, + 0.2429012507200241, + 0.23243370652198792, + 0.1478126496076584, + 0.03641147166490555, + 0.0306304469704628 + ], + "credit_map": [ + 0.021092241629958153, + 0.0133553147315979, + 0.015550785697996616, + 0.015713728964328766, + 0.015016639605164528, + 0.008824008516967297, + 0.014718395657837391, + 0.013811500743031502, + 0.008520211093127728, + 0.004955708980560303, + 0.004573936108499765, + 0.004633817821741104, + 0.007447332609444857, + 0.004949490074068308, + 0.006573626771569252, + 0.005777872633188963, + 0.006376416422426701, + 0.004257818218320608, + 0.005883070640265942, + 0.00794882234185934, + 0.012708356603980064, + 0.007271362468600273, + 0.011918443255126476, + 0.010793598368763924, + 0.002826258074492216, + 0.002780361333861947, + 0.003465291578322649, + 0.003474850207567215, + 0.005723278038203716, + 0.006275080610066652, + 0.006973419804126024, + 0.004502801690250635, + 0.0055901287123560905, + 0.005586186423897743, + 0.003629807149991393, + 0.005055566783994436, + 0.005577920936048031, + 0.003547217231243849, + 0.007133896928280592, + 0.004995869472622871, + 0.012869871221482754, + 0.006355883087962866, + 0.058767057955265045, + 0.009683502838015556, + 0.012771275825798512, + 0.020338816568255424, + 0.017031632363796234, + 0.10508322715759277, + 0.01773851364850998, + 0.011477469466626644, + 0.017921287566423416, + 0.01183661725372076, + 0.013115604408085346, + 0.04225964471697807, + 0.02296803705394268, + 0.09511654078960419, + 0.008734505623579025, + 0.009301052428781986, + 0.04066550359129906, + 0.016303585842251778, + 0.014672488905489445, + 0.01779978722333908, + 0.011766359210014343, + 0.02856936864554882, + 0.003983724862337112, + 0.004564796574413776, + 0.003131823381409049, + 0.0031564224045723677, + 0.006294245831668377, + 0.006683752406388521, + 0.003539696568623185, + 0.005057009868323803, + 0.003160082036629319, + 0.005935203284025192, + 0.004286631941795349, + 0.006767882965505123, + 0.0024941281881183386, + 0.0033718261402100325, + 0.0020917560905218124, + 0.00252293748781085 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 6, + 7 + ], + "selected_reward": -1.701535701751709 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 living room with a large floor-to-ceiling window?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera shot a long shot starting in the distance and slowly moving forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera angle slightly overhead?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a mother sitting on a beige rug wearing an off-white knit sweater with long brown hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the mother pick up a white T-shirt and bring it to her nose while closing her eyes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04868452623486519 + ], + "opposite_logprob_tokens": [ + -3.048684597015381 + ], + "logprob_reward": -0.04868452623486519, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04868452623486519, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1833234429359436, + 0.03552008047699928, + 0.08918149769306183, + 0.05030703544616699, + 0.030417660251259804, + 0.27071064710617065, + 0.09760106354951859, + 0.10656951367855072, + 0.0936143547296524, + 0.04275468364357948 + ], + "credit_map": [ + 0.024368828162550926, + 0.016004864126443863, + 0.015484964475035667, + 0.016971677541732788, + 0.034502312541007996, + 0.02716914936900139, + 0.024570154026150703, + 0.024251483380794525, + 0.0071515911258757114, + 0.0030097495764493942, + 0.00287421396933496, + 0.003540554316714406, + 0.004240408074110746, + 0.004091737791895866, + 0.005833669565618038, + 0.004778154660016298, + 0.005333309527486563, + 0.004489109851419926, + 0.007614900358021259, + 0.006702801212668419, + 0.016497429460287094, + 0.012503450736403465, + 0.01955198310315609, + 0.016488512977957726, + 0.004715282935649157, + 0.0032638879492878914, + 0.004745388403534889, + 0.004480801988393068, + 0.009486132301390171, + 0.005911730229854584, + 0.010704081505537033, + 0.0069997296668589115, + 0.002925707958638668, + 0.003205286804586649, + 0.0026853892486542463, + 0.004368228372186422, + 0.004880634136497974, + 0.003206818364560604, + 0.004361115861684084, + 0.004784479737281799, + 0.01414744183421135, + 0.005493992939591408, + 0.08221261203289032, + 0.010573584586381912, + 0.00905019324272871, + 0.016173403710126877, + 0.016795210540294647, + 0.11626420170068741, + 0.008673851378262043, + 0.00484518613666296, + 0.007234631571918726, + 0.005467081442475319, + 0.0049666292034089565, + 0.017516089603304863, + 0.007896614260971546, + 0.041000980883836746, + 0.007759849075227976, + 0.0076161655597388744, + 0.02603934518992901, + 0.009362676180899143, + 0.012757774442434311, + 0.01325839664787054, + 0.010722866281867027, + 0.019052449613809586, + 0.009118541143834591, + 0.010090386494994164, + 0.00813673622906208, + 0.008480486460030079, + 0.014185619540512562, + 0.01914834789931774, + 0.012531260959804058, + 0.011922982521355152, + 0.00371177913621068, + 0.004730087239295244, + 0.0043206545524299145, + 0.005778120364993811, + 0.004861864261329174, + 0.010367185808718204, + 0.00303636328317225, + 0.005948629230260849 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 7 + ], + "selected_reward": -0.04868452623486519 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 living room with a large floor-to-ceiling window?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera shot a long shot starting in the distance and slowly moving forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera angle slightly overhead?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a mother sitting on a beige rug wearing an off-white knit sweater with long brown hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the mother pick up a white T-shirt and bring it to her nose while closing her eyes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0046457452699542046 + ], + "opposite_logprob_tokens": [ + -5.379645824432373 + ], + "logprob_reward": -0.0046457452699542046, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.0046457452699542046, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10801587998867035, + 0.09098772704601288, + 0.03862703591585159, + 0.0293879434466362, + 0.03450009599328041, + 0.14275310933589935, + 0.31198474764823914, + 0.057132985442876816, + 0.14802439510822296, + 0.0385860949754715 + ], + "credit_map": [ + 0.011480510234832764, + 0.012726888991892338, + 0.010469382628798485, + 0.010275614447891712, + 0.021834054961800575, + 0.01322176493704319, + 0.015124144963920116, + 0.012883512303233147, + 0.0075686960481107235, + 0.006941067986190319, + 0.007712292950600386, + 0.008571235463023186, + 0.01820213906466961, + 0.011876373551785946, + 0.019783584401011467, + 0.01033234503120184, + 0.0041993106715381145, + 0.0029651508666574955, + 0.004077916964888573, + 0.006172785069793463, + 0.00780227966606617, + 0.0034701344557106495, + 0.004838451277464628, + 0.005101003218442202, + 0.00283444719389081, + 0.002166921040043235, + 0.002501774812117219, + 0.002285428810864687, + 0.005538359750062227, + 0.003584148595109582, + 0.00602307915687561, + 0.004453785251826048, + 0.0030565483029931784, + 0.002487925812602043, + 0.003477358026430011, + 0.005055239889770746, + 0.005886747967451811, + 0.003929731901735067, + 0.0044683669693768024, + 0.006138178054243326, + 0.008061896078288555, + 0.005315147805958986, + 0.02910243719816208, + 0.010230626910924911, + 0.018137682229280472, + 0.011461464688181877, + 0.01576680690050125, + 0.04467704892158508, + 0.028152380138635635, + 0.012156874872744083, + 0.022172946482896805, + 0.016609465703368187, + 0.015457911416888237, + 0.05739886686205864, + 0.027737101539969444, + 0.13229919970035553, + 0.004791143350303173, + 0.004880635533481836, + 0.01022491417825222, + 0.007043312769383192, + 0.0074970233254134655, + 0.009736246429383755, + 0.005230270326137543, + 0.007729441858828068, + 0.011999592185020447, + 0.018667560070753098, + 0.012454405426979065, + 0.012359962798655033, + 0.025422455742955208, + 0.03223208338022232, + 0.014176902361214161, + 0.020711423829197884, + 0.004536060616374016, + 0.004760295152664185, + 0.003928740043193102, + 0.0052439128048717976, + 0.005904579069465399, + 0.0061690122820436954, + 0.0027183559723198414, + 0.005325141828507185 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 8, + 5 + ], + "selected_reward": -0.0046457452699542046 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 living room with a large floor-to-ceiling window?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera shot a long shot starting in the distance and slowly moving forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera angle slightly overhead?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a mother sitting on a beige rug wearing an off-white knit sweater with long brown hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the mother pick up a white T-shirt and bring it to her nose while closing her eyes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.048670560121536255 + ], + "opposite_logprob_tokens": [ + -3.048670530319214 + ], + "logprob_reward": -0.048670560121536255, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.048670560121536255, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14138390123844147, + 0.03190961480140686, + 0.03280654549598694, + 0.033405851572752, + 0.035217758268117905, + 0.37598341703414917, + 0.07039307057857513, + 0.10796219855546951, + 0.10698744654655457, + 0.06395023316144943 + ], + "credit_map": [ + 0.009187712334096432, + 0.012206544168293476, + 0.013967737555503845, + 0.01549215242266655, + 0.031497564166784286, + 0.01814008317887783, + 0.023059362545609474, + 0.01783275604248047, + 0.0031616161577403545, + 0.0023858302738517523, + 0.0032065194100141525, + 0.002521984977647662, + 0.00578620471060276, + 0.005000966601073742, + 0.005187373608350754, + 0.004659120924770832, + 0.0025176473427563906, + 0.002373484428972006, + 0.0030642934143543243, + 0.004069455433636904, + 0.005084204953163862, + 0.004006907343864441, + 0.005642736330628395, + 0.006047816947102547, + 0.006645312998443842, + 0.003368882229551673, + 0.002825259929522872, + 0.0032936539500951767, + 0.003858076874166727, + 0.00447048107162118, + 0.005343683063983917, + 0.0036005007568746805, + 0.004159811418503523, + 0.005069083534181118, + 0.003528695087879896, + 0.005436261184513569, + 0.0038999831303954124, + 0.0030543983448296785, + 0.005089063197374344, + 0.00498046213760972, + 0.018824484199285507, + 0.008259577676653862, + 0.11381076276302338, + 0.015584454871714115, + 0.016026955097913742, + 0.023002302274107933, + 0.022440271452069283, + 0.15803460776805878, + 0.006342428736388683, + 0.007050635758787394, + 0.0052121481858193874, + 0.01012842170894146, + 0.008748062886297703, + 0.007783773820847273, + 0.009435711428523064, + 0.015691887587308884, + 0.007032718043774366, + 0.00942053273320198, + 0.023112153634428978, + 0.016342228278517723, + 0.013470705598592758, + 0.013308538123965263, + 0.008512160740792751, + 0.01676316186785698, + 0.008902285248041153, + 0.01707141101360321, + 0.013412633910775185, + 0.01248441357165575, + 0.01952269859611988, + 0.016114166006445885, + 0.009218745864927769, + 0.010261091403663158, + 0.005041375290602446, + 0.006151554174721241, + 0.008685432374477386, + 0.010609561577439308, + 0.0060497005470097065, + 0.010195339098572731, + 0.00466138543561101, + 0.012555881403386593 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 7 + ], + "selected_reward": -0.048670560121536255 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "No", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "long shot, starting in the distance and slowly moving forward, captured from a slightly overhead angle, captures a tranquil living room bathed in the golden afternoon sun. Warm light slants in through a large floor-to-ceiling window, casting long, soft shadows across the polished oak floor. Tiny particles of dust drifting in the air shimmer in the beams of light. At the center of the frame, a mother sits comfortably on a soft beige rug. Her long, softly wavy brown hair is casually tied back, and she wears a loose, off-white knit sweater. Before her lies a small pile of freshly washed, neatly folded baby clothes in soft shades of pale yellow, sky blue, and pure white. A gentle, contented smile graces her face, her eyes filled with maternal love. She reaches out her slender fingers and delicately strokes the soft cotton of a pale yellow baby bodysuit, her movements gentle and cherishing. Then, she picked up a small white T-shirt and, steadily and calmly, brought it to her nose. She closed her eyes, as if inhaling the comforting, fresh scent of sunlight and baby laundry soap. The entire scene, bathed in warm, soft tones, was shrouded in a tranquil, healing atmosphere, perfectly embodying the warmth and comfort of home.", + "seed": 2922, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000056/group_14/index_3036_seed_2922_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": -0.4729597649537027, + "vlm_logprob": -0.4729597649537027, + "vlm_yes_no_margin": 2.299999988079071, + "vlm_token_credit": -0.4729597649537027, + "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 living room with a large floor-to-ceiling window?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera shot a long shot starting in the distance and slowly moving forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera angle slightly overhead?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a mother sitting on a beige rug wearing an off-white knit sweater with long brown hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the mother pick up a white T-shirt and bring it to her nose while closing her eyes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.36542779207229614 + ], + "opposite_logprob_tokens": [ + -1.2404277324676514 + ], + "logprob_reward": -0.36542779207229614, + "yes_no_margin_reward": 0.8749999403953552, + "token_credit_reward": -0.36542779207229614, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1716928631067276, + 0.05530566722154617, + 0.0697546973824501, + 0.061391424387693405, + 0.14227312803268433, + 0.1413727104663849, + 0.09703235328197479, + 0.06222137436270714, + 0.03039965219795704, + 0.16855613887310028 + ], + "credit_map": [ + 0.016781890764832497, + 0.01926364190876484, + 0.015909312292933464, + 0.024710215628147125, + 0.027804376557469368, + 0.022612404078245163, + 0.02322223223745823, + 0.021388808265328407, + 0.005411927122622728, + 0.0063327704556286335, + 0.0053390199318528175, + 0.010981399565935135, + 0.0068291304633021355, + 0.006501601543277502, + 0.00725732184946537, + 0.00665249302983284, + 0.00943836197257042, + 0.007809637114405632, + 0.01612633466720581, + 0.009928586892783642, + 0.007928869687020779, + 0.006976280827075243, + 0.005283589009195566, + 0.00626303581520915, + 0.007547828834503889, + 0.004338033497333527, + 0.011418846435844898, + 0.008327292278409004, + 0.007123574614524841, + 0.009862332604825497, + 0.006327866110950708, + 0.006445650011301041, + 0.012336622923612595, + 0.008000639267265797, + 0.01855423115193844, + 0.020997868850827217, + 0.013983655720949173, + 0.017634393647313118, + 0.024533838033676147, + 0.02623187005519867, + 0.018550382927060127, + 0.007649486884474754, + 0.027270358055830002, + 0.00956786423921585, + 0.016916926950216293, + 0.02046550251543522, + 0.01666588895022869, + 0.0242863018065691, + 0.007623393554240465, + 0.0036163069307804108, + 0.029592221602797508, + 0.0071919565089046955, + 0.006836493033915758, + 0.02839078940451145, + 0.0069677336141467094, + 0.00681346608325839, + 0.004830881487578154, + 0.006283655297011137, + 0.004776646848767996, + 0.0066578807309269905, + 0.011782839894294739, + 0.00797054823487997, + 0.009060555137693882, + 0.010858368128538132, + 0.003474624129012227, + 0.002795403590425849, + 0.004916450008749962, + 0.0027820810209959745, + 0.00426762318238616, + 0.004948885645717382, + 0.003354327054694295, + 0.0038602573331445456, + 0.02066420204937458, + 0.007786948699504137, + 0.013065258041024208, + 0.01108322199434042, + 0.026044370606541634, + 0.047610197216272354, + 0.020346077159047127, + 0.02195585146546364 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 4 + ], + "selected_reward": -0.36542779207229614 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 living room with a large floor-to-ceiling window?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera shot a long shot starting in the distance and slowly moving forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera angle slightly overhead?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a mother sitting on a beige rug wearing an off-white knit sweater with long brown hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the mother pick up a white T-shirt and bring it to her nose while closing her eyes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"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": [ + -1.9103370904922485 + ], + "opposite_logprob_tokens": [ + -0.16033713519573212 + ], + "logprob_reward": -1.9103370904922485, + "yes_no_margin_reward": -1.75, + "token_credit_reward": -1.9103370904922485, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.24282602965831757, + 0.08073524385690689, + 0.04461615905165672, + 0.06616134196519852, + 0.060777124017477036, + 0.04302018880844116, + 0.10925381630659103, + 0.08890537172555923, + 0.09930282831192017, + 0.16440194845199585 + ], + "credit_map": [ + 0.044028304517269135, + 0.018931955099105835, + 0.05692620575428009, + 0.03078790381550789, + 0.027457065880298615, + 0.029526198282837868, + 0.019313249737024307, + 0.015855127945542336, + 0.0067407819442451, + 0.015316671691834927, + 0.006291660014539957, + 0.01528562605381012, + 0.009557745419442654, + 0.011449049226939678, + 0.009112291038036346, + 0.006981419865041971, + 0.005634821020066738, + 0.004793862346559763, + 0.009419726207852364, + 0.005386024247854948, + 0.005572160705924034, + 0.005245013628154993, + 0.004352436866611242, + 0.004212112631648779, + 0.006856522988528013, + 0.003964703064411879, + 0.021121617406606674, + 0.006057241000235081, + 0.006095029413700104, + 0.01238174457103014, + 0.004797478672116995, + 0.004887007642537355, + 0.005093035288155079, + 0.0036334923934191465, + 0.007732966914772987, + 0.008576246909797192, + 0.006931586656719446, + 0.007839897647500038, + 0.00975434947758913, + 0.01121554896235466, + 0.010415851138532162, + 0.005019977688789368, + 0.008927823975682259, + 0.0035251574590802193, + 0.004726314451545477, + 0.005596758797764778, + 0.002030629198998213, + 0.0027776765637099743, + 0.01467085164040327, + 0.007950757630169392, + 0.028821410611271858, + 0.008274608291685581, + 0.007595287635922432, + 0.02093205600976944, + 0.011044974438846111, + 0.009963864460587502, + 0.012609095312654972, + 0.00941847637295723, + 0.011676139198243618, + 0.013508410193026066, + 0.01294010691344738, + 0.01287190429866314, + 0.009286005049943924, + 0.006595234386622906, + 0.014308905228972435, + 0.010336264036595821, + 0.020311731845140457, + 0.010469737462699413, + 0.009369591251015663, + 0.02093985117971897, + 0.007078398950397968, + 0.006488344632089138, + 0.028389453887939453, + 0.012546396814286709, + 0.025459246709942818, + 0.01440854649990797, + 0.0175248384475708, + 0.04272531345486641, + 0.013207610696554184, + 0.010140529833734035 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 6 + ], + "selected_reward": -1.9103370904922485 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 living room with a large floor-to-ceiling window?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera shot a long shot starting in the distance and slowly moving forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera angle slightly overhead?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a mother sitting on a beige rug wearing an off-white knit sweater with long brown hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the mother pick up a white T-shirt and bring it to her nose while closing her eyes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03815329819917679 + ], + "opposite_logprob_tokens": [ + -3.2881534099578857 + ], + "logprob_reward": -0.03815329819917679, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03815329819917679, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3729626536369324, + 0.12842239439487457, + 0.0553879477083683, + 0.06276584416627884, + 0.03764296695590019, + 0.024211768060922623, + 0.08605606108903885, + 0.10412976890802383, + 0.05609219893813133, + 0.07232842594385147 + ], + "credit_map": [ + 0.07153132557868958, + 0.03550093621015549, + 0.09185709804296494, + 0.04852484539151192, + 0.035862356424331665, + 0.04035019874572754, + 0.022729668766260147, + 0.02660622075200081, + 0.012076687067747116, + 0.02105439268052578, + 0.010824262164533138, + 0.026736054569482803, + 0.015951791778206825, + 0.019860757514834404, + 0.011928188614547253, + 0.009990254417061806, + 0.0051537323743104935, + 0.007364037912338972, + 0.01045523677021265, + 0.006379588972777128, + 0.00837977509945631, + 0.008070265874266624, + 0.004106823820620775, + 0.0054784840904176235, + 0.006061505060642958, + 0.0055992500856518745, + 0.014117804355919361, + 0.006513499189168215, + 0.00850206520408392, + 0.010372677817940712, + 0.0058074817061424255, + 0.005791556555777788, + 0.004746398888528347, + 0.002192520536482334, + 0.005056431051343679, + 0.0046489848755300045, + 0.005255243740975857, + 0.005565359257161617, + 0.004335660487413406, + 0.005842366721481085, + 0.0027612948324531317, + 0.001799187040887773, + 0.0048210471868515015, + 0.0022259794641286135, + 0.0028239423409104347, + 0.0047540743835270405, + 0.0018896209076046944, + 0.0031366199254989624, + 0.008816550485789776, + 0.0060690767131745815, + 0.02617502398788929, + 0.0047670756466686726, + 0.0058775595389306545, + 0.02383856102824211, + 0.005015213508158922, + 0.0054969931952655315, + 0.013285444118082523, + 0.011231164447963238, + 0.012443461455404758, + 0.018094701692461967, + 0.01568746753036976, + 0.012532512657344341, + 0.012977605685591698, + 0.007877412252128124, + 0.007113209459930658, + 0.006155593320727348, + 0.00947237852960825, + 0.007089224178344011, + 0.0060356417670845985, + 0.01041922252625227, + 0.0049897003918886185, + 0.004817228298634291, + 0.011886823922395706, + 0.007268334273248911, + 0.011252039112150669, + 0.0056462096981704235, + 0.008453779853880405, + 0.016765788197517395, + 0.0058403038419783115, + 0.005215142387896776 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 7 + ], + "selected_reward": -0.03815329819917679 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 living room with a large floor-to-ceiling window?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera shot a long shot starting in the distance and slowly moving forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera angle slightly overhead?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a mother sitting on a beige rug wearing an off-white knit sweater with long brown hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the mother pick up a white T-shirt and bring it to her nose while closing her eyes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0022050845436751842 + ], + "opposite_logprob_tokens": [ + -6.127204895019531 + ], + "logprob_reward": -0.0022050845436751842, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.0022050845436751842, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3729941248893738, + 0.1173349991440773, + 0.07301211357116699, + 0.06594784557819366, + 0.04467790201306343, + 0.030909711495041847, + 0.04589100927114487, + 0.09667215496301651, + 0.049201276153326035, + 0.1033589243888855 + ], + "credit_map": [ + 0.08495256304740906, + 0.030961541458964348, + 0.1014237031340599, + 0.035339951515197754, + 0.025616612285375595, + 0.029681798070669174, + 0.040765341371297836, + 0.024252597242593765, + 0.010867475531995296, + 0.020299559459090233, + 0.010317525826394558, + 0.02353980392217636, + 0.013219289481639862, + 0.015460008755326271, + 0.011997141875326633, + 0.011634190566837788, + 0.00754172308370471, + 0.007908905856311321, + 0.019184257835149765, + 0.0101820919662714, + 0.007475797552615404, + 0.007800253573805094, + 0.007194384001195431, + 0.005724698770791292, + 0.0076564010232687, + 0.003987843636423349, + 0.021213870495557785, + 0.006413497030735016, + 0.005836439318954945, + 0.01016179658472538, + 0.005449919495731592, + 0.005228078458458185, + 0.005058330483734608, + 0.004166784230619669, + 0.006888825446367264, + 0.006255843676626682, + 0.005395854823291302, + 0.005406303331255913, + 0.0051020667888224125, + 0.006403894163668156, + 0.004872339311987162, + 0.0026503289118409157, + 0.00596089381724596, + 0.0035107939038425684, + 0.004671250004321337, + 0.004104863852262497, + 0.002084719017148018, + 0.0030545222107321024, + 0.003970629535615444, + 0.003857426578179002, + 0.009714676067233086, + 0.00477656489238143, + 0.004846720024943352, + 0.007979745976626873, + 0.005081136245280504, + 0.0056641073897480965, + 0.008963664062321186, + 0.008928898721933365, + 0.011032250709831715, + 0.019223947077989578, + 0.010822734795510769, + 0.014384799636900425, + 0.015408431179821491, + 0.007907428778707981, + 0.00653547840192914, + 0.004885491449385881, + 0.009966697543859482, + 0.004751401953399181, + 0.005544610321521759, + 0.008402658626437187, + 0.004557251464575529, + 0.004557684529572725, + 0.018931586295366287, + 0.009543362073600292, + 0.016960682347416878, + 0.007569567766040564, + 0.012024562805891037, + 0.02453637309372425, + 0.006682387553155422, + 0.007110407575964928 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 9 + ], + "selected_reward": -0.0022050845436751842 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 living room with a large floor-to-ceiling window?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera shot a long shot starting in the distance and slowly moving forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera angle slightly overhead?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a mother sitting on a beige rug wearing an off-white knit sweater with long brown hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the mother pick up a white T-shirt and bring it to her nose while closing her eyes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04867555946111679 + ], + "opposite_logprob_tokens": [ + -3.048675537109375 + ], + "logprob_reward": -0.04867555946111679, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04867555946111679, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.185728058218956, + 0.04621003195643425, + 0.0456719845533371, + 0.13942919671535492, + 0.04516896605491638, + 0.030035249888896942, + 0.05575244128704071, + 0.24878497421741486, + 0.08228371292352676, + 0.12093535810709 + ], + "credit_map": [ + 0.04646334424614906, + 0.009005647152662277, + 0.05805420130491257, + 0.011485099792480469, + 0.013730209320783615, + 0.016582066193223, + 0.019031545147299767, + 0.01137594971805811, + 0.004591842647641897, + 0.005296669900417328, + 0.004757129121571779, + 0.007292795926332474, + 0.007141450420022011, + 0.006763810757547617, + 0.0059399413876235485, + 0.00442639272660017, + 0.004788023419678211, + 0.0030579983722418547, + 0.01114983856678009, + 0.0052683912217617035, + 0.004911442287266254, + 0.006434199400246143, + 0.0051830122247338295, + 0.0048790848813951015, + 0.016628924757242203, + 0.007178697735071182, + 0.044559329748153687, + 0.01177289430052042, + 0.012153051793575287, + 0.0263131745159626, + 0.010916728526353836, + 0.009906414896249771, + 0.006119911093264818, + 0.0031882156617939472, + 0.007929024286568165, + 0.005676768720149994, + 0.005480040796101093, + 0.006019013933837414, + 0.003989746794104576, + 0.006766248028725386, + 0.004181082360446453, + 0.0030659569893032312, + 0.003855140646919608, + 0.0032120472751557827, + 0.004227153491228819, + 0.0044214739464223385, + 0.0026018512435257435, + 0.004470545798540115, + 0.006178696174174547, + 0.004989203531295061, + 0.005915668793022633, + 0.0069768186658620834, + 0.004950488917529583, + 0.00856755394488573, + 0.009178880602121353, + 0.00899513904005289, + 0.02423107996582985, + 0.0231375340372324, + 0.023847229778766632, + 0.05291011556982994, + 0.031185468658804893, + 0.03348172828555107, + 0.04022502526640892, + 0.019766809418797493, + 0.010738006792962551, + 0.00906831119209528, + 0.015806671231985092, + 0.01029040478169918, + 0.009565775282680988, + 0.01250170823186636, + 0.007532898336648941, + 0.006779942195862532, + 0.021639715880155563, + 0.014691434800624847, + 0.017464371398091316, + 0.013003231957554817, + 0.01653493382036686, + 0.016890347003936768, + 0.01083407737314701, + 0.009877249598503113 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 3 + ], + "selected_reward": -0.04867555946111679 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "No", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "long shot, starting in the distance and slowly moving forward, captured from a slightly overhead angle, captures a tranquil living room bathed in the golden afternoon sun. Warm light slants in through a large floor-to-ceiling window, casting long, soft shadows across the polished oak floor. Tiny particles of dust drifting in the air shimmer in the beams of light. At the center of the frame, a mother sits comfortably on a soft beige rug. Her long, softly wavy brown hair is casually tied back, and she wears a loose, off-white knit sweater. Before her lies a small pile of freshly washed, neatly folded baby clothes in soft shades of pale yellow, sky blue, and pure white. A gentle, contented smile graces her face, her eyes filled with maternal love. She reaches out her slender fingers and delicately strokes the soft cotton of a pale yellow baby bodysuit, her movements gentle and cherishing. Then, she picked up a small white T-shirt and, steadily and calmly, brought it to her nose. She closed her eyes, as if inhaling the comforting, fresh scent of sunlight and baby laundry soap. The entire scene, bathed in warm, soft tones, was shrouded in a tranquil, healing atmosphere, perfectly embodying the warmth and comfort of home.", + "seed": 2923, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000056/group_14/index_3036_seed_2923_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": -0.4678442723117769, + "vlm_logprob": -0.4678442723117769, + "vlm_yes_no_margin": 2.25, + "vlm_token_credit": -0.4678442723117769, + "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 living room with a large floor-to-ceiling window?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera shot a long shot starting in the distance and slowly moving forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera angle slightly overhead?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a mother sitting on a beige rug wearing an off-white knit sweater with long brown hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the mother pick up a white T-shirt and bring it to her nose while closing her eyes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2435816526412964 + ], + "opposite_logprob_tokens": [ + -1.6185816526412964 + ], + "logprob_reward": -0.2435816526412964, + "yes_no_margin_reward": 1.375, + "token_credit_reward": -0.2435816526412964, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.670714795589447, + 0.05515323579311371, + 0.025168420746922493, + 0.006538973655551672, + 0.04695085063576698, + 0.03495235741138458, + 0.028200218454003334, + 0.017334165051579475, + 0.0144977867603302, + 0.10048920661211014 + ], + "credit_map": [ + 0.23940409719944, + 0.019601473584771156, + 0.28195902705192566, + 0.03261713311076164, + 0.015425394289195538, + 0.03948316350579262, + 0.0128987617790699, + 0.029325714334845543, + 0.0070589808747172356, + 0.003874013898894191, + 0.006126972381025553, + 0.006697788834571838, + 0.004120001569390297, + 0.012549478560686111, + 0.004427960608154535, + 0.010298039764165878, + 0.0029706228524446487, + 0.0018156558508053422, + 0.00250554783269763, + 0.002558489330112934, + 0.0022226276341825724, + 0.003785251174122095, + 0.0033704398665577173, + 0.005939785856753588, + 0.0008836498600430787, + 0.0007245675660669804, + 0.0006692129536531866, + 0.0006084413616918027, + 0.000996690010651946, + 0.0008308622636832297, + 0.0007531841984018683, + 0.0010723654413595796, + 0.01712973415851593, + 0.001764672459103167, + 0.004612178076058626, + 0.0016757538542151451, + 0.0015433194348588586, + 0.014474036172032356, + 0.00259331869892776, + 0.0031578352209180593, + 0.00400969386100769, + 0.001688184100203216, + 0.012495101429522038, + 0.00207881024107337, + 0.0013681257842108607, + 0.002737439936026931, + 0.0022029175888746977, + 0.008372083306312561, + 0.002485173987224698, + 0.0024339656811207533, + 0.0041868435218930244, + 0.003111637895926833, + 0.002441534772515297, + 0.004755599889904261, + 0.0035176218952983618, + 0.0052678389474749565, + 0.0016481054481118917, + 0.0012367632007226348, + 0.002519340254366398, + 0.0014059353852644563, + 0.0014756993623450398, + 0.0024855455849319696, + 0.001929571502842009, + 0.0046332054771482944, + 0.0014906759606674314, + 0.0015499643050134182, + 0.0014251882676035166, + 0.0011619293363764882, + 0.0011450429446995258, + 0.0031379712745547295, + 0.0024024518206715584, + 0.002184563083574176, + 0.008220384828746319, + 0.007307541090995073, + 0.008001442067325115, + 0.006692721974104643, + 0.009025522507727146, + 0.026492921635508537, + 0.014068273827433586, + 0.020680399611592293 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 1 + ], + "selected_reward": -0.2435816526412964 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 living room with a large floor-to-ceiling window?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera shot a long shot starting in the distance and slowly moving forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera angle slightly overhead?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a mother sitting on a beige rug wearing an off-white knit sweater with long brown hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the mother pick up a white T-shirt and bring it to her nose while closing her eyes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"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": [ + -2.017792224884033 + ], + "opposite_logprob_tokens": [ + -0.14279219508171082 + ], + "logprob_reward": -2.017792224884033, + "yes_no_margin_reward": -1.875, + "token_credit_reward": -2.017792224884033, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.27342021465301514, + 0.040239911526441574, + 0.07239324599504471, + 0.05429278686642647, + 0.25173765420913696, + 0.10984988510608673, + 0.023930950090289116, + 0.025819025933742523, + 0.07446479797363281, + 0.07385145127773285 + ], + "credit_map": [ + 0.09464778751134872, + 0.009064325131475925, + 0.1060141995549202, + 0.01599784381687641, + 0.0090029863640666, + 0.01872880570590496, + 0.008939187973737717, + 0.011025100015103817, + 0.00862907338887453, + 0.003634055843576789, + 0.005189607851207256, + 0.0048232958652079105, + 0.002462435979396105, + 0.006077214609831572, + 0.004392201546579599, + 0.0050320313312113285, + 0.008819136768579483, + 0.005601777695119381, + 0.007321132346987724, + 0.00803426280617714, + 0.006348724011331797, + 0.011697839014232159, + 0.0091937854886055, + 0.015376594848930836, + 0.013648346066474915, + 0.00404551113024354, + 0.0038750784005969763, + 0.0037915969733148813, + 0.0049877772107720375, + 0.016140686348080635, + 0.004189277999103069, + 0.003614516230300069, + 0.09120294451713562, + 0.009290493093430996, + 0.013347448781132698, + 0.007323433645069599, + 0.00740068918094039, + 0.09809616208076477, + 0.011453483253717422, + 0.013623026199638844, + 0.014666150324046612, + 0.005383425857871771, + 0.0404522567987442, + 0.007651933468878269, + 0.003793714800849557, + 0.006838814355432987, + 0.006591797340661287, + 0.02447180263698101, + 0.002760169096291065, + 0.002528874669224024, + 0.0026418017223477364, + 0.004087930079549551, + 0.0028645212296396494, + 0.002287918236106634, + 0.003511559683829546, + 0.0032481763046234846, + 0.0029081953689455986, + 0.0024947188794612885, + 0.004315318539738655, + 0.0026522905100136995, + 0.0022490203846246004, + 0.0032825705129653215, + 0.0029841584619134665, + 0.004932754673063755, + 0.00830467976629734, + 0.006878267508000135, + 0.009749257005751133, + 0.008662812411785126, + 0.00909897219389677, + 0.008315667510032654, + 0.012888241559267044, + 0.010566906072199345, + 0.008153621107339859, + 0.007261280901730061, + 0.014019764959812164, + 0.007584030739963055, + 0.004583748523145914, + 0.014164519496262074, + 0.009471405297517776, + 0.00861309003084898 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 5 + ], + "selected_reward": -2.017792224884033 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 living room with a large floor-to-ceiling window?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera shot a long shot starting in the distance and slowly moving forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera angle slightly overhead?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a mother sitting on a beige rug wearing an off-white knit sweater with long brown hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the mother pick up a white T-shirt and bring it to her nose while closing her eyes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.038164082914590836 + ], + "opposite_logprob_tokens": [ + -3.2881641387939453 + ], + "logprob_reward": -0.038164082914590836, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.038164082914590836, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.46366962790489197, + 0.029111864045262337, + 0.0605030320584774, + 0.062043484300374985, + 0.21836625039577484, + 0.054167550057172775, + 0.04045794904232025, + 0.019695160910487175, + 0.026432733982801437, + 0.02555231936275959 + ], + "credit_map": [ + 0.16226989030838013, + 0.014847072772681713, + 0.19136342406272888, + 0.022552339360117912, + 0.010841773822903633, + 0.03084489330649376, + 0.011402166448533535, + 0.019548097625374794, + 0.0046115354634821415, + 0.0033815118949860334, + 0.003318733535706997, + 0.003906716126948595, + 0.00189454504288733, + 0.003896063193678856, + 0.0035507434513419867, + 0.004552014172077179, + 0.007269374094903469, + 0.004891344346106052, + 0.0073472983203828335, + 0.007643323391675949, + 0.004952670074999332, + 0.00777948135510087, + 0.007384391967207193, + 0.013235144317150116, + 0.02142772078514099, + 0.0033493144437670708, + 0.0040102675557136536, + 0.0026174536906182766, + 0.0022910453844815493, + 0.022433491423726082, + 0.0029720263555645943, + 0.0029421616345643997, + 0.07445657253265381, + 0.008364156819880009, + 0.010934446007013321, + 0.007173309102654457, + 0.006545546464622021, + 0.08931082487106323, + 0.009627736173570156, + 0.011953650042414665, + 0.007095810025930405, + 0.002810510341078043, + 0.018245428800582886, + 0.003459990257397294, + 0.0026980387046933174, + 0.004562646150588989, + 0.005505701061338186, + 0.009789425879716873, + 0.0029801924247294664, + 0.0027183976490050554, + 0.0049931262619793415, + 0.005061090923845768, + 0.0037924170028418303, + 0.007174604572355747, + 0.00535323191434145, + 0.008384885266423225, + 0.0017680461751297116, + 0.0015624492662027478, + 0.002385467756539583, + 0.002240645233541727, + 0.0019348656060174108, + 0.00399136682972312, + 0.0027929276693612337, + 0.003019393188878894, + 0.004121269099414349, + 0.0022665986325591803, + 0.003712979843840003, + 0.0022829726804047823, + 0.0022002309560775757, + 0.006896737031638622, + 0.002176904119551182, + 0.002775041153654456, + 0.002876967191696167, + 0.0026908447034657, + 0.004890987649559975, + 0.0027067505288869143, + 0.0016800157027319074, + 0.004820914939045906, + 0.0032062875106930733, + 0.0026795531157404184 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 3 + ], + "selected_reward": -0.038164082914590836 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 living room with a large floor-to-ceiling window?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera shot a long shot starting in the distance and slowly moving forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera angle slightly overhead?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a mother sitting on a beige rug wearing an off-white knit sweater with long brown hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the mother pick up a white T-shirt and bring it to her nose while closing her eyes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00596374599263072 + ], + "opposite_logprob_tokens": [ + -5.1309638023376465 + ], + "logprob_reward": -0.00596374599263072, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.00596374599263072, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3564988076686859, + 0.03815050423145294, + 0.039015889167785645, + 0.05003326013684273, + 0.24239276349544525, + 0.07314839214086533, + 0.07101684808731079, + 0.029775941744446754, + 0.01952829211950302, + 0.08043932914733887 + ], + "credit_map": [ + 0.12500602006912231, + 0.013306302949786186, + 0.1423942595720291, + 0.021939976140856743, + 0.008703640662133694, + 0.019765254110097885, + 0.007919580675661564, + 0.01746373623609543, + 0.005446264985948801, + 0.004001632798463106, + 0.003328372025862336, + 0.005687755532562733, + 0.0027879327535629272, + 0.004863032139837742, + 0.0061682043597102165, + 0.005867306608706713, + 0.005190279800444841, + 0.004639472346752882, + 0.0025534993037581444, + 0.005560374353080988, + 0.0028748984914273024, + 0.00572004821151495, + 0.0045592328533530235, + 0.007918081246316433, + 0.008438232354819775, + 0.005217320751398802, + 0.004274947568774223, + 0.004549176897853613, + 0.006418644450604916, + 0.009769855067133904, + 0.00560741638764739, + 0.005757657345384359, + 0.08300872147083282, + 0.00954638235270977, + 0.012135682627558708, + 0.008393342606723309, + 0.007579186465591192, + 0.09838579595088959, + 0.009871315211057663, + 0.013472320511937141, + 0.0075103663839399815, + 0.004532173275947571, + 0.011691807769238949, + 0.010263444855809212, + 0.005657237488776445, + 0.010204864665865898, + 0.014552326872944832, + 0.008736161515116692, + 0.005911981221288443, + 0.00667285080999136, + 0.009162827394902706, + 0.008358199149370193, + 0.00906269159168005, + 0.01010378822684288, + 0.009972447529435158, + 0.011772050522267818, + 0.002921850187703967, + 0.0033228546380996704, + 0.0038602459244430065, + 0.004510888829827309, + 0.003774910233914852, + 0.0031392881646752357, + 0.004472238477319479, + 0.00377366179600358, + 0.002811894752085209, + 0.001647604745812714, + 0.0027112143579870462, + 0.001679050736129284, + 0.001957247033715248, + 0.004713172093033791, + 0.002132059307768941, + 0.001876046764664352, + 0.00770858908072114, + 0.006994717288762331, + 0.010958723723888397, + 0.006899281404912472, + 0.007369156461209059, + 0.013329803012311459, + 0.009877578355371952, + 0.017301466315984726 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 9 + ], + "selected_reward": -0.00596374599263072 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 living room with a large floor-to-ceiling window?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera shot a long shot starting in the distance and slowly moving forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera angle slightly overhead?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a mother sitting on a beige rug wearing an off-white knit sweater with long brown hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the mother pick up a white T-shirt and bring it to her nose while closing her eyes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03371965512633324 + ], + "opposite_logprob_tokens": [ + -3.408719539642334 + ], + "logprob_reward": -0.03371965512633324, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03371965512633324, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14443156123161316, + 0.03214295953512192, + 0.11542032659053802, + 0.08427449315786362, + 0.15492822229862213, + 0.18339304625988007, + 0.0376143641769886, + 0.08462433516979218, + 0.06248687952756882, + 0.10068383812904358 + ], + "credit_map": [ + 0.04805831238627434, + 0.0050647747702896595, + 0.05473503842949867, + 0.00917113944888115, + 0.00571793457493186, + 0.010177591815590858, + 0.004525796975940466, + 0.006980970501899719, + 0.005195369943976402, + 0.0030354768969118595, + 0.004505658522248268, + 0.0036576881539076567, + 0.0033173367846757174, + 0.005619429051876068, + 0.002343046246096492, + 0.004468950442969799, + 0.014121460728347301, + 0.009356898255646229, + 0.010894089005887508, + 0.013955015689134598, + 0.009827353991568089, + 0.017012182623147964, + 0.015267465263605118, + 0.02498585358262062, + 0.026503920555114746, + 0.003911044914275408, + 0.005902791395783424, + 0.005635604728013277, + 0.003239539684727788, + 0.028325166553258896, + 0.004383569583296776, + 0.006372856441885233, + 0.05181233584880829, + 0.006010507233440876, + 0.00815324205905199, + 0.006667835172265768, + 0.005516334902495146, + 0.060205258429050446, + 0.008168774656951427, + 0.008393936790525913, + 0.021340180188417435, + 0.008567340672016144, + 0.06586317718029022, + 0.016727834939956665, + 0.00709553575143218, + 0.011946583166718483, + 0.011031895875930786, + 0.04082050174474716, + 0.003508766880258918, + 0.0031428781803697348, + 0.0048170085065066814, + 0.005370094440877438, + 0.004254304803907871, + 0.005047001410275698, + 0.005680730100721121, + 0.005793578457087278, + 0.009304535575211048, + 0.011384178884327412, + 0.013246322050690651, + 0.014562785625457764, + 0.00862713623791933, + 0.008036147803068161, + 0.012452209368348122, + 0.007011018227785826, + 0.00782535970211029, + 0.006354637909680605, + 0.010701669380068779, + 0.006795793771743774, + 0.006620119325816631, + 0.007579303812235594, + 0.009381399489939213, + 0.007228593807667494, + 0.010845447890460491, + 0.010439286939799786, + 0.018207764253020287, + 0.012023829855024815, + 0.0065611875616014, + 0.01892576552927494, + 0.012729238718748093, + 0.010951327160000801 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 4, + 0 + ], + "selected_reward": -0.03371965512633324 + } + ], + "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/0000056/group_15/index_tensor([4599, 4599])_group_15_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000056/group_15/index_tensor([4599, 4599])_group_15_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..c42ec4468c1dafe25dfc8b04cf82d83690f2bbdd --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000056/group_15/index_tensor([4599, 4599])_group_15_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "slow, macro zoom-in shot glides through crystal-clear turquoise water, gradually approaching an intimate display of meticulous detail. Focus is on a pair of vibrant orange and white-striped Ocellaris clownfish engaged in parental care. Their bodies are a bright, saturated orange, marked by three distinct white vertical stripes outlined in black and purposefully positioned. They hover over a small group of shimmering, translucent amber eggs, firmly nestled on the algae-covered rocky base of their host anemone. A single fish, positioned directly above, diligently fanns its pectoral fins in steady, rhythmic motions, generating gentle, life-sustaining currents over its delicate claws. The magnificent anemone's surrounding tentacles, fleshy green with soft purple tips, sway gently and hypnotically in the clear water. Bright sunlight filters through the surface, creating a dynamic play of light and shadow and casting dancing caustic patterns that make the individual eggs shimmer. The background blends into a deep blue, soft-focus abyss, enhancing the tranquil and focused atmosphere of this captivating scene of aquatic life.", + "seed": 25670, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000056/group_15/index_4599_seed_25670_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": -0.056311243865638974, + "vlm_logprob": -0.056311243865638974, + "vlm_yes_no_margin": 3.575, + "vlm_token_credit": -0.056311243865638974, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a slow macro zoom-in gliding through crystal-clear turquoise water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are two Ocellaris clownfish visible with bright orange bodies and three distinct white vertical stripes outlined in black?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a single fish positioned directly above the eggs, fanning its pectoral fins in steady, rhythmic motions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene featuring a host anemone with fleshy green tentacles and soft purple tips swaying gently?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does bright sunlight filter through the surface creating dancing caustic patterns that make the translucent amber eggs shimmer?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.18240921199321747 + ], + "opposite_logprob_tokens": [ + -1.8074091672897339 + ], + "logprob_reward": -0.18240921199321747, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.18240921199321747, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1175619438290596, + 0.1958511769771576, + 0.06089353933930397, + 0.1609332263469696, + 0.1664987951517105, + 0.03529299423098564, + 0.1111689954996109, + 0.08809289336204529, + 0.03901591897010803, + 0.02469058521091938 + ], + "credit_map": [ + 0.012983313761651516, + 0.0129763875156641, + 0.013236623257398605, + 0.013678978197276592, + 0.015547883696854115, + 0.01676969788968563, + 0.016924425959587097, + 0.015444615855813026, + 0.013153194449841976, + 0.01583814062178135, + 0.022586818784475327, + 0.036098770797252655, + 0.01581592671573162, + 0.015799308195710182, + 0.05610540881752968, + 0.02045358344912529, + 0.007947152480483055, + 0.006764223799109459, + 0.00548349367454648, + 0.007245427463203669, + 0.006913057528436184, + 0.009125337935984135, + 0.011275111697614193, + 0.006139727775007486, + 0.01741291768848896, + 0.007042708806693554, + 0.027899129316210747, + 0.013854984194040298, + 0.006886914372444153, + 0.06813153624534607, + 0.007712770253419876, + 0.01199224404990673, + 0.02137608267366886, + 0.011522479355335236, + 0.01171598769724369, + 0.013394780457019806, + 0.012497062794864178, + 0.05391748994588852, + 0.012937268242239952, + 0.029137635603547096, + 0.0044069113209843636, + 0.00172481220215559, + 0.004406381398439407, + 0.002643186366185546, + 0.0017860395601019263, + 0.013876432552933693, + 0.0025762680452317, + 0.0038729573134332895, + 0.006133703980594873, + 0.0022854909766465425, + 0.0382658913731575, + 0.00497377710416913, + 0.005603397730737925, + 0.0061010257340967655, + 0.004248300101608038, + 0.04355740174651146, + 0.016764795407652855, + 0.0068414960987865925, + 0.00852446909993887, + 0.01271031703799963, + 0.008696457371115685, + 0.011495680548250675, + 0.01031007245182991, + 0.012749594636261463, + 0.0039031757041811943, + 0.003656745655462146, + 0.004214308224618435, + 0.008027389645576477, + 0.0034127498511224985, + 0.004153468646109104, + 0.005693765357136726, + 0.005954311694949865, + 0.0026962452102452517, + 0.004273629281669855, + 0.0028335077222436666, + 0.00231701647862792, + 0.0032846657559275627, + 0.0031276042573153973, + 0.002710087224841118, + 0.003447824390605092 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 4, + 3 + ], + "selected_reward": -0.18240921199321747 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a slow macro zoom-in gliding through crystal-clear turquoise water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are two Ocellaris clownfish visible with bright orange bodies and three distinct white vertical stripes outlined in black?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a single fish positioned directly above the eggs, fanning its pectoral fins in steady, rhythmic motions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene featuring a host anemone with fleshy green tentacles and soft purple tips swaying gently?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does bright sunlight filter through the surface creating dancing caustic patterns that make the translucent amber eggs shimmer?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00980407278984785 + ], + "opposite_logprob_tokens": [ + -4.6348042488098145 + ], + "logprob_reward": -0.00980407278984785, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.00980407278984785, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.055768344551324844, + 0.14700676500797272, + 0.11211100220680237, + 0.23319292068481445, + 0.05814818665385246, + 0.1286899894475937, + 0.05981484428048134, + 0.12997344136238098, + 0.023497402667999268, + 0.051797036081552505 + ], + "credit_map": [ + 0.005494787823408842, + 0.007190715055912733, + 0.008160942234098911, + 0.006895286496728659, + 0.005811101756989956, + 0.005960484500974417, + 0.007789733819663525, + 0.008465292863547802, + 0.011026187799870968, + 0.023037321865558624, + 0.015291825868189335, + 0.02128259465098381, + 0.017013544216752052, + 0.00996681023389101, + 0.03216877207159996, + 0.017219699919223785, + 0.011209926567971706, + 0.012607414275407791, + 0.010439001023769379, + 0.015963630750775337, + 0.014918405562639236, + 0.014967150054872036, + 0.02065921016037464, + 0.011346258223056793, + 0.02431950531899929, + 0.009309651330113411, + 0.049189917743206024, + 0.020060861483216286, + 0.009688309393823147, + 0.0948842391371727, + 0.010952357202768326, + 0.014788070693612099, + 0.009639274328947067, + 0.0044754850678145885, + 0.004022258799523115, + 0.005862902384251356, + 0.003921717870980501, + 0.016474876552820206, + 0.00588912982493639, + 0.007862545549869537, + 0.012387273833155632, + 0.003040481125935912, + 0.014913509599864483, + 0.006400156766176224, + 0.005504618398845196, + 0.04308473318815231, + 0.005609138403087854, + 0.03775007650256157, + 0.006099820137023926, + 0.002323825377970934, + 0.0148426229134202, + 0.004764694720506668, + 0.0033550411462783813, + 0.004041808657348156, + 0.004712387453764677, + 0.019674643874168396, + 0.037777453660964966, + 0.01049413438886404, + 0.014777084812521935, + 0.011054626666009426, + 0.009421474300324917, + 0.008244074881076813, + 0.010865169577300549, + 0.027339430525898933, + 0.004046352114528418, + 0.002895493758842349, + 0.0025828657671809196, + 0.0025923147331923246, + 0.0023558586835861206, + 0.003566168015822768, + 0.0022260923869907856, + 0.0032322602346539497, + 0.005330898333340883, + 0.007833483628928661, + 0.005669211037456989, + 0.006107468623667955, + 0.005899486131966114, + 0.005523462314158678, + 0.008282315917313099, + 0.007150714285671711 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 1, + 7 + ], + "selected_reward": -0.00980407278984785 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a slow macro zoom-in gliding through crystal-clear turquoise water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are two Ocellaris clownfish visible with bright orange bodies and three distinct white vertical stripes outlined in black?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a single fish positioned directly above the eggs, fanning its pectoral fins in steady, rhythmic motions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene featuring a host anemone with fleshy green tentacles and soft purple tips swaying gently?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does bright sunlight filter through the surface creating dancing caustic patterns that make the translucent amber eggs shimmer?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"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": [ + 0.09604286402463913, + 0.08035522699356079, + 0.050320208072662354, + 0.3437947630882263, + 0.06415470689535141, + 0.0891847312450409, + 0.08878031373023987, + 0.10948756337165833, + 0.04969174787402153, + 0.028187904506921768 + ], + "credit_map": [ + 0.008703161031007767, + 0.008328492753207684, + 0.013907931745052338, + 0.011198154650628567, + 0.010113411583006382, + 0.014156362041831017, + 0.01279788464307785, + 0.0168374702334404, + 0.010809900239109993, + 0.013804182410240173, + 0.008231284096837044, + 0.011163954623043537, + 0.008825914934277534, + 0.008096667937934399, + 0.01068513747304678, + 0.008738183416426182, + 0.004198000300675631, + 0.005378485657274723, + 0.006604538299143314, + 0.006989119574427605, + 0.0051008956506848335, + 0.004131497349590063, + 0.012186079286038876, + 0.005731591954827309, + 0.0369114875793457, + 0.011345228180289268, + 0.07165685296058655, + 0.02841782383620739, + 0.0136339096352458, + 0.1450100839138031, + 0.0168930571526289, + 0.01992630772292614, + 0.008912306278944016, + 0.0036567854695022106, + 0.003961128648370504, + 0.004287049639970064, + 0.0031508926767855883, + 0.024868112057447433, + 0.0039236219599843025, + 0.011394809931516647, + 0.011288640089333057, + 0.00306308432482183, + 0.01968958228826523, + 0.005778579972684383, + 0.003513153176754713, + 0.03196556121110916, + 0.0057348329573869705, + 0.008151295594871044, + 0.01927083171904087, + 0.006345556583255529, + 0.015421350486576557, + 0.011356762610375881, + 0.006912478245794773, + 0.006873775739222765, + 0.01061052642762661, + 0.01198902539908886, + 0.01343544665724039, + 0.01234315987676382, + 0.01648406870663166, + 0.01155350636690855, + 0.0074585480615496635, + 0.013216222636401653, + 0.010669752955436707, + 0.024326859042048454, + 0.005559999495744705, + 0.005830780137330294, + 0.006262433249503374, + 0.009716998785734177, + 0.005556663032621145, + 0.004272338468581438, + 0.006708210334181786, + 0.005784322042018175, + 0.0025739429984241724, + 0.0032157849054783583, + 0.0034843431785702705, + 0.003179228864610195, + 0.002925523789599538, + 0.0035969314631074667, + 0.003636357607319951, + 0.005575789604336023 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 7, + 0 + ], + "selected_reward": -0.06203708052635193 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a slow macro zoom-in gliding through crystal-clear turquoise water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are two Ocellaris clownfish visible with bright orange bodies and three distinct white vertical stripes outlined in black?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a single fish positioned directly above the eggs, fanning its pectoral fins in steady, rhythmic motions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene featuring a host anemone with fleshy green tentacles and soft purple tips swaying gently?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does bright sunlight filter through the surface creating dancing caustic patterns that make the translucent amber eggs shimmer?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020571796223521233 + ], + "opposite_logprob_tokens": [ + -3.895571708679199 + ], + "logprob_reward": -0.020571796223521233, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020571796223521233, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11189320683479309, + 0.12405943870544434, + 0.04138757288455963, + 0.321093887090683, + 0.08897692710161209, + 0.03995472565293312, + 0.12068582326173782, + 0.10438744723796844, + 0.016111990436911583, + 0.03144898638129234 + ], + "credit_map": [ + 0.013493835926055908, + 0.011627749539911747, + 0.010698250494897366, + 0.0126215610653162, + 0.015109773725271225, + 0.016391681507229805, + 0.019530411809682846, + 0.012419946491718292, + 0.01013405155390501, + 0.010643147863447666, + 0.014181488193571568, + 0.02112068422138691, + 0.011298207566142082, + 0.011346276849508286, + 0.03278305009007454, + 0.012552527710795403, + 0.00500495545566082, + 0.003293100278824568, + 0.00477656489238143, + 0.005668183322995901, + 0.003412854392081499, + 0.006040401756763458, + 0.008608601056039333, + 0.004582912195473909, + 0.034408777952194214, + 0.011258033104240894, + 0.06611290574073792, + 0.026675978675484657, + 0.012415950186550617, + 0.1356557309627533, + 0.014614605344831944, + 0.01995190791785717, + 0.014537409879267216, + 0.004712249618023634, + 0.0066732726991176605, + 0.005379803478717804, + 0.005098257213830948, + 0.030603714287281036, + 0.00541576137766242, + 0.01655646041035652, + 0.004867747891694307, + 0.0024289842694997787, + 0.007983820512890816, + 0.002918178914114833, + 0.002164720557630062, + 0.01191479992121458, + 0.0032765434589236975, + 0.0043999310582876205, + 0.010006789118051529, + 0.0033234215807169676, + 0.038488052785396576, + 0.005562388803809881, + 0.006091668736189604, + 0.0061536976136267185, + 0.004641327075660229, + 0.046418480575084686, + 0.024008983746170998, + 0.007690129801630974, + 0.015616499818861485, + 0.005600423086434603, + 0.005970298312604427, + 0.008425094187259674, + 0.006032510660588741, + 0.031043512746691704, + 0.002675181720405817, + 0.0016663389978930354, + 0.0017732366686686873, + 0.002041150350123644, + 0.001799197169020772, + 0.0022019874304533005, + 0.001740821753628552, + 0.00221407413482666, + 0.003981808200478554, + 0.005244256928563118, + 0.004149682819843292, + 0.0023950061295181513, + 0.0025286031886935234, + 0.003229334484785795, + 0.0028775972314178944, + 0.007042696699500084 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 1, + 6 + ], + "selected_reward": -0.020571796223521233 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a slow macro zoom-in gliding through crystal-clear turquoise water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are two Ocellaris clownfish visible with bright orange bodies and three distinct white vertical stripes outlined in black?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a single fish positioned directly above the eggs, fanning its pectoral fins in steady, rhythmic motions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene featuring a host anemone with fleshy green tentacles and soft purple tips swaying gently?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does bright sunlight filter through the surface creating dancing caustic patterns that make the translucent amber eggs shimmer?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"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": [ + 0.1398889422416687, + 0.18188895285129547, + 0.05555339157581329, + 0.12435628473758698, + 0.12615589797496796, + 0.12841376662254333, + 0.05858788266777992, + 0.11294393986463547, + 0.02681337669491768, + 0.04539761692285538 + ], + "credit_map": [ + 0.014066803269088268, + 0.01511379424482584, + 0.01425441075116396, + 0.012807591818273067, + 0.02029719389975071, + 0.021302470937371254, + 0.022636983543634415, + 0.019409693777561188, + 0.013541934080421925, + 0.01549548003822565, + 0.02022353559732437, + 0.03373781219124794, + 0.015328087843954563, + 0.015045179054141045, + 0.049738939851522446, + 0.01877797767519951, + 0.005939497146755457, + 0.006114724092185497, + 0.006188756786286831, + 0.006840398069471121, + 0.006491938140243292, + 0.007356858812272549, + 0.01018966268748045, + 0.006431554444134235, + 0.013233807869255543, + 0.008522333577275276, + 0.018726343289017677, + 0.00996137410402298, + 0.008008823730051517, + 0.05018523707985878, + 0.006686840672045946, + 0.009031526744365692, + 0.019857265055179596, + 0.009531986899673939, + 0.009427428245544434, + 0.007710730191320181, + 0.007830427028238773, + 0.040940649807453156, + 0.0071481638588011265, + 0.023709258064627647, + 0.013738106936216354, + 0.004903674125671387, + 0.011018182151019573, + 0.0065369014628231525, + 0.004719207063317299, + 0.048377875238657, + 0.005593580659478903, + 0.033526238054037094, + 0.009805211797356606, + 0.0026992252096533775, + 0.011901559308171272, + 0.004552118014544249, + 0.003111752215772867, + 0.003921179100871086, + 0.0037676929496228695, + 0.018829142674803734, + 0.026471760123968124, + 0.008076619356870651, + 0.015569031238555908, + 0.010219709947705269, + 0.006601949222385883, + 0.008492142893373966, + 0.007906247861683369, + 0.029606478288769722, + 0.004682551138103008, + 0.0030618398450315, + 0.0030590726528316736, + 0.004621745087206364, + 0.0028579363133758307, + 0.002418682910501957, + 0.003266747575253248, + 0.0028448011726140976, + 0.005778432358056307, + 0.009091545827686787, + 0.005253390874713659, + 0.004503641743212938, + 0.005904967896640301, + 0.005559802521020174, + 0.004417574964463711, + 0.004888261202722788 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 5 + ], + "selected_reward": -0.006734057795256376 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "slow, macro zoom-in shot glides through crystal-clear turquoise water, gradually approaching an intimate display of meticulous detail. Focus is on a pair of vibrant orange and white-striped Ocellaris clownfish engaged in parental care. Their bodies are a bright, saturated orange, marked by three distinct white vertical stripes outlined in black and purposefully positioned. They hover over a small group of shimmering, translucent amber eggs, firmly nestled on the algae-covered rocky base of their host anemone. A single fish, positioned directly above, diligently fanns its pectoral fins in steady, rhythmic motions, generating gentle, life-sustaining currents over its delicate claws. The magnificent anemone's surrounding tentacles, fleshy green with soft purple tips, sway gently and hypnotically in the clear water. Bright sunlight filters through the surface, creating a dynamic play of light and shadow and casting dancing caustic patterns that make the individual eggs shimmer. The background blends into a deep blue, soft-focus abyss, enhancing the tranquil and focused atmosphere of this captivating scene of aquatic life.", + "seed": 25671, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000056/group_15/index_4599_seed_25671_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": -0.02791490196250379, + "vlm_logprob": -0.02791490196250379, + "vlm_yes_no_margin": 4.0, + "vlm_token_credit": -0.02791490196250379, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a slow macro zoom-in gliding through crystal-clear turquoise water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are two Ocellaris clownfish visible with bright orange bodies and three distinct white vertical stripes outlined in black?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a single fish positioned directly above the eggs, fanning its pectoral fins in steady, rhythmic motions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene featuring a host anemone with fleshy green tentacles and soft purple tips swaying gently?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does bright sunlight filter through the surface creating dancing caustic patterns that make the translucent amber eggs shimmer?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06381858885288239 + ], + "opposite_logprob_tokens": [ + -2.8138186931610107 + ], + "logprob_reward": -0.06381858885288239, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06381858885288239, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13774539530277252, + 0.5064688324928284, + 0.06747344881296158, + 0.04768991097807884, + 0.09368307143449783, + 0.022614460438489914, + 0.03499791771173477, + 0.020993730053305626, + 0.029339946806430817, + 0.038993287831544876 + ], + "credit_map": [ + 0.011230269446969032, + 0.009345890022814274, + 0.01218799315392971, + 0.03212732449173927, + 0.022802822291851044, + 0.022369757294654846, + 0.014426390640437603, + 0.013254974037408829, + 0.0481787770986557, + 0.03155779466032982, + 0.021266834810376167, + 0.04995187744498253, + 0.21418310701847076, + 0.06399629265069962, + 0.025526855140924454, + 0.051807329058647156, + 0.009099467657506466, + 0.005658769514411688, + 0.003722164547070861, + 0.01586676947772503, + 0.006392276845872402, + 0.012896481901407242, + 0.008672628551721573, + 0.005164895206689835, + 0.004446051083505154, + 0.00582539476454258, + 0.003951037302613258, + 0.008472874760627747, + 0.008442888036370277, + 0.005422645714133978, + 0.006829716730862856, + 0.00429930817335844, + 0.007233737502247095, + 0.010027079842984676, + 0.007714850828051567, + 0.020346811041235924, + 0.012226310558617115, + 0.014927846379578114, + 0.013766313903033733, + 0.0074401358142495155, + 0.001751649659126997, + 0.003273916430771351, + 0.0030149887315928936, + 0.0034504670184105635, + 0.0022496788296848536, + 0.0028412260580807924, + 0.0031152439769357443, + 0.002917291596531868, + 0.002673857845366001, + 0.0032252739183604717, + 0.003917464520782232, + 0.005296868272125721, + 0.005273634102195501, + 0.0036089408677071333, + 0.006538090296089649, + 0.004463790450245142, + 0.0013846223009750247, + 0.002485782140865922, + 0.0017492579063400626, + 0.002833389211446047, + 0.0031136174220591784, + 0.0023368657566607, + 0.004459149669855833, + 0.0026310475077480078, + 0.0020417640917003155, + 0.0025129858404397964, + 0.001916753244586289, + 0.0038375265430659056, + 0.007370653096586466, + 0.0029109984170645475, + 0.006244397256523371, + 0.002504871692508459, + 0.0027991794049739838, + 0.004189817700535059, + 0.0026802178472280502, + 0.00676371157169342, + 0.007876786403357983, + 0.0032896012999117374, + 0.00878691766411066, + 0.002607059897854924 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 4 + ], + "selected_reward": -0.06381858885288239 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a slow macro zoom-in gliding through crystal-clear turquoise water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are two Ocellaris clownfish visible with bright orange bodies and three distinct white vertical stripes outlined in black?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a single fish positioned directly above the eggs, fanning its pectoral fins in steady, rhythmic motions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene featuring a host anemone with fleshy green tentacles and soft purple tips swaying gently?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does bright sunlight filter through the surface creating dancing caustic patterns that make the translucent amber eggs shimmer?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03371838480234146 + ], + "opposite_logprob_tokens": [ + -3.4087183475494385 + ], + "logprob_reward": -0.03371838480234146, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03371838480234146, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.38060665130615234, + 0.16085192561149597, + 0.03354758024215698, + 0.05406579747796059, + 0.05775279551744461, + 0.09723462164402008, + 0.08285900950431824, + 0.04778958857059479, + 0.05408143624663353, + 0.03121059387922287 + ], + "credit_map": [ + 0.02642451785504818, + 0.017247077077627182, + 0.01723043993115425, + 0.0717334970831871, + 0.05420966446399689, + 0.13330689072608948, + 0.040722381323575974, + 0.019732175394892693, + 0.016582239419221878, + 0.009867899119853973, + 0.004871805664151907, + 0.023585133254528046, + 0.07046224176883698, + 0.0216878242790699, + 0.008747249841690063, + 0.005047534126788378, + 0.003933901432901621, + 0.0029117087833583355, + 0.0020250603556632996, + 0.008503545075654984, + 0.0031630476005375385, + 0.004517319146543741, + 0.005829791072756052, + 0.002663204912096262, + 0.005108140874654055, + 0.0072256759740412235, + 0.004504697863012552, + 0.006827529985457659, + 0.006814796477556229, + 0.00855043064802885, + 0.006764501333236694, + 0.008270022459328175, + 0.004382136277854443, + 0.007065571844577789, + 0.005703178234398365, + 0.007169470191001892, + 0.01020984910428524, + 0.0057538170367479324, + 0.009016618132591248, + 0.008452151902019978, + 0.007730043493211269, + 0.012574920430779457, + 0.00647077988833189, + 0.016244802623987198, + 0.015697650611400604, + 0.011265610344707966, + 0.019765496253967285, + 0.0074853147380054, + 0.006209459155797958, + 0.00847986526787281, + 0.006169130560010672, + 0.013553541153669357, + 0.019125765189528465, + 0.006947980262339115, + 0.015517565421760082, + 0.006855697836726904, + 0.0032883957028388977, + 0.004620115272700787, + 0.003778827842324972, + 0.005734877660870552, + 0.011259926483035088, + 0.004638241603970528, + 0.01039391290396452, + 0.0040752943605184555, + 0.005323471967130899, + 0.006849116180092096, + 0.005090833175927401, + 0.0071218679659068584, + 0.00726077426224947, + 0.006563854869455099, + 0.010259818285703659, + 0.005611698143184185, + 0.0026090831961482763, + 0.0027743875980377197, + 0.002510560443624854, + 0.005077475216239691, + 0.005991063080728054, + 0.002771176164969802, + 0.0073153181001544, + 0.0021615291479974985 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 5 + ], + "selected_reward": -0.03371838480234146 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a slow macro zoom-in gliding through crystal-clear turquoise water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are two Ocellaris clownfish visible with bright orange bodies and three distinct white vertical stripes outlined in black?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a single fish positioned directly above the eggs, fanning its pectoral fins in steady, rhythmic motions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene featuring a host anemone with fleshy green tentacles and soft purple tips swaying gently?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does bright sunlight filter through the surface creating dancing caustic patterns that make the translucent amber eggs shimmer?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"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": [ + 0.1549226939678192, + 0.5264687538146973, + 0.015792110934853554, + 0.05019363388419151, + 0.07560799270868301, + 0.01851966232061386, + 0.05688442289829254, + 0.03926611319184303, + 0.03643040359020233, + 0.025914231315255165 + ], + "credit_map": [ + 0.012035724706947803, + 0.007873454131186008, + 0.007566593121737242, + 0.03230952098965645, + 0.02019614540040493, + 0.05370066687464714, + 0.013063573278486729, + 0.008177014999091625, + 0.0465611070394516, + 0.03130928426980972, + 0.019161520525813103, + 0.05638297647237778, + 0.22333848476409912, + 0.07368289679288864, + 0.026136256754398346, + 0.049896229058504105, + 0.0021537079010158777, + 0.0011525554582476616, + 0.0010709589114412665, + 0.0031576664187014103, + 0.0016442181076854467, + 0.0028819353319704533, + 0.002410846995189786, + 0.0013202194822952151, + 0.00584526639431715, + 0.005037290975451469, + 0.003877580864354968, + 0.010194091126322746, + 0.005515551194548607, + 0.004864853341132402, + 0.009066903032362461, + 0.005792092066258192, + 0.006747774314135313, + 0.007948390208184719, + 0.006828860379755497, + 0.015066619031131268, + 0.009786086156964302, + 0.00924012903124094, + 0.010867669247090816, + 0.009122463874518871, + 0.0015167762758210301, + 0.002710512839257717, + 0.0016548777930438519, + 0.002895255805924535, + 0.002465086290612817, + 0.002306773094460368, + 0.0025326525792479515, + 0.002437727525830269, + 0.0042349062860012054, + 0.005737980827689171, + 0.004473455715924501, + 0.0071132369339466095, + 0.013565605506300926, + 0.005364023614674807, + 0.01123403012752533, + 0.005161182023584843, + 0.0030118273571133614, + 0.004710841923952103, + 0.002442175755277276, + 0.0055572339333593845, + 0.007422447204589844, + 0.003385538700968027, + 0.00920553132891655, + 0.0035305151250213385, + 0.001980811357498169, + 0.0039455099031329155, + 0.003037832211703062, + 0.0056426674127578735, + 0.006289721932262182, + 0.0028113957960158587, + 0.009996293112635612, + 0.002726170467212796, + 0.0020721491891890764, + 0.002449058461934328, + 0.004133131355047226, + 0.0035064266994595528, + 0.004154989030212164, + 0.0022978249471634626, + 0.003957634791731834, + 0.0033430145122110844 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 4 + ], + "selected_reward": -0.029793808236718178 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a slow macro zoom-in gliding through crystal-clear turquoise water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are two Ocellaris clownfish visible with bright orange bodies and three distinct white vertical stripes outlined in black?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a single fish positioned directly above the eggs, fanning its pectoral fins in steady, rhythmic motions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene featuring a host anemone with fleshy green tentacles and soft purple tips swaying gently?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does bright sunlight filter through the surface creating dancing caustic patterns that make the translucent amber eggs shimmer?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008631067350506783 + ], + "opposite_logprob_tokens": [ + -4.758631229400635 + ], + "logprob_reward": -0.008631067350506783, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008631067350506783, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10777594894170761, + 0.5210603475570679, + 0.03318368270993233, + 0.027127113193273544, + 0.02850351668894291, + 0.08113409578800201, + 0.060468610376119614, + 0.04770959913730621, + 0.05169247090816498, + 0.04134465754032135 + ], + "credit_map": [ + 0.010643444024026394, + 0.008138200268149376, + 0.008084611035883427, + 0.015399145893752575, + 0.012766629457473755, + 0.03648228570818901, + 0.007403368595987558, + 0.008858262561261654, + 0.051826607435941696, + 0.035402409732341766, + 0.02652258798480034, + 0.0359220877289772, + 0.21332767605781555, + 0.06645700335502625, + 0.03135548532009125, + 0.06024647504091263, + 0.003281639888882637, + 0.004617826547473669, + 0.0031419694423675537, + 0.004617992788553238, + 0.004212190862745047, + 0.005599324591457844, + 0.00308908149600029, + 0.004623654298484325, + 0.0027682450599968433, + 0.002764670643955469, + 0.002518516266718507, + 0.004792625550180674, + 0.004135199822485447, + 0.003335032146424055, + 0.004665857180953026, + 0.0021469644270837307, + 0.002680353354662657, + 0.00394123699516058, + 0.002274627098813653, + 0.003413101425394416, + 0.004551466088742018, + 0.0045293522998690605, + 0.004063799977302551, + 0.0030495799146592617, + 0.006275198422372341, + 0.013329262845218182, + 0.008836637251079082, + 0.0120774507522583, + 0.008104858919978142, + 0.011197629384696484, + 0.010729605332016945, + 0.010583457536995411, + 0.004616736900061369, + 0.007375472225248814, + 0.006260885391384363, + 0.0077704694122076035, + 0.010995268821716309, + 0.007547624409198761, + 0.010137088596820831, + 0.005765065550804138, + 0.0037278032395988703, + 0.005240006372332573, + 0.00409749336540699, + 0.006648968439549208, + 0.007293089758604765, + 0.004673622082918882, + 0.010080548003315926, + 0.005948066245764494, + 0.004193013999611139, + 0.005756776314228773, + 0.004323266446590424, + 0.0061383736319839954, + 0.010968906804919243, + 0.006137982476502657, + 0.009073033928871155, + 0.0051011135801672935, + 0.0030052640940994024, + 0.004841803573071957, + 0.004067182540893555, + 0.006253054831176996, + 0.006749902386218309, + 0.003934057429432869, + 0.009039017371833324, + 0.0034543746151030064 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 5 + ], + "selected_reward": -0.008631067350506783 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a slow macro zoom-in gliding through crystal-clear turquoise water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are two Ocellaris clownfish visible with bright orange bodies and three distinct white vertical stripes outlined in black?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a single fish positioned directly above the eggs, fanning its pectoral fins in steady, rhythmic motions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene featuring a host anemone with fleshy green tentacles and soft purple tips swaying gently?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does bright sunlight filter through the surface creating dancing caustic patterns that make the translucent amber eggs shimmer?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0036126605700701475 + ], + "opposite_logprob_tokens": [ + -5.628612518310547 + ], + "logprob_reward": -0.0036126605700701475, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.0036126605700701475, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.29024451971054077, + 0.45956534147262573, + 0.09845568239688873, + 0.022895539179444313, + 0.031230151653289795, + 0.027320681139826775, + 0.020979728549718857, + 0.015448623336851597, + 0.0227326862514019, + 0.011127054691314697 + ], + "credit_map": [ + 0.02157478965818882, + 0.012221763841807842, + 0.01218172162771225, + 0.048938773572444916, + 0.05072327330708504, + 0.11190183460712433, + 0.018088964745402336, + 0.01461341604590416, + 0.04523199796676636, + 0.03149225190281868, + 0.022379782050848007, + 0.03163176402449608, + 0.18905583024024963, + 0.0596868135035038, + 0.027490796521306038, + 0.05259612575173378, + 0.013392169959843159, + 0.007762807421386242, + 0.005581406410783529, + 0.026335688307881355, + 0.00970036257058382, + 0.014313443563878536, + 0.014024377800524235, + 0.007345431484282017, + 0.0025138105265796185, + 0.0017149552004411817, + 0.0022843526676297188, + 0.0033505528699606657, + 0.0030411162879318, + 0.002348104026168585, + 0.0031932697165757418, + 0.004449377302080393, + 0.0025823977775871754, + 0.003427067305892706, + 0.0020202372688800097, + 0.006209562998265028, + 0.0035971193574368954, + 0.0063944305293262005, + 0.004268278367817402, + 0.002731058280915022, + 0.002144184662029147, + 0.0033000630792230368, + 0.0028016939759254456, + 0.00535078439861536, + 0.003235571552067995, + 0.0035682914312928915, + 0.004086761735379696, + 0.0028333307709544897, + 0.002062693703919649, + 0.0020715605933219194, + 0.0025380076840519905, + 0.003185397945344448, + 0.0030846945010125637, + 0.002215755172073841, + 0.0033087448682636023, + 0.0025128747802227736, + 0.0009384594159200788, + 0.001682052738033235, + 0.001338062109425664, + 0.002351987175643444, + 0.0019214068306609988, + 0.0014345262898132205, + 0.004017271567136049, + 0.0017648582579568028, + 0.0018094904953613877, + 0.0017363918013870716, + 0.0019965381361544132, + 0.002924948465079069, + 0.004898033104836941, + 0.0024201825726777315, + 0.004405676387250423, + 0.0025414247065782547, + 0.0008795856265351176, + 0.0012363161658868194, + 0.0009789946489036083, + 0.0017621503211557865, + 0.002035808051005006, + 0.0010894215665757656, + 0.0022870227694511414, + 0.0008577560656704009 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 2 + ], + "selected_reward": -0.0036126605700701475 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "slow, macro zoom-in shot glides through crystal-clear turquoise water, gradually approaching an intimate display of meticulous detail. Focus is on a pair of vibrant orange and white-striped Ocellaris clownfish engaged in parental care. Their bodies are a bright, saturated orange, marked by three distinct white vertical stripes outlined in black and purposefully positioned. They hover over a small group of shimmering, translucent amber eggs, firmly nestled on the algae-covered rocky base of their host anemone. A single fish, positioned directly above, diligently fanns its pectoral fins in steady, rhythmic motions, generating gentle, life-sustaining currents over its delicate claws. The magnificent anemone's surrounding tentacles, fleshy green with soft purple tips, sway gently and hypnotically in the clear water. Bright sunlight filters through the surface, creating a dynamic play of light and shadow and casting dancing caustic patterns that make the individual eggs shimmer. The background blends into a deep blue, soft-focus abyss, enhancing the tranquil and focused atmosphere of this captivating scene of aquatic life.", + "seed": 25672, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000056/group_15/index_4599_seed_25672_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": -0.02381310174241662, + "vlm_logprob": -0.02381310174241662, + "vlm_yes_no_margin": 4.0, + "vlm_token_credit": -0.02381310174241662, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a slow macro zoom-in gliding through crystal-clear turquoise water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are two Ocellaris clownfish visible with bright orange bodies and three distinct white vertical stripes outlined in black?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a single fish positioned directly above the eggs, fanning its pectoral fins in steady, rhythmic motions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene featuring a host anemone with fleshy green tentacles and soft purple tips swaying gently?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does bright sunlight filter through the surface creating dancing caustic patterns that make the translucent amber eggs shimmer?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03595853969454765 + ], + "opposite_logprob_tokens": [ + -3.4109585285186768 + ], + "logprob_reward": -0.03595853969454765, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03595853969454765, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20344795286655426, + 0.32942137122154236, + 0.08542932569980621, + 0.054305531084537506, + 0.018003400415182114, + 0.13828490674495697, + 0.06529321521520615, + 0.03857873007655144, + 0.038047369569540024, + 0.02918812818825245 + ], + "credit_map": [ + 0.01189293060451746, + 0.018199803307652473, + 0.012210293672978878, + 0.05256197229027748, + 0.013245539739727974, + 0.02185702510178089, + 0.02556918002665043, + 0.047911208122968674, + 0.02172686904668808, + 0.05350293964147568, + 0.03111867979168892, + 0.05389881134033203, + 0.03944423049688339, + 0.030683930963277817, + 0.05858800187706947, + 0.04045791178941727, + 0.008732647635042667, + 0.00800230074673891, + 0.005400489550083876, + 0.01829441264271736, + 0.0064212786965072155, + 0.01805698685348034, + 0.013263430446386337, + 0.007257775869220495, + 0.003994336351752281, + 0.00823179166764021, + 0.005529606249183416, + 0.008103176020085812, + 0.005931448191404343, + 0.004892780911177397, + 0.00842640083283186, + 0.009195992723107338, + 0.0010957227787002921, + 0.0017947013257071376, + 0.0013359935255721211, + 0.0047041079960763454, + 0.0010094402823597193, + 0.0012413709191605449, + 0.0016685323789715767, + 0.005153532139956951, + 0.011563722975552082, + 0.018349308520555496, + 0.013748863711953163, + 0.01686246320605278, + 0.022471603006124496, + 0.011823257431387901, + 0.018550503998994827, + 0.024915194138884544, + 0.007789654191583395, + 0.010621280409395695, + 0.00847680028527975, + 0.005833130329847336, + 0.006329866126179695, + 0.005710574798285961, + 0.008721202611923218, + 0.01181070413440466, + 0.0029676640406250954, + 0.004700550343841314, + 0.004042126704007387, + 0.004745895508676767, + 0.00479561323300004, + 0.004443247802555561, + 0.004940787795931101, + 0.007942846044898033, + 0.0038660175632685423, + 0.005542743019759655, + 0.003395341569557786, + 0.0041363779455423355, + 0.005004470702260733, + 0.0054268604144454, + 0.00438833748921752, + 0.0062872241251170635, + 0.002397012198343873, + 0.006530881859362125, + 0.0031700495164841413, + 0.0028017559088766575, + 0.003363477298989892, + 0.005246322136372328, + 0.0024783879052847624, + 0.003200240433216095 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 5 + ], + "selected_reward": -0.03595853969454765 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a slow macro zoom-in gliding through crystal-clear turquoise water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are two Ocellaris clownfish visible with bright orange bodies and three distinct white vertical stripes outlined in black?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a single fish positioned directly above the eggs, fanning its pectoral fins in steady, rhythmic motions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene featuring a host anemone with fleshy green tentacles and soft purple tips swaying gently?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does bright sunlight filter through the surface creating dancing caustic patterns that make the translucent amber eggs shimmer?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011088889092206955 + ], + "opposite_logprob_tokens": [ + -4.511088848114014 + ], + "logprob_reward": -0.011088889092206955, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011088889092206955, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.32308200001716614, + 0.16353611648082733, + 0.08355629444122314, + 0.020554738119244576, + 0.01257251389324665, + 0.07448988407850266, + 0.11652757972478867, + 0.09011026471853256, + 0.056007079780101776, + 0.05956359580159187 + ], + "credit_map": [ + 0.018773959949612617, + 0.033948469907045364, + 0.019248250871896744, + 0.08153674006462097, + 0.02353566512465477, + 0.033590223640203476, + 0.04696139320731163, + 0.06548728048801422, + 0.011763767339289188, + 0.029644355177879333, + 0.01597687415778637, + 0.026228664442896843, + 0.01899084448814392, + 0.01499666552990675, + 0.025899384170770645, + 0.020035559311509132, + 0.010586028918623924, + 0.006364848464727402, + 0.004380485508590937, + 0.010879232548177242, + 0.006874353624880314, + 0.02102186530828476, + 0.012145829387009144, + 0.011303654871881008, + 0.0011495904764160514, + 0.004196211230009794, + 0.001993767451494932, + 0.003270139219239354, + 0.002615797333419323, + 0.002044284949079156, + 0.003014229005202651, + 0.002270718337967992, + 0.0009138777386397123, + 0.0019053880823776126, + 0.0016154691111296415, + 0.0021844094153493643, + 0.0010918286861851811, + 0.0014401093358173966, + 0.0016427469672635198, + 0.0017786851385608315, + 0.005075566004961729, + 0.01227821409702301, + 0.007641549687832594, + 0.00971921719610691, + 0.0068892366252839565, + 0.00695456936955452, + 0.01341933012008667, + 0.012512199580669403, + 0.008675811812281609, + 0.01798452064394951, + 0.012356957420706749, + 0.01481756940484047, + 0.014326328411698341, + 0.012309794314205647, + 0.015384807251393795, + 0.020671796053647995, + 0.006812267936766148, + 0.020992062985897064, + 0.005387437529861927, + 0.009282884187996387, + 0.011351346969604492, + 0.008615533821284771, + 0.017155278474092484, + 0.01051345281302929, + 0.004610131029039621, + 0.008466440252959728, + 0.0043115015141665936, + 0.007090453989803791, + 0.009376820176839828, + 0.005161135457456112, + 0.010277658700942993, + 0.006712936796247959, + 0.004969973582774401, + 0.010730531997978687, + 0.005604370031505823, + 0.005118452478200197, + 0.007536169141530991, + 0.013342072255909443, + 0.007189020048826933, + 0.00507300766184926 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 6 + ], + "selected_reward": -0.011088889092206955 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a slow macro zoom-in gliding through crystal-clear turquoise water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are two Ocellaris clownfish visible with bright orange bodies and three distinct white vertical stripes outlined in black?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a single fish positioned directly above the eggs, fanning its pectoral fins in steady, rhythmic motions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene featuring a host anemone with fleshy green tentacles and soft purple tips swaying gently?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does bright sunlight filter through the surface creating dancing caustic patterns that make the translucent amber eggs shimmer?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03369268402457237 + ], + "opposite_logprob_tokens": [ + -3.4086925983428955 + ], + "logprob_reward": -0.03369268402457237, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03369268402457237, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3447244465351105, + 0.13364095985889435, + 0.11592404544353485, + 0.022454386577010155, + 0.024007121101021767, + 0.1299162656068802, + 0.050110816955566406, + 0.039728887379169464, + 0.09614279866218567, + 0.043350327759981155 + ], + "credit_map": [ + 0.021617524325847626, + 0.03749958053231239, + 0.019252313300967216, + 0.09176476299762726, + 0.023372594267129898, + 0.03164754807949066, + 0.05390917509794235, + 0.06566092371940613, + 0.009102964773774147, + 0.022810978814959526, + 0.013852222822606564, + 0.021537793800234795, + 0.016024557873606682, + 0.01288717519491911, + 0.02119038812816143, + 0.016234861686825752, + 0.008067375980317593, + 0.007579050026834011, + 0.004941991064697504, + 0.013964996673166752, + 0.008131034672260284, + 0.02472548931837082, + 0.021251201629638672, + 0.027262892574071884, + 0.0013002606574445963, + 0.00444116210564971, + 0.001671557780355215, + 0.00410766527056694, + 0.0023908379953354597, + 0.0018960494780912995, + 0.003436659462749958, + 0.003210191149264574, + 0.001669674296863377, + 0.0030192406848073006, + 0.001767392153851688, + 0.004765408579260111, + 0.002932013012468815, + 0.002151638036593795, + 0.0037583971861749887, + 0.003943353425711393, + 0.00734195439144969, + 0.0175629872828722, + 0.01757628470659256, + 0.018171997740864754, + 0.021358221769332886, + 0.010504763573408127, + 0.018664062023162842, + 0.018735969439148903, + 0.003650384023785591, + 0.008386154659092426, + 0.0048827799037098885, + 0.008077376522123814, + 0.0051364717073738575, + 0.0042941756546497345, + 0.008902606554329395, + 0.00678086094558239, + 0.0032111380714923143, + 0.006833389401435852, + 0.0035372197162359953, + 0.005474776960909367, + 0.004379057791084051, + 0.004020386841148138, + 0.0054981824941933155, + 0.006774731911718845, + 0.008797060698270798, + 0.015816109254956245, + 0.011545402929186821, + 0.010044402442872524, + 0.013557069934904575, + 0.010363400913774967, + 0.018624959513545036, + 0.0073943790048360825, + 0.004234463907778263, + 0.006000392138957977, + 0.005425967741757631, + 0.003981858026236296, + 0.0041465130634605885, + 0.010462213307619095, + 0.004115358926355839, + 0.00498355645686388 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 5 + ], + "selected_reward": -0.03369268402457237 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a slow macro zoom-in gliding through crystal-clear turquoise water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are two Ocellaris clownfish visible with bright orange bodies and three distinct white vertical stripes outlined in black?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a single fish positioned directly above the eggs, fanning its pectoral fins in steady, rhythmic motions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene featuring a host anemone with fleshy green tentacles and soft purple tips swaying gently?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does bright sunlight filter through the surface creating dancing caustic patterns that make the translucent amber eggs shimmer?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.033687956631183624 + ], + "opposite_logprob_tokens": [ + -3.4086880683898926 + ], + "logprob_reward": -0.033687956631183624, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.033687956631183624, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.6833611130714417, + 0.08032882958650589, + 0.03538241237401962, + 0.017330152913928032, + 0.012660762295126915, + 0.04103085771203041, + 0.024826722219586372, + 0.0442243255674839, + 0.03569979593157768, + 0.025154944509267807 + ], + "credit_map": [ + 0.04233913496136665, + 0.07653395086526871, + 0.03833474963903427, + 0.16714945435523987, + 0.04902099072933197, + 0.06918565183877945, + 0.1096344143152237, + 0.1311628222465515, + 0.005818682722747326, + 0.014155601151287556, + 0.008226478472352028, + 0.01177418977022171, + 0.009547511115670204, + 0.008029157295823097, + 0.012015336193144321, + 0.010761880315840244, + 0.0018919871654361486, + 0.002824890660122037, + 0.0017407345585525036, + 0.011585360392928123, + 0.0018773293122649193, + 0.004089050926268101, + 0.0085686556994915, + 0.002804406685754657, + 0.0013712249929085374, + 0.0017007797723636031, + 0.0018433007644489408, + 0.0032728847581893206, + 0.0012225386453792453, + 0.0019304042216390371, + 0.0019044858636334538, + 0.004084537271410227, + 0.0007248225156217813, + 0.001240753335878253, + 0.0009384452132508159, + 0.002910532057285309, + 0.001302700606174767, + 0.0009932236280292273, + 0.0020494197960942984, + 0.002500867238268256, + 0.0029504396952688694, + 0.005579074844717979, + 0.006866276264190674, + 0.006062912289053202, + 0.0046200379729270935, + 0.003958297427743673, + 0.005377139896154404, + 0.005616684444248676, + 0.0020272033289074898, + 0.004523195326328278, + 0.0027014922816306353, + 0.0037309492472559214, + 0.0018352725310251117, + 0.002412372035905719, + 0.0031128216069191694, + 0.004483419470489025, + 0.006192962173372507, + 0.005856362171471119, + 0.005964319221675396, + 0.006159478332847357, + 0.0030197182204574347, + 0.006278563756495714, + 0.00456302659586072, + 0.006189899053424597, + 0.0033836867660284042, + 0.004987084772437811, + 0.0038697076961398125, + 0.00317763932980597, + 0.008580176159739494, + 0.0034863390028476715, + 0.004791800864040852, + 0.0034233643673360348, + 0.0020780768245458603, + 0.0048867445439100266, + 0.003025254001840949, + 0.002406647428870201, + 0.0027301183436065912, + 0.00457508722320199, + 0.0028800375293940306, + 0.0025729823391884565 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 7 + ], + "selected_reward": -0.033687956631183624 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a slow macro zoom-in gliding through crystal-clear turquoise water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are two Ocellaris clownfish visible with bright orange bodies and three distinct white vertical stripes outlined in black?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a single fish positioned directly above the eggs, fanning its pectoral fins in steady, rhythmic motions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene featuring a host anemone with fleshy green tentacles and soft purple tips swaying gently?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does bright sunlight filter through the surface creating dancing caustic patterns that make the translucent amber eggs shimmer?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"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": [ + 0.2076612412929535, + 0.16885150969028473, + 0.06167205050587654, + 0.049208760261535645, + 0.03289605304598808, + 0.18791238963603973, + 0.08129797875881195, + 0.051232170313596725, + 0.038365498185157776, + 0.12090232968330383 + ], + "credit_map": [ + 0.01375510636717081, + 0.020463770255446434, + 0.014747736044228077, + 0.04151245951652527, + 0.0177551731467247, + 0.012077213265001774, + 0.046387240290641785, + 0.04096255078911781, + 0.013011977076530457, + 0.028932085260748863, + 0.018285950645804405, + 0.029691826552152634, + 0.01726635918021202, + 0.01506175845861435, + 0.023742854595184326, + 0.02285870723426342, + 0.008990062400698662, + 0.0067907036282122135, + 0.003336670808494091, + 0.010124456137418747, + 0.004187745042145252, + 0.011953615583479404, + 0.006389997899532318, + 0.009898796677589417, + 0.00272270361892879, + 0.01179476547986269, + 0.0027794588822871447, + 0.011008152738213539, + 0.0036775320768356323, + 0.0033167654182761908, + 0.006544360890984535, + 0.007365020923316479, + 0.002015484729781747, + 0.004105965606868267, + 0.003065802389755845, + 0.006780010182410479, + 0.0023649667855352163, + 0.0031882154289633036, + 0.003754020668566227, + 0.0076215872541069984, + 0.014588710851967335, + 0.023307103663682938, + 0.021968448534607887, + 0.023011600598692894, + 0.03125997632741928, + 0.016458235681056976, + 0.026332316920161247, + 0.03098599798977375, + 0.008030466735363007, + 0.013585801236331463, + 0.01019380521029234, + 0.00849895179271698, + 0.0061437818221747875, + 0.007346805650740862, + 0.013776610605418682, + 0.013721761293709278, + 0.004572716075927019, + 0.01005940418690443, + 0.004184318706393242, + 0.005672162864357233, + 0.007483573630452156, + 0.0053653959184885025, + 0.007744139060378075, + 0.00615046126767993, + 0.0038995174691081047, + 0.004836362786591053, + 0.004471655003726482, + 0.005295769777148962, + 0.005475983954966068, + 0.003855096409097314, + 0.004578300751745701, + 0.005952812265604734, + 0.010556504130363464, + 0.02193501777946949, + 0.012139544822275639, + 0.010312945581972599, + 0.010387574322521687, + 0.030065275728702545, + 0.010747987776994705, + 0.014757486060261726 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 1 + ], + "selected_reward": -0.004637439269572496 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "slow, macro zoom-in shot glides through crystal-clear turquoise water, gradually approaching an intimate display of meticulous detail. Focus is on a pair of vibrant orange and white-striped Ocellaris clownfish engaged in parental care. Their bodies are a bright, saturated orange, marked by three distinct white vertical stripes outlined in black and purposefully positioned. They hover over a small group of shimmering, translucent amber eggs, firmly nestled on the algae-covered rocky base of their host anemone. A single fish, positioned directly above, diligently fanns its pectoral fins in steady, rhythmic motions, generating gentle, life-sustaining currents over its delicate claws. The magnificent anemone's surrounding tentacles, fleshy green with soft purple tips, sway gently and hypnotically in the clear water. Bright sunlight filters through the surface, creating a dynamic play of light and shadow and casting dancing caustic patterns that make the individual eggs shimmer. The background blends into a deep blue, soft-focus abyss, enhancing the tranquil and focused atmosphere of this captivating scene of aquatic life.", + "seed": 25673, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000056/group_15/index_4599_seed_25673_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": -0.04111365498974919, + "vlm_logprob": -0.04111365498974919, + "vlm_yes_no_margin": 3.574999952316284, + "vlm_token_credit": -0.04111365498974919, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a slow macro zoom-in gliding through crystal-clear turquoise water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are two Ocellaris clownfish visible with bright orange bodies and three distinct white vertical stripes outlined in black?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a single fish positioned directly above the eggs, fanning its pectoral fins in steady, rhythmic motions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene featuring a host anemone with fleshy green tentacles and soft purple tips swaying gently?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does bright sunlight filter through the surface creating dancing caustic patterns that make the translucent amber eggs shimmer?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02017325721681118 + ], + "opposite_logprob_tokens": [ + -4.020173072814941 + ], + "logprob_reward": -0.02017325721681118, + "yes_no_margin_reward": 3.999999761581421, + "token_credit_reward": -0.02017325721681118, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2489883303642273, + 0.038197606801986694, + 0.06136307865381241, + 0.017088787630200386, + 0.054517149925231934, + 0.04165590927004814, + 0.0209492314606905, + 0.0976436659693718, + 0.3259851336479187, + 0.09361106157302856 + ], + "credit_map": [ + 0.030843650922179222, + 0.05890173092484474, + 0.013385695405304432, + 0.03055237978696823, + 0.014738280326128006, + 0.025495925918221474, + 0.05365130677819252, + 0.021419337019324303, + 0.00304676266387105, + 0.0019906654488295317, + 0.003064897609874606, + 0.006542867515236139, + 0.0030115507543087006, + 0.0030784078408032656, + 0.010185085237026215, + 0.007277366239577532, + 0.003276756266131997, + 0.002772314241155982, + 0.0031312676146626472, + 0.0088806776329875, + 0.003685691626742482, + 0.003243781626224518, + 0.019328927621245384, + 0.01704365573823452, + 0.0013250436168164015, + 0.0013101594522595406, + 0.002923010615631938, + 0.0025268541648983955, + 0.0024379543028771877, + 0.001848556217737496, + 0.0026259676087647676, + 0.00209124106913805, + 0.0036598951555788517, + 0.0035778447054326534, + 0.002281363820657134, + 0.01799778826534748, + 0.00320003111846745, + 0.0026802578940987587, + 0.009101520292460918, + 0.012018444016575813, + 0.005130745004862547, + 0.006296470761299133, + 0.004196709021925926, + 0.005306060891598463, + 0.004633848089724779, + 0.004039517603814602, + 0.006589997094124556, + 0.005462558474391699, + 0.0025668195448815823, + 0.0021062749437987804, + 0.0019415633287280798, + 0.0029985029250383377, + 0.0020713789854198694, + 0.0020806966349482536, + 0.003092068014666438, + 0.004091924987733364, + 0.00821650493890047, + 0.008253198117017746, + 0.00869449321180582, + 0.01936633139848709, + 0.017772987484931946, + 0.00897511001676321, + 0.01277745608240366, + 0.013587582856416702, + 0.022381393238902092, + 0.017968351021409035, + 0.016224531456828117, + 0.1210094690322876, + 0.012545143254101276, + 0.0851740688085556, + 0.021199475973844528, + 0.029482677578926086, + 0.01252669095993042, + 0.01069146953523159, + 0.01044443715363741, + 0.016480162739753723, + 0.01182052306830883, + 0.01673371158540249, + 0.008129608817398548, + 0.006784460041671991 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 7 + ], + "selected_reward": -0.02017325721681118 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a slow macro zoom-in gliding through crystal-clear turquoise water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are two Ocellaris clownfish visible with bright orange bodies and three distinct white vertical stripes outlined in black?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a single fish positioned directly above the eggs, fanning its pectoral fins in steady, rhythmic motions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene featuring a host anemone with fleshy green tentacles and soft purple tips swaying gently?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does bright sunlight filter through the surface creating dancing caustic patterns that make the translucent amber eggs shimmer?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04307881370186806 + ], + "opposite_logprob_tokens": [ + -3.168078899383545 + ], + "logprob_reward": -0.04307881370186806, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04307881370186806, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2773423194885254, + 0.16581429541110992, + 0.04895234480500221, + 0.021514788269996643, + 0.013639221899211407, + 0.04553117975592613, + 0.030475186184048653, + 0.06998530775308609, + 0.09983513504266739, + 0.22691018879413605 + ], + "credit_map": [ + 0.03484395518898964, + 0.06376436352729797, + 0.01116705872118473, + 0.031107239425182343, + 0.016512928530573845, + 0.016707811504602432, + 0.076459139585495, + 0.026779847219586372, + 0.01070292666554451, + 0.0057576741091907024, + 0.0071808756329119205, + 0.023305106908082962, + 0.009226324036717415, + 0.008128768764436245, + 0.055092837661504745, + 0.04641980677843094, + 0.002478641225025058, + 0.0027835830114781857, + 0.00448884442448616, + 0.011390659026801586, + 0.003956709522753954, + 0.004430950619280338, + 0.013263603672385216, + 0.006159357726573944, + 0.002154978923499584, + 0.0021079466678202152, + 0.002776648383587599, + 0.0030327411368489265, + 0.00212886487133801, + 0.002768081845715642, + 0.004343128297477961, + 0.002202398842200637, + 0.0020542938727885485, + 0.0018083814065903425, + 0.00132531498093158, + 0.002568619791418314, + 0.0011770743876695633, + 0.0010694031370803714, + 0.001991748344153166, + 0.0016443877248093486, + 0.004598295781761408, + 0.009257044643163681, + 0.00380725204013288, + 0.007179096806794405, + 0.005225653760135174, + 0.0035960653331130743, + 0.0063402955420315266, + 0.005527474917471409, + 0.003187835216522217, + 0.002729187486693263, + 0.002921829931437969, + 0.003712317207828164, + 0.0032268688082695007, + 0.0046277581714093685, + 0.006258708890527487, + 0.003810683498159051, + 0.005443695466965437, + 0.006548845674842596, + 0.006075546145439148, + 0.010435700416564941, + 0.012936893850564957, + 0.00731898657977581, + 0.01153695397078991, + 0.009688697755336761, + 0.007691497448831797, + 0.005964938551187515, + 0.005553063470870256, + 0.03618646785616875, + 0.0038480423390865326, + 0.021486202254891396, + 0.006034419406205416, + 0.013070505112409592, + 0.014641360379755497, + 0.02591525949537754, + 0.02092023752629757, + 0.0480167530477047, + 0.04084111750125885, + 0.023101087659597397, + 0.030621333047747612, + 0.02285304293036461 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 1 + ], + "selected_reward": -0.04307881370186806 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a slow macro zoom-in gliding through crystal-clear turquoise water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are two Ocellaris clownfish visible with bright orange bodies and three distinct white vertical stripes outlined in black?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a single fish positioned directly above the eggs, fanning its pectoral fins in steady, rhythmic motions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene featuring a host anemone with fleshy green tentacles and soft purple tips swaying gently?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does bright sunlight filter through the surface creating dancing caustic patterns that make the translucent amber eggs shimmer?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.048648420721292496 + ], + "opposite_logprob_tokens": [ + -3.0486483573913574 + ], + "logprob_reward": -0.048648420721292496, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.048648420721292496, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23340199887752533, + 0.16339921951293945, + 0.05900176987051964, + 0.02500181272625923, + 0.030192825943231583, + 0.03463619202375412, + 0.08090002834796906, + 0.14438043534755707, + 0.13008099794387817, + 0.09900470077991486 + ], + "credit_map": [ + 0.0273303110152483, + 0.0472547747194767, + 0.010179517790675163, + 0.02932610735297203, + 0.01290902215987444, + 0.014666067436337471, + 0.06741093844175339, + 0.024325259029865265, + 0.010649080388247967, + 0.006322003435343504, + 0.007494283374398947, + 0.023088352754712105, + 0.008471441455185413, + 0.006763374898582697, + 0.054243333637714386, + 0.04636735841631889, + 0.003766513429582119, + 0.0042257895693182945, + 0.00543339503929019, + 0.00846023391932249, + 0.005623573902994394, + 0.008653493598103523, + 0.011731632053852081, + 0.011107143014669418, + 0.0023451398592442274, + 0.002357403514906764, + 0.0035593663342297077, + 0.003066758159548044, + 0.002712630433961749, + 0.003104269038885832, + 0.004683544859290123, + 0.003172700060531497, + 0.0034662827383726835, + 0.003100314876064658, + 0.0017148058395832777, + 0.0076778363436460495, + 0.002652785973623395, + 0.002052782569080591, + 0.0042029195465147495, + 0.005325096193701029, + 0.003955863881856203, + 0.005988355726003647, + 0.0030474571976810694, + 0.00465160608291626, + 0.0033474431838840246, + 0.0028237267397344112, + 0.005305539816617966, + 0.00551619753241539, + 0.008884038776159286, + 0.010095112025737762, + 0.009394369088113308, + 0.012156151235103607, + 0.009059702046215534, + 0.010357401333749294, + 0.01121184416115284, + 0.009741407819092274, + 0.013302434235811234, + 0.012875779531896114, + 0.01406418439000845, + 0.024108389392495155, + 0.028743823990225792, + 0.012350521050393581, + 0.022784549742937088, + 0.01615075394511223, + 0.008423029445111752, + 0.008051450364291668, + 0.008104745298624039, + 0.04170013591647148, + 0.006144143640995026, + 0.026155654340982437, + 0.012417672201991081, + 0.01908416673541069, + 0.007918206043541431, + 0.010970956645905972, + 0.013586705550551414, + 0.01371140405535698, + 0.01639392040669918, + 0.013873630203306675, + 0.013507673516869545, + 0.009042198769748211 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 7 + ], + "selected_reward": -0.048648420721292496 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a slow macro zoom-in gliding through crystal-clear turquoise water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are two Ocellaris clownfish visible with bright orange bodies and three distinct white vertical stripes outlined in black?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a single fish positioned directly above the eggs, fanning its pectoral fins in steady, rhythmic motions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene featuring a host anemone with fleshy green tentacles and soft purple tips swaying gently?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does bright sunlight filter through the surface creating dancing caustic patterns that make the translucent amber eggs shimmer?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0890311747789383 + ], + "opposite_logprob_tokens": [ + -2.464031219482422 + ], + "logprob_reward": -0.0890311747789383, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.0890311747789383, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2711801826953888, + 0.15854313969612122, + 0.07748402655124664, + 0.020206958055496216, + 0.05815983936190605, + 0.06485734134912491, + 0.07668919861316681, + 0.06911119073629379, + 0.17274358868598938, + 0.031024573370814323 + ], + "credit_map": [ + 0.03379351273179054, + 0.061710748821496964, + 0.013007891364395618, + 0.032385144382715225, + 0.016972405835986137, + 0.02222543954849243, + 0.06591037660837173, + 0.025174638256430626, + 0.010474225506186485, + 0.005618880968540907, + 0.007035431917756796, + 0.021511835977435112, + 0.00920499674975872, + 0.006954637356102467, + 0.05227313190698624, + 0.04546999931335449, + 0.005381837021559477, + 0.0063370103016495705, + 0.008529912680387497, + 0.014658719301223755, + 0.008713106624782085, + 0.011648816056549549, + 0.013212078250944614, + 0.009002545848488808, + 0.0014793367590755224, + 0.0018361912807449698, + 0.002688314998522401, + 0.0028054777067154646, + 0.0022155242040753365, + 0.0025742794387042522, + 0.004009344149380922, + 0.002598489634692669, + 0.005374228581786156, + 0.007032366935163736, + 0.0025595661718398333, + 0.013105800375342369, + 0.009360288269817829, + 0.004373269155621529, + 0.00858929194509983, + 0.007765024434775114, + 0.007149926852434874, + 0.012914703227579594, + 0.00656107859686017, + 0.009633921086788177, + 0.006892826408147812, + 0.004643883090466261, + 0.010018449276685715, + 0.00704254861921072, + 0.007376184221357107, + 0.009736764244735241, + 0.010236595757305622, + 0.011587931774556637, + 0.008553056977689266, + 0.00956379808485508, + 0.008932516910135746, + 0.010702352970838547, + 0.006746088154613972, + 0.009213821031153202, + 0.006161753553897142, + 0.008692238479852676, + 0.013530333526432514, + 0.00669121602550149, + 0.01090033259242773, + 0.007175405975431204, + 0.011740492656826973, + 0.00944507122039795, + 0.008750458247959614, + 0.06224881857633591, + 0.008818656206130981, + 0.04573149234056473, + 0.011794177815318108, + 0.014214429073035717, + 0.0023743468336760998, + 0.0023597616236656904, + 0.0032471742015331984, + 0.006462836172431707, + 0.0034994371235370636, + 0.003493262454867363, + 0.005267806816846132, + 0.004319948144257069 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 1 + ], + "selected_reward": -0.0890311747789383 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a slow macro zoom-in gliding through crystal-clear turquoise water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are two Ocellaris clownfish visible with bright orange bodies and three distinct white vertical stripes outlined in black?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a single fish positioned directly above the eggs, fanning its pectoral fins in steady, rhythmic motions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene featuring a host anemone with fleshy green tentacles and soft purple tips swaying gently?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does bright sunlight filter through the surface creating dancing caustic patterns that make the translucent amber eggs shimmer?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004636608529835939 + ], + "opposite_logprob_tokens": [ + -5.379636764526367 + ], + "logprob_reward": -0.004636608529835939, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004636608529835939, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.44310274720191956, + 0.11716856807470322, + 0.029376547783613205, + 0.019205274060368538, + 0.07539312541484833, + 0.01633318141102791, + 0.036674708127975464, + 0.018407072871923447, + 0.19964757561683655, + 0.04469119384884834 + ], + "credit_map": [ + 0.053672511130571365, + 0.09456606209278107, + 0.018548011779785156, + 0.05327089875936508, + 0.024714350700378418, + 0.02406141720712185, + 0.1292424201965332, + 0.04502704739570618, + 0.007442929781973362, + 0.005141448229551315, + 0.0051439134404063225, + 0.016604028642177582, + 0.0060952650383114815, + 0.007487268187105656, + 0.036701884120702744, + 0.0325518399477005, + 0.0018537793075665832, + 0.0020756875164806843, + 0.0030606593936681747, + 0.005648690741509199, + 0.002922038547694683, + 0.003990913275629282, + 0.005105033051222563, + 0.004719748627394438, + 0.001734446152113378, + 0.002465495141223073, + 0.0021942502353340387, + 0.002958429278805852, + 0.002484281314536929, + 0.002374522155150771, + 0.0026936912909150124, + 0.0023001586087048054, + 0.00471865851432085, + 0.0034461619798094034, + 0.002593619516119361, + 0.0271581020206213, + 0.004933544900268316, + 0.0033580614253878593, + 0.012951069511473179, + 0.016233906149864197, + 0.0018525191117078066, + 0.0038181080017238855, + 0.0010981005616486073, + 0.00234910543076694, + 0.002553964499384165, + 0.0011771005811169744, + 0.0021084605250507593, + 0.0013758239801973104, + 0.0032930492889136076, + 0.004186187870800495, + 0.0051547568291425705, + 0.0055281720124185085, + 0.003790887538343668, + 0.004605108872056007, + 0.004905419424176216, + 0.005211126524955034, + 0.0016353779938071966, + 0.0025135278701782227, + 0.002033787779510021, + 0.00226611434482038, + 0.0028422961477190256, + 0.0028083843644708395, + 0.002583759604021907, + 0.001723824068903923, + 0.01336968969553709, + 0.011385541409254074, + 0.009537560865283012, + 0.07405252009630203, + 0.0065993256866931915, + 0.04550880938768387, + 0.012804599478840828, + 0.026389513164758682, + 0.004732154309749603, + 0.004081102553755045, + 0.003788145026192069, + 0.010348217561841011, + 0.005771499127149582, + 0.007898062467575073, + 0.004865782335400581, + 0.00320623186416924 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 1 + ], + "selected_reward": -0.004636608529835939 + } + ], + "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/0000056/group_2/index_tensor([5858, 5858])_group_2_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000056/group_2/index_tensor([5858, 5858])_group_2_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..038e32e688f3fcf2dfa24cab62e6540bf4976f94 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000056/group_2/index_tensor([5858, 5858])_group_2_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "medium close-up shot from a level angle, its composition steady and composed, transports us to a night of relentless, cold rain. At the center of the frame, a solitary young woman leans limply against a damp, old brick wall in a narrow back alley. Her long, dark hair is completely soaked by the rain, strands plastered to her cold-white cheeks and neck. She wears a dark gray, slightly oversized hooded sweatshirt, the fabric dappled in the dim light with varying shades of wetness. Her shoulders slump, her gaze fixed vacantly forward, her eyes reddened as large, uncontrollable tears slip silently from her hollow irises. They flow down the contours of her cheeks, mingling with the steadily falling icy raindrops before pooling at her chin and dripping onto the lapels of her dark garment. A lone street lamp at the far end of the alleyway provides the sole source of light, its dim yellow light fragmented by the dense curtain of rain, casting long, twisting shadows across the slippery, mossy brick walls and the waterlogged stone pavement. The painting's dominant hue is a cold, austere blue-gray, with only the faint light bringing a barely perceptible hint of warmth, which only serves to accentuate the character's loneliness and coldness, creating an atmosphere of oppression and profound sadness, forgotten by the world.", + "seed": 11650, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000056/group_2/index_5858_seed_11650_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": -0.04511115181958303, + "vlm_logprob": -0.04511115181958303, + "vlm_yes_no_margin": 4.3, + "vlm_token_credit": -0.04511115181958303, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium close-up of a young woman leaning against a damp brick wall in a narrow alley?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a dark gray, slightly oversized hooded sweatshirt with wet fabric?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are large tears visible flowing from the woman's eyes down her cheeks and dripping onto her garment?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by a single dim yellow street lamp at the far end of the alley?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the overall color palette dominated by cold blue-gray tones with a dim yellow light source?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.027136124670505524 + ], + "opposite_logprob_tokens": [ + -4.152136325836182 + ], + "logprob_reward": -0.027136124670505524, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.027136124670505524, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3586808145046234, + 0.11130458116531372, + 0.04957738518714905, + 0.039940640330314636, + 0.050958551466464996, + 0.056916967034339905, + 0.06079168990254402, + 0.06936318427324295, + 0.11973419785499573, + 0.08273197710514069 + ], + "credit_map": [ + 0.02589399181306362, + 0.03544295206665993, + 0.0289856418967247, + 0.04366408288478851, + 0.05070081353187561, + 0.05113697424530983, + 0.06730776280164719, + 0.05554860830307007, + 0.019408514723181725, + 0.011668743565678596, + 0.013051068410277367, + 0.010411581955850124, + 0.012774666771292686, + 0.018142394721508026, + 0.010941848158836365, + 0.014905759133398533, + 0.0064325639978051186, + 0.005266149993985891, + 0.0043487404473125935, + 0.008848869241774082, + 0.00522456131875515, + 0.005741382483392954, + 0.00581291364505887, + 0.007902203127741814, + 0.003915954381227493, + 0.00399907398968935, + 0.0033568223007023335, + 0.0049339281395077705, + 0.0053198449313640594, + 0.006569466553628445, + 0.005582399200648069, + 0.006263150367885828, + 0.003371353028342128, + 0.004899791907519102, + 0.004058581776916981, + 0.006602808367460966, + 0.009955891408026218, + 0.007041302043944597, + 0.006337902974337339, + 0.008690920658409595, + 0.006435681600123644, + 0.006039110012352467, + 0.0046447450295090675, + 0.007320568896830082, + 0.0077347103506326675, + 0.007069074083119631, + 0.008912508375942707, + 0.008760571479797363, + 0.006720028817653656, + 0.008094199933111668, + 0.005715601146221161, + 0.012130817398428917, + 0.005212417803704739, + 0.007861869409680367, + 0.006026481278240681, + 0.009030274115502834, + 0.005649172700941563, + 0.011262696236371994, + 0.007164090871810913, + 0.012957491911947727, + 0.005663610063493252, + 0.00885900016874075, + 0.00647432915866375, + 0.011332795023918152, + 0.019007718190550804, + 0.013694440945982933, + 0.015617058612406254, + 0.01686517335474491, + 0.011797484941780567, + 0.012392458505928516, + 0.013703091070055962, + 0.016656778752803802, + 0.00936946365982294, + 0.009980859234929085, + 0.007265331689268351, + 0.015172540210187435, + 0.012511273846030235, + 0.009418272413313389, + 0.006723274942487478, + 0.012290959246456623 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 1 + ], + "selected_reward": -0.027136124670505524 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium close-up of a young woman leaning against a damp brick wall in a narrow alley?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a dark gray, slightly oversized hooded sweatshirt with wet fabric?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are large tears visible flowing from the woman's eyes down her cheeks and dripping onto her garment?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by a single dim yellow street lamp at the far end of the alley?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the overall color palette dominated by cold blue-gray tones with a dim yellow light source?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014232086017727852 + ], + "opposite_logprob_tokens": [ + -4.264232158660889 + ], + "logprob_reward": -0.014232086017727852, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014232086017727852, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3235756754875183, + 0.100493885576725, + 0.045443736016750336, + 0.0669839158654213, + 0.045474376529455185, + 0.0598575584590435, + 0.03695010766386986, + 0.06265268474817276, + 0.16662169992923737, + 0.09194628894329071 + ], + "credit_map": [ + 0.025039762258529663, + 0.03152279555797577, + 0.019951054826378822, + 0.037999123334884644, + 0.05331698805093765, + 0.04687664285302162, + 0.060275185853242874, + 0.04859413951635361, + 0.011381248012185097, + 0.010210291482508183, + 0.010961335152387619, + 0.009264063090085983, + 0.013804386369884014, + 0.019075531512498856, + 0.01186260487884283, + 0.013934439048171043, + 0.004855061415582895, + 0.004151035565882921, + 0.005269078072160482, + 0.004151205066591501, + 0.007913039065897465, + 0.005971385166049004, + 0.0061668651178479195, + 0.006966070272028446, + 0.008042523637413979, + 0.010079750791192055, + 0.006978709250688553, + 0.011639700271189213, + 0.005384236574172974, + 0.008747383020818233, + 0.006288053002208471, + 0.009823568165302277, + 0.0039129434153437614, + 0.00392115069553256, + 0.0038723822217434645, + 0.0068303742446005344, + 0.0076769934967160225, + 0.006570258177816868, + 0.005457994993776083, + 0.0072322869673371315, + 0.0050802938640117645, + 0.004626039415597916, + 0.0042893425561487675, + 0.010884013026952744, + 0.007090728729963303, + 0.00836691353470087, + 0.008866488002240658, + 0.010653747245669365, + 0.0031607290729880333, + 0.003467750735580921, + 0.004109777510166168, + 0.005074596963822842, + 0.004742512013763189, + 0.005813772324472666, + 0.004983353894203901, + 0.00559761980548501, + 0.00542746065184474, + 0.012113106437027454, + 0.006750375498086214, + 0.01136639155447483, + 0.004633299075067043, + 0.008070599287748337, + 0.0046098013408482075, + 0.009681657887995243, + 0.023992763832211494, + 0.021366415545344353, + 0.02076978050172329, + 0.023231755942106247, + 0.02127389796078205, + 0.015384319238364697, + 0.017493778839707375, + 0.02310900017619133, + 0.013013695366680622, + 0.01265160646289587, + 0.009809731505811214, + 0.013250843621790409, + 0.013968361541628838, + 0.009371032938361168, + 0.007053299807012081, + 0.012827727943658829 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 1 + ], + "selected_reward": -0.014232086017727852 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium close-up of a young woman leaning against a damp brick wall in a narrow alley?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a dark gray, slightly oversized hooded sweatshirt with wet fabric?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are large tears visible flowing from the woman's eyes down her cheeks and dripping onto her garment?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by a single dim yellow street lamp at the far end of the alley?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the overall color palette dominated by cold blue-gray tones with a dim yellow light source?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1604054719209671 + ], + "opposite_logprob_tokens": [ + -1.9104055166244507 + ], + "logprob_reward": -0.1604054719209671, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.1604054719209671, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.330800324678421, + 0.0984933003783226, + 0.04779946431517601, + 0.07877751439809799, + 0.059050120413303375, + 0.07340514659881592, + 0.04046516865491867, + 0.047309376299381256, + 0.1110261008143425, + 0.11287349462509155 + ], + "credit_map": [ + 0.03936571627855301, + 0.02827325277030468, + 0.027521410956978798, + 0.04004155844449997, + 0.03783220425248146, + 0.055213361978530884, + 0.054044198244810104, + 0.048508595675230026, + 0.011400272138416767, + 0.008795206435024738, + 0.011622698046267033, + 0.01393780019134283, + 0.007918200455605984, + 0.01807182841002941, + 0.011560795828700066, + 0.015186494216322899, + 0.006503075361251831, + 0.004238571971654892, + 0.006771307438611984, + 0.005366926547139883, + 0.004165273625403643, + 0.007379539776593447, + 0.00502832280471921, + 0.008346444927155972, + 0.008903609588742256, + 0.010664899833500385, + 0.00930008850991726, + 0.012372090481221676, + 0.0062457965686917305, + 0.010913163423538208, + 0.008291791193187237, + 0.012086072005331516, + 0.00705668656155467, + 0.005692885257303715, + 0.0065413410775363445, + 0.009903127327561378, + 0.006593610160052776, + 0.007693000137805939, + 0.005932010244578123, + 0.009637461975216866, + 0.008501963689923286, + 0.006640973035246134, + 0.006668021436780691, + 0.01178937777876854, + 0.008197655901312828, + 0.010005947202444077, + 0.00922032818198204, + 0.012380875647068024, + 0.004493302665650845, + 0.004540489986538887, + 0.004600982181727886, + 0.005825812462717295, + 0.004368978552520275, + 0.0057056560181081295, + 0.005083924159407616, + 0.005846024956554174, + 0.004612178076058626, + 0.009233121760189533, + 0.005048099439591169, + 0.009090965613722801, + 0.0036546150222420692, + 0.0055401865392923355, + 0.003051202278584242, + 0.007079008035361767, + 0.013397851958870888, + 0.013688628561794758, + 0.013031327165663242, + 0.01567178964614868, + 0.01603076606988907, + 0.010619177483022213, + 0.013192604295909405, + 0.015393964014947414, + 0.01833844929933548, + 0.01704448089003563, + 0.013219420798122883, + 0.015742158517241478, + 0.014300509355962276, + 0.01020174939185381, + 0.008211316540837288, + 0.015815408900380135 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 8 + ], + "selected_reward": -0.1604054719209671 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium close-up of a young woman leaning against a damp brick wall in a narrow alley?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a dark gray, slightly oversized hooded sweatshirt with wet fabric?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are large tears visible flowing from the woman's eyes down her cheeks and dripping onto her garment?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by a single dim yellow street lamp at the far end of the alley?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the overall color palette dominated by cold blue-gray tones with a dim yellow light source?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023282833397388458 + ], + "opposite_logprob_tokens": [ + -3.773282766342163 + ], + "logprob_reward": -0.023282833397388458, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023282833397388458, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2540103495121002, + 0.22961793839931488, + 0.15142835676670074, + 0.09724400192499161, + 0.05169091373682022, + 0.03729585185647011, + 0.025833958759903908, + 0.05291883274912834, + 0.04108650982379913, + 0.058873265981674194 + ], + "credit_map": [ + 0.05184093862771988, + 0.02723570354282856, + 0.033294856548309326, + 0.033341411501169205, + 0.017137575894594193, + 0.025850629433989525, + 0.020801689475774765, + 0.04450757056474686, + 0.03656596690416336, + 0.02585081197321415, + 0.022006727755069733, + 0.04634047672152519, + 0.018589433282613754, + 0.028999481350183487, + 0.019252009689807892, + 0.03201304376125336, + 0.027293149381875992, + 0.018252864480018616, + 0.022381896153092384, + 0.02359188161790371, + 0.011671816930174828, + 0.015354807488620281, + 0.010868394747376442, + 0.022013531997799873, + 0.016207821667194366, + 0.011668309569358826, + 0.014180553145706654, + 0.014924256131052971, + 0.007348580751568079, + 0.011521019041538239, + 0.007844383828341961, + 0.013549075461924076, + 0.009721039794385433, + 0.005397881846874952, + 0.006513508968055248, + 0.009495750069618225, + 0.003941302187740803, + 0.004513096529990435, + 0.0040276204235851765, + 0.008080716244876385, + 0.005698131863027811, + 0.0049334377981722355, + 0.003489232389256358, + 0.007972330786287785, + 0.0032940225210040808, + 0.0033282663207501173, + 0.0031136476900428534, + 0.00546678202226758, + 0.0037722152192145586, + 0.003559679025784135, + 0.0029475814662873745, + 0.004040585830807686, + 0.0021020013373345137, + 0.0034826614428311586, + 0.0023784905206412077, + 0.003550743218511343, + 0.006016440223902464, + 0.008941706269979477, + 0.007050482090562582, + 0.010762539692223072, + 0.003140190150588751, + 0.00536776939406991, + 0.0034813580568879843, + 0.008158346638083458, + 0.007708804681897163, + 0.004879410844296217, + 0.0061108688823878765, + 0.005181867629289627, + 0.00237572961486876, + 0.00605225982144475, + 0.0033068072516471148, + 0.0054707652889192104, + 0.011088003404438496, + 0.008845620788633823, + 0.007045453879982233, + 0.008187704719603062, + 0.004467006307095289, + 0.00592928659170866, + 0.004681580234318972, + 0.008628608658909798 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.023282833397388458 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium close-up of a young woman leaning against a damp brick wall in a narrow alley?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a dark gray, slightly oversized hooded sweatshirt with wet fabric?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are large tears visible flowing from the woman's eyes down her cheeks and dripping onto her garment?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by a single dim yellow street lamp at the far end of the alley?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the overall color palette dominated by cold blue-gray tones with a dim yellow light source?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0004992430913262069 + ], + "opposite_logprob_tokens": [ + -7.625499248504639 + ], + "logprob_reward": -0.0004992430913262069, + "yes_no_margin_reward": 7.625, + "token_credit_reward": -0.0004992430913262069, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2695721685886383, + 0.16219541430473328, + 0.08846532553434372, + 0.06041524186730385, + 0.04852374643087387, + 0.04327230155467987, + 0.05489042401313782, + 0.07459092885255814, + 0.12369313836097717, + 0.07438124716281891 + ], + "credit_map": [ + 0.0363718681037426, + 0.027341602370142937, + 0.029696805402636528, + 0.03938844054937363, + 0.0298890620470047, + 0.03266188129782677, + 0.03473971411585808, + 0.03948282450437546, + 0.029378080740571022, + 0.017859185114502907, + 0.016945568844676018, + 0.02735326997935772, + 0.014142360538244247, + 0.021278144791722298, + 0.013438942842185497, + 0.021799873560667038, + 0.017595142126083374, + 0.010715141892433167, + 0.011651045642793179, + 0.011530270799994469, + 0.008655015379190445, + 0.009237119928002357, + 0.007268748711794615, + 0.011812850832939148, + 0.009546429850161076, + 0.005986588075757027, + 0.010112653486430645, + 0.007120213471353054, + 0.005628639832139015, + 0.007308001630008221, + 0.006252342835068703, + 0.008460375480353832, + 0.006809057667851448, + 0.005749597679823637, + 0.005102664232254028, + 0.008990667760372162, + 0.00437213946133852, + 0.0060053435154259205, + 0.004774438217282295, + 0.006719843950122595, + 0.005088387988507748, + 0.005928763188421726, + 0.003681478789076209, + 0.009390492923557758, + 0.0038280764129012823, + 0.005428466480225325, + 0.0031958946492522955, + 0.006730744615197182, + 0.007191602606326342, + 0.0085239727050066, + 0.005347152706235647, + 0.007291763089597225, + 0.005118953995406628, + 0.008065490052103996, + 0.0050834049470722675, + 0.008268091827630997, + 0.00849185697734356, + 0.01277630403637886, + 0.007915961556136608, + 0.013888283632695675, + 0.00568181648850441, + 0.00809310469776392, + 0.005887843668460846, + 0.01185576617717743, + 0.02207130752503872, + 0.014392202720046043, + 0.01803857833147049, + 0.018757272511720657, + 0.00906385574489832, + 0.01111716590821743, + 0.008808746933937073, + 0.021444013342261314, + 0.00914975255727768, + 0.010756755247712135, + 0.0074560451321303844, + 0.011192508973181248, + 0.011928633786737919, + 0.007800664287060499, + 0.005989637225866318, + 0.010107253678143024 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 8 + ], + "selected_reward": -0.0004992430913262069 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "medium close-up shot from a level angle, its composition steady and composed, transports us to a night of relentless, cold rain. At the center of the frame, a solitary young woman leans limply against a damp, old brick wall in a narrow back alley. Her long, dark hair is completely soaked by the rain, strands plastered to her cold-white cheeks and neck. She wears a dark gray, slightly oversized hooded sweatshirt, the fabric dappled in the dim light with varying shades of wetness. Her shoulders slump, her gaze fixed vacantly forward, her eyes reddened as large, uncontrollable tears slip silently from her hollow irises. They flow down the contours of her cheeks, mingling with the steadily falling icy raindrops before pooling at her chin and dripping onto the lapels of her dark garment. A lone street lamp at the far end of the alleyway provides the sole source of light, its dim yellow light fragmented by the dense curtain of rain, casting long, twisting shadows across the slippery, mossy brick walls and the waterlogged stone pavement. The painting's dominant hue is a cold, austere blue-gray, with only the faint light bringing a barely perceptible hint of warmth, which only serves to accentuate the character's loneliness and coldness, creating an atmosphere of oppression and profound sadness, forgotten by the world.", + "seed": 11651, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000056/group_2/index_5858_seed_11651_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": -0.027412121364613996, + "vlm_logprob": -0.027412121364613996, + "vlm_yes_no_margin": 4.85, + "vlm_token_credit": -0.027412121364613996, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium close-up of a young woman leaning against a damp brick wall in a narrow alley?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a dark gray, slightly oversized hooded sweatshirt with wet fabric?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are large tears visible flowing from the woman's eyes down her cheeks and dripping onto her garment?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by a single dim yellow street lamp at the far end of the alley?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the overall color palette dominated by cold blue-gray tones with a dim yellow light source?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.013581610284745693 + ], + "opposite_logprob_tokens": [ + -4.8885817527771 + ], + "logprob_reward": -0.013581610284745693, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.013581610284745693, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3795315623283386, + 0.09249469637870789, + 0.09393163025379181, + 0.11642560362815857, + 0.04225834831595421, + 0.045090071856975555, + 0.05208175256848335, + 0.09145578742027283, + 0.04357380419969559, + 0.04315676912665367 + ], + "credit_map": [ + 0.03731754049658775, + 0.03290582075715065, + 0.04200514778494835, + 0.02758914604783058, + 0.07068970054388046, + 0.05670155584812164, + 0.05312329903244972, + 0.05919938534498215, + 0.007856207899749279, + 0.008214976638555527, + 0.009671635925769806, + 0.006569381337612867, + 0.0172192994505167, + 0.015673711895942688, + 0.012806805782020092, + 0.014482683502137661, + 0.007397986948490143, + 0.00615908345207572, + 0.006157358177006245, + 0.007495196536183357, + 0.024121394380927086, + 0.01414254866540432, + 0.015680283308029175, + 0.012777778320014477, + 0.009013393893837929, + 0.006092479452490807, + 0.00980247464030981, + 0.009120662696659565, + 0.028581898659467697, + 0.016802463680505753, + 0.020504888147115707, + 0.0165073424577713, + 0.005929643753916025, + 0.003432192839682102, + 0.007120417896658182, + 0.003835210809484124, + 0.004552104510366917, + 0.005847522988915443, + 0.005480330903083086, + 0.006060921121388674, + 0.005286780185997486, + 0.0028760177083313465, + 0.005462430417537689, + 0.0034808837808668613, + 0.006515685003250837, + 0.0037974785082042217, + 0.008610676042735577, + 0.009060120210051537, + 0.0072494689375162125, + 0.004965461790561676, + 0.004690699279308319, + 0.003180725034326315, + 0.008419792167842388, + 0.006275732070207596, + 0.008452477864921093, + 0.00884739775210619, + 0.008388957008719444, + 0.008495568297803402, + 0.009018989279866219, + 0.007854921743273735, + 0.01883009634912014, + 0.011175323277711868, + 0.015755267813801765, + 0.01193666085600853, + 0.004034433513879776, + 0.0032922846730798483, + 0.005215928424149752, + 0.0025787651538848877, + 0.004097374621778727, + 0.0037942056078463793, + 0.009385114535689354, + 0.011175693944096565, + 0.003895781934261322, + 0.004792726133018732, + 0.005021287593990564, + 0.002861201763153076, + 0.005428673699498177, + 0.004868966992944479, + 0.007306484505534172, + 0.008981646969914436 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 2 + ], + "selected_reward": -0.013581610284745693 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium close-up of a young woman leaning against a damp brick wall in a narrow alley?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a dark gray, slightly oversized hooded sweatshirt with wet fabric?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are large tears visible flowing from the woman's eyes down her cheeks and dripping onto her garment?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by a single dim yellow street lamp at the far end of the alley?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the overall color palette dominated by cold blue-gray tones with a dim yellow light source?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006753476802259684 + ], + "opposite_logprob_tokens": [ + -5.006753444671631 + ], + "logprob_reward": -0.006753476802259684, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006753476802259684, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3971710205078125, + 0.09307299554347992, + 0.10555267333984375, + 0.11480005830526352, + 0.05146057903766632, + 0.03570852801203728, + 0.05380802974104881, + 0.07742568850517273, + 0.03526638075709343, + 0.035734035074710846 + ], + "credit_map": [ + 0.038014356046915054, + 0.015789762139320374, + 0.03237688168883324, + 0.0351775586605072, + 0.09472277760505676, + 0.06634862720966339, + 0.056243497878313065, + 0.058497555553913116, + 0.010446938686072826, + 0.006220497190952301, + 0.007336408365517855, + 0.00912176538258791, + 0.017801523208618164, + 0.013960658572614193, + 0.014832758344709873, + 0.013352450914680958, + 0.010158328339457512, + 0.006529206410050392, + 0.0073369331657886505, + 0.010121292434632778, + 0.02559066005051136, + 0.013027074746787548, + 0.017755161970853806, + 0.015034015290439129, + 0.009780186228454113, + 0.00630566431209445, + 0.009264137595891953, + 0.009289302863180637, + 0.0257582925260067, + 0.017902854830026627, + 0.0186432134360075, + 0.017856411635875702, + 0.004060231149196625, + 0.0016932559665292501, + 0.005543417762964964, + 0.004993510898202658, + 0.01157047413289547, + 0.008480224758386612, + 0.007995469495654106, + 0.007123994641005993, + 0.004788251128047705, + 0.0027519187424331903, + 0.004818023182451725, + 0.002829183591529727, + 0.004495812114328146, + 0.0026669458020478487, + 0.00689248600974679, + 0.006465907208621502, + 0.007088904269039631, + 0.005209168419241905, + 0.0045367698185145855, + 0.003459544852375984, + 0.00845401082187891, + 0.007248363923281431, + 0.008398281410336494, + 0.009412981569766998, + 0.006385772954672575, + 0.004958339035511017, + 0.005663134157657623, + 0.008076556958258152, + 0.017060158774256706, + 0.011084775440394878, + 0.01334443874657154, + 0.010852514766156673, + 0.005075313150882721, + 0.001934480736963451, + 0.003924925811588764, + 0.001704318099655211, + 0.004272953141480684, + 0.004288868512958288, + 0.006551912985742092, + 0.00751360971480608, + 0.00404940452426672, + 0.0029987001325935125, + 0.003862002631649375, + 0.00224397424608469, + 0.0046146404929459095, + 0.0037704145070165396, + 0.006461449433118105, + 0.007733450271189213 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 2 + ], + "selected_reward": -0.006753476802259684 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium close-up of a young woman leaning against a damp brick wall in a narrow alley?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a dark gray, slightly oversized hooded sweatshirt with wet fabric?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are large tears visible flowing from the woman's eyes down her cheeks and dripping onto her garment?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by a single dim yellow street lamp at the far end of the alley?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the overall color palette dominated by cold blue-gray tones with a dim yellow light source?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10032056272029877 + ], + "opposite_logprob_tokens": [ + -2.35032057762146 + ], + "logprob_reward": -0.10032056272029877, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10032056272029877, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.29685497283935547, + 0.16257326304912567, + 0.08488521724939346, + 0.07976003736257553, + 0.11878565698862076, + 0.04623910412192345, + 0.08245198428630829, + 0.05814342200756073, + 0.03520820289850235, + 0.03509820997714996 + ], + "credit_map": [ + 0.033315930515527725, + 0.016750816255807877, + 0.024229571223258972, + 0.023887217044830322, + 0.05988362804055214, + 0.042620450258255005, + 0.052440814673900604, + 0.043726515024900436, + 0.01265642512589693, + 0.011461913585662842, + 0.015815800055861473, + 0.012519381940364838, + 0.02929179184138775, + 0.03874710574746132, + 0.0175310205668211, + 0.024549808353185654, + 0.009061257354915142, + 0.004852667450904846, + 0.010836482048034668, + 0.007716984488070011, + 0.0117124542593956, + 0.014178117737174034, + 0.012555979192256927, + 0.013971269130706787, + 0.006635990459471941, + 0.004084216896444559, + 0.004632390569895506, + 0.007574793882668018, + 0.01701441965997219, + 0.01267806813120842, + 0.014994526281952858, + 0.01214562077075243, + 0.01311599463224411, + 0.004662306047976017, + 0.0081587303429842, + 0.004078188445419073, + 0.007636921014636755, + 0.008234511129558086, + 0.03200414776802063, + 0.040894851088523865, + 0.006503700278699398, + 0.0035657731350511312, + 0.006987997330725193, + 0.0034185745753347874, + 0.004946477711200714, + 0.005411752499639988, + 0.007390338461846113, + 0.008014483377337456, + 0.00995190255343914, + 0.006773833185434341, + 0.0072392006404697895, + 0.004416004288941622, + 0.008256287313997746, + 0.007490447256714106, + 0.01713009737432003, + 0.021194200962781906, + 0.005462585482746363, + 0.005421972833573818, + 0.005422018002718687, + 0.005757021252065897, + 0.010890922509133816, + 0.006942918989807367, + 0.010619075037539005, + 0.0076268999837338924, + 0.005819234997034073, + 0.0031640706583857536, + 0.0035802845377475023, + 0.001754725817590952, + 0.0030880027916282415, + 0.0036074125673621893, + 0.006694246083498001, + 0.007500221021473408, + 0.004390025977045298, + 0.003652649698778987, + 0.003911259584128857, + 0.0026946179568767548, + 0.004208994098007679, + 0.003627947997301817, + 0.005732222460210323, + 0.006880490574985743 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 4 + ], + "selected_reward": -0.10032056272029877 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium close-up of a young woman leaning against a damp brick wall in a narrow alley?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a dark gray, slightly oversized hooded sweatshirt with wet fabric?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are large tears visible flowing from the woman's eyes down her cheeks and dripping onto her garment?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by a single dim yellow street lamp at the far end of the alley?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the overall color palette dominated by cold blue-gray tones with a dim yellow light source?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"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": [ + 0.27019360661506653, + 0.1404425948858261, + 0.09662000089883804, + 0.13686828315258026, + 0.09459105879068375, + 0.02655497007071972, + 0.09424381703138351, + 0.07290367037057877, + 0.02386510744690895, + 0.04371688514947891 + ], + "credit_map": [ + 0.03405631706118584, + 0.0203398484736681, + 0.02627958171069622, + 0.015947630628943443, + 0.021213466301560402, + 0.025258129462599754, + 0.05582655221223831, + 0.07127208262681961, + 0.019878922030329704, + 0.009344453923404217, + 0.014231872744858265, + 0.0064194765873253345, + 0.008185045793652534, + 0.013707651756703854, + 0.03040483221411705, + 0.038270339369773865, + 0.010381561703979969, + 0.008844814263284206, + 0.01025462057441473, + 0.005377833731472492, + 0.014017265290021896, + 0.014152923598885536, + 0.01592852734029293, + 0.01766246370971203, + 0.014451700262725353, + 0.017014384269714355, + 0.015104677528142929, + 0.010481972247362137, + 0.025802303105592728, + 0.015282226726412773, + 0.021199798211455345, + 0.017531229183077812, + 0.01203366369009018, + 0.00901605561375618, + 0.009083818644285202, + 0.0044871242716908455, + 0.005751467775553465, + 0.007404157891869545, + 0.02039446122944355, + 0.026420313864946365, + 0.0033829074818640947, + 0.0013541387161239982, + 0.002804545219987631, + 0.0014264535857364535, + 0.0032540049869567156, + 0.0031543313525617123, + 0.00512426532804966, + 0.006054322235286236, + 0.010889332741498947, + 0.008542599156498909, + 0.007686912082135677, + 0.005041160620748997, + 0.011957861483097076, + 0.00842046458274126, + 0.01973973959684372, + 0.02196574956178665, + 0.010168486274778843, + 0.008574781939387321, + 0.008347047492861748, + 0.006011921912431717, + 0.011411537416279316, + 0.00982165802270174, + 0.009970448911190033, + 0.008597788400948048, + 0.0036959387362003326, + 0.0015299636870622635, + 0.003048602258786559, + 0.0014464746927842498, + 0.0020205206237733364, + 0.002367081819102168, + 0.004260614980012178, + 0.005495910998433828, + 0.005168662406504154, + 0.0022376927081495523, + 0.004476310219615698, + 0.0016484168590977788, + 0.002843838185071945, + 0.004198407754302025, + 0.010030227713286877, + 0.013113327324390411 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 3 + ], + "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 shot a medium close-up of a young woman leaning against a damp brick wall in a narrow alley?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a dark gray, slightly oversized hooded sweatshirt with wet fabric?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are large tears visible flowing from the woman's eyes down her cheeks and dripping onto her garment?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by a single dim yellow street lamp at the far end of the alley?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the overall color palette dominated by cold blue-gray tones with a dim yellow light source?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00034397884155623615 + ], + "opposite_logprob_tokens": [ + -8.000344276428223 + ], + "logprob_reward": -0.00034397884155623615, + "yes_no_margin_reward": 8.0, + "token_credit_reward": -0.00034397884155623615, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18662239611148834, + 0.1451423317193985, + 0.07754217833280563, + 0.12982173264026642, + 0.0664580687880516, + 0.07786015421152115, + 0.08487989753484726, + 0.13269104063510895, + 0.04463913291692734, + 0.05434306338429451 + ], + "credit_map": [ + 0.017206376418471336, + 0.02536635287106037, + 0.0252067968249321, + 0.015301189385354519, + 0.029743576422333717, + 0.02348054014146328, + 0.02267345041036606, + 0.027644123882055283, + 0.012555427849292755, + 0.007907544262707233, + 0.017201239243149757, + 0.009612898342311382, + 0.025196455419063568, + 0.02900170348584652, + 0.01974347047507763, + 0.023923590779304504, + 0.006335598416626453, + 0.005600187461823225, + 0.00891796126961708, + 0.005411204881966114, + 0.00962644349783659, + 0.01497593428939581, + 0.01177543867379427, + 0.014899412170052528, + 0.011792132630944252, + 0.011934791691601276, + 0.011739752255380154, + 0.009834561496973038, + 0.028247080743312836, + 0.018422476947307587, + 0.020396051928400993, + 0.017454877495765686, + 0.007940969429910183, + 0.0059344582259655, + 0.004718876909464598, + 0.004202856682240963, + 0.008818196132779121, + 0.008379870094358921, + 0.011335842311382294, + 0.015127000398933887, + 0.010647375136613846, + 0.008857215754687786, + 0.0098279332742095, + 0.004657316487282515, + 0.010872973129153252, + 0.009835540316998959, + 0.01231878250837326, + 0.010843015275895596, + 0.008830917999148369, + 0.006808815989643335, + 0.0089388657361269, + 0.006986341904848814, + 0.016477085649967194, + 0.011125984601676464, + 0.014270051382482052, + 0.011441834270954132, + 0.017272986471652985, + 0.01556207612156868, + 0.015536348335444927, + 0.010565337724983692, + 0.02062433771789074, + 0.016013355925679207, + 0.02031397633254528, + 0.01680261269211769, + 0.0062865917570889, + 0.0025984779931604862, + 0.005647115409374237, + 0.0027298452332615852, + 0.0039027470629662275, + 0.005393770523369312, + 0.007928735576570034, + 0.010151849128305912, + 0.00430124020203948, + 0.005237397737801075, + 0.007046939339488745, + 0.003791330149397254, + 0.004902427084743977, + 0.008171903900802135, + 0.00903413537889719, + 0.011857685633003712 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 7 + ], + "selected_reward": -0.00034397884155623615 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "medium close-up shot from a level angle, its composition steady and composed, transports us to a night of relentless, cold rain. At the center of the frame, a solitary young woman leans limply against a damp, old brick wall in a narrow back alley. Her long, dark hair is completely soaked by the rain, strands plastered to her cold-white cheeks and neck. She wears a dark gray, slightly oversized hooded sweatshirt, the fabric dappled in the dim light with varying shades of wetness. Her shoulders slump, her gaze fixed vacantly forward, her eyes reddened as large, uncontrollable tears slip silently from her hollow irises. They flow down the contours of her cheeks, mingling with the steadily falling icy raindrops before pooling at her chin and dripping onto the lapels of her dark garment. A lone street lamp at the far end of the alleyway provides the sole source of light, its dim yellow light fragmented by the dense curtain of rain, casting long, twisting shadows across the slippery, mossy brick walls and the waterlogged stone pavement. The painting's dominant hue is a cold, austere blue-gray, with only the faint light bringing a barely perceptible hint of warmth, which only serves to accentuate the character's loneliness and coldness, creating an atmosphere of oppression and profound sadness, forgotten by the world.", + "seed": 11652, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000056/group_2/index_5858_seed_11652_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": -0.05778506291098893, + "vlm_logprob": -0.05778506291098893, + "vlm_yes_no_margin": 4.125, + "vlm_token_credit": -0.05778506291098893, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium close-up of a young woman leaning against a damp brick wall in a narrow alley?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a dark gray, slightly oversized hooded sweatshirt with wet fabric?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are large tears visible flowing from the woman's eyes down her cheeks and dripping onto her garment?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by a single dim yellow street lamp at the far end of the alley?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the overall color palette dominated by cold blue-gray tones with a dim yellow light source?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05052271485328674 + ], + "opposite_logprob_tokens": [ + -3.300522804260254 + ], + "logprob_reward": -0.05052271485328674, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.05052271485328674, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4071601629257202, + 0.15724126994609833, + 0.12905098497867584, + 0.031875114887952805, + 0.03786902129650116, + 0.04246170446276665, + 0.04852421581745148, + 0.07469162344932556, + 0.03214553743600845, + 0.038980379700660706 + ], + "credit_map": [ + 0.04048578441143036, + 0.041850101202726364, + 0.05474086105823517, + 0.0726478174328804, + 0.04516225308179855, + 0.038675129413604736, + 0.06319399923086166, + 0.05040426179766655, + 0.01801137626171112, + 0.015130214393138885, + 0.022255489602684975, + 0.02106359973549843, + 0.016000812873244286, + 0.01779770851135254, + 0.02655729651451111, + 0.02042476087808609, + 0.018206903710961342, + 0.012364179827272892, + 0.019694233313202858, + 0.018004046753048897, + 0.011442088522017002, + 0.014735315926373005, + 0.018077973276376724, + 0.016526244580745697, + 0.006063677836209536, + 0.003365860553458333, + 0.004546750336885452, + 0.0042722029611468315, + 0.003291668836027384, + 0.0019379639998078346, + 0.004332056734710932, + 0.004064931999891996, + 0.003903439035639167, + 0.0031070602126419544, + 0.004911983851343393, + 0.005838189739733934, + 0.004621006548404694, + 0.004052862524986267, + 0.006541391834616661, + 0.004893086384981871, + 0.0041419160552322865, + 0.004025696776807308, + 0.005021972116082907, + 0.006385090295225382, + 0.004753974266350269, + 0.006198767106980085, + 0.0065556797198951244, + 0.005378606729209423, + 0.003843213664367795, + 0.00481503177434206, + 0.005672868341207504, + 0.008080510422587395, + 0.0061114453710615635, + 0.0041999551467597485, + 0.008955623023211956, + 0.006845570635050535, + 0.005553883500397205, + 0.008634060621261597, + 0.00927931722253561, + 0.013590327464044094, + 0.00950624980032444, + 0.007567727006971836, + 0.01061384379863739, + 0.009946217760443687, + 0.0033174650743603706, + 0.002112326445057988, + 0.005132582038640976, + 0.004292450379580259, + 0.004198126029223204, + 0.0030102927703410387, + 0.0054727704264223576, + 0.004609524738043547, + 0.005121568683534861, + 0.004188130609691143, + 0.005570810753852129, + 0.006643776316195726, + 0.004395806230604649, + 0.003667292883619666, + 0.004904548637568951, + 0.004488446284085512 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.05052271485328674 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium close-up of a young woman leaning against a damp brick wall in a narrow alley?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a dark gray, slightly oversized hooded sweatshirt with wet fabric?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are large tears visible flowing from the woman's eyes down her cheeks and dripping onto her garment?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by a single dim yellow street lamp at the far end of the alley?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the overall color palette dominated by cold blue-gray tones with a dim yellow light source?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01822757162153721 + ], + "opposite_logprob_tokens": [ + -4.018227577209473 + ], + "logprob_reward": -0.01822757162153721, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.01822757162153721, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.328372061252594, + 0.17283795773983002, + 0.11284670978784561, + 0.06256561726331711, + 0.06054571270942688, + 0.04192609712481499, + 0.054296042770147324, + 0.06557851284742355, + 0.04327185079455376, + 0.057759422808885574 + ], + "credit_map": [ + 0.027452688664197922, + 0.026609187945723534, + 0.04006354138255119, + 0.05261247605085373, + 0.03632933646440506, + 0.03029365837574005, + 0.06451864540576935, + 0.0504925437271595, + 0.0289152879267931, + 0.018335910513997078, + 0.022243738174438477, + 0.02438277192413807, + 0.01783577911555767, + 0.015038547106087208, + 0.025864072144031525, + 0.020221853628754616, + 0.010564461350440979, + 0.00996360369026661, + 0.018200477585196495, + 0.014043708331882954, + 0.012954615987837315, + 0.015042775310575962, + 0.01686922088265419, + 0.015207852236926556, + 0.00470429752022028, + 0.007869139313697815, + 0.006841040216386318, + 0.012196958996355534, + 0.006742180325090885, + 0.007781690452247858, + 0.008453316986560822, + 0.007976996712386608, + 0.004606890492141247, + 0.005312092136591673, + 0.005936274770647287, + 0.010801135562360287, + 0.007599732372909784, + 0.007006355561316013, + 0.010527968406677246, + 0.00875526387244463, + 0.006679947953671217, + 0.0038417743053287268, + 0.007732228375971317, + 0.005087121855467558, + 0.003653082298114896, + 0.0034404818434268236, + 0.005364942830055952, + 0.006126522086560726, + 0.010658427141606808, + 0.005223626736551523, + 0.006876584142446518, + 0.008959912694990635, + 0.004296646453440189, + 0.003531775204464793, + 0.007108148653060198, + 0.007640928495675325, + 0.010754372924566269, + 0.00915115512907505, + 0.008179872296750546, + 0.01044868491590023, + 0.004255663603544235, + 0.008378912694752216, + 0.007215575780719519, + 0.007194284815341234, + 0.007067408878356218, + 0.0034566447138786316, + 0.005083227530121803, + 0.005432297475636005, + 0.0046390569768846035, + 0.002999246120452881, + 0.007481215987354517, + 0.007112756837159395, + 0.007337280549108982, + 0.007406747434288263, + 0.007865204475820065, + 0.009105011820793152, + 0.006531507708132267, + 0.004160893149673939, + 0.008520077913999557, + 0.006832706276327372 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.01822757162153721 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium close-up of a young woman leaning against a damp brick wall in a narrow alley?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a dark gray, slightly oversized hooded sweatshirt with wet fabric?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are large tears visible flowing from the woman's eyes down her cheeks and dripping onto her garment?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by a single dim yellow street lamp at the far end of the alley?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the overall color palette dominated by cold blue-gray tones with a dim yellow light source?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.20160219073295593 + ], + "opposite_logprob_tokens": [ + -1.7016022205352783 + ], + "logprob_reward": -0.20160219073295593, + "yes_no_margin_reward": 1.5, + "token_credit_reward": -0.20160219073295593, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21995699405670166, + 0.19532543420791626, + 0.12600654363632202, + 0.0892394557595253, + 0.09443778544664383, + 0.059134069830179214, + 0.07818740606307983, + 0.03819756582379341, + 0.055057309567928314, + 0.04445742443203926 + ], + "credit_map": [ + 0.022317642346024513, + 0.017368178814649582, + 0.03051065094769001, + 0.02752416580915451, + 0.02936280518770218, + 0.030733829364180565, + 0.03580622747540474, + 0.02633349597454071, + 0.024584323167800903, + 0.018322572112083435, + 0.03254718706011772, + 0.0255818460136652, + 0.01781112141907215, + 0.022504983469843864, + 0.029556628316640854, + 0.024416765198111534, + 0.01572747901082039, + 0.010887951590120792, + 0.020154614001512527, + 0.013966076076030731, + 0.015492442063987255, + 0.015646694228053093, + 0.01913154311478138, + 0.01499974261969328, + 0.006981527898460627, + 0.011116010136902332, + 0.011769922450184822, + 0.017150331288576126, + 0.009469819255173206, + 0.011390216648578644, + 0.010568330064415932, + 0.010793296620249748, + 0.0069431220181286335, + 0.009333374910056591, + 0.012635987251996994, + 0.016438135877251625, + 0.01094847358763218, + 0.009582769125699997, + 0.015784455463290215, + 0.012771469540894032, + 0.011804829351603985, + 0.005117456894367933, + 0.009033088572323322, + 0.007945368066430092, + 0.004738629795610905, + 0.005086356308311224, + 0.008104244247078896, + 0.007304093800485134, + 0.01623145304620266, + 0.007680314593017101, + 0.013020513579249382, + 0.012370388954877853, + 0.005124120507389307, + 0.006132161244750023, + 0.008512476459145546, + 0.009115977212786674, + 0.0024311773013323545, + 0.004743431694805622, + 0.005358942784368992, + 0.00675620511174202, + 0.0032262203749269247, + 0.0053708809427917, + 0.005212182179093361, + 0.005098529160022736, + 0.012569946236908436, + 0.006037641782313585, + 0.0067034559324383736, + 0.0067442720755934715, + 0.004234694875776768, + 0.004352655727416277, + 0.00674327090382576, + 0.00767136737704277, + 0.0059259748086333275, + 0.003278488991782069, + 0.006486647296696901, + 0.004909425508230925, + 0.006122502963989973, + 0.004202351905405521, + 0.006954870652407408, + 0.006577162072062492 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.20160219073295593 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium close-up of a young woman leaning against a damp brick wall in a narrow alley?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a dark gray, slightly oversized hooded sweatshirt with wet fabric?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are large tears visible flowing from the woman's eyes down her cheeks and dripping onto her garment?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by a single dim yellow street lamp at the far end of the alley?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the overall color palette dominated by cold blue-gray tones with a dim yellow light source?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018182622268795967 + ], + "opposite_logprob_tokens": [ + -4.018182754516602 + ], + "logprob_reward": -0.018182622268795967, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018182622268795967, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16369304060935974, + 0.09709816426038742, + 0.22327390313148499, + 0.14469803869724274, + 0.06092358008027077, + 0.037435729056596756, + 0.08176544308662415, + 0.057175394147634506, + 0.0924772396683693, + 0.041459422558546066 + ], + "credit_map": [ + 0.025960369035601616, + 0.015146151185035706, + 0.01733229123055935, + 0.027663404121994972, + 0.012520966120064259, + 0.020865125581622124, + 0.021607931703329086, + 0.022596802562475204, + 0.011642000637948513, + 0.011145362630486488, + 0.014903281815350056, + 0.015398566611111164, + 0.007647462654858828, + 0.01311871875077486, + 0.00981502141803503, + 0.01342774834483862, + 0.021125715225934982, + 0.03208272531628609, + 0.025769002735614777, + 0.043184224516153336, + 0.01452516671270132, + 0.03795315697789192, + 0.020821353420615196, + 0.027812553569674492, + 0.018122587352991104, + 0.01549900695681572, + 0.018073244020342827, + 0.028662605211138725, + 0.008465786464512348, + 0.018001852557063103, + 0.0169018916785717, + 0.020971063524484634, + 0.007686985190957785, + 0.006257330067455769, + 0.007599904201924801, + 0.011371695436537266, + 0.00425665732473135, + 0.009892560541629791, + 0.00548193696886301, + 0.00837650801986456, + 0.00465744361281395, + 0.004386585205793381, + 0.005356291774660349, + 0.007501173298805952, + 0.0020748598035424948, + 0.0035681838635355234, + 0.004779311828315258, + 0.00511188292875886, + 0.01450788602232933, + 0.010600103996694088, + 0.011106766760349274, + 0.01286319363862276, + 0.006521888542920351, + 0.007489683106541634, + 0.008408835157752037, + 0.010267090052366257, + 0.0038623579312115908, + 0.008503730408847332, + 0.007307596039026976, + 0.011772439815104008, + 0.003886425867676735, + 0.008005228824913502, + 0.0066207596100866795, + 0.007216858211904764, + 0.0133364237844944, + 0.011271713301539421, + 0.012942786328494549, + 0.017483791336417198, + 0.006432714872062206, + 0.009771325625479221, + 0.009107791818678379, + 0.012130700051784515, + 0.003997426945716143, + 0.005239699501544237, + 0.005754785146564245, + 0.008218038827180862, + 0.004142834804952145, + 0.003386828815564513, + 0.005072755739092827, + 0.005647053010761738 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 3 + ], + "selected_reward": -0.018182622268795967 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium close-up of a young woman leaning against a damp brick wall in a narrow alley?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a dark gray, slightly oversized hooded sweatshirt with wet fabric?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are large tears visible flowing from the woman's eyes down her cheeks and dripping onto her garment?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by a single dim yellow street lamp at the far end of the alley?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the overall color palette dominated by cold blue-gray tones with a dim yellow light source?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.000390215078368783 + ], + "opposite_logprob_tokens": [ + -7.87539005279541 + ], + "logprob_reward": -0.000390215078368783, + "yes_no_margin_reward": 7.875, + "token_credit_reward": -0.000390215078368783, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.42254722118377686, + 0.0739595964550972, + 0.09229563921689987, + 0.047783106565475464, + 0.06409182399511337, + 0.0427122637629509, + 0.06354960054159164, + 0.06072879582643509, + 0.08302143216133118, + 0.04931051284074783 + ], + "credit_map": [ + 0.0477740652859211, + 0.04559953510761261, + 0.0573500320315361, + 0.08866680413484573, + 0.03710068017244339, + 0.041745297610759735, + 0.04596790298819542, + 0.05834294483065605, + 0.007044659927487373, + 0.006842232774943113, + 0.011176498606801033, + 0.012930812314152718, + 0.008595882914960384, + 0.006467362400144339, + 0.011644723825156689, + 0.009257420897483826, + 0.011780854314565659, + 0.0110926553606987, + 0.013232503086328506, + 0.015862079337239265, + 0.00738182570785284, + 0.010988517664372921, + 0.011159996502101421, + 0.010797207243740559, + 0.0055580441839993, + 0.005660949274897575, + 0.0065187555737793446, + 0.006786393467336893, + 0.005292793270200491, + 0.005845497362315655, + 0.006254254840314388, + 0.0058664195239543915, + 0.008230977691709995, + 0.006721190642565489, + 0.007283156272023916, + 0.008172895759344101, + 0.006766441743820906, + 0.008595694787800312, + 0.009698070585727692, + 0.008623388595879078, + 0.005990809760987759, + 0.00449673505499959, + 0.007177482359111309, + 0.006174252834171057, + 0.003829441498965025, + 0.003015035530552268, + 0.006041056476533413, + 0.005987452808767557, + 0.013684180565178394, + 0.00855687353760004, + 0.006954146083444357, + 0.007203306537121534, + 0.005914062261581421, + 0.0050863344222307205, + 0.008021606132388115, + 0.00812908448278904, + 0.016038089990615845, + 0.0076522063463926315, + 0.008525594137609005, + 0.0080336919054389, + 0.003407249692827463, + 0.004385556560009718, + 0.005618006456643343, + 0.0070684002712368965, + 0.009108435362577438, + 0.009772391058504581, + 0.012006966397166252, + 0.018431873992085457, + 0.006018004380166531, + 0.008745853789150715, + 0.008251174353063107, + 0.010686732828617096, + 0.008544809184968472, + 0.0049776239320635796, + 0.005652793683111668, + 0.007397288456559181, + 0.004919931758195162, + 0.004776686429977417, + 0.006665907334536314, + 0.0063754720613360405 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 8 + ], + "selected_reward": -0.000390215078368783 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "medium close-up shot from a level angle, its composition steady and composed, transports us to a night of relentless, cold rain. At the center of the frame, a solitary young woman leans limply against a damp, old brick wall in a narrow back alley. Her long, dark hair is completely soaked by the rain, strands plastered to her cold-white cheeks and neck. She wears a dark gray, slightly oversized hooded sweatshirt, the fabric dappled in the dim light with varying shades of wetness. Her shoulders slump, her gaze fixed vacantly forward, her eyes reddened as large, uncontrollable tears slip silently from her hollow irises. They flow down the contours of her cheeks, mingling with the steadily falling icy raindrops before pooling at her chin and dripping onto the lapels of her dark garment. A lone street lamp at the far end of the alleyway provides the sole source of light, its dim yellow light fragmented by the dense curtain of rain, casting long, twisting shadows across the slippery, mossy brick walls and the waterlogged stone pavement. The painting's dominant hue is a cold, austere blue-gray, with only the faint light bringing a barely perceptible hint of warmth, which only serves to accentuate the character's loneliness and coldness, creating an atmosphere of oppression and profound sadness, forgotten by the world.", + "seed": 11653, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000056/group_2/index_5858_seed_11653_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": -0.04480297958943993, + "vlm_logprob": -0.04480297958943993, + "vlm_yes_no_margin": 4.4, + "vlm_token_credit": -0.04480297958943993, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium close-up of a young woman leaning against a damp brick wall in a narrow alley?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a dark gray, slightly oversized hooded sweatshirt with wet fabric?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are large tears visible flowing from the woman's eyes down her cheeks and dripping onto her garment?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by a single dim yellow street lamp at the far end of the alley?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the overall color palette dominated by cold blue-gray tones with a dim yellow light source?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02584923803806305 + ], + "opposite_logprob_tokens": [ + -4.150849342346191 + ], + "logprob_reward": -0.02584923803806305, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.02584923803806305, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3906453549861908, + 0.1259802132844925, + 0.056679900735616684, + 0.029530474916100502, + 0.07392245531082153, + 0.14233899116516113, + 0.06762731075286865, + 0.05205994471907616, + 0.03147463500499725, + 0.02974071353673935 + ], + "credit_map": [ + 0.029465027153491974, + 0.06907704472541809, + 0.04099968075752258, + 0.03250474855303764, + 0.05089753866195679, + 0.05512710288167, + 0.05319284275174141, + 0.05938141047954559, + 0.012512890622019768, + 0.01689915359020233, + 0.01806824840605259, + 0.008280700072646141, + 0.016187332570552826, + 0.016949888318777084, + 0.018042299896478653, + 0.01903969794511795, + 0.006886272691190243, + 0.006616414524614811, + 0.007066831458359957, + 0.006471091881394386, + 0.0057275486178696156, + 0.005018826574087143, + 0.01045189332216978, + 0.008441020734608173, + 0.005930942948907614, + 0.002621858613565564, + 0.004283561371266842, + 0.0027221154887229204, + 0.003298272145912051, + 0.0025143104139715433, + 0.003694827202707529, + 0.004464586731046438, + 0.0076166605576872826, + 0.00571027398109436, + 0.007793624885380268, + 0.0066727157682180405, + 0.010706025175750256, + 0.013747293502092361, + 0.011507249437272549, + 0.010168607346713543, + 0.012976962141692638, + 0.007819065824151039, + 0.01702253893017769, + 0.007700737565755844, + 0.020987609401345253, + 0.039109453558921814, + 0.01570088416337967, + 0.021021729335188866, + 0.005387819372117519, + 0.012434860691428185, + 0.006825444754213095, + 0.00721451872959733, + 0.007893805392086506, + 0.010569244623184204, + 0.00878617912530899, + 0.008515440858900547, + 0.0053714304231107235, + 0.007410263177007437, + 0.007090020924806595, + 0.005345900077372789, + 0.005425778683274984, + 0.008507132530212402, + 0.0050391582772135735, + 0.007870263420045376, + 0.003417255124077201, + 0.004560955334454775, + 0.004307831637561321, + 0.0031343826558440924, + 0.0036360847298055887, + 0.004004239570349455, + 0.003909197170287371, + 0.004504688549786806, + 0.0032958057709038258, + 0.005305520258843899, + 0.004762697499245405, + 0.00370990764349699, + 0.002975460374727845, + 0.0031481843907386065, + 0.002780681708827615, + 0.0037624549586325884 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 1 + ], + "selected_reward": -0.02584923803806305 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium close-up of a young woman leaning against a damp brick wall in a narrow alley?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a dark gray, slightly oversized hooded sweatshirt with wet fabric?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are large tears visible flowing from the woman's eyes down her cheeks and dripping onto her garment?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by a single dim yellow street lamp at the far end of the alley?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the overall color palette dominated by cold blue-gray tones with a dim yellow light source?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011102446354925632 + ], + "opposite_logprob_tokens": [ + -4.511102676391602 + ], + "logprob_reward": -0.011102446354925632, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011102446354925632, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2821347713470459, + 0.1548999547958374, + 0.0702860876917839, + 0.03078857809305191, + 0.07950029522180557, + 0.13558387756347656, + 0.15164843201637268, + 0.035461220890283585, + 0.02605689875781536, + 0.03363987058401108 + ], + "credit_map": [ + 0.02238994836807251, + 0.04291125759482384, + 0.028149722144007683, + 0.021513383835554123, + 0.04505600035190582, + 0.044141996651887894, + 0.03773857280611992, + 0.040233880281448364, + 0.018570711836218834, + 0.028662346303462982, + 0.020535456016659737, + 0.01119864173233509, + 0.018560774624347687, + 0.01957371085882187, + 0.016530156135559082, + 0.021268166601657867, + 0.00958926510065794, + 0.010472109541296959, + 0.009895384311676025, + 0.007661930751055479, + 0.006641683168709278, + 0.005976149346679449, + 0.010328189469873905, + 0.009721377864480019, + 0.00536158075556159, + 0.003276723437011242, + 0.0039912862703204155, + 0.00336680025793612, + 0.0032039235811680555, + 0.003997684922069311, + 0.0030927613843232393, + 0.004497816786170006, + 0.00751712080091238, + 0.005943540018051863, + 0.007026747800409794, + 0.008255834691226482, + 0.011120003648102283, + 0.016786526888608932, + 0.011291990987956524, + 0.011558525264263153, + 0.013157631270587444, + 0.013647060841321945, + 0.01754421554505825, + 0.008805662393569946, + 0.01712975837290287, + 0.0334305614233017, + 0.01175510324537754, + 0.020113881677389145, + 0.01133431401103735, + 0.008229801431298256, + 0.01048772968351841, + 0.013073389418423176, + 0.027237415313720703, + 0.03792835399508476, + 0.021398471668362617, + 0.021958952769637108, + 0.003188114147633314, + 0.0033440443221479654, + 0.0039755357429385185, + 0.0042105671018362045, + 0.0052339243702590466, + 0.006665784865617752, + 0.004140020813792944, + 0.004703233018517494, + 0.0041123563423752785, + 0.0021772936452180147, + 0.003827346721664071, + 0.003085747128352523, + 0.0029572814237326384, + 0.0032851113937795162, + 0.002920757047832012, + 0.0036910055205225945, + 0.00444447249174118, + 0.0037515172734856606, + 0.003802055260166526, + 0.0030761132948100567, + 0.00417289137840271, + 0.004943277221173048, + 0.004828977398574352, + 0.004620568361133337 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 6 + ], + "selected_reward": -0.011102446354925632 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium close-up of a young woman leaning against a damp brick wall in a narrow alley?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a dark gray, slightly oversized hooded sweatshirt with wet fabric?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are large tears visible flowing from the woman's eyes down her cheeks and dripping onto her garment?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by a single dim yellow street lamp at the far end of the alley?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the overall color palette dominated by cold blue-gray tones with a dim yellow light source?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.16038241982460022 + ], + "opposite_logprob_tokens": [ + -1.9103823900222778 + ], + "logprob_reward": -0.16038241982460022, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.16038241982460022, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.30912575125694275, + 0.08711777627468109, + 0.07540000230073929, + 0.0341314822435379, + 0.23622667789459229, + 0.08316972106695175, + 0.06269866228103638, + 0.055326417088508606, + 0.02613421343266964, + 0.030669305473566055 + ], + "credit_map": [ + 0.027677081525325775, + 0.06414433568716049, + 0.03557029739022255, + 0.02335250936448574, + 0.0412982776761055, + 0.041930776089429855, + 0.03209845721721649, + 0.043054014444351196, + 0.010986598208546638, + 0.008737297728657722, + 0.01242405641824007, + 0.013561305589973927, + 0.01006481796503067, + 0.008646814152598381, + 0.010817324742674828, + 0.011879565194249153, + 0.008440548554062843, + 0.017084022983908653, + 0.009982880204916, + 0.009027776308357716, + 0.0061973100528120995, + 0.006830787286162376, + 0.007047906517982483, + 0.010788767598569393, + 0.0041328431107103825, + 0.002615983597934246, + 0.003796094562858343, + 0.0035761615727096796, + 0.004547016229480505, + 0.004596562124788761, + 0.005287236534059048, + 0.005579586140811443, + 0.023854942992329597, + 0.012281043455004692, + 0.019914045929908752, + 0.021947888657450676, + 0.03648928180336952, + 0.0461505725979805, + 0.04102766886353493, + 0.034561265259981155, + 0.01065317913889885, + 0.0061129420064389706, + 0.012744059786200523, + 0.008721948601305485, + 0.01024344377219677, + 0.011483979411423206, + 0.011639312840998173, + 0.011570862494409084, + 0.005221965257078409, + 0.0036692828871309757, + 0.005955131258815527, + 0.006562212016433477, + 0.00917273759841919, + 0.014388218522071838, + 0.008954857476055622, + 0.008774260990321636, + 0.006069565191864967, + 0.005322284065186977, + 0.006220625247806311, + 0.005173988174647093, + 0.006311346776783466, + 0.012377789244055748, + 0.005787075962871313, + 0.008063748478889465, + 0.004705055151134729, + 0.00385535741224885, + 0.003878396237269044, + 0.003071243641898036, + 0.002122696489095688, + 0.00242980825714767, + 0.00230040168389678, + 0.0037712575867772102, + 0.004567052237689495, + 0.0021997110452502966, + 0.004048261325806379, + 0.002951269270852208, + 0.004299506079405546, + 0.0038355442229658365, + 0.004381234757602215, + 0.004386727698147297 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 1 + ], + "selected_reward": -0.16038241982460022 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium close-up of a young woman leaning against a damp brick wall in a narrow alley?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a dark gray, slightly oversized hooded sweatshirt with wet fabric?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are large tears visible flowing from the woman's eyes down her cheeks and dripping onto her garment?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by a single dim yellow street lamp at the far end of the alley?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the overall color palette dominated by cold blue-gray tones with a dim yellow light source?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026335861533880234 + ], + "opposite_logprob_tokens": [ + -3.6513359546661377 + ], + "logprob_reward": -0.026335861533880234, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026335861533880234, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10119971632957458, + 0.12266658991575241, + 0.44220539927482605, + 0.1716422140598297, + 0.04293477162718773, + 0.03970684856176376, + 0.034231241792440414, + 0.015179787762463093, + 0.010381687432527542, + 0.019851822406053543 + ], + "credit_map": [ + 0.013187753967940807, + 0.01985955238342285, + 0.01583017408847809, + 0.012922678142786026, + 0.007254301104694605, + 0.008071020245552063, + 0.009607117623090744, + 0.014467115513980389, + 0.012793654575943947, + 0.02527048997581005, + 0.017677931115031242, + 0.019586484879255295, + 0.007244979031383991, + 0.008714834228157997, + 0.008387740701436996, + 0.02299046888947487, + 0.040782466530799866, + 0.1304204761981964, + 0.058706555515527725, + 0.048534926027059555, + 0.030818289145827293, + 0.0385083369910717, + 0.035674117505550385, + 0.05876021832227707, + 0.018449192866683006, + 0.05186689272522926, + 0.018093258142471313, + 0.015038169920444489, + 0.014555521309375763, + 0.015206349082291126, + 0.016454730182886124, + 0.021978095173835754, + 0.0048002093099057674, + 0.009373705834150314, + 0.004794662818312645, + 0.005865576211363077, + 0.005019326228648424, + 0.0034134897869080305, + 0.00392100540921092, + 0.005746795795857906, + 0.0038937004283070564, + 0.003253389149904251, + 0.005793900229036808, + 0.004320189356803894, + 0.004606989677995443, + 0.008131058886647224, + 0.0032575998920947313, + 0.006450023967772722, + 0.003943526651710272, + 0.009753962978720665, + 0.004497769754379988, + 0.003830929985269904, + 0.0026214017998427153, + 0.002585401525720954, + 0.0024711082223802805, + 0.00452714366838336, + 0.0013586533023044467, + 0.0022752583026885986, + 0.00182991789188236, + 0.0017294369172304869, + 0.0015713372267782688, + 0.003169546602293849, + 0.0012621149653568864, + 0.001983522204682231, + 0.0011904884595423937, + 0.0012867662589997053, + 0.0014307881938293576, + 0.0015360208926722407, + 0.0014088063035160303, + 0.001007995568215847, + 0.0011461907997727394, + 0.0013746307231485844, + 0.0020809932611882687, + 0.0018845396116375923, + 0.0029716072604060173, + 0.00311745866201818, + 0.0015952093526721, + 0.002860610606148839, + 0.0018333167536184192, + 0.0035080888774245977 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 3, + 1 + ], + "selected_reward": -0.026335861533880234 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium close-up of a young woman leaning against a damp brick wall in a narrow alley?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a dark gray, slightly oversized hooded sweatshirt with wet fabric?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are large tears visible flowing from the woman's eyes down her cheeks and dripping onto her garment?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by a single dim yellow street lamp at the far end of the alley?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the overall color palette dominated by cold blue-gray tones with a dim yellow light source?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0003449321957305074 + ], + "opposite_logprob_tokens": [ + -8.000345230102539 + ], + "logprob_reward": -0.0003449321957305074, + "yes_no_margin_reward": 8.0, + "token_credit_reward": -0.0003449321957305074, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23691552877426147, + 0.07098009437322617, + 0.20920543372631073, + 0.0961202010512352, + 0.0865534320473671, + 0.06703881919384003, + 0.08795598894357681, + 0.0718260109424591, + 0.04023022577166557, + 0.03317425027489662 + ], + "credit_map": [ + 0.02671738527715206, + 0.04299021139740944, + 0.03445614129304886, + 0.0236051045358181, + 0.022377923130989075, + 0.030964665114879608, + 0.021058214828372, + 0.03474590554833412, + 0.007611255627125502, + 0.008674860931932926, + 0.009024309925734997, + 0.013503740541636944, + 0.005728306248784065, + 0.0065402621403336525, + 0.007080591283738613, + 0.012816771864891052, + 0.021300917491316795, + 0.05676696449518204, + 0.031138187274336815, + 0.02336493507027626, + 0.014156169258058071, + 0.01767486147582531, + 0.018354225903749466, + 0.0264491718262434, + 0.013541493564844131, + 0.025165585801005363, + 0.010623430833220482, + 0.007855904288589954, + 0.008257300592958927, + 0.00956899207085371, + 0.01000323798507452, + 0.011104261502623558, + 0.009495711885392666, + 0.009163892827928066, + 0.008279385045170784, + 0.009105641394853592, + 0.011862508952617645, + 0.012956740334630013, + 0.014173452742397785, + 0.011516098864376545, + 0.006545620504766703, + 0.00560015020892024, + 0.009519001469016075, + 0.0070064617320895195, + 0.0075665866024792194, + 0.011871960014104843, + 0.009661216288805008, + 0.009267818182706833, + 0.006883503869175911, + 0.013609781861305237, + 0.007395218592137098, + 0.008759541437029839, + 0.01323719136416912, + 0.015269596129655838, + 0.011652376502752304, + 0.011148778721690178, + 0.005169950425624847, + 0.006268164608627558, + 0.007286699954420328, + 0.006283463444560766, + 0.009793715551495552, + 0.019019370898604393, + 0.007478346582502127, + 0.010526305064558983, + 0.005207651760429144, + 0.005955985281616449, + 0.005820684600621462, + 0.004200756549835205, + 0.005130165722221136, + 0.004413717892020941, + 0.004476593341678381, + 0.00502467155456543, + 0.0029912188183516264, + 0.003682367969304323, + 0.003736666403710842, + 0.0045947544276714325, + 0.004551880527287722, + 0.0035255493130534887, + 0.00564356055110693, + 0.004448249004781246 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 3 + ], + "selected_reward": -0.0003449321957305074 + } + ], + "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/0000056/group_3/index_tensor([5644, 5644])_group_3_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000056/group_3/index_tensor([5644, 5644])_group_3_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..77cd11e6a67b87b41365769677d8b87a0f8aedae --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000056/group_3/index_tensor([5644, 5644])_group_3_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "dynamic shot, shot from a low, upward angle, depicts the camera in a slow, powerful arc, capturing the subject in a meticulously rendered 3D style. Before the majestic Water Curtain Cave in the Flower-Fruit Mountain, the Monkey King, Sun Wukong, stands in a steady horse stance, legs slightly bent. He clad in intricately carved golden chainmail armor, its plates engraved with intricate cloud patterns that reflect the surrounding light. Atop his head, a phoenix-wing purple-gold crown, two long pheasant feathers flutter majestically with his movements. His golden hair, rendered in physical rendering, is distinct and shimmers with a silken sheen. His fiery eyes radiate an unyielding golden light, his facial expression stern and filled with fighting spirit. He clutches the legendary golden cudgel in both hands. The cudgel's shaft doesn't simply glow; instead, it emanates a fiery, molten-like energy from within. Golden hoops, inlaid at each end, appear and disappear beneath the surge of energy. Behind him, a vast waterfall cascades down from a towering cliff like the Milky Way, forming a vast curtain of water. Countless water droplets splash through the air, refracted by the light of the golden cudgel, creating a rainbow of colors. With a low roar, Sun Wukong thrust his golden cudgel in a perfect golden arc from the right rear to the left front. The cudgel slices through the air with a deep, powerful whistling sound, and the resulting air currents even momentarily distort the surrounding waterfall. His movements are imbued with explosive power and control. The cudgel's blazing light serves as the scene's primary source of light, illuminating every detail of his armor and the damp rocks, creating a mythical, epic atmosphere filled with power and tension.", + "seed": 26609, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000056/group_3/index_5644_seed_26609_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": -0.04916837997734547, + "vlm_logprob": -0.04916837997734547, + "vlm_yes_no_margin": 3.35, + "vlm_token_credit": -0.04916837997734547, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject Sun Wukong (the Monkey King) depicted in a steady horse stance with legs slightly bent?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is Sun Wukong wearing intricately carved golden chainmail armor with cloud patterns and a phoenix-wing purple-gold crown with two long pheasant feathers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene set before the Water Curtain Cave in the Flower-Fruit Mountain with a vast waterfall cascading down a towering cliff?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does Sun Wukong thrust a golden cudgel in a perfect golden arc from the right rear to the left front, causing the waterfall to distort?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot filmed from a low, upward angle with the camera moving in a slow, powerful arc?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.12979555130004883 + ], + "opposite_logprob_tokens": [ + -2.129795551300049 + ], + "logprob_reward": -0.12979555130004883, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.12979555130004883, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07411764562129974, + 0.06768126040697098, + 0.12123297899961472, + 0.2115434855222702, + 0.041799288243055344, + 0.06651866436004639, + 0.10300479829311371, + 0.09583240002393723, + 0.10058654099702835, + 0.11768291145563126 + ], + "credit_map": [ + 0.008474960923194885, + 0.011531193740665913, + 0.0054375967010855675, + 0.007430434226989746, + 0.008876398205757141, + 0.010894863866269588, + 0.008460411801934242, + 0.013011790812015533, + 0.006109409034252167, + 0.005078161600977182, + 0.0076774051412940025, + 0.005600553471595049, + 0.005365034099668264, + 0.012406990863382816, + 0.004802575334906578, + 0.020641125738620758, + 0.018072383478283882, + 0.011802466586232185, + 0.008289499208331108, + 0.0157465897500515, + 0.012870965525507927, + 0.017035769298672676, + 0.010628987103700638, + 0.026786327362060547, + 0.05331934988498688, + 0.013527167029678822, + 0.03696748986840248, + 0.011950642801821232, + 0.013666968792676926, + 0.04036429151892662, + 0.009437131695449352, + 0.032310448586940765, + 0.003511861665174365, + 0.006559478584676981, + 0.0049648890271782875, + 0.006357777863740921, + 0.005435661878436804, + 0.0043319109827280045, + 0.0061031365767121315, + 0.004534571431577206, + 0.00923294946551323, + 0.006614225450903177, + 0.007599970791488886, + 0.008465939201414585, + 0.009240723215043545, + 0.007727478165179491, + 0.007223399356007576, + 0.01041397638618946, + 0.020644977688789368, + 0.007513777818530798, + 0.0124250752851367, + 0.011034569703042507, + 0.012431896291673183, + 0.016299065202474594, + 0.00977271143347025, + 0.012882730923593044, + 0.019100677222013474, + 0.007274797651916742, + 0.010828169994056225, + 0.015063246712088585, + 0.010076187551021576, + 0.009062163531780243, + 0.010407294146716595, + 0.01401986088603735, + 0.010946697555482388, + 0.014943698421120644, + 0.00951208546757698, + 0.014193840324878693, + 0.01307876966893673, + 0.011654927395284176, + 0.013635177165269852, + 0.012621342204511166, + 0.0159130971878767, + 0.01689363270998001, + 0.012879034504294395, + 0.010223894380033016, + 0.027069535106420517, + 0.010793727822601795, + 0.014673732221126556, + 0.009236256591975689 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 2, + 9 + ], + "selected_reward": -0.12979555130004883 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject Sun Wukong (the Monkey King) depicted in a steady horse stance with legs slightly bent?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is Sun Wukong wearing intricately carved golden chainmail armor with cloud patterns and a phoenix-wing purple-gold crown with two long pheasant feathers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene set before the Water Curtain Cave in the Flower-Fruit Mountain with a vast waterfall cascading down a towering cliff?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does Sun Wukong thrust a golden cudgel in a perfect golden arc from the right rear to the left front, causing the waterfall to distort?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot filmed from a low, upward angle with the camera moving in a slow, powerful arc?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011103389784693718 + ], + "opposite_logprob_tokens": [ + -4.51110315322876 + ], + "logprob_reward": -0.011103389784693718, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011103389784693718, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1085435003042221, + 0.07982656359672546, + 0.055147770792245865, + 0.309114933013916, + 0.038946766406297684, + 0.06397634744644165, + 0.10362750291824341, + 0.06574161350727081, + 0.07780283689498901, + 0.09727221727371216 + ], + "credit_map": [ + 0.012366828508675098, + 0.014988134615123272, + 0.013082698918879032, + 0.009075569920241833, + 0.011131448671221733, + 0.01834789104759693, + 0.010336548089981079, + 0.019214369356632233, + 0.0070207202807068825, + 0.005429481156170368, + 0.010588245466351509, + 0.0063309804536402225, + 0.005160611122846603, + 0.014611761085689068, + 0.005668119993060827, + 0.02501663751900196, + 0.0070542991161346436, + 0.0076596918515861034, + 0.00349420472048223, + 0.006352028343826532, + 0.0056767244823277, + 0.008131256327033043, + 0.004822168964892626, + 0.011957386508584023, + 0.08761069178581238, + 0.017509358003735542, + 0.06038763001561165, + 0.015474666841328144, + 0.015413605608046055, + 0.057077884674072266, + 0.011027971282601357, + 0.04461308568716049, + 0.005525273736566305, + 0.004635569639503956, + 0.004593546036630869, + 0.006541214883327484, + 0.0041499026119709015, + 0.003909134306013584, + 0.004102937411516905, + 0.005489180330187082, + 0.010135997086763382, + 0.0072690085507929325, + 0.007320288568735123, + 0.008846743032336235, + 0.008990981616079807, + 0.0063505326397717, + 0.0066734435968101025, + 0.008389340713620186, + 0.016287419945001602, + 0.007824678905308247, + 0.01094629243016243, + 0.01176886260509491, + 0.01333362516015768, + 0.019300511106848717, + 0.010758141987025738, + 0.013407954946160316, + 0.009859825484454632, + 0.005783724132925272, + 0.0064774020574986935, + 0.011366842314600945, + 0.0074684531427919865, + 0.006708973087370396, + 0.008282983675599098, + 0.009793396107852459, + 0.009713349863886833, + 0.011733204126358032, + 0.007198184262961149, + 0.012863311916589737, + 0.010393680073320866, + 0.007100888527929783, + 0.00969026517122984, + 0.00910994317382574, + 0.01147898193448782, + 0.014385225251317024, + 0.010798966512084007, + 0.008163689635694027, + 0.024469854310154915, + 0.007944059558212757, + 0.012386975809931755, + 0.007644443307071924 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 6 + ], + "selected_reward": -0.011103389784693718 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject Sun Wukong (the Monkey King) depicted in a steady horse stance with legs slightly bent?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is Sun Wukong wearing intricately carved golden chainmail armor with cloud patterns and a phoenix-wing purple-gold crown with two long pheasant feathers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene set before the Water Curtain Cave in the Flower-Fruit Mountain with a vast waterfall cascading down a towering cliff?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does Sun Wukong thrust a golden cudgel in a perfect golden arc from the right rear to the left front, causing the waterfall to distort?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot filmed from a low, upward angle with the camera moving in a slow, powerful arc?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04865863919258118 + ], + "opposite_logprob_tokens": [ + -3.048658609390259 + ], + "logprob_reward": -0.04865863919258118, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04865863919258118, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.051621291786432266, + 0.18469956517219543, + 0.08461914211511612, + 0.1853608340024948, + 0.05060219392180443, + 0.04310877248644829, + 0.21019092202186584, + 0.11675245314836502, + 0.046396009624004364, + 0.026648767292499542 + ], + "credit_map": [ + 0.008199367672204971, + 0.00791203510016203, + 0.00411161407828331, + 0.004941393621265888, + 0.005769264884293079, + 0.007470753975212574, + 0.005685256328433752, + 0.007531612180173397, + 0.02030595764517784, + 0.009472220204770565, + 0.028757626190781593, + 0.011989710852503777, + 0.00799141451716423, + 0.03577004745602608, + 0.009662728756666183, + 0.06074985861778259, + 0.01666082628071308, + 0.0075042336247861385, + 0.006282464135438204, + 0.010638518258929253, + 0.005896181333810091, + 0.012340277433395386, + 0.0064186411909759045, + 0.018877997994422913, + 0.04262292757630348, + 0.012525323778390884, + 0.028853558003902435, + 0.013438092544674873, + 0.013557994738221169, + 0.0349433571100235, + 0.00899853277951479, + 0.030421042814850807, + 0.005984180141240358, + 0.0057957787066698074, + 0.005121816415339708, + 0.0072132982313632965, + 0.0052428655326366425, + 0.005887800361961126, + 0.006639487575739622, + 0.008716966956853867, + 0.008596179075539112, + 0.004066805355250835, + 0.005726594477891922, + 0.0044275205582380295, + 0.004784011282026768, + 0.004577035550028086, + 0.004447530955076218, + 0.006483097560703754, + 0.03806143254041672, + 0.01574532687664032, + 0.023633120581507683, + 0.024882204830646515, + 0.025268474593758583, + 0.03462249040603638, + 0.021563541144132614, + 0.026414332911372185, + 0.021299665793776512, + 0.009218909777700901, + 0.012684700079262257, + 0.017770390957593918, + 0.0118330093100667, + 0.012716379016637802, + 0.014614931307733059, + 0.016614466905593872, + 0.005951111204922199, + 0.004806423559784889, + 0.004178561270236969, + 0.005530828610062599, + 0.005501783452928066, + 0.0052231065928936005, + 0.0071987877599895, + 0.008005402982234955, + 0.00348352431319654, + 0.0032155581284314394, + 0.002625829540193081, + 0.0019601641688495874, + 0.005883628968149424, + 0.0033905087038874626, + 0.0032401839271187782, + 0.002849370241165161 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 3, + 1 + ], + "selected_reward": -0.04865863919258118 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject Sun Wukong (the Monkey King) depicted in a steady horse stance with legs slightly bent?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is Sun Wukong wearing intricately carved golden chainmail armor with cloud patterns and a phoenix-wing purple-gold crown with two long pheasant feathers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene set before the Water Curtain Cave in the Flower-Fruit Mountain with a vast waterfall cascading down a towering cliff?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does Sun Wukong thrust a golden cudgel in a perfect golden arc from the right rear to the left front, causing the waterfall to distort?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot filmed from a low, upward angle with the camera moving in a slow, powerful arc?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.038099825382232666 + ], + "opposite_logprob_tokens": [ + -3.288099765777588 + ], + "logprob_reward": -0.038099825382232666, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.038099825382232666, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10108010470867157, + 0.08474951982498169, + 0.05793635547161102, + 0.19625070691108704, + 0.08508675545454025, + 0.09564439952373505, + 0.09214712679386139, + 0.049904126673936844, + 0.15483587980270386, + 0.08236504346132278 + ], + "credit_map": [ + 0.015455406159162521, + 0.013356628827750683, + 0.012116353027522564, + 0.008438619785010815, + 0.008905738592147827, + 0.015854623168706894, + 0.009141417220234871, + 0.01781131885945797, + 0.01850094273686409, + 0.009210537187755108, + 0.0067144157364964485, + 0.008930109441280365, + 0.008447009138762951, + 0.011225966736674309, + 0.008800279349088669, + 0.01292025949805975, + 0.008224648423492908, + 0.007519283331930637, + 0.003983824979513884, + 0.008484948426485062, + 0.0059018563479185104, + 0.007791373413056135, + 0.005279946140944958, + 0.010750474408268929, + 0.04324890300631523, + 0.014573815278708935, + 0.027674345299601555, + 0.01748119294643402, + 0.014039136469364166, + 0.030984152108430862, + 0.011758591048419476, + 0.03649057075381279, + 0.010851647704839706, + 0.009629842825233936, + 0.008803240954875946, + 0.012892818078398705, + 0.009259392507374287, + 0.008595943450927734, + 0.010955166071653366, + 0.014098708517849445, + 0.013786256313323975, + 0.009117537178099155, + 0.01008655782788992, + 0.011491279117763042, + 0.011853507719933987, + 0.011743479408323765, + 0.01081826351583004, + 0.01674751378595829, + 0.015818972140550613, + 0.0074738385155797005, + 0.010594392195343971, + 0.010964987799525261, + 0.010516430251300335, + 0.01119998749345541, + 0.011783784255385399, + 0.013794733211398125, + 0.008387885987758636, + 0.004932247567921877, + 0.005238895770162344, + 0.006967199966311455, + 0.005232150666415691, + 0.005442918743938208, + 0.005560346879065037, + 0.008142481558024883, + 0.013895956799387932, + 0.029109811410307884, + 0.013995765708386898, + 0.02581479586660862, + 0.02387744002044201, + 0.011490637436509132, + 0.018802614882588387, + 0.017848854884505272, + 0.010100153274834156, + 0.00920011568814516, + 0.007956380024552345, + 0.006809618324041367, + 0.015777532011270523, + 0.011507213115692139, + 0.009633409790694714, + 0.011380620300769806 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 8, + 0 + ], + "selected_reward": -0.038099825382232666 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject Sun Wukong (the Monkey King) depicted in a steady horse stance with legs slightly bent?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is Sun Wukong wearing intricately carved golden chainmail armor with cloud patterns and a phoenix-wing purple-gold crown with two long pheasant feathers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene set before the Water Curtain Cave in the Flower-Fruit Mountain with a vast waterfall cascading down a towering cliff?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does Sun Wukong thrust a golden cudgel in a perfect golden arc from the right rear to the left front, causing the waterfall to distort?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot filmed from a low, upward angle with the camera moving in a slow, powerful arc?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018184494227170944 + ], + "opposite_logprob_tokens": [ + -4.018184661865234 + ], + "logprob_reward": -0.018184494227170944, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018184494227170944, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08059673756361008, + 0.11111394315958023, + 0.06728857755661011, + 0.21996858716011047, + 0.05557892471551895, + 0.06921903789043427, + 0.08555512875318527, + 0.07590631395578384, + 0.10980430990457535, + 0.12496845424175262 + ], + "credit_map": [ + 0.013025617226958275, + 0.009135352447628975, + 0.010777123272418976, + 0.007742063608020544, + 0.006727056577801704, + 0.012743453495204449, + 0.006431582849472761, + 0.01401448529213667, + 0.008583075366914272, + 0.007343426812440157, + 0.01907954551279545, + 0.006104866974055767, + 0.008380454033613205, + 0.020461415871977806, + 0.00867423601448536, + 0.032486919313669205, + 0.006723849102854729, + 0.008688660338521004, + 0.004344845190644264, + 0.009336454793810844, + 0.00802332442253828, + 0.009412436746060848, + 0.006148115731775761, + 0.014610894955694675, + 0.058285851031541824, + 0.011999247595667839, + 0.0402955636382103, + 0.010022473521530628, + 0.011544475331902504, + 0.04436194524168968, + 0.008101601153612137, + 0.03535742685198784, + 0.009682153351604939, + 0.00406241649761796, + 0.007953891530632973, + 0.00581477303057909, + 0.0070963832549750805, + 0.006646701600402594, + 0.008740570396184921, + 0.005582030862569809, + 0.011877999641001225, + 0.006583655718713999, + 0.009431605227291584, + 0.006964996922761202, + 0.007377429399639368, + 0.007832173258066177, + 0.007609148044139147, + 0.011542028747498989, + 0.018853621557354927, + 0.007566316984593868, + 0.011630446650087833, + 0.00941428728401661, + 0.008860251866281033, + 0.01209985464811325, + 0.00802771095186472, + 0.009102638810873032, + 0.015895597636699677, + 0.006236911751329899, + 0.008727083913981915, + 0.011631513945758343, + 0.006901517976075411, + 0.006282768212258816, + 0.008642340078949928, + 0.011588579043745995, + 0.012094155885279179, + 0.017697954550385475, + 0.0103759141638875, + 0.015103567391633987, + 0.014945684932172298, + 0.010429203510284424, + 0.014893744140863419, + 0.014264081604778767, + 0.014930974692106247, + 0.017581790685653687, + 0.013671264052391052, + 0.008948354050517082, + 0.03263787552714348, + 0.011004219762980938, + 0.015599293634295464, + 0.010594678111374378 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 9, + 1 + ], + "selected_reward": -0.018184494227170944 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "dynamic shot, shot from a low, upward angle, depicts the camera in a slow, powerful arc, capturing the subject in a meticulously rendered 3D style. Before the majestic Water Curtain Cave in the Flower-Fruit Mountain, the Monkey King, Sun Wukong, stands in a steady horse stance, legs slightly bent. He clad in intricately carved golden chainmail armor, its plates engraved with intricate cloud patterns that reflect the surrounding light. Atop his head, a phoenix-wing purple-gold crown, two long pheasant feathers flutter majestically with his movements. His golden hair, rendered in physical rendering, is distinct and shimmers with a silken sheen. His fiery eyes radiate an unyielding golden light, his facial expression stern and filled with fighting spirit. He clutches the legendary golden cudgel in both hands. The cudgel's shaft doesn't simply glow; instead, it emanates a fiery, molten-like energy from within. Golden hoops, inlaid at each end, appear and disappear beneath the surge of energy. Behind him, a vast waterfall cascades down from a towering cliff like the Milky Way, forming a vast curtain of water. Countless water droplets splash through the air, refracted by the light of the golden cudgel, creating a rainbow of colors. With a low roar, Sun Wukong thrust his golden cudgel in a perfect golden arc from the right rear to the left front. The cudgel slices through the air with a deep, powerful whistling sound, and the resulting air currents even momentarily distort the surrounding waterfall. His movements are imbued with explosive power and control. The cudgel's blazing light serves as the scene's primary source of light, illuminating every detail of his armor and the damp rocks, creating a mythical, epic atmosphere filled with power and tension.", + "seed": 26610, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000056/group_3/index_5644_seed_26610_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": -0.034361819922924045, + "vlm_logprob": -0.034361819922924045, + "vlm_yes_no_margin": 3.9, + "vlm_token_credit": -0.034361819922924045, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject Sun Wukong (the Monkey King) depicted in a steady horse stance with legs slightly bent?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is Sun Wukong wearing intricately carved golden chainmail armor with cloud patterns and a phoenix-wing purple-gold crown with two long pheasant feathers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene set before the Water Curtain Cave in the Flower-Fruit Mountain with a vast waterfall cascading down a towering cliff?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does Sun Wukong thrust a golden cudgel in a perfect golden arc from the right rear to the left front, causing the waterfall to distort?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot filmed from a low, upward angle with the camera moving in a slow, powerful arc?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11572161316871643 + ], + "opposite_logprob_tokens": [ + -2.2407217025756836 + ], + "logprob_reward": -0.11572161316871643, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11572161316871643, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21797160804271698, + 0.08342112600803375, + 0.05375925824046135, + 0.05287201330065727, + 0.1059952825307846, + 0.06318735331296921, + 0.06434259563684464, + 0.09020144492387772, + 0.06963934749364853, + 0.19861000776290894 + ], + "credit_map": [ + 0.03395452722907066, + 0.03137272223830223, + 0.028576157987117767, + 0.018747380003333092, + 0.03278040140867233, + 0.019801510497927666, + 0.025875432416796684, + 0.026863476261496544, + 0.01784934289753437, + 0.013738017529249191, + 0.011127544566988945, + 0.006691355723887682, + 0.007037870120257139, + 0.006583775859326124, + 0.007563731633126736, + 0.012829494662582874, + 0.011159422807395458, + 0.011652710847556591, + 0.006858753971755505, + 0.0037405493203550577, + 0.005946963094174862, + 0.003945119213312864, + 0.00512275006622076, + 0.005332989618182182, + 0.012369226664304733, + 0.008576121181249619, + 0.00588953634724021, + 0.0043167066760361195, + 0.005714741535484791, + 0.0050897118635475636, + 0.005568965338170528, + 0.005347004160284996, + 0.01033577136695385, + 0.011089347302913666, + 0.015588188543915749, + 0.013017726130783558, + 0.013016236014664173, + 0.015144885517656803, + 0.011153471656143665, + 0.016649655997753143, + 0.008960513398051262, + 0.006972843315452337, + 0.014000404626131058, + 0.006440260447561741, + 0.006730333436280489, + 0.005944095551967621, + 0.006286193151026964, + 0.007852710783481598, + 0.006896097678691149, + 0.008544242940843105, + 0.010606318712234497, + 0.006334112491458654, + 0.007688164710998535, + 0.009629037231206894, + 0.00813672412186861, + 0.0065079000778496265, + 0.009912483394145966, + 0.008617941290140152, + 0.01293316762894392, + 0.012142528779804707, + 0.008792075328528881, + 0.014131952077150345, + 0.014644289389252663, + 0.009027001447975636, + 0.007782617583870888, + 0.008210238069295883, + 0.008601834066212177, + 0.008919266052544117, + 0.010270913131535053, + 0.009590801782906055, + 0.008768093772232533, + 0.00749559048563242, + 0.023845955729484558, + 0.01103845052421093, + 0.07667573541402817, + 0.014133239164948463, + 0.01049114391207695, + 0.04526343569159508, + 0.007717362605035305, + 0.009444684721529484 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 4 + ], + "selected_reward": -0.11572161316871643 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject Sun Wukong (the Monkey King) depicted in a steady horse stance with legs slightly bent?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is Sun Wukong wearing intricately carved golden chainmail armor with cloud patterns and a phoenix-wing purple-gold crown with two long pheasant feathers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene set before the Water Curtain Cave in the Flower-Fruit Mountain with a vast waterfall cascading down a towering cliff?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does Sun Wukong thrust a golden cudgel in a perfect golden arc from the right rear to the left front, causing the waterfall to distort?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot filmed from a low, upward angle with the camera moving in a slow, powerful arc?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008663685992360115 + ], + "opposite_logprob_tokens": [ + -4.758663654327393 + ], + "logprob_reward": -0.008663685992360115, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008663685992360115, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1433778703212738, + 0.07619700580835342, + 0.0362757071852684, + 0.026738567277789116, + 0.05259048566222191, + 0.028779126703739166, + 0.04004404693841934, + 0.034186605364084244, + 0.05098061263561249, + 0.5108299255371094 + ], + "credit_map": [ + 0.01717827469110489, + 0.016005609184503555, + 0.019933145493268967, + 0.014346860349178314, + 0.02836318500339985, + 0.013910703361034393, + 0.018657127395272255, + 0.014982975088059902, + 0.017542637884616852, + 0.011344279162585735, + 0.011040236800909042, + 0.005718495696783066, + 0.006025239825248718, + 0.006627379450947046, + 0.004900025203824043, + 0.012998714111745358, + 0.007051743101328611, + 0.008362754248082638, + 0.004990982823073864, + 0.003067502286285162, + 0.004005788825452328, + 0.0025850709062069654, + 0.0031873465050011873, + 0.003024518024176359, + 0.00427663279697299, + 0.0030083423480391502, + 0.0052147372625768185, + 0.002439643256366253, + 0.0024341472890228033, + 0.0023531396873295307, + 0.0023269993253052235, + 0.004684926941990852, + 0.005363321863114834, + 0.006091855466365814, + 0.009022912010550499, + 0.007361878640949726, + 0.006887940689921379, + 0.005355888977646828, + 0.0070575145073235035, + 0.005449181422591209, + 0.0032128647435456514, + 0.0027997090946882963, + 0.006229711230844259, + 0.0026841508224606514, + 0.003418529639020562, + 0.0031572412699460983, + 0.0034653174225240946, + 0.0038116038776934147, + 0.0038923074025660753, + 0.005921973381191492, + 0.006345176137983799, + 0.00464678555727005, + 0.005372202023863792, + 0.004961279686540365, + 0.005533705931156874, + 0.0033706205431371927, + 0.004322993103414774, + 0.0046910736709833145, + 0.0062685078009963036, + 0.0033238213509321213, + 0.003323429962620139, + 0.004685630556195974, + 0.003604393219575286, + 0.003966758027672768, + 0.004756706301122904, + 0.005536519456654787, + 0.0055154915899038315, + 0.0057491702027618885, + 0.008991640992462635, + 0.006778063252568245, + 0.0071697230450809, + 0.006483302917331457, + 0.07996524125337601, + 0.021905090659856796, + 0.21226437389850616, + 0.037594206631183624, + 0.02532636746764183, + 0.10286960005760193, + 0.013917754404246807, + 0.016987347975373268 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 1 + ], + "selected_reward": -0.008663685992360115 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject Sun Wukong (the Monkey King) depicted in a steady horse stance with legs slightly bent?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is Sun Wukong wearing intricately carved golden chainmail armor with cloud patterns and a phoenix-wing purple-gold crown with two long pheasant feathers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene set before the Water Curtain Cave in the Flower-Fruit Mountain with a vast waterfall cascading down a towering cliff?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does Sun Wukong thrust a golden cudgel in a perfect golden arc from the right rear to the left front, causing the waterfall to distort?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot filmed from a low, upward angle with the camera moving in a slow, powerful arc?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008649386465549469 + ], + "opposite_logprob_tokens": [ + -4.7586493492126465 + ], + "logprob_reward": -0.008649386465549469, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008649386465549469, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15279585123062134, + 0.14622746407985687, + 0.08513055741786957, + 0.04735107719898224, + 0.0519908182322979, + 0.03162062540650368, + 0.053224314004182816, + 0.0529978908598423, + 0.07195939868688583, + 0.3067021071910858 + ], + "credit_map": [ + 0.020224807783961296, + 0.020805127918720245, + 0.025306684896349907, + 0.016773726791143417, + 0.02137373574078083, + 0.015110024251043797, + 0.01664733700454235, + 0.016554374247789383, + 0.02978631667792797, + 0.017343636602163315, + 0.027183284983038902, + 0.012475098483264446, + 0.011225984431803226, + 0.01409124955534935, + 0.012390324845910072, + 0.021731562912464142, + 0.021109431982040405, + 0.02151002548635006, + 0.010500328615307808, + 0.005787010304629803, + 0.009099319577217102, + 0.0064156632870435715, + 0.005470620468258858, + 0.00523814931511879, + 0.007866536267101765, + 0.007011054549366236, + 0.008266393095254898, + 0.0037157228216528893, + 0.00625373562797904, + 0.003509719157591462, + 0.00400244165211916, + 0.006725470535457134, + 0.008372022770345211, + 0.0071784136816859245, + 0.007895338349044323, + 0.00454607093706727, + 0.006470996420830488, + 0.0060640075244009495, + 0.004913428798317909, + 0.006550535559654236, + 0.0032721871975809336, + 0.003048809478059411, + 0.007034722715616226, + 0.0030143491458147764, + 0.0045783184468746185, + 0.003273304784670472, + 0.003627409227192402, + 0.003771525574848056, + 0.004579600878059864, + 0.007875432260334492, + 0.007276242598891258, + 0.007260117679834366, + 0.00784696452319622, + 0.005426772870123386, + 0.009020393714308739, + 0.003938784822821617, + 0.008772151544690132, + 0.0065688337199389935, + 0.01029638946056366, + 0.006437957286834717, + 0.004787010606378317, + 0.005492712836712599, + 0.0044706473127007484, + 0.006172179244458675, + 0.007869222201406956, + 0.009528694674372673, + 0.005181918852031231, + 0.010235072113573551, + 0.013205958530306816, + 0.008478489704430103, + 0.010197819210588932, + 0.007262213155627251, + 0.04225052520632744, + 0.012569970451295376, + 0.12756012380123138, + 0.022540856152772903, + 0.014431318268179893, + 0.06895197182893753, + 0.008487270213663578, + 0.009910034947097301 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 1 + ], + "selected_reward": -0.008649386465549469 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject Sun Wukong (the Monkey King) depicted in a steady horse stance with legs slightly bent?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is Sun Wukong wearing intricately carved golden chainmail armor with cloud patterns and a phoenix-wing purple-gold crown with two long pheasant feathers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene set before the Water Curtain Cave in the Flower-Fruit Mountain with a vast waterfall cascading down a towering cliff?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does Sun Wukong thrust a golden cudgel in a perfect golden arc from the right rear to the left front, causing the waterfall to distort?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot filmed from a low, upward angle with the camera moving in a slow, powerful arc?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018193624913692474 + ], + "opposite_logprob_tokens": [ + -4.01819372177124 + ], + "logprob_reward": -0.018193624913692474, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018193624913692474, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1317642331123352, + 0.052683956921100616, + 0.06280108541250229, + 0.059808652848005295, + 0.06014324352145195, + 0.0538855604827404, + 0.058499716222286224, + 0.05669594928622246, + 0.15649323165416718, + 0.3072243630886078 + ], + "credit_map": [ + 0.02099103480577469, + 0.02592935971915722, + 0.019551394507288933, + 0.01245593186467886, + 0.018111838027834892, + 0.012143796309828758, + 0.011715194210410118, + 0.010865681804716587, + 0.01440897025167942, + 0.008038565516471863, + 0.006759766489267349, + 0.003684855764731765, + 0.004953841213136911, + 0.0046925791539251804, + 0.003588692517951131, + 0.006556690204888582, + 0.013338138349354267, + 0.01819729432463646, + 0.00734844570979476, + 0.004061995539814234, + 0.00782015174627304, + 0.004746954422444105, + 0.0036262222565710545, + 0.0036618830636143684, + 0.0076784105040133, + 0.007210408803075552, + 0.008226193487644196, + 0.006240644492208958, + 0.007869429886341095, + 0.00460337707772851, + 0.005580685567110777, + 0.012399503029882908, + 0.00880691409111023, + 0.006799611728638411, + 0.008377378806471825, + 0.006791314575821161, + 0.005731218494474888, + 0.008098583668470383, + 0.0077661569230258465, + 0.0077720643021166325, + 0.005584798753261566, + 0.00606291089206934, + 0.009504886344075203, + 0.005410952027887106, + 0.008456870913505554, + 0.004794917535036802, + 0.007955830544233322, + 0.00611438974738121, + 0.00533332210034132, + 0.010002459399402142, + 0.00835067592561245, + 0.006689268629997969, + 0.008699338883161545, + 0.004678444471210241, + 0.00964681338518858, + 0.005099394358694553, + 0.009267115965485573, + 0.005685962270945311, + 0.008457985706627369, + 0.00750709930434823, + 0.005785524379462004, + 0.006615104153752327, + 0.006107947323471308, + 0.007269212510436773, + 0.018801821395754814, + 0.026677047833800316, + 0.021642807871103287, + 0.029211195185780525, + 0.014673829078674316, + 0.014839732088148594, + 0.016343751922249794, + 0.014303041622042656, + 0.035492222756147385, + 0.014940799213945866, + 0.12500786781311035, + 0.024213051423430443, + 0.014333276078104973, + 0.07260595262050629, + 0.010366150178015232, + 0.010265037417411804 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 0 + ], + "selected_reward": -0.018193624913692474 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject Sun Wukong (the Monkey King) depicted in a steady horse stance with legs slightly bent?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is Sun Wukong wearing intricately carved golden chainmail armor with cloud patterns and a phoenix-wing purple-gold crown with two long pheasant feathers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene set before the Water Curtain Cave in the Flower-Fruit Mountain with a vast waterfall cascading down a towering cliff?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does Sun Wukong thrust a golden cudgel in a perfect golden arc from the right rear to the left front, causing the waterfall to distort?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot filmed from a low, upward angle with the camera moving in a slow, powerful arc?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02058078907430172 + ], + "opposite_logprob_tokens": [ + -3.895580768585205 + ], + "logprob_reward": -0.02058078907430172, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.02058078907430172, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1101640984416008, + 0.04904869198799133, + 0.03664568439126015, + 0.04429396614432335, + 0.1096462830901146, + 0.1133660301566124, + 0.2180219441652298, + 0.11804118752479553, + 0.07550586014986038, + 0.12526623904705048 + ], + "credit_map": [ + 0.015743674710392952, + 0.017512476071715355, + 0.01612805388867855, + 0.01084181573241949, + 0.016006331890821457, + 0.012781547382473946, + 0.010560324415564537, + 0.01058987807482481, + 0.01209970097988844, + 0.007023494690656662, + 0.006740881595760584, + 0.003899490227922797, + 0.005275313276797533, + 0.005321684293448925, + 0.004419118165969849, + 0.00426901038736105, + 0.006710847374051809, + 0.007916982285678387, + 0.004552510567009449, + 0.002603149274364114, + 0.005371031817048788, + 0.0028362891171127558, + 0.003300311043858528, + 0.0033545601181685925, + 0.008393092080950737, + 0.005490281153470278, + 0.006384766660630703, + 0.003916766494512558, + 0.0042587760835886, + 0.0036544823087751865, + 0.004045594017952681, + 0.008150205947458744, + 0.00891075935214758, + 0.013288435526192188, + 0.025704046711325645, + 0.012843811884522438, + 0.01518084667623043, + 0.012020301073789597, + 0.012182437814772129, + 0.009515642188489437, + 0.01785656251013279, + 0.012037728913128376, + 0.023473823443055153, + 0.008622391149401665, + 0.014471574686467648, + 0.010796824470162392, + 0.01275828666985035, + 0.013348845764994621, + 0.02893245592713356, + 0.028866561129689217, + 0.04878128692507744, + 0.011795718222856522, + 0.015709316357970238, + 0.04596211016178131, + 0.013596934266388416, + 0.02437755838036537, + 0.019488178193569183, + 0.010802166536450386, + 0.022927459329366684, + 0.011641240678727627, + 0.007674516644328833, + 0.013754847459495068, + 0.011080862954258919, + 0.020671909675002098, + 0.006876036059111357, + 0.00955845508724451, + 0.008760010823607445, + 0.010986215434968472, + 0.012400143779814243, + 0.007262180559337139, + 0.010972350835800171, + 0.00869046337902546, + 0.02225319854915142, + 0.009043536148965359, + 0.04818086698651314, + 0.009854420088231564, + 0.00681461812928319, + 0.0198055487126112, + 0.00405124481767416, + 0.005262807011604309 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 9, + 7 + ], + "selected_reward": -0.02058078907430172 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "dynamic shot, shot from a low, upward angle, depicts the camera in a slow, powerful arc, capturing the subject in a meticulously rendered 3D style. Before the majestic Water Curtain Cave in the Flower-Fruit Mountain, the Monkey King, Sun Wukong, stands in a steady horse stance, legs slightly bent. He clad in intricately carved golden chainmail armor, its plates engraved with intricate cloud patterns that reflect the surrounding light. Atop his head, a phoenix-wing purple-gold crown, two long pheasant feathers flutter majestically with his movements. His golden hair, rendered in physical rendering, is distinct and shimmers with a silken sheen. His fiery eyes radiate an unyielding golden light, his facial expression stern and filled with fighting spirit. He clutches the legendary golden cudgel in both hands. The cudgel's shaft doesn't simply glow; instead, it emanates a fiery, molten-like energy from within. Golden hoops, inlaid at each end, appear and disappear beneath the surge of energy. Behind him, a vast waterfall cascades down from a towering cliff like the Milky Way, forming a vast curtain of water. Countless water droplets splash through the air, refracted by the light of the golden cudgel, creating a rainbow of colors. With a low roar, Sun Wukong thrust his golden cudgel in a perfect golden arc from the right rear to the left front. The cudgel slices through the air with a deep, powerful whistling sound, and the resulting air currents even momentarily distort the surrounding waterfall. His movements are imbued with explosive power and control. The cudgel's blazing light serves as the scene's primary source of light, illuminating every detail of his armor and the damp rocks, creating a mythical, epic atmosphere filled with power and tension.", + "seed": 26611, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000056/group_3/index_5644_seed_26611_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": -0.06054368540644646, + "vlm_logprob": -0.06054368540644646, + "vlm_yes_no_margin": 3.824999952316284, + "vlm_token_credit": -0.06054368540644646, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject Sun Wukong (the Monkey King) depicted in a steady horse stance with legs slightly bent?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is Sun Wukong wearing intricately carved golden chainmail armor with cloud patterns and a phoenix-wing purple-gold crown with two long pheasant feathers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene set before the Water Curtain Cave in the Flower-Fruit Mountain with a vast waterfall cascading down a towering cliff?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does Sun Wukong thrust a golden cudgel in a perfect golden arc from the right rear to the left front, causing the waterfall to distort?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot filmed from a low, upward angle with the camera moving in a slow, powerful arc?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2550136148929596 + ], + "opposite_logprob_tokens": [ + -1.5050135850906372 + ], + "logprob_reward": -0.2550136148929596, + "yes_no_margin_reward": 1.25, + "token_credit_reward": -0.2550136148929596, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0368383526802063, + 0.006800835486501455, + 0.005650307983160019, + 0.007845438085496426, + 0.010266106575727463, + 0.014142749831080437, + 0.006810375954955816, + 0.01915990188717842, + 0.886704683303833, + 0.005781215149909258 + ], + "credit_map": [ + 0.003907753154635429, + 0.00499856797978282, + 0.004329696763306856, + 0.00703295087441802, + 0.005110116209834814, + 0.0037769756745547056, + 0.004686350934207439, + 0.002995941322296858, + 0.0006569812539964914, + 0.0010158178629353642, + 0.0008368175476789474, + 0.0010015479056164622, + 0.0007421594928018749, + 0.0009239192586392164, + 0.0008234266424551606, + 0.0008001657552085817, + 0.000483120878925547, + 0.0008662125328555703, + 0.0006598895997740328, + 0.0005172766395844519, + 0.0011185907060280442, + 0.0006463952595368028, + 0.000671150628477335, + 0.000687671999912709, + 0.0009693035972304642, + 0.001328315818682313, + 0.0009881986770778894, + 0.0008186859195120633, + 0.0011450255988165736, + 0.0008329209522344172, + 0.0007855957373976707, + 0.0009773920755833387, + 0.0007680439157411456, + 0.0017791388090699911, + 0.0010278683621436357, + 0.0011826030677184463, + 0.0017678384901955724, + 0.0009434204548597336, + 0.0019164339173585176, + 0.0008807588601484895, + 0.0013555627083405852, + 0.0022495866287499666, + 0.0013584840344265103, + 0.0013529194984585047, + 0.001957722706720233, + 0.0021851283963769674, + 0.0017663430189713836, + 0.0019170019077137113, + 0.0011212974786758423, + 0.0008660076418891549, + 0.0011513177305459976, + 0.0009277656790800393, + 0.0008027172880247235, + 0.0005486080190166831, + 0.0008973394287750125, + 0.0004953225143253803, + 0.0013664100551977754, + 0.0018049391219392419, + 0.0028743937145918608, + 0.002592082368209958, + 0.0026123481802642345, + 0.002685168292373419, + 0.002317574340850115, + 0.002906986279413104, + 0.11333167552947998, + 0.07091087102890015, + 0.3245058059692383, + 0.22955937683582306, + 0.03910369426012039, + 0.03640443831682205, + 0.05314580723643303, + 0.019743038341403008, + 0.0009153016726486385, + 0.0008135103853419423, + 0.0006998428725637496, + 0.0006462035235017538, + 0.0006293203332461417, + 0.0007173069752752781, + 0.0008305953233502805, + 0.0005291340639814734 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 7 + ], + "selected_reward": -0.2550136148929596 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject Sun Wukong (the Monkey King) depicted in a steady horse stance with legs slightly bent?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is Sun Wukong wearing intricately carved golden chainmail armor with cloud patterns and a phoenix-wing purple-gold crown with two long pheasant feathers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene set before the Water Curtain Cave in the Flower-Fruit Mountain with a vast waterfall cascading down a towering cliff?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does Sun Wukong thrust a golden cudgel in a perfect golden arc from the right rear to the left front, causing the waterfall to distort?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot filmed from a low, upward angle with the camera moving in a slow, powerful arc?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00980218406766653 + ], + "opposite_logprob_tokens": [ + -4.634802341461182 + ], + "logprob_reward": -0.00980218406766653, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.00980218406766653, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.165335550904274, + 0.0391654334962368, + 0.016746561974287033, + 0.022634122520685196, + 0.009222929365932941, + 0.026183510199189186, + 0.01857299730181694, + 0.011133987456560135, + 0.6688514947891235, + 0.02215340919792652 + ], + "credit_map": [ + 0.018529092893004417, + 0.0217959713190794, + 0.025381218641996384, + 0.026524927467107773, + 0.027322372421622276, + 0.015523528680205345, + 0.01870168000459671, + 0.011556758545339108, + 0.004129274748265743, + 0.007297567557543516, + 0.008018435910344124, + 0.0037963292561471462, + 0.003904149867594242, + 0.005040987394750118, + 0.002999458461999893, + 0.003979234024882317, + 0.002126537961885333, + 0.0020269150845706463, + 0.002514367690309882, + 0.001211289083585143, + 0.002896140329539776, + 0.002417335519567132, + 0.001976778730750084, + 0.0015771972248330712, + 0.0029106908477842808, + 0.0037904311902821064, + 0.002988852560520172, + 0.00200229836627841, + 0.0033598828595131636, + 0.002924572443589568, + 0.0021118216682225466, + 0.0025455737486481667, + 0.0007367061916738749, + 0.001172213931567967, + 0.0010165354469791055, + 0.0010788682848215103, + 0.0013813109835609794, + 0.0014301525661721826, + 0.0011107907630503178, + 0.0012963514309376478, + 0.003616626374423504, + 0.004250881262123585, + 0.003002692712470889, + 0.002533821389079094, + 0.0028892094269394875, + 0.005060473922640085, + 0.0022928291000425816, + 0.0025369783397763968, + 0.002509278943762183, + 0.0023504088167101145, + 0.002545412629842758, + 0.0018734935438260436, + 0.0026464955881237984, + 0.0023497198708355427, + 0.002341860206797719, + 0.001956328982487321, + 0.0011438056826591492, + 0.0014171592192724347, + 0.002092601964250207, + 0.0016097348416224122, + 0.0016923637595027685, + 0.001111735007725656, + 0.0010011543054133654, + 0.0010654331417754292, + 0.08381010591983795, + 0.05322016030550003, + 0.24281935393810272, + 0.1742188036441803, + 0.029519418254494667, + 0.02751295454800129, + 0.04120761901140213, + 0.016543108969926834, + 0.0025819360744208097, + 0.0028040932957082987, + 0.0017311163246631622, + 0.002980130724608898, + 0.0028231688775122166, + 0.002969560446217656, + 0.0041769882664084435, + 0.0020864170510321856 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 1 + ], + "selected_reward": -0.00980218406766653 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject Sun Wukong (the Monkey King) depicted in a steady horse stance with legs slightly bent?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is Sun Wukong wearing intricately carved golden chainmail armor with cloud patterns and a phoenix-wing purple-gold crown with two long pheasant feathers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene set before the Water Curtain Cave in the Flower-Fruit Mountain with a vast waterfall cascading down a towering cliff?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does Sun Wukong thrust a golden cudgel in a perfect golden arc from the right rear to the left front, causing the waterfall to distort?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot filmed from a low, upward angle with the camera moving in a slow, powerful arc?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00864654965698719 + ], + "opposite_logprob_tokens": [ + -4.758646488189697 + ], + "logprob_reward": -0.00864654965698719, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.00864654965698719, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06770701706409454, + 0.02213353104889393, + 0.012509578838944435, + 0.0066451262682676315, + 0.008458560332655907, + 0.043912436813116074, + 0.009544074535369873, + 0.025057164952158928, + 0.7952507734298706, + 0.008781702257692814 + ], + "credit_map": [ + 0.006292374338954687, + 0.005143031012266874, + 0.009981704875826836, + 0.00961826927959919, + 0.019399188458919525, + 0.006642439868301153, + 0.007067305035889149, + 0.003562710015103221, + 0.0025869361124932766, + 0.0042400117963552475, + 0.003956766799092293, + 0.0023943879641592503, + 0.0024313731119036674, + 0.0025713329669088125, + 0.0018591278931125998, + 0.0020935973152518272, + 0.0014836420305073261, + 0.002187654608860612, + 0.0012490306980907917, + 0.0010656497906893492, + 0.0020038061775267124, + 0.0026861466467380524, + 0.0009973293635994196, + 0.000836320105008781, + 0.0008255004649981856, + 0.0007841357146389782, + 0.00102633994538337, + 0.0006237993948161602, + 0.0008638465078547597, + 0.0008072664495557547, + 0.0007628041785210371, + 0.0009514344856142998, + 0.0007741026929579675, + 0.0012637156760320067, + 0.0009586134110577404, + 0.0007783706532791257, + 0.0009549336973577738, + 0.0011213079560548067, + 0.0009677351918071508, + 0.0016397822182625532, + 0.004942660219967365, + 0.007377047557383776, + 0.0050736539997160435, + 0.004392215050756931, + 0.005787402857095003, + 0.00811330508440733, + 0.004817180335521698, + 0.003408970544114709, + 0.001357052125968039, + 0.0013210498727858067, + 0.0013722960138693452, + 0.0015105947386473417, + 0.001116299768909812, + 0.0010312579106539488, + 0.0009745471179485321, + 0.0008609772776253521, + 0.0024988448712974787, + 0.0029021899681538343, + 0.004806447308510542, + 0.0031818232964724302, + 0.0030953187961131334, + 0.0031080530025064945, + 0.0026052037719637156, + 0.002859285566955805, + 0.1007741317152977, + 0.06345828622579575, + 0.28986310958862305, + 0.2068975418806076, + 0.03491450846195221, + 0.03261202946305275, + 0.04837499186396599, + 0.01835622824728489, + 0.0009679817594587803, + 0.0011911618057638407, + 0.0007558609358966351, + 0.0012455721152946353, + 0.0009652163134887815, + 0.0012243156088516116, + 0.0016103644156828523, + 0.0008212297107093036 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 5 + ], + "selected_reward": -0.00864654965698719 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject Sun Wukong (the Monkey King) depicted in a steady horse stance with legs slightly bent?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is Sun Wukong wearing intricately carved golden chainmail armor with cloud patterns and a phoenix-wing purple-gold crown with two long pheasant feathers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene set before the Water Curtain Cave in the Flower-Fruit Mountain with a vast waterfall cascading down a towering cliff?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does Sun Wukong thrust a golden cudgel in a perfect golden arc from the right rear to the left front, causing the waterfall to distort?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot filmed from a low, upward angle with the camera moving in a slow, powerful arc?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011075095273554325 + ], + "opposite_logprob_tokens": [ + -4.511075019836426 + ], + "logprob_reward": -0.011075095273554325, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011075095273554325, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05551727116107941, + 0.011997920460999012, + 0.007444706279784441, + 0.006353931967169046, + 0.0073450710624456406, + 0.008475357666611671, + 0.010532142594456673, + 0.007218459155410528, + 0.8772999048233032, + 0.007815291173756123 + ], + "credit_map": [ + 0.0065352413803339005, + 0.008166014216840267, + 0.005490802228450775, + 0.010785196907818317, + 0.009690511971712112, + 0.00530828395858407, + 0.005443553440272808, + 0.004097668454051018, + 0.0013245146255940199, + 0.0018910656217485666, + 0.0014587676851078868, + 0.0017754447180777788, + 0.0014554245863109827, + 0.001552502391859889, + 0.0015575457364320755, + 0.0009826553286984563, + 0.0007474976591765881, + 0.0009136878070421517, + 0.0010302204173058271, + 0.0007777500431984663, + 0.0009734783670864999, + 0.0011162101291120052, + 0.0007681691786274314, + 0.0011176930274814367, + 0.000619235448539257, + 0.0008980704005807638, + 0.0009886046173051, + 0.0006663204985670745, + 0.0010316623374819756, + 0.0005611319793388247, + 0.0007115225889720023, + 0.0008773842127993703, + 0.0006285856361500919, + 0.0008095008670352399, + 0.0010044216178357601, + 0.0009455165127292275, + 0.0008252511615864933, + 0.0007794067496433854, + 0.0010176623472943902, + 0.0013347259955480695, + 0.0009808441391214728, + 0.0014952878700569272, + 0.000900299521163106, + 0.000783318595495075, + 0.0010803837794810534, + 0.001920508686453104, + 0.0005618188879452646, + 0.000752896536141634, + 0.0010820258175954223, + 0.0013758266577497125, + 0.0012983555207028985, + 0.0016281778225675225, + 0.0018446146277710795, + 0.001162359374575317, + 0.0010773983085528016, + 0.0010633844649419188, + 0.0009534292039461434, + 0.0009066924685612321, + 0.0014093266800045967, + 0.0005809991271235049, + 0.0011491526383906603, + 0.0007625236175954342, + 0.000667180516757071, + 0.0007891553686931729, + 0.11176960170269012, + 0.07011320441961288, + 0.3204328417778015, + 0.227554589509964, + 0.03854445740580559, + 0.035958703607320786, + 0.05288881063461304, + 0.02003762312233448, + 0.0006708310102112591, + 0.001042405259795487, + 0.0006516343564726412, + 0.0012233606539666653, + 0.0008260803879238665, + 0.0009598410106264055, + 0.0015991065884009004, + 0.0008420318481512368 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 1 + ], + "selected_reward": -0.011075095273554325 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject Sun Wukong (the Monkey King) depicted in a steady horse stance with legs slightly bent?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is Sun Wukong wearing intricately carved golden chainmail armor with cloud patterns and a phoenix-wing purple-gold crown with two long pheasant feathers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene set before the Water Curtain Cave in the Flower-Fruit Mountain with a vast waterfall cascading down a towering cliff?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does Sun Wukong thrust a golden cudgel in a perfect golden arc from the right rear to the left front, causing the waterfall to distort?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot filmed from a low, upward angle with the camera moving in a slow, powerful arc?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018180983141064644 + ], + "opposite_logprob_tokens": [ + -4.018180847167969 + ], + "logprob_reward": -0.018180983141064644, + "yes_no_margin_reward": 3.999999761581421, + "token_credit_reward": -0.018180983141064644, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.012894632294774055, + 0.016071243211627007, + 0.0038963472470641136, + 0.003682018956169486, + 0.003837268566712737, + 0.013692907057702541, + 0.005396706983447075, + 0.004167178180068731, + 0.9318882822990417, + 0.0044734119437634945 + ], + "credit_map": [ + 0.0012911949306726456, + 0.0011442552786320448, + 0.0019696583040058613, + 0.0013806120259687304, + 0.0036903689615428448, + 0.0010585777927190065, + 0.0015039980644360185, + 0.000855966005474329, + 0.001680075190961361, + 0.0031594682950526476, + 0.002994077280163765, + 0.001638473360799253, + 0.0018032423686236143, + 0.002212900435552001, + 0.001360380440019071, + 0.0012226274702697992, + 0.00040899746818467975, + 0.0005786276888102293, + 0.000499432033393532, + 0.0003767685848288238, + 0.0007355684647336602, + 0.00038335908902809024, + 0.000494569365400821, + 0.0004190248146187514, + 0.0003818704281002283, + 0.0006753492634743452, + 0.00035929575096815825, + 0.0003836951800622046, + 0.0007946879486553371, + 0.00035627122269943357, + 0.0003665984550025314, + 0.0003642508527263999, + 0.0003195515018887818, + 0.0007322350284084678, + 0.0004333599063102156, + 0.00034165853867307305, + 0.0004854461585637182, + 0.00047760308370925486, + 0.0005817766068503261, + 0.0004656379751395434, + 0.0011737896129488945, + 0.0025333792436867952, + 0.0012608994729816914, + 0.0015501449815928936, + 0.001967240124940872, + 0.0030438993126153946, + 0.0011499555548653007, + 0.00101359817199409, + 0.0006587586249224842, + 0.0008330178097821772, + 0.0008040106622502208, + 0.0006906375056132674, + 0.0006542750634253025, + 0.000487005861941725, + 0.0007935073226690292, + 0.00047549381270073354, + 0.0004425508086569607, + 0.00045299704652279615, + 0.0007164007402025163, + 0.0006541525945067406, + 0.00048464510473422706, + 0.0005697011365555227, + 0.000424237921833992, + 0.0004224928270559758, + 0.1191127747297287, + 0.07449762523174286, + 0.34109804034233093, + 0.24089589715003967, + 0.04085814207792282, + 0.0384247861802578, + 0.05595220997929573, + 0.02104876935482025, + 0.0006452660891227424, + 0.0007632131455466151, + 0.0005073647480458021, + 0.0005171024240553379, + 0.0005404729163274169, + 0.0004084814863745123, + 0.0007014396251179278, + 0.0003900716546922922 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 1, + 5 + ], + "selected_reward": -0.018180983141064644 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "dynamic shot, shot from a low, upward angle, depicts the camera in a slow, powerful arc, capturing the subject in a meticulously rendered 3D style. Before the majestic Water Curtain Cave in the Flower-Fruit Mountain, the Monkey King, Sun Wukong, stands in a steady horse stance, legs slightly bent. He clad in intricately carved golden chainmail armor, its plates engraved with intricate cloud patterns that reflect the surrounding light. Atop his head, a phoenix-wing purple-gold crown, two long pheasant feathers flutter majestically with his movements. His golden hair, rendered in physical rendering, is distinct and shimmers with a silken sheen. His fiery eyes radiate an unyielding golden light, his facial expression stern and filled with fighting spirit. He clutches the legendary golden cudgel in both hands. The cudgel's shaft doesn't simply glow; instead, it emanates a fiery, molten-like energy from within. Golden hoops, inlaid at each end, appear and disappear beneath the surge of energy. Behind him, a vast waterfall cascades down from a towering cliff like the Milky Way, forming a vast curtain of water. Countless water droplets splash through the air, refracted by the light of the golden cudgel, creating a rainbow of colors. With a low roar, Sun Wukong thrust his golden cudgel in a perfect golden arc from the right rear to the left front. The cudgel slices through the air with a deep, powerful whistling sound, and the resulting air currents even momentarily distort the surrounding waterfall. His movements are imbued with explosive power and control. The cudgel's blazing light serves as the scene's primary source of light, illuminating every detail of his armor and the damp rocks, creating a mythical, epic atmosphere filled with power and tension.", + "seed": 26612, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000056/group_3/index_5644_seed_26612_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": -0.03180571813136339, + "vlm_logprob": -0.03180571813136339, + "vlm_yes_no_margin": 4.1, + "vlm_token_credit": -0.03180571813136339, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject Sun Wukong (the Monkey King) depicted in a steady horse stance with legs slightly bent?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is Sun Wukong wearing intricately carved golden chainmail armor with cloud patterns and a phoenix-wing purple-gold crown with two long pheasant feathers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene set before the Water Curtain Cave in the Flower-Fruit Mountain with a vast waterfall cascading down a towering cliff?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does Sun Wukong thrust a golden cudgel in a perfect golden arc from the right rear to the left front, causing the waterfall to distort?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot filmed from a low, upward angle with the camera moving in a slow, powerful arc?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11603543907403946 + ], + "opposite_logprob_tokens": [ + -2.2410354614257812 + ], + "logprob_reward": -0.11603543907403946, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11603543907403946, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06755667924880981, + 0.06496582180261612, + 0.18594089150428772, + 0.11599163711071014, + 0.07595644146203995, + 0.13949178159236908, + 0.15364879369735718, + 0.10533711314201355, + 0.05492831766605377, + 0.036182552576065063 + ], + "credit_map": [ + 0.00891716219484806, + 0.007835302501916885, + 0.007427333854138851, + 0.006184577010571957, + 0.009810748510062695, + 0.010178238153457642, + 0.008219219744205475, + 0.0089840954169631, + 0.008758369833230972, + 0.011125599034130573, + 0.01020762138068676, + 0.004859852604568005, + 0.00725172832608223, + 0.009263943880796432, + 0.006853324826806784, + 0.0066453842446208, + 0.020488610491156578, + 0.01911788247525692, + 0.025551697239279747, + 0.008730366826057434, + 0.014766690321266651, + 0.018796756863594055, + 0.013338479213416576, + 0.0651504173874855, + 0.013011306524276733, + 0.023061100393533707, + 0.011501970700919628, + 0.011875990778207779, + 0.016540829092264175, + 0.014492031186819077, + 0.01546810194849968, + 0.010040304623544216, + 0.00580513384193182, + 0.004525394178926945, + 0.012964868918061256, + 0.004893644712865353, + 0.005614329595118761, + 0.026559066027402878, + 0.005100958049297333, + 0.010493041947484016, + 0.009941210970282555, + 0.009713688865303993, + 0.036692727357149124, + 0.013543868437409401, + 0.01264915894716978, + 0.029347339645028114, + 0.01708752103149891, + 0.010516279377043247, + 0.01062954030930996, + 0.028323587030172348, + 0.05578736960887909, + 0.008541921153664589, + 0.005833100993186235, + 0.02911411225795746, + 0.0062819900922477245, + 0.0091371675953269, + 0.009387117810547352, + 0.010519259609282017, + 0.022950399667024612, + 0.014195206575095654, + 0.011909594759345055, + 0.012736990116536617, + 0.011722173541784286, + 0.011916371993720531, + 0.005601389799267054, + 0.006328838411718607, + 0.012099990621209145, + 0.00524164829403162, + 0.004767016042023897, + 0.009233428165316582, + 0.004922571126371622, + 0.006733432412147522, + 0.003942562732845545, + 0.004069048445671797, + 0.004427042789757252, + 0.004397199954837561, + 0.004457904491573572, + 0.005331669934093952, + 0.005632605869323015, + 0.003924518823623657 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 6, + 5 + ], + "selected_reward": -0.11603543907403946 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject Sun Wukong (the Monkey King) depicted in a steady horse stance with legs slightly bent?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is Sun Wukong wearing intricately carved golden chainmail armor with cloud patterns and a phoenix-wing purple-gold crown with two long pheasant feathers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene set before the Water Curtain Cave in the Flower-Fruit Mountain with a vast waterfall cascading down a towering cliff?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does Sun Wukong thrust a golden cudgel in a perfect golden arc from the right rear to the left front, causing the waterfall to distort?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot filmed from a low, upward angle with the camera moving in a slow, powerful arc?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"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": [ + 0.13034728169441223, + 0.06941032409667969, + 0.2233426868915558, + 0.08277049660682678, + 0.07940586656332016, + 0.09474030882120132, + 0.18513213098049164, + 0.06608840078115463, + 0.031447604298591614, + 0.037314873188734055 + ], + "credit_map": [ + 0.01706842891871929, + 0.014216035604476929, + 0.00910244695842266, + 0.009665332734584808, + 0.022779030725359917, + 0.03324781358242035, + 0.013661219738423824, + 0.01060696691274643, + 0.008732950314879417, + 0.006966331973671913, + 0.010420585982501507, + 0.0034837680868804455, + 0.0074323974549770355, + 0.023251915350556374, + 0.004034775774925947, + 0.005087603814899921, + 0.021321173757314682, + 0.019753064960241318, + 0.031144145876169205, + 0.011816917918622494, + 0.021080678328871727, + 0.028854208067059517, + 0.013052469119429588, + 0.07632002979516983, + 0.019743304699659348, + 0.00538653414696455, + 0.0077303447760641575, + 0.0031371896620839834, + 0.005462035536766052, + 0.030233442783355713, + 0.004933341406285763, + 0.006144300568848848, + 0.007626085076481104, + 0.006824796553701162, + 0.014768005348742008, + 0.0030317450873553753, + 0.003528103930875659, + 0.0303377415984869, + 0.003623219672590494, + 0.009666170924901962, + 0.010770199820399284, + 0.006596606690436602, + 0.009845110587775707, + 0.008542044088244438, + 0.007908472791314125, + 0.024917852133512497, + 0.013198746368288994, + 0.012961284257471561, + 0.011562430299818516, + 0.0447850339114666, + 0.07161854207515717, + 0.009008441120386124, + 0.006517665460705757, + 0.02683831751346588, + 0.0038421056233346462, + 0.010959595441818237, + 0.007322090677917004, + 0.010803178884088993, + 0.012775566428899765, + 0.004101387690752745, + 0.00340632488951087, + 0.015885064378380775, + 0.003795091761276126, + 0.007999698631465435, + 0.004176041577011347, + 0.004157341551035643, + 0.003238476812839508, + 0.0034149540588259697, + 0.003595141926780343, + 0.005944066680967808, + 0.0034589958377182484, + 0.0034625872503966093, + 0.004197163041681051, + 0.004241307731717825, + 0.004715942312031984, + 0.005324993282556534, + 0.004050979390740395, + 0.007662925869226456, + 0.0048292819410562515, + 0.0022922800853848457 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 6, + 0 + ], + "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 subject Sun Wukong (the Monkey King) depicted in a steady horse stance with legs slightly bent?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is Sun Wukong wearing intricately carved golden chainmail armor with cloud patterns and a phoenix-wing purple-gold crown with two long pheasant feathers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene set before the Water Curtain Cave in the Flower-Fruit Mountain with a vast waterfall cascading down a towering cliff?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does Sun Wukong thrust a golden cudgel in a perfect golden arc from the right rear to the left front, causing the waterfall to distort?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot filmed from a low, upward angle with the camera moving in a slow, powerful arc?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006745661608874798 + ], + "opposite_logprob_tokens": [ + -5.0067458152771 + ], + "logprob_reward": -0.006745661608874798, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006745661608874798, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05613576993346214, + 0.053397588431835175, + 0.23584377765655518, + 0.09152960777282715, + 0.019991660490632057, + 0.037817906588315964, + 0.33642101287841797, + 0.1132945790886879, + 0.034207534044981, + 0.02136056497693062 + ], + "credit_map": [ + 0.009925088845193386, + 0.005047052167356014, + 0.006894710939377546, + 0.0035231567453593016, + 0.004789175000041723, + 0.015763550996780396, + 0.005077606998383999, + 0.005115436855703592, + 0.00913741160184145, + 0.0033099346328526735, + 0.009263182990252972, + 0.00258641573600471, + 0.0034429992083460093, + 0.01967836357653141, + 0.002779975300654769, + 0.003199308877810836, + 0.021474694833159447, + 0.020864542573690414, + 0.041282009333372116, + 0.01096523180603981, + 0.016155345365405083, + 0.02457791194319725, + 0.012765658088028431, + 0.08775839954614639, + 0.014481879770755768, + 0.006410188041627407, + 0.01313104759901762, + 0.004516616463661194, + 0.004674696829169989, + 0.03637631982564926, + 0.005142013542354107, + 0.0067968545481562614, + 0.0020224000327289104, + 0.0014353987062349916, + 0.0028523614164441824, + 0.0014974813675507903, + 0.0016360380686819553, + 0.0056575387716293335, + 0.0026762159541249275, + 0.002214227570220828, + 0.003251558868214488, + 0.0034480879548937082, + 0.009323528967797756, + 0.002990337088704109, + 0.0024912841618061066, + 0.010667127557098866, + 0.0020431759767234325, + 0.0036028088070452213, + 0.02479678764939308, + 0.060709673911333084, + 0.12218821048736572, + 0.01610454171895981, + 0.011060652323067188, + 0.07206249982118607, + 0.006355730816721916, + 0.023142941296100616, + 0.01019327063113451, + 0.01241415273398161, + 0.03673137351870537, + 0.00873871985822916, + 0.005840482655912638, + 0.0110549321398139, + 0.005066663026809692, + 0.023254992440342903, + 0.0026985967997461557, + 0.005446505732834339, + 0.009405006654560566, + 0.0030725491233170033, + 0.004128088243305683, + 0.00450361892580986, + 0.0026741954497992992, + 0.0022789763752371073, + 0.002357230521738529, + 0.0025437623262405396, + 0.002546519273892045, + 0.0025472333654761314, + 0.0026664743199944496, + 0.0034251995384693146, + 0.002931544091552496, + 0.002342604799196124 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 2, + 7 + ], + "selected_reward": -0.006745661608874798 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject Sun Wukong (the Monkey King) depicted in a steady horse stance with legs slightly bent?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is Sun Wukong wearing intricately carved golden chainmail armor with cloud patterns and a phoenix-wing purple-gold crown with two long pheasant feathers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene set before the Water Curtain Cave in the Flower-Fruit Mountain with a vast waterfall cascading down a towering cliff?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does Sun Wukong thrust a golden cudgel in a perfect golden arc from the right rear to the left front, causing the waterfall to distort?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot filmed from a low, upward angle with the camera moving in a slow, powerful arc?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01253888476639986 + ], + "opposite_logprob_tokens": [ + -4.387538909912109 + ], + "logprob_reward": -0.01253888476639986, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.01253888476639986, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05369184911251068, + 0.09688679128885269, + 0.22370660305023193, + 0.08896418660879135, + 0.09588468074798584, + 0.0421869195997715, + 0.10795751959085464, + 0.13564980030059814, + 0.1184755265712738, + 0.036596089601516724 + ], + "credit_map": [ + 0.007852795533835888, + 0.004041626118123531, + 0.0077124242670834064, + 0.00437160162255168, + 0.006281782872974873, + 0.0126486262306571, + 0.0050844792276620865, + 0.005698514636605978, + 0.010286879725754261, + 0.003787437453866005, + 0.024154357612133026, + 0.00333467498421669, + 0.0034397628623992205, + 0.03866559639573097, + 0.0032702037133276463, + 0.00994787272065878, + 0.024001019075512886, + 0.01650111936032772, + 0.05917613208293915, + 0.009884589351713657, + 0.011212536133825779, + 0.01829410158097744, + 0.009399515576660633, + 0.07523758709430695, + 0.011681538075208664, + 0.007257611490786076, + 0.01327199675142765, + 0.005348764359951019, + 0.006270088721066713, + 0.03153509274125099, + 0.006823509931564331, + 0.006775587797164917, + 0.010851181112229824, + 0.009288365021348, + 0.01616872288286686, + 0.006068377289921045, + 0.005987038370221853, + 0.03163490816950798, + 0.005981468595564365, + 0.009904609061777592, + 0.003208879381418228, + 0.003188047092407942, + 0.004077218938618898, + 0.004458533134311438, + 0.004094592295587063, + 0.010868801735341549, + 0.006795212626457214, + 0.005495637189596891, + 0.012431486509740353, + 0.028216062113642693, + 0.008888944052159786, + 0.0038277183193713427, + 0.005546375643461943, + 0.03717048093676567, + 0.003034283872693777, + 0.00884217955172062, + 0.01224375981837511, + 0.02207198739051819, + 0.04377664625644684, + 0.010726459324359894, + 0.009396634064614773, + 0.016053957864642143, + 0.006135819014161825, + 0.01524453517049551, + 0.01117035560309887, + 0.010237005539238453, + 0.02229260466992855, + 0.010679375380277634, + 0.008744793944060802, + 0.01840704306960106, + 0.010496136732399464, + 0.0264482069760561, + 0.004315417725592852, + 0.0038832458667457104, + 0.005426208488643169, + 0.004491109866648912, + 0.005014464259147644, + 0.004372864495962858, + 0.005197793245315552, + 0.0038949837908148766 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 7, + 8 + ], + "selected_reward": -0.01253888476639986 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject Sun Wukong (the Monkey King) depicted in a steady horse stance with legs slightly bent?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is Sun Wukong wearing intricately carved golden chainmail armor with cloud patterns and a phoenix-wing purple-gold crown with two long pheasant feathers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene set before the Water Curtain Cave in the Flower-Fruit Mountain with a vast waterfall cascading down a towering cliff?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does Sun Wukong thrust a golden cudgel in a perfect golden arc from the right rear to the left front, causing the waterfall to distort?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot filmed from a low, upward angle with the camera moving in a slow, powerful arc?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016062501817941666 + ], + "opposite_logprob_tokens": [ + -4.1410627365112305 + ], + "logprob_reward": -0.016062501817941666, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016062501817941666, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.057296909391880035, + 0.040618959814310074, + 0.19064605236053467, + 0.08617036044597626, + 0.09599386900663376, + 0.04757387563586235, + 0.4052574634552002, + 0.023377129808068275, + 0.031712181866168976, + 0.021353276446461678 + ], + "credit_map": [ + 0.007635166868567467, + 0.006871099583804607, + 0.0059437742456793785, + 0.005085841752588749, + 0.00823698379099369, + 0.013159334659576416, + 0.005620652809739113, + 0.004744051489979029, + 0.0037617890629917383, + 0.0037443346809595823, + 0.010695738717913628, + 0.0023961374536156654, + 0.0016909807454794645, + 0.013852504082024097, + 0.0020764553919434547, + 0.002401015255600214, + 0.025125961750745773, + 0.012263619340956211, + 0.05969656631350517, + 0.008287327364087105, + 0.006933624390512705, + 0.012154879048466682, + 0.006825654301792383, + 0.05935840308666229, + 0.020299876108765602, + 0.008350946940481663, + 0.0064498442225158215, + 0.004732250701636076, + 0.006500476039946079, + 0.028104238212108612, + 0.004961876664310694, + 0.006770845036953688, + 0.011757122352719307, + 0.009888265281915665, + 0.01465091947466135, + 0.006060544867068529, + 0.005896442104130983, + 0.030063925310969353, + 0.007625914178788662, + 0.010050729848444462, + 0.006303737871348858, + 0.0037562113720923662, + 0.008481607772409916, + 0.004922991618514061, + 0.00412457762286067, + 0.0071554058231413364, + 0.006712527014315128, + 0.006116815377026796, + 0.02820459194481373, + 0.07970721274614334, + 0.15040630102157593, + 0.019586214795708656, + 0.013505254872143269, + 0.07896483689546585, + 0.008481224067509174, + 0.026401786133646965, + 0.0014750533737242222, + 0.006382132414728403, + 0.00766092911362648, + 0.0016652187332510948, + 0.0013602388789877295, + 0.0022138827480375767, + 0.0011754470178857446, + 0.0014442260144278407, + 0.0030689644627273083, + 0.003169284900650382, + 0.0031582703813910484, + 0.0035087093710899353, + 0.005026542581617832, + 0.005611827597022057, + 0.0038088650908321142, + 0.0043597170151770115, + 0.0026557513047009706, + 0.00257304310798645, + 0.0024314620532095432, + 0.0024124665651470423, + 0.0026435204781591892, + 0.003471998730674386, + 0.0023379321210086346, + 0.0028271013870835304 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 2, + 4 + ], + "selected_reward": -0.016062501817941666 + } + ], + "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/0000056/group_4/index_tensor([5204, 5204])_group_4_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000056/group_4/index_tensor([5204, 5204])_group_4_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..4f144ff47a1123a1e81d285e4f5a71162a2f17fd --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000056/group_4/index_tensor([5204, 5204])_group_4_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "medium shot begins at a slight upward angle, then slowly and steadily zooms out. At the center of the frame is a general clad in ancient armor, standing on a vast, empty battlefield at dusk. His armor, forged from a blend of bronze and black iron, is covered in the scars of swords and the rust of time. The beast totems carved on his shoulder plates still shone fiercely in the dim light. His weathered face, sunken eyes and furrowed brow, is filled with despair and resignation, his beard matted with dried blood and dust. He jerks his head back, the veins in his neck bulging with the sheer force. His mouth opens wide, and he lets out a silent yet heartbreaking roar into the low, leaden clouds. His fists clench, his body trembling with the outpouring of emotion. Beneath his feet lies cracked, scorched earth, strewn with broken flags, shattered weapons, and unclaimed helmets. The last blood-red rays of sunset struggled to pierce through the clouds, gilding the outline of his armor with a poignant golden edge and casting a long, distorted shadow behind him. As the camera continuously and steadily zoomed out, the general's figure gradually shrank across the vast battlefield, ultimately becoming a tiny, lonely black dot on the charred earth. The entire scene was filled with a sense of epic tragedy and the endless desolation of a hero's final journey.", + "seed": 3160, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000056/group_4/index_5204_seed_3160_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": -0.11460041459649802, + "vlm_logprob": -0.11460041459649802, + "vlm_yes_no_margin": 3.225, + "vlm_token_credit": -0.11460041459649802, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on a vast, empty battlefield at dusk with cracked, scorched earth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a general wearing ancient armor made of bronze and black iron visible at the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the general's armor feature beast totems on the shoulder plates and show signs of sword scars and rust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the general perform the action of jerking his head back with bulging neck veins and opening his mouth wide?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start at a slight upward angle and slowly, steadily zoom out until the general becomes a tiny dot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01113333459943533 + ], + "opposite_logprob_tokens": [ + -4.761133193969727 + ], + "logprob_reward": -0.01113333459943533, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.01113333459943533, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.29455578327178955, + 0.11933483928442001, + 0.05254702270030975, + 0.114889957010746, + 0.05603398010134697, + 0.07014773786067963, + 0.06070508062839508, + 0.07813060283660889, + 0.08446291089057922, + 0.06919211149215698 + ], + "credit_map": [ + 0.10256975144147873, + 0.03204036131501198, + 0.022490886971354485, + 0.014985034242272377, + 0.016235971823334694, + 0.019501876085996628, + 0.013723434880375862, + 0.07300848513841629, + 0.025555860251188278, + 0.0145645746961236, + 0.021660521626472473, + 0.01192272175103426, + 0.008263343945145607, + 0.011103975586593151, + 0.008209487423300743, + 0.018054351210594177, + 0.01271395105868578, + 0.003977084066718817, + 0.0029193744994699955, + 0.006133465562015772, + 0.007827715016901493, + 0.004616711288690567, + 0.006570515688508749, + 0.007788205984979868, + 0.028021641075611115, + 0.009905867278575897, + 0.006568523123860359, + 0.006818384863436222, + 0.00787423737347126, + 0.0394834540784359, + 0.009256473742425442, + 0.006961369886994362, + 0.012263888493180275, + 0.003806299762800336, + 0.004214297980070114, + 0.0051308865658938885, + 0.006000427063554525, + 0.013965866528451443, + 0.007001855876296759, + 0.003650454105809331, + 0.012237603776156902, + 0.005238373298197985, + 0.005045364145189524, + 0.006841192953288555, + 0.009897272102534771, + 0.017087677493691444, + 0.007749903481453657, + 0.00605034688487649, + 0.010298674926161766, + 0.004525524098426104, + 0.005838051903992891, + 0.005869221407920122, + 0.007894993759691715, + 0.012342089787125587, + 0.007463200483471155, + 0.006473318208009005, + 0.014336569234728813, + 0.004728547763079405, + 0.005468115676194429, + 0.00618717772886157, + 0.006118030287325382, + 0.022186284884810448, + 0.009265214204788208, + 0.009840653277933598, + 0.014859351329505444, + 0.005468930583447218, + 0.006755429320037365, + 0.007572007365524769, + 0.007734136655926704, + 0.024006055667996407, + 0.010223761200904846, + 0.007843234576284885, + 0.007094353437423706, + 0.005905728321522474, + 0.007895266637206078, + 0.011381072923541069, + 0.005482579581439495, + 0.0136820487678051, + 0.009136703796684742, + 0.008614351972937584 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 3 + ], + "selected_reward": -0.01113333459943533 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on a vast, empty battlefield at dusk with cracked, scorched earth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a general wearing ancient armor made of bronze and black iron visible at the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the general's armor feature beast totems on the shoulder plates and show signs of sword scars and rust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the general perform the action of jerking his head back with bulging neck veins and opening his mouth wide?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start at a slight upward angle and slowly, steadily zoom out until the general becomes a tiny dot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008651631884276867 + ], + "opposite_logprob_tokens": [ + -4.7586517333984375 + ], + "logprob_reward": -0.008651631884276867, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008651631884276867, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2588901221752167, + 0.1283043473958969, + 0.1003640666604042, + 0.11560685932636261, + 0.0733347088098526, + 0.05625619366765022, + 0.05902988091111183, + 0.041351933032274246, + 0.09454543888568878, + 0.0723164826631546 + ], + "credit_map": [ + 0.08653012663125992, + 0.024146802723407745, + 0.02344615012407303, + 0.01645127683877945, + 0.016194889321923256, + 0.018176225945353508, + 0.01845102198421955, + 0.055493634194135666, + 0.046541113406419754, + 0.014852198772132397, + 0.008333154022693634, + 0.004717661999166012, + 0.0075666396878659725, + 0.008246823213994503, + 0.005939994938671589, + 0.03210674971342087, + 0.023126620799303055, + 0.007405911106616259, + 0.00799340195953846, + 0.014770523644983768, + 0.013260986655950546, + 0.007638061419129372, + 0.010043442249298096, + 0.01612512581050396, + 0.02736016735434532, + 0.006066570989787579, + 0.005368104670196772, + 0.006833996623754501, + 0.009517024271190166, + 0.04129926487803459, + 0.011799325235188007, + 0.007362410891801119, + 0.01677568629384041, + 0.004798963665962219, + 0.003436286700889468, + 0.00590551970526576, + 0.007764697074890137, + 0.020867886021733284, + 0.009350443258881569, + 0.00443523284047842, + 0.008930759504437447, + 0.004319547675549984, + 0.0033912041690200567, + 0.006467835512012243, + 0.008050389587879181, + 0.013300072401762009, + 0.007120710331946611, + 0.004675676580518484, + 0.010548621416091919, + 0.0040292213670909405, + 0.004794244188815355, + 0.005787572357803583, + 0.007199559826403856, + 0.013081417419016361, + 0.007185752969235182, + 0.0064034899696707726, + 0.007473799400031567, + 0.0028976378962397575, + 0.00319093419238925, + 0.003926364704966545, + 0.004153500311076641, + 0.008082576096057892, + 0.007327655330300331, + 0.004299463238567114, + 0.018459876999258995, + 0.006957170087844133, + 0.006692104507237673, + 0.0071509466506540775, + 0.007876199670135975, + 0.0302419476211071, + 0.009499751962721348, + 0.007667445577681065, + 0.007166732102632523, + 0.006324835121631622, + 0.0075923725962638855, + 0.01428171619772911, + 0.0068330420181155205, + 0.010177697986364365, + 0.009663601405918598, + 0.010276482440531254 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 3 + ], + "selected_reward": -0.008651631884276867 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on a vast, empty battlefield at dusk with cracked, scorched earth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a general wearing ancient armor made of bronze and black iron visible at the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the general's armor feature beast totems on the shoulder plates and show signs of sword scars and rust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the general perform the action of jerking his head back with bulging neck veins and opening his mouth wide?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start at a slight upward angle and slowly, steadily zoom out until the general becomes a tiny dot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014209755696356297 + ], + "opposite_logprob_tokens": [ + -4.264209747314453 + ], + "logprob_reward": -0.014209755696356297, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014209755696356297, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.28349778056144714, + 0.17533598840236664, + 0.10760774463415146, + 0.04906601458787918, + 0.058038122951984406, + 0.05970318242907524, + 0.06652556359767914, + 0.04072624817490578, + 0.0865481048822403, + 0.07295125722885132 + ], + "credit_map": [ + 0.10578470677137375, + 0.027415646240115166, + 0.01678493060171604, + 0.012337918393313885, + 0.015760090202093124, + 0.021030103787779808, + 0.01642749272286892, + 0.06795688718557358, + 0.06842252612113953, + 0.021565284579992294, + 0.006621812004595995, + 0.006669085007160902, + 0.009664976969361305, + 0.010615445673465729, + 0.006950356997549534, + 0.04482649266719818, + 0.021232658997178078, + 0.006665671709924936, + 0.011483828537166119, + 0.01913304068148136, + 0.013640481047332287, + 0.006913122721016407, + 0.011546251364052296, + 0.016992690041661263, + 0.008768226020038128, + 0.0028651291504502296, + 0.003888204460963607, + 0.005904053803533316, + 0.006447564344853163, + 0.008653928525745869, + 0.007380285300314426, + 0.005158626474440098, + 0.011891110800206661, + 0.003944981377571821, + 0.0037601771764457226, + 0.004774459172040224, + 0.008615021593868732, + 0.014371910132467747, + 0.006785208825021982, + 0.0038952562026679516, + 0.009843339212238789, + 0.004244564101099968, + 0.003485144581645727, + 0.007059135939925909, + 0.00890403427183628, + 0.014494877308607101, + 0.006849717348814011, + 0.004822366870939732, + 0.009914477355778217, + 0.0052001141011714935, + 0.007552864495664835, + 0.0071897548623383045, + 0.009265557862818241, + 0.010923977941274643, + 0.00861172191798687, + 0.007867099717259407, + 0.007243562955409288, + 0.002516514156013727, + 0.0029955622740089893, + 0.003919622860848904, + 0.004586778115481138, + 0.008182307705283165, + 0.006909061688929796, + 0.004372836090624332, + 0.017076535150408745, + 0.0058707972057163715, + 0.005835267715156078, + 0.006558486260473728, + 0.0069375247694551945, + 0.02766021341085434, + 0.009080671705305576, + 0.0075286072678864, + 0.0076391552574932575, + 0.006564820185303688, + 0.007263139821588993, + 0.014360584318637848, + 0.00684062484651804, + 0.011232237331569195, + 0.00906034279614687, + 0.009990348480641842 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.014209755696356297 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on a vast, empty battlefield at dusk with cracked, scorched earth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a general wearing ancient armor made of bronze and black iron visible at the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the general's armor feature beast totems on the shoulder plates and show signs of sword scars and rust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the general perform the action of jerking his head back with bulging neck veins and opening his mouth wide?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start at a slight upward angle and slowly, steadily zoom out until the general becomes a tiny dot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.3134599030017853 + ], + "opposite_logprob_tokens": [ + -1.313459873199463 + ], + "logprob_reward": -0.3134599030017853, + "yes_no_margin_reward": 1.0, + "token_credit_reward": -0.3134599030017853, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09398285299539566, + 0.16432660818099976, + 0.05803860351443291, + 0.13871079683303833, + 0.05645168945193291, + 0.07150980830192566, + 0.08846350759267807, + 0.06876260042190552, + 0.16991165280342102, + 0.08984187990427017 + ], + "credit_map": [ + 0.012985789217054844, + 0.011365188285708427, + 0.011470875702798367, + 0.010979831218719482, + 0.011207068338990211, + 0.014589653350412846, + 0.011260642670094967, + 0.01012380514293909, + 0.05987241491675377, + 0.019009780138731003, + 0.014421799220144749, + 0.007521781604737043, + 0.008920950815081596, + 0.009625740349292755, + 0.006098664365708828, + 0.038855474442243576, + 0.010267907753586769, + 0.003683909308165312, + 0.005109625868499279, + 0.008498864248394966, + 0.008417795412242413, + 0.004877464380115271, + 0.00932474434375763, + 0.00785829033702612, + 0.02900531515479088, + 0.00878194347023964, + 0.008030414581298828, + 0.012378624640405178, + 0.01343885064125061, + 0.04169650003314018, + 0.015823950991034508, + 0.009555204771459103, + 0.0063090017065405846, + 0.0047486647963523865, + 0.005788098555058241, + 0.006955892313271761, + 0.011938346549868584, + 0.006290392018854618, + 0.010374333709478378, + 0.004046956077218056, + 0.00992569886147976, + 0.005164338741451502, + 0.004488289821892977, + 0.006354636047035456, + 0.016156328842043877, + 0.015104947611689568, + 0.008170811459422112, + 0.006144754122942686, + 0.015071478672325611, + 0.006234080996364355, + 0.007334758993238211, + 0.008656314574182034, + 0.01321870181709528, + 0.018702197819948196, + 0.01001470722258091, + 0.009231270290911198, + 0.00968154426664114, + 0.004448865540325642, + 0.005853762850165367, + 0.009066645056009293, + 0.009754812344908714, + 0.008725151419639587, + 0.013407586142420769, + 0.007824234664440155, + 0.034003276377916336, + 0.012433444149792194, + 0.010601450689136982, + 0.012062918394804, + 0.012816572561860085, + 0.05842838063836098, + 0.01650463417172432, + 0.013060973957180977, + 0.010213046334683895, + 0.007639908231794834, + 0.0085316626355052, + 0.015901997685432434, + 0.007939922623336315, + 0.017490806058049202, + 0.011145866475999355, + 0.010978667996823788 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 1, + 3 + ], + "selected_reward": -0.3134599030017853 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on a vast, empty battlefield at dusk with cracked, scorched earth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a general wearing ancient armor made of bronze and black iron visible at the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the general's armor feature beast totems on the shoulder plates and show signs of sword scars and rust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the general perform the action of jerking his head back with bulging neck veins and opening his mouth wide?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start at a slight upward angle and slowly, steadily zoom out until the general becomes a tiny dot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2255474478006363 + ], + "opposite_logprob_tokens": [ + -1.600547432899475 + ], + "logprob_reward": -0.2255474478006363, + "yes_no_margin_reward": 1.375, + "token_credit_reward": -0.2255474478006363, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10714530944824219, + 0.042538903653621674, + 0.07826629281044006, + 0.11400488018989563, + 0.05085994303226471, + 0.09639498591423035, + 0.08419755101203918, + 0.1020113006234169, + 0.19493424892425537, + 0.1296466439962387 + ], + "credit_map": [ + 0.033876627683639526, + 0.008435624651610851, + 0.005257417447865009, + 0.006377746816724539, + 0.009587906301021576, + 0.007109893951565027, + 0.013011193834245205, + 0.02348889783024788, + 0.004342475440353155, + 0.0025720202829688787, + 0.004441495053470135, + 0.0052090054377913475, + 0.01001035887748003, + 0.0033518928103148937, + 0.007609417662024498, + 0.005002240650355816, + 0.013926603831350803, + 0.005154305603355169, + 0.005456169601529837, + 0.012000269256532192, + 0.013475162908434868, + 0.00683400547131896, + 0.010333397425711155, + 0.011086375452578068, + 0.026939164847135544, + 0.005606131162494421, + 0.005434887018054724, + 0.007613154128193855, + 0.011869696900248528, + 0.03851839900016785, + 0.012230799533426762, + 0.005792646203190088, + 0.007199530489742756, + 0.004414689727127552, + 0.004937725607305765, + 0.006539271213114262, + 0.007630195003002882, + 0.007496885489672422, + 0.007837320677936077, + 0.0048043266870081425, + 0.014650215394794941, + 0.00672198785468936, + 0.0056898933835327625, + 0.007852853275835514, + 0.019252896308898926, + 0.022683503106236458, + 0.011091132648289204, + 0.00845250766724348, + 0.009510398842394352, + 0.005727531388401985, + 0.007418540772050619, + 0.009933039546012878, + 0.012866327539086342, + 0.012198644690215588, + 0.0148979676887393, + 0.01164509728550911, + 0.010181126184761524, + 0.008134887553751469, + 0.009471623227000237, + 0.013593916781246662, + 0.01782810688018799, + 0.011517095379531384, + 0.01882185786962509, + 0.01246268767863512, + 0.03568962216377258, + 0.017135299742221832, + 0.016599446535110474, + 0.0152616947889328, + 0.0155832814052701, + 0.06046032905578613, + 0.019566373899579048, + 0.014638201333582401, + 0.014225049875676632, + 0.011377321556210518, + 0.01247925590723753, + 0.01979454793035984, + 0.016386376693844795, + 0.020672790706157684, + 0.020464876666665077, + 0.014246433041989803 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 3 + ], + "selected_reward": -0.2255474478006363 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "medium shot begins at a slight upward angle, then slowly and steadily zooms out. At the center of the frame is a general clad in ancient armor, standing on a vast, empty battlefield at dusk. His armor, forged from a blend of bronze and black iron, is covered in the scars of swords and the rust of time. The beast totems carved on his shoulder plates still shone fiercely in the dim light. His weathered face, sunken eyes and furrowed brow, is filled with despair and resignation, his beard matted with dried blood and dust. He jerks his head back, the veins in his neck bulging with the sheer force. His mouth opens wide, and he lets out a silent yet heartbreaking roar into the low, leaden clouds. His fists clench, his body trembling with the outpouring of emotion. Beneath his feet lies cracked, scorched earth, strewn with broken flags, shattered weapons, and unclaimed helmets. The last blood-red rays of sunset struggled to pierce through the clouds, gilding the outline of his armor with a poignant golden edge and casting a long, distorted shadow behind him. As the camera continuously and steadily zoomed out, the general's figure gradually shrank across the vast battlefield, ultimately becoming a tiny, lonely black dot on the charred earth. The entire scene was filled with a sense of epic tragedy and the endless desolation of a hero's final journey.", + "seed": 3161, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000056/group_4/index_5204_seed_3161_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": -0.03326022336259484, + "vlm_logprob": -0.03326022336259484, + "vlm_yes_no_margin": 4.225, + "vlm_token_credit": -0.03326022336259484, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on a vast, empty battlefield at dusk with cracked, scorched earth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a general wearing ancient armor made of bronze and black iron visible at the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the general's armor feature beast totems on the shoulder plates and show signs of sword scars and rust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the general perform the action of jerking his head back with bulging neck veins and opening his mouth wide?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start at a slight upward angle and slowly, steadily zoom out until the general becomes a tiny dot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008780678734183311 + ], + "opposite_logprob_tokens": [ + -5.133780479431152 + ], + "logprob_reward": -0.008780678734183311, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.008780678734183311, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14784970879554749, + 0.10952354222536087, + 0.08560363203287125, + 0.0749417319893837, + 0.08466188609600067, + 0.10402114689350128, + 0.09993326663970947, + 0.08284059911966324, + 0.10089949518442154, + 0.10972501337528229 + ], + "credit_map": [ + 0.01983349770307541, + 0.013860546983778477, + 0.015581770800054073, + 0.01641068235039711, + 0.012271689251065254, + 0.02253490872681141, + 0.016628092154860497, + 0.03072851151227951, + 0.014608084224164486, + 0.009549172595143318, + 0.01232180092483759, + 0.018108349293470383, + 0.011083950288593769, + 0.008089159615337849, + 0.018546724691987038, + 0.01721630059182644, + 0.00840840209275484, + 0.008676779456436634, + 0.018057875335216522, + 0.009583670645952225, + 0.008234011009335518, + 0.008326815441250801, + 0.012064781039953232, + 0.012251297943294048, + 0.012998034246265888, + 0.008388062939047813, + 0.01299903355538845, + 0.005361123476177454, + 0.008046742528676987, + 0.01074292790144682, + 0.006870071403682232, + 0.009535742923617363, + 0.014270108193159103, + 0.011972436681389809, + 0.009033255279064178, + 0.009388715028762817, + 0.010896270163357258, + 0.008586474694311619, + 0.01176224835216999, + 0.008752377703785896, + 0.0076550706289708614, + 0.008282209746539593, + 0.025179920718073845, + 0.013027210719883442, + 0.007568469271063805, + 0.011667451821267605, + 0.012697500176727772, + 0.017943309620022774, + 0.011903831735253334, + 0.007668221835047007, + 0.02027852274477482, + 0.011894883587956429, + 0.010604679584503174, + 0.011005247943103313, + 0.010281113907694817, + 0.016296764835715294, + 0.008365411311388016, + 0.012341978028416634, + 0.008694829419255257, + 0.010302006267011166, + 0.012311004102230072, + 0.010550422593951225, + 0.009895826689898968, + 0.010379128158092499, + 0.007154627237468958, + 0.018092669546604156, + 0.014383172616362572, + 0.012254228815436363, + 0.01087202038615942, + 0.014654087834060192, + 0.010358810424804688, + 0.013129879720509052, + 0.009615880437195301, + 0.01220772322267294, + 0.012178328819572926, + 0.015795107930898666, + 0.01687481813132763, + 0.015219821594655514, + 0.015492566861212254, + 0.012340767309069633 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 1 + ], + "selected_reward": -0.008780678734183311 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on a vast, empty battlefield at dusk with cracked, scorched earth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a general wearing ancient armor made of bronze and black iron visible at the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the general's armor feature beast totems on the shoulder plates and show signs of sword scars and rust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the general perform the action of jerking his head back with bulging neck veins and opening his mouth wide?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start at a slight upward angle and slowly, steadily zoom out until the general becomes a tiny dot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006751463748514652 + ], + "opposite_logprob_tokens": [ + -5.006751537322998 + ], + "logprob_reward": -0.006751463748514652, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006751463748514652, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13207776844501495, + 0.16060908138751984, + 0.14964300394058228, + 0.09031663835048676, + 0.09158723056316376, + 0.05313144624233246, + 0.050869252532720566, + 0.0673748329281807, + 0.07808582484722137, + 0.12630489468574524 + ], + "credit_map": [ + 0.02064545825123787, + 0.01222437433898449, + 0.014993304386734962, + 0.016746865585446358, + 0.012933488935232162, + 0.01600923016667366, + 0.019675856456160545, + 0.018849197775125504, + 0.02544666826725006, + 0.015047106891870499, + 0.018406469374895096, + 0.019868802279233932, + 0.012539892457425594, + 0.013718726113438606, + 0.029556546360254288, + 0.026024870574474335, + 0.014504944905638695, + 0.009320743381977081, + 0.020242203027009964, + 0.018196256831288338, + 0.015563001856207848, + 0.010938627645373344, + 0.03661274537444115, + 0.024264473468065262, + 0.005901198368519545, + 0.00520078232511878, + 0.007574582472443581, + 0.011959165334701538, + 0.01518424041569233, + 0.006174034904688597, + 0.0263706985861063, + 0.011951930820941925, + 0.008346821181476116, + 0.00795926246792078, + 0.013252416625618935, + 0.01377140637487173, + 0.013657167553901672, + 0.00805255863815546, + 0.015194609761238098, + 0.01135298702865839, + 0.0045355819165706635, + 0.005365313030779362, + 0.0071395039558410645, + 0.007562585640698671, + 0.005401081405580044, + 0.0046300808899104595, + 0.006720829289406538, + 0.011776471510529518, + 0.004623470362275839, + 0.005335227120667696, + 0.008444179780781269, + 0.006478581111878157, + 0.004959448240697384, + 0.0059301289729774, + 0.00625710841268301, + 0.008841108530759811, + 0.0072348774410784245, + 0.012430493719875813, + 0.006190901156514883, + 0.008469657972455025, + 0.009034198708832264, + 0.008891543373465538, + 0.0065527004189789295, + 0.008570467121899128, + 0.005668019875884056, + 0.017839323729276657, + 0.007065242156386375, + 0.014166569337248802, + 0.006583886686712503, + 0.01452429499477148, + 0.006959257181733847, + 0.005279233679175377, + 0.010021730326116085, + 0.016439728438854218, + 0.01386246643960476, + 0.024410860612988472, + 0.015694329515099525, + 0.016966693103313446, + 0.016618860885500908, + 0.012290231883525848 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 2, + 0 + ], + "selected_reward": -0.006751463748514652 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on a vast, empty battlefield at dusk with cracked, scorched earth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a general wearing ancient armor made of bronze and black iron visible at the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the general's armor feature beast totems on the shoulder plates and show signs of sword scars and rust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the general perform the action of jerking his head back with bulging neck veins and opening his mouth wide?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start at a slight upward angle and slowly, steadily zoom out until the general becomes a tiny dot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016093002632260323 + ], + "opposite_logprob_tokens": [ + -4.141092777252197 + ], + "logprob_reward": -0.016093002632260323, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016093002632260323, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10865606367588043, + 0.1730119287967682, + 0.15187305212020874, + 0.04716050252318382, + 0.1078668162226677, + 0.09122630953788757, + 0.06216159462928772, + 0.06583192944526672, + 0.08910898119211197, + 0.10310279577970505 + ], + "credit_map": [ + 0.01766384020447731, + 0.011856711469590664, + 0.016634531319141388, + 0.01012491062283516, + 0.011946098878979683, + 0.014381143264472485, + 0.011772950179874897, + 0.01427588053047657, + 0.015925606712698936, + 0.013660482130944729, + 0.022539181634783745, + 0.02931797504425049, + 0.026490245014429092, + 0.012269788421690464, + 0.027386406436562538, + 0.02542225830256939, + 0.02181258425116539, + 0.01615397073328495, + 0.017696045339107513, + 0.016346819698810577, + 0.01327457930892706, + 0.015027103014290333, + 0.02967410907149315, + 0.021887848153710365, + 0.0037137751933187246, + 0.00369330495595932, + 0.004949384368956089, + 0.005941743962466717, + 0.007624086923897266, + 0.0033995085395872593, + 0.011687000282108784, + 0.006151699926704168, + 0.01404572930186987, + 0.01046739611774683, + 0.01694214902818203, + 0.015103612095117569, + 0.011368868872523308, + 0.009008429013192654, + 0.017591649666428566, + 0.013338985852897167, + 0.008839205838739872, + 0.007747945841401815, + 0.010891125537455082, + 0.013953234069049358, + 0.011794837191700935, + 0.008324098773300648, + 0.014894497580826283, + 0.014781368896365166, + 0.007435225881636143, + 0.004740891046822071, + 0.012491290457546711, + 0.008115706965327263, + 0.005642247386276722, + 0.006787450052797794, + 0.006105831358581781, + 0.010842952877283096, + 0.006991751957684755, + 0.01093896571546793, + 0.007440797984600067, + 0.007195703219622374, + 0.00968281738460064, + 0.008894593454897404, + 0.005061805713921785, + 0.009625489823520184, + 0.006521380506455898, + 0.010479414835572243, + 0.012335246428847313, + 0.01487021055072546, + 0.008215365931391716, + 0.01308311615139246, + 0.011238139122724533, + 0.012366104871034622, + 0.009784799069166183, + 0.01051880232989788, + 0.009646519087255001, + 0.02095334418118, + 0.01194604579359293, + 0.013689948245882988, + 0.013433399610221386, + 0.013129942119121552 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 2, + 0 + ], + "selected_reward": -0.016093002632260323 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on a vast, empty battlefield at dusk with cracked, scorched earth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a general wearing ancient armor made of bronze and black iron visible at the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the general's armor feature beast totems on the shoulder plates and show signs of sword scars and rust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the general perform the action of jerking his head back with bulging neck veins and opening his mouth wide?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start at a slight upward angle and slowly, steadily zoom out until the general becomes a tiny dot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.12704962491989136 + ], + "opposite_logprob_tokens": [ + -2.127049684524536 + ], + "logprob_reward": -0.12704962491989136, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.12704962491989136, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14093513786792755, + 0.2526226043701172, + 0.05760873854160309, + 0.029927168041467667, + 0.09431079775094986, + 0.06507585197687149, + 0.03850635141134262, + 0.10900074243545532, + 0.13178297877311707, + 0.08022958040237427 + ], + "credit_map": [ + 0.020154083147644997, + 0.013341957703232765, + 0.01349677424877882, + 0.011982761323451996, + 0.01353384368121624, + 0.0351787693798542, + 0.01787785440683365, + 0.015369103290140629, + 0.05465307831764221, + 0.020716341212391853, + 0.02507857047021389, + 0.022236786782741547, + 0.023703988641500473, + 0.022875240072607994, + 0.031875643879175186, + 0.05148298293352127, + 0.00793335773050785, + 0.00828647892922163, + 0.0063635557889938354, + 0.007278659380972385, + 0.0057404604740440845, + 0.006800745148211718, + 0.00808688998222351, + 0.007118595764040947, + 0.0033702640794217587, + 0.002332785865291953, + 0.003640207462012768, + 0.004041345790028572, + 0.0035552815534174442, + 0.0029421821236610413, + 0.006370170041918755, + 0.003674933221191168, + 0.0051068332977592945, + 0.005062634125351906, + 0.012711659073829651, + 0.018456723541021347, + 0.012353514321148396, + 0.00796644389629364, + 0.019550103694200516, + 0.013102885335683823, + 0.005928520578891039, + 0.0052442122250795364, + 0.008330068551003933, + 0.01058995258063078, + 0.008413614705204964, + 0.006008072756230831, + 0.011047537438571453, + 0.009513873606920242, + 0.0035464982502162457, + 0.003779452061280608, + 0.004814756102859974, + 0.006897210143506527, + 0.00574337737634778, + 0.0035332334227859974, + 0.005342686548829079, + 0.004849142394959927, + 0.010565079748630524, + 0.016916127875447273, + 0.011451677419245243, + 0.014714398421347141, + 0.012488496489822865, + 0.013588626869022846, + 0.013394373469054699, + 0.015881966799497604, + 0.010931962169706821, + 0.015184099785983562, + 0.0174259040504694, + 0.02345721609890461, + 0.015793126076459885, + 0.0190370362251997, + 0.013298592530190945, + 0.01665506511926651, + 0.005980238318443298, + 0.008802159689366817, + 0.010411528870463371, + 0.015077491290867329, + 0.010139829479157925, + 0.009814955294132233, + 0.010979210026562214, + 0.00902416929602623 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 8 + ], + "selected_reward": -0.12704962491989136 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on a vast, empty battlefield at dusk with cracked, scorched earth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a general wearing ancient armor made of bronze and black iron visible at the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the general's armor feature beast totems on the shoulder plates and show signs of sword scars and rust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the general perform the action of jerking his head back with bulging neck veins and opening his mouth wide?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start at a slight upward angle and slowly, steadily zoom out until the general becomes a tiny dot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007626346778124571 + ], + "opposite_logprob_tokens": [ + -4.882626533508301 + ], + "logprob_reward": -0.007626346778124571, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007626346778124571, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10217349976301193, + 0.07371088117361069, + 0.05650259926915169, + 0.04557885602116585, + 0.08860061317682266, + 0.060682620853185654, + 0.039075203239917755, + 0.10847438871860504, + 0.14103133976459503, + 0.2841699421405792 + ], + "credit_map": [ + 0.01935870386660099, + 0.007044115103781223, + 0.013383972458541393, + 0.009348793886601925, + 0.009646670892834663, + 0.024723511189222336, + 0.010178091935813427, + 0.008489638566970825, + 0.02094615064561367, + 0.008158640936017036, + 0.005630312021821737, + 0.0043198345229029655, + 0.005005185026675463, + 0.007637572009116411, + 0.006231514271348715, + 0.015781665220856667, + 0.0071001662872731686, + 0.006032015662640333, + 0.00841359980404377, + 0.006166909355670214, + 0.005991941783577204, + 0.00633954256772995, + 0.008753526955842972, + 0.007704896852374077, + 0.007227474357932806, + 0.004983135499060154, + 0.004583751317113638, + 0.005297035910189152, + 0.005810943897813559, + 0.0044302623718976974, + 0.006241188384592533, + 0.007005064282566309, + 0.009177591651678085, + 0.007705747149884701, + 0.015638405457139015, + 0.01264673750847578, + 0.007002863567322493, + 0.006734501104801893, + 0.011346876621246338, + 0.01834789477288723, + 0.006477645598351955, + 0.006005217786878347, + 0.008593013510107994, + 0.007642701268196106, + 0.00709925452247262, + 0.0063698855228722095, + 0.008355029858648777, + 0.010139871388673782, + 0.0031718574464321136, + 0.003796230535954237, + 0.005256700329482555, + 0.0073352474719285965, + 0.0044718594290316105, + 0.0037110645789653063, + 0.006072345655411482, + 0.005259897094219923, + 0.009867831133306026, + 0.020655740052461624, + 0.011349601671099663, + 0.015628980472683907, + 0.014095385558903217, + 0.014775767922401428, + 0.009941439144313335, + 0.012159640900790691, + 0.009256357327103615, + 0.02437780424952507, + 0.016629450023174286, + 0.027204420417547226, + 0.01231638714671135, + 0.025239204987883568, + 0.014818653464317322, + 0.01118906494230032, + 0.025727082043886185, + 0.031760990619659424, + 0.031169675290584564, + 0.04496510326862335, + 0.034468211233615875, + 0.037211861461400986, + 0.050209399312734604, + 0.028657646849751472 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 7 + ], + "selected_reward": -0.007626346778124571 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "medium shot begins at a slight upward angle, then slowly and steadily zooms out. At the center of the frame is a general clad in ancient armor, standing on a vast, empty battlefield at dusk. His armor, forged from a blend of bronze and black iron, is covered in the scars of swords and the rust of time. The beast totems carved on his shoulder plates still shone fiercely in the dim light. His weathered face, sunken eyes and furrowed brow, is filled with despair and resignation, his beard matted with dried blood and dust. He jerks his head back, the veins in his neck bulging with the sheer force. His mouth opens wide, and he lets out a silent yet heartbreaking roar into the low, leaden clouds. His fists clench, his body trembling with the outpouring of emotion. Beneath his feet lies cracked, scorched earth, strewn with broken flags, shattered weapons, and unclaimed helmets. The last blood-red rays of sunset struggled to pierce through the clouds, gilding the outline of his armor with a poignant golden edge and casting a long, distorted shadow behind him. As the camera continuously and steadily zoomed out, the general's figure gradually shrank across the vast battlefield, ultimately becoming a tiny, lonely black dot on the charred earth. The entire scene was filled with a sense of epic tragedy and the endless desolation of a hero's final journey.", + "seed": 3162, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000056/group_4/index_5204_seed_3162_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": -0.015016143396496772, + "vlm_logprob": -0.015016143396496772, + "vlm_yes_no_margin": 4.325, + "vlm_token_credit": -0.015016143396496772, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on a vast, empty battlefield at dusk with cracked, scorched earth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a general wearing ancient armor made of bronze and black iron visible at the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the general's armor feature beast totems on the shoulder plates and show signs of sword scars and rust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the general perform the action of jerking his head back with bulging neck veins and opening his mouth wide?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start at a slight upward angle and slowly, steadily zoom out until the general becomes a tiny dot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009944183751940727 + ], + "opposite_logprob_tokens": [ + -5.009943962097168 + ], + "logprob_reward": -0.009944183751940727, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.009944183751940727, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23008431494235992, + 0.06368711590766907, + 0.12248454988002777, + 0.05532195791602135, + 0.058167099952697754, + 0.08573044091463089, + 0.08460231125354767, + 0.08793945610523224, + 0.10704641789197922, + 0.10493627190589905 + ], + "credit_map": [ + 0.020026059821248055, + 0.021116193383932114, + 0.03346569091081619, + 0.03819732740521431, + 0.03527277708053589, + 0.025750208646059036, + 0.02697608806192875, + 0.02927996963262558, + 0.0060449386946856976, + 0.006577068008482456, + 0.007140930742025375, + 0.007804405875504017, + 0.015048682689666748, + 0.005953623913228512, + 0.006702585145831108, + 0.008414880372583866, + 0.014587180688977242, + 0.013403501361608505, + 0.02364186756312847, + 0.012241359800100327, + 0.010961097665131092, + 0.012557553127408028, + 0.013626119121909142, + 0.02146587148308754, + 0.006068017799407244, + 0.004900026135146618, + 0.0060836332850158215, + 0.008634145371615887, + 0.006218788679689169, + 0.005144288297742605, + 0.011364680714905262, + 0.006908375304192305, + 0.008770718239247799, + 0.008676182478666306, + 0.006533547304570675, + 0.005510763265192509, + 0.00759689649567008, + 0.006827095523476601, + 0.008504798635840416, + 0.005747092887759209, + 0.006540164817124605, + 0.007251868024468422, + 0.010025170631706715, + 0.013283882290124893, + 0.011776193976402283, + 0.008079804480075836, + 0.019169596955180168, + 0.009603766724467278, + 0.011481410823762417, + 0.009373991750180721, + 0.016890857368707657, + 0.007505620364099741, + 0.010150679387152195, + 0.010445144027471542, + 0.008449788205325603, + 0.010304822586476803, + 0.008360221050679684, + 0.009013276547193527, + 0.012660848908126354, + 0.012001829221844673, + 0.01188686117529869, + 0.010148254223167896, + 0.011894146911799908, + 0.011974011547863483, + 0.011117572896182537, + 0.013059979304671288, + 0.01268745306879282, + 0.013070540502667427, + 0.01174123678356409, + 0.013770167715847492, + 0.0161199439316988, + 0.015479530207812786, + 0.00784717034548521, + 0.008784966543316841, + 0.013790728524327278, + 0.010935911908745766, + 0.012752319686114788, + 0.013333672657608986, + 0.018439602106809616, + 0.01905190758407116 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 8 + ], + "selected_reward": -0.009944183751940727 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on a vast, empty battlefield at dusk with cracked, scorched earth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a general wearing ancient armor made of bronze and black iron visible at the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the general's armor feature beast totems on the shoulder plates and show signs of sword scars and rust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the general perform the action of jerking his head back with bulging neck veins and opening his mouth wide?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start at a slight upward angle and slowly, steadily zoom out until the general becomes a tiny dot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01820603385567665 + ], + "opposite_logprob_tokens": [ + -4.0182061195373535 + ], + "logprob_reward": -0.01820603385567665, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.01820603385567665, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13980060815811157, + 0.07532164454460144, + 0.1672808676958084, + 0.11433187872171402, + 0.08475634455680847, + 0.20627696812152863, + 0.06417689472436905, + 0.044368427246809006, + 0.07322926074266434, + 0.0304571446031332 + ], + "credit_map": [ + 0.014595132321119308, + 0.012255399487912655, + 0.015133609063923359, + 0.022946391254663467, + 0.026642154902219772, + 0.014032627455890179, + 0.018581870943307877, + 0.015613420866429806, + 0.0056757996790111065, + 0.007952465675771236, + 0.006017951760441065, + 0.010992787778377533, + 0.015151012688875198, + 0.0074995579198002815, + 0.009508425369858742, + 0.012523646466434002, + 0.010576154105365276, + 0.014083638787269592, + 0.013317926786839962, + 0.03625849261879921, + 0.019969608634710312, + 0.013882274739444256, + 0.03697390854358673, + 0.022218873724341393, + 0.0055513205006718636, + 0.00783555954694748, + 0.010851601138710976, + 0.02258506417274475, + 0.018809355795383453, + 0.011335361748933792, + 0.024670617654919624, + 0.012692990712821484, + 0.0037801063153892756, + 0.005560462828725576, + 0.006621618755161762, + 0.012627187184989452, + 0.02011672407388687, + 0.008574851788580418, + 0.018424760550260544, + 0.009050633758306503, + 0.023178130388259888, + 0.021242910996079445, + 0.017576610669493675, + 0.030308295041322708, + 0.02985779568552971, + 0.018477002158761024, + 0.041790921241045, + 0.02384530007839203, + 0.007999836467206478, + 0.007396929431706667, + 0.006723866332322359, + 0.00874421838670969, + 0.011371933855116367, + 0.006552026141434908, + 0.008892033249139786, + 0.006496052257716656, + 0.0023212204687297344, + 0.0037752706557512283, + 0.0046718730591237545, + 0.008323898538947105, + 0.006292705424129963, + 0.004602116998285055, + 0.008609255775809288, + 0.005772084929049015, + 0.007742776535451412, + 0.010898171924054623, + 0.008900451473891735, + 0.010403322987258434, + 0.007839229889214039, + 0.00792047381401062, + 0.011665529571473598, + 0.007859302684664726, + 0.002210908103734255, + 0.0031901688780635595, + 0.004580991808325052, + 0.0034587064292281866, + 0.004134258721023798, + 0.0031502607744187117, + 0.005266143474727869, + 0.004465707577764988 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 2, + 0 + ], + "selected_reward": -0.01820603385567665 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on a vast, empty battlefield at dusk with cracked, scorched earth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a general wearing ancient armor made of bronze and black iron visible at the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the general's armor feature beast totems on the shoulder plates and show signs of sword scars and rust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the general perform the action of jerking his head back with bulging neck veins and opening his mouth wide?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start at a slight upward angle and slowly, steadily zoom out until the general becomes a tiny dot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"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": [ + 0.241362527012825, + 0.06360534578561783, + 0.11034784466028214, + 0.06701907515525818, + 0.06689074635505676, + 0.11841394007205963, + 0.07083024084568024, + 0.04607591778039932, + 0.09924352914094925, + 0.11621079593896866 + ], + "credit_map": [ + 0.027578365057706833, + 0.023880084976553917, + 0.028459202498197556, + 0.03886181116104126, + 0.046397820115089417, + 0.023072833195328712, + 0.02647760696709156, + 0.026634788140654564, + 0.004840068984776735, + 0.008297238498926163, + 0.00515818502753973, + 0.0077705164439976215, + 0.013680445030331612, + 0.005965632852166891, + 0.007726719602942467, + 0.010166540741920471, + 0.009940775111317635, + 0.011423670686781406, + 0.010265608318150043, + 0.02255743369460106, + 0.015059981495141983, + 0.009673520922660828, + 0.020469533279538155, + 0.01095732394605875, + 0.004082196392118931, + 0.004348086193203926, + 0.006130406633019447, + 0.008916628547012806, + 0.00813213549554348, + 0.007083808537572622, + 0.007537391968071461, + 0.020788419991731644, + 0.0038141515105962753, + 0.005638621747493744, + 0.008614004589617252, + 0.010468017309904099, + 0.009247524663805962, + 0.006200628820806742, + 0.015017635188996792, + 0.007890159264206886, + 0.015525075607001781, + 0.013931632041931152, + 0.009737620130181313, + 0.0135872857645154, + 0.020039275288581848, + 0.011103525757789612, + 0.020545240491628647, + 0.013944289647042751, + 0.009988684207201004, + 0.008435841649770737, + 0.008250189945101738, + 0.007473009172827005, + 0.010689159855246544, + 0.007213148754090071, + 0.010618902742862701, + 0.008161306381225586, + 0.005020574200898409, + 0.004905987065285444, + 0.004880543332546949, + 0.007330010179430246, + 0.006604946684092283, + 0.004932524636387825, + 0.006375035271048546, + 0.006026299670338631, + 0.010790209285914898, + 0.012445563450455666, + 0.010834332555532455, + 0.012249878607690334, + 0.01140973623842001, + 0.011781435459852219, + 0.019223343580961227, + 0.010509027168154716, + 0.008569302037358284, + 0.008986513130366802, + 0.014665159396827221, + 0.014090496115386486, + 0.018086301162838936, + 0.012992754578590393, + 0.024485649541020393, + 0.014334622770547867 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 9 + ], + "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 scene set on a vast, empty battlefield at dusk with cracked, scorched earth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a general wearing ancient armor made of bronze and black iron visible at the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the general's armor feature beast totems on the shoulder plates and show signs of sword scars and rust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the general perform the action of jerking his head back with bulging neck veins and opening his mouth wide?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start at a slight upward angle and slowly, steadily zoom out until the general becomes a tiny dot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023295411840081215 + ], + "opposite_logprob_tokens": [ + -3.7732954025268555 + ], + "logprob_reward": -0.023295411840081215, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023295411840081215, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10986261069774628, + 0.4063539206981659, + 0.042275018990039825, + 0.03829125687479973, + 0.030617251992225647, + 0.08616530895233154, + 0.09917087852954865, + 0.05853896588087082, + 0.05925286188721657, + 0.06947194039821625 + ], + "credit_map": [ + 0.014485475607216358, + 0.015080827288329601, + 0.015371913090348244, + 0.014037322252988815, + 0.014702830463647842, + 0.011044556275010109, + 0.01187495980411768, + 0.013264715671539307, + 0.030689623206853867, + 0.03758881613612175, + 0.02878388576209545, + 0.03777654469013214, + 0.11698605865240097, + 0.03876742720603943, + 0.05615125223994255, + 0.05961030349135399, + 0.00255821505561471, + 0.0038436364848166704, + 0.004406946711242199, + 0.00858624093234539, + 0.005368155427277088, + 0.003995803184807301, + 0.008619001135230064, + 0.004897015169262886, + 0.0029149996116757393, + 0.0032600939739495516, + 0.0037574598100036383, + 0.004634204786270857, + 0.006814388558268547, + 0.0039125243201851845, + 0.008095895871520042, + 0.004901684820652008, + 0.0025980693753808737, + 0.002965508261695504, + 0.004416018258780241, + 0.004423712845891714, + 0.004863088484853506, + 0.0027238214388489723, + 0.004174362868070602, + 0.004452669061720371, + 0.02145364135503769, + 0.016806356608867645, + 0.007758013438433409, + 0.005473763216286898, + 0.009954043664038181, + 0.010832574218511581, + 0.006335962563753128, + 0.007550940848886967, + 0.008201819844543934, + 0.00995924137532711, + 0.011732209473848343, + 0.0176017414778471, + 0.010287175886332989, + 0.009043879806995392, + 0.020737938582897186, + 0.011606867425143719, + 0.004082180093973875, + 0.004989347420632839, + 0.005821008235216141, + 0.01044608186930418, + 0.007781366817653179, + 0.006584919523447752, + 0.011102267540991306, + 0.007731787394732237, + 0.0036091364454478025, + 0.004452432505786419, + 0.005728606600314379, + 0.01185312494635582, + 0.008010041899979115, + 0.005937675945460796, + 0.012635353021323681, + 0.00702648377045989, + 0.0038409957196563482, + 0.004794151522219181, + 0.008685213513672352, + 0.008272649720311165, + 0.010628613643348217, + 0.007710525766015053, + 0.017189806327223778, + 0.008349977433681488 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 6 + ], + "selected_reward": -0.023295411840081215 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on a vast, empty battlefield at dusk with cracked, scorched earth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a general wearing ancient armor made of bronze and black iron visible at the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the general's armor feature beast totems on the shoulder plates and show signs of sword scars and rust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the general perform the action of jerking his head back with bulging neck veins and opening his mouth wide?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start at a slight upward angle and slowly, steadily zoom out until the general becomes a tiny dot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012535941787064075 + ], + "opposite_logprob_tokens": [ + -4.38753604888916 + ], + "logprob_reward": -0.012535941787064075, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012535941787064075, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1432773619890213, + 0.04785360023379326, + 0.05894197151064873, + 0.04163569211959839, + 0.05106332153081894, + 0.10884002596139908, + 0.11295779794454575, + 0.08841781318187714, + 0.11711688339710236, + 0.22989556193351746 + ], + "credit_map": [ + 0.01313599944114685, + 0.01499883458018303, + 0.016702819615602493, + 0.02206483483314514, + 0.025243660435080528, + 0.012924788519740105, + 0.019390244036912918, + 0.018816182389855385, + 0.00397915206849575, + 0.005119952838867903, + 0.006216112524271011, + 0.004772542044520378, + 0.008951405994594097, + 0.004301031120121479, + 0.0055642398074269295, + 0.008949163369834423, + 0.00540191400796175, + 0.006160170771181583, + 0.0069818152114748955, + 0.008661272935569286, + 0.008898128755390644, + 0.006012017838656902, + 0.007696203887462616, + 0.009130444377660751, + 0.007658515591174364, + 0.00630097696557641, + 0.0037608512211591005, + 0.0044486080296337605, + 0.004368592984974384, + 0.004674666095525026, + 0.004315777216106653, + 0.006107701454311609, + 0.006283178459852934, + 0.00513009587302804, + 0.00815623439848423, + 0.008273828774690628, + 0.006698098499327898, + 0.004813233856111765, + 0.0062146116979420185, + 0.005494039040058851, + 0.019573817029595375, + 0.014766199514269829, + 0.009947141632437706, + 0.011035321280360222, + 0.01786011829972267, + 0.01146275270730257, + 0.013651905581355095, + 0.010542770847678185, + 0.015323610045015812, + 0.011292323470115662, + 0.016576392576098442, + 0.0168851837515831, + 0.010212709195911884, + 0.01183561235666275, + 0.012622513808310032, + 0.01820945553481579, + 0.006813414860516787, + 0.009790929965674877, + 0.008306476287543774, + 0.011475540697574615, + 0.01644669473171234, + 0.012206966057419777, + 0.013898147270083427, + 0.009479634463787079, + 0.009388594888150692, + 0.011221643537282944, + 0.011915454640984535, + 0.018649639561772346, + 0.015772568061947823, + 0.012661715038120747, + 0.02393496036529541, + 0.013572306372225285, + 0.018155278638005257, + 0.01791447214782238, + 0.022874167189002037, + 0.030871037393808365, + 0.040688104927539825, + 0.02432674542069435, + 0.047544341534376144, + 0.027521412819623947 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 8 + ], + "selected_reward": -0.012535941787064075 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "medium shot begins at a slight upward angle, then slowly and steadily zooms out. At the center of the frame is a general clad in ancient armor, standing on a vast, empty battlefield at dusk. His armor, forged from a blend of bronze and black iron, is covered in the scars of swords and the rust of time. The beast totems carved on his shoulder plates still shone fiercely in the dim light. His weathered face, sunken eyes and furrowed brow, is filled with despair and resignation, his beard matted with dried blood and dust. He jerks his head back, the veins in his neck bulging with the sheer force. His mouth opens wide, and he lets out a silent yet heartbreaking roar into the low, leaden clouds. His fists clench, his body trembling with the outpouring of emotion. Beneath his feet lies cracked, scorched earth, strewn with broken flags, shattered weapons, and unclaimed helmets. The last blood-red rays of sunset struggled to pierce through the clouds, gilding the outline of his armor with a poignant golden edge and casting a long, distorted shadow behind him. As the camera continuously and steadily zoomed out, the general's figure gradually shrank across the vast battlefield, ultimately becoming a tiny, lonely black dot on the charred earth. The entire scene was filled with a sense of epic tragedy and the endless desolation of a hero's final journey.", + "seed": 3163, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000056/group_4/index_5204_seed_3163_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": -0.09416124345734715, + "vlm_logprob": -0.09416124345734715, + "vlm_yes_no_margin": 3.449999976158142, + "vlm_token_credit": -0.09416124345734715, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on a vast, empty battlefield at dusk with cracked, scorched earth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a general wearing ancient armor made of bronze and black iron visible at the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the general's armor feature beast totems on the shoulder plates and show signs of sword scars and rust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the general perform the action of jerking his head back with bulging neck veins and opening his mouth wide?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start at a slight upward angle and slowly, steadily zoom out until the general becomes a tiny dot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011839713901281357 + ], + "opposite_logprob_tokens": [ + -4.761839866638184 + ], + "logprob_reward": -0.011839713901281357, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.011839713901281357, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10405569523572922, + 0.067890964448452, + 0.02392624132335186, + 0.046096064150333405, + 0.11455100029706955, + 0.06400331854820251, + 0.0645686462521553, + 0.19301867485046387, + 0.11592429876327515, + 0.20596502721309662 + ], + "credit_map": [ + 0.01533996220678091, + 0.01472215075045824, + 0.013021097518503666, + 0.010498961433768272, + 0.014398548752069473, + 0.006904074922204018, + 0.015686793252825737, + 0.01348411850631237, + 0.009083821438252926, + 0.01011406909674406, + 0.012049377895891666, + 0.005535501055419445, + 0.009880349971354008, + 0.0048914444632828236, + 0.010927137918770313, + 0.005409269127994776, + 0.003201778745278716, + 0.0031194547191262245, + 0.004019348882138729, + 0.002164360135793686, + 0.004102820530533791, + 0.0022161193192005157, + 0.003306545550003648, + 0.0017958162352442741, + 0.00843049120157957, + 0.005440977402031422, + 0.005319696385413408, + 0.0029337164014577866, + 0.005853235721588135, + 0.008737103082239628, + 0.003259537508711219, + 0.006121310405433178, + 0.01810753345489502, + 0.013441399671137333, + 0.015624145977199078, + 0.011414038017392159, + 0.011927928775548935, + 0.012238796800374985, + 0.012134471908211708, + 0.019662702456116676, + 0.009011533111333847, + 0.011843016371130943, + 0.005876889452338219, + 0.007891927845776081, + 0.009139806032180786, + 0.0067404224537312984, + 0.006643050815910101, + 0.0068566822446882725, + 0.006219654809683561, + 0.007350700907409191, + 0.007791764102876186, + 0.01705778017640114, + 0.006166402250528336, + 0.0047684055753052235, + 0.005309749394655228, + 0.009904197417199612, + 0.010789979249238968, + 0.024738401174545288, + 0.013641418889164925, + 0.016765514388680458, + 0.025386618450284004, + 0.05798877403140068, + 0.015920037403702736, + 0.0277879498898983, + 0.008882579393684864, + 0.011857004836201668, + 0.018675806000828743, + 0.024482501670718193, + 0.007860197685658932, + 0.019467730075120926, + 0.012290491722524166, + 0.012407997623085976, + 0.015692388638854027, + 0.0194455124437809, + 0.034266263246536255, + 0.06592623889446259, + 0.008823608048260212, + 0.01470175664871931, + 0.02395336888730526, + 0.023155920207500458 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 7, + 8 + ], + "selected_reward": -0.011839713901281357 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on a vast, empty battlefield at dusk with cracked, scorched earth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a general wearing ancient armor made of bronze and black iron visible at the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the general's armor feature beast totems on the shoulder plates and show signs of sword scars and rust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the general perform the action of jerking his head back with bulging neck veins and opening his mouth wide?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start at a slight upward angle and slowly, steadily zoom out until the general becomes a tiny dot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00595983536913991 + ], + "opposite_logprob_tokens": [ + -5.130959987640381 + ], + "logprob_reward": -0.00595983536913991, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.00595983536913991, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1609858274459839, + 0.22486136853694916, + 0.052776362746953964, + 0.0619729608297348, + 0.13576963543891907, + 0.05716703459620476, + 0.04145926609635353, + 0.07126403599977493, + 0.04966484382748604, + 0.14407867193222046 + ], + "credit_map": [ + 0.019300196319818497, + 0.015516621060669422, + 0.01764351688325405, + 0.017587440088391304, + 0.027339721098542213, + 0.013936455361545086, + 0.02859252318739891, + 0.021069351583719254, + 0.027775006368756294, + 0.0255239550024271, + 0.02077832818031311, + 0.02013707160949707, + 0.03397587686777115, + 0.015169755555689335, + 0.0584588460624218, + 0.02304251119494438, + 0.007399362046271563, + 0.004544791299849749, + 0.005378767382353544, + 0.0068459417670965195, + 0.011162867769598961, + 0.002930340589955449, + 0.007772878278046846, + 0.006741409655660391, + 0.00867935735732317, + 0.010193286463618279, + 0.004918624181300402, + 0.0049827927723526955, + 0.010368933901190758, + 0.012974242679774761, + 0.005229391623288393, + 0.0046263341791927814, + 0.0171841811388731, + 0.021726591512560844, + 0.020039664581418037, + 0.013844363391399384, + 0.009403337724506855, + 0.01724798046052456, + 0.012550371699035168, + 0.02377314493060112, + 0.005512596108019352, + 0.005306972190737724, + 0.00817625317722559, + 0.011868729256093502, + 0.007007094565778971, + 0.005042748991400003, + 0.006459616124629974, + 0.007793024647980928, + 0.004875166807323694, + 0.006441131234169006, + 0.004383462946861982, + 0.005797448102384806, + 0.004325887653976679, + 0.006833926774561405, + 0.004603995010256767, + 0.004198246169835329, + 0.006407300941646099, + 0.006559348665177822, + 0.010887231677770615, + 0.008882434107363224, + 0.00867274310439825, + 0.0071479566395282745, + 0.010336458683013916, + 0.012370562180876732, + 0.003562633180990815, + 0.004378911107778549, + 0.007294624112546444, + 0.013071252964437008, + 0.0030814071651548147, + 0.004973296541720629, + 0.006590423174202442, + 0.0067122941836714745, + 0.011694408021867275, + 0.011630794033408165, + 0.021964386105537415, + 0.04329011216759682, + 0.00869575422257185, + 0.009775910526514053, + 0.01795451156795025, + 0.01907278597354889 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 9 + ], + "selected_reward": -0.00595983536913991 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on a vast, empty battlefield at dusk with cracked, scorched earth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a general wearing ancient armor made of bronze and black iron visible at the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the general's armor feature beast totems on the shoulder plates and show signs of sword scars and rust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the general perform the action of jerking his head back with bulging neck veins and opening his mouth wide?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start at a slight upward angle and slowly, steadily zoom out until the general becomes a tiny dot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012551246210932732 + ], + "opposite_logprob_tokens": [ + -4.387551307678223 + ], + "logprob_reward": -0.012551246210932732, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012551246210932732, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15095356106758118, + 0.13778534531593323, + 0.052802249789237976, + 0.053885478526353836, + 0.09942510724067688, + 0.10411136597394943, + 0.04517161101102829, + 0.1408892273902893, + 0.14543747901916504, + 0.06953855603933334 + ], + "credit_map": [ + 0.02183310128748417, + 0.021447185426950455, + 0.02045212686061859, + 0.009888388216495514, + 0.02420693077147007, + 0.017488563433289528, + 0.01659729890525341, + 0.019039971753954887, + 0.021527035161852837, + 0.013414192944765091, + 0.009603881277143955, + 0.01364053599536419, + 0.02720564976334572, + 0.008587527088820934, + 0.030525492504239082, + 0.013281023129820824, + 0.00929252803325653, + 0.005020019598305225, + 0.00499118585139513, + 0.003851791610941291, + 0.012597125954926014, + 0.0037232411559671164, + 0.006544125732034445, + 0.006782229524105787, + 0.006498122122138739, + 0.008859884925186634, + 0.006214112043380737, + 0.004992952570319176, + 0.006098301615566015, + 0.008755638264119625, + 0.004341090098023415, + 0.00812537595629692, + 0.012289806269109249, + 0.011340800672769547, + 0.022211235016584396, + 0.008550549857318401, + 0.010950115509331226, + 0.010266038589179516, + 0.010399559512734413, + 0.01341700553894043, + 0.011512720957398415, + 0.008654194884002209, + 0.012875712476670742, + 0.02107163518667221, + 0.01184975914657116, + 0.008831111714243889, + 0.010342155583202839, + 0.018974076956510544, + 0.004717962350696325, + 0.006931825075298548, + 0.0047482033260166645, + 0.009378056973218918, + 0.005134173668920994, + 0.004050677176564932, + 0.004962288308888674, + 0.005248421803116798, + 0.009918528608977795, + 0.015721380710601807, + 0.013917830772697926, + 0.013663985766470432, + 0.018519608303904533, + 0.03187626972794533, + 0.01553359068930149, + 0.021738039329648018, + 0.00941956602036953, + 0.010902789421379566, + 0.026821807026863098, + 0.035253919661045074, + 0.011610125191509724, + 0.012761658057570457, + 0.01699013076722622, + 0.021677477285265923, + 0.008365591987967491, + 0.006160149350762367, + 0.009270431473851204, + 0.013258563354611397, + 0.007380570750683546, + 0.006289841141551733, + 0.009960638359189034, + 0.008852766826748848 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 7 + ], + "selected_reward": -0.012551246210932732 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on a vast, empty battlefield at dusk with cracked, scorched earth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a general wearing ancient armor made of bronze and black iron visible at the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the general's armor feature beast totems on the shoulder plates and show signs of sword scars and rust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the general perform the action of jerking his head back with bulging neck veins and opening his mouth wide?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start at a slight upward angle and slowly, steadily zoom out until the general becomes a tiny dot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.12704406678676605 + ], + "opposite_logprob_tokens": [ + -2.1270439624786377 + ], + "logprob_reward": -0.12704406678676605, + "yes_no_margin_reward": 1.9999998807907104, + "token_credit_reward": -0.12704406678676605, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.172981396317482, + 0.0802890956401825, + 0.05021532252430916, + 0.07283870130777359, + 0.1622808575630188, + 0.1038007140159607, + 0.05452956631779671, + 0.10263031721115112, + 0.13868115842342377, + 0.06175293028354645 + ], + "credit_map": [ + 0.019225480034947395, + 0.0166185162961483, + 0.024761665612459183, + 0.016055112704634666, + 0.02760067954659462, + 0.014469774439930916, + 0.03389981389045715, + 0.020350340753793716, + 0.009777417406439781, + 0.009047935716807842, + 0.007144903298467398, + 0.008909719064831734, + 0.012269559316337109, + 0.004937524441629648, + 0.02076796069741249, + 0.007434061728417873, + 0.00649147666990757, + 0.003930267877876759, + 0.004444419872015715, + 0.005083328578621149, + 0.013927136547863483, + 0.003193910699337721, + 0.008634821511805058, + 0.004509955178946257, + 0.011989417485892773, + 0.007803649175912142, + 0.008412805385887623, + 0.006754281930625439, + 0.015759969130158424, + 0.008136991411447525, + 0.006132693495601416, + 0.007848886772990227, + 0.01891138032078743, + 0.028158361092209816, + 0.019604727625846863, + 0.016215788200497627, + 0.014929729513823986, + 0.02050560899078846, + 0.01986571028828621, + 0.02408953197300434, + 0.010419179685413837, + 0.015227869153022766, + 0.008155464194715023, + 0.01200027298182249, + 0.018116917461156845, + 0.011121966876089573, + 0.01884773187339306, + 0.009911306202411652, + 0.0056720091961324215, + 0.006641766056418419, + 0.006720112636685371, + 0.011804279871284962, + 0.0045096478424966335, + 0.004816190805286169, + 0.006249507889151573, + 0.008116047829389572, + 0.012032921425998211, + 0.008452470414340496, + 0.011775624006986618, + 0.01337892934679985, + 0.013348015956580639, + 0.012790754437446594, + 0.016906460747122765, + 0.013945134356617928, + 0.0077859582379460335, + 0.01046749483793974, + 0.025326216593384743, + 0.03704182058572769, + 0.010756999254226685, + 0.014100993983447552, + 0.017613599076867104, + 0.015588055364787579, + 0.004936717450618744, + 0.00675224931910634, + 0.009281879290938377, + 0.012682147324085236, + 0.004197283182293177, + 0.0067105283960700035, + 0.008829749189317226, + 0.008362368680536747 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 8 + ], + "selected_reward": -0.12704406678676605 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on a vast, empty battlefield at dusk with cracked, scorched earth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a general wearing ancient armor made of bronze and black iron visible at the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the general's armor feature beast totems on the shoulder plates and show signs of sword scars and rust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the general perform the action of jerking his head back with bulging neck veins and opening his mouth wide?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start at a slight upward angle and slowly, steadily zoom out until the general becomes a tiny dot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.3134113550186157 + ], + "opposite_logprob_tokens": [ + -1.3134113550186157 + ], + "logprob_reward": -0.3134113550186157, + "yes_no_margin_reward": 1.0, + "token_credit_reward": -0.3134113550186157, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07885202020406723, + 0.03794824331998825, + 0.03646709397435188, + 0.055311862379312515, + 0.18744109570980072, + 0.053720176219940186, + 0.07195418328046799, + 0.14322347939014435, + 0.22141453623771667, + 0.11366727203130722 + ], + "credit_map": [ + 0.010466545820236206, + 0.007740308064967394, + 0.010990969836711884, + 0.006673905998468399, + 0.01354078110307455, + 0.006485294550657272, + 0.01374566275626421, + 0.009208555333316326, + 0.005326578393578529, + 0.005154687911272049, + 0.007795526646077633, + 0.004058920778334141, + 0.005408049561083317, + 0.0025961219798773527, + 0.004600147716701031, + 0.0030082145240157843, + 0.004985807463526726, + 0.0032749890815466642, + 0.007508073467761278, + 0.004397036042064428, + 0.0053433822467923164, + 0.002127445535734296, + 0.004479100927710533, + 0.004351260140538216, + 0.0077710882760584354, + 0.006484690122306347, + 0.010707460343837738, + 0.005121553782373667, + 0.0081765316426754, + 0.005358031019568443, + 0.0050398982129991055, + 0.006652613636106253, + 0.024826746433973312, + 0.024684522300958633, + 0.029056157916784286, + 0.017750505357980728, + 0.015159505419433117, + 0.02361244335770607, + 0.020283859223127365, + 0.032067377120256424, + 0.0068357400596141815, + 0.008489051833748817, + 0.0048916395753622055, + 0.005765419919043779, + 0.00853759702295065, + 0.007227018941193819, + 0.006136914249509573, + 0.005836797878146172, + 0.0077190157026052475, + 0.008835804648697376, + 0.008077076636254787, + 0.016387030482292175, + 0.007075909990817308, + 0.007262521889060736, + 0.00597026664763689, + 0.010626563802361488, + 0.00917741097509861, + 0.015774555504322052, + 0.01384684257209301, + 0.020340289920568466, + 0.015588981099426746, + 0.02548171579837799, + 0.026451071724295616, + 0.016562627628445625, + 0.013128897175192833, + 0.01742403581738472, + 0.03729527071118355, + 0.06077340990304947, + 0.01827235333621502, + 0.023661106824874878, + 0.030314527451992035, + 0.020544955506920815, + 0.009123431518673897, + 0.010897048749029636, + 0.01387768518179655, + 0.026214882731437683, + 0.008734332397580147, + 0.011525830253958702, + 0.017514213919639587, + 0.015779858455061913 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 4, + 7 + ], + "selected_reward": -0.3134113550186157 + } + ], + "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/0000056/group_5/index_tensor([1652, 1652])_group_5_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000056/group_5/index_tensor([1652, 1652])_group_5_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..ba0b50883e44235fa4cf9831d12aff047235a0de --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000056/group_5/index_tensor([1652, 1652])_group_5_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "static, medium shot, at eye level, captures a man in his thirties with neatly cropped brown hair, who initially speaks directly to the camera with an engaging and confident expression. He appears smartly casual in a dark blue fine-knit Merino wool sweater layered over a crisp white collared shirt. The setting, a modern, minimalist home office, is bathed in soft, diffuse light from an unseen sidelight, creating a gentle highlight on his features and a clean, professional aura. With a smooth, deliberate movement, his right hand dips beneath the frame before rising again, clutching a slim, modern smartphone with a dark glass screen and polished metal edges. His gaze momentarily wanders from the camera, his head tilted slightly, looking down at the device with a look of quiet satisfaction. Then, turning back to the camera, a proud smile spreading across his features, he purposefully tilts the phone forward, presenting the screen directly to the viewer in a gesture of apparent ostentation, creating an air of evocative elegance and persuasiveness.", + "seed": 8373, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000056/group_5/index_1652_seed_8373_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": -0.017978928983211517, + "vlm_logprob": -0.017978928983211517, + "vlm_yes_no_margin": 4.649999952316284, + "vlm_token_credit": -0.017978928983211517, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static medium shot at eye level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject a man in his thirties with neatly cropped brown hair wearing a dark blue fine-knit sweater over a white collared shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the setting appear to be a modern, minimalist home office with soft, diffuse sidelight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the man dip his right hand beneath the frame to reveal a slim smartphone with a dark glass screen and polished metal edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the man tilt the phone forward to present the screen directly to the viewer with a proud smile?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04913531243801117 + ], + "opposite_logprob_tokens": [ + -4.549135208129883 + ], + "logprob_reward": -0.04913531243801117, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.04913531243801117, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15059888362884521, + 0.22529754042625427, + 0.04901473596692085, + 0.0915612205862999, + 0.08821490406990051, + 0.03381738439202309, + 0.032800156623125076, + 0.08740701526403427, + 0.06219560652971268, + 0.17909258604049683 + ], + "credit_map": [ + 0.03487864136695862, + 0.01070873998105526, + 0.03604548051953316, + 0.01083275955170393, + 0.014263453893363476, + 0.013431369327008724, + 0.012271938845515251, + 0.018166471272706985, + 0.06763287633657455, + 0.010320469737052917, + 0.08351785689592361, + 0.013357137329876423, + 0.014105581678450108, + 0.011101782321929932, + 0.009465743787586689, + 0.015796080231666565, + 0.007466178387403488, + 0.005535311996936798, + 0.008271492086350918, + 0.006228061858564615, + 0.005586343817412853, + 0.003232960356399417, + 0.006386923603713512, + 0.006307457108050585, + 0.01609020307660103, + 0.009084523655474186, + 0.018447954207658768, + 0.008082415908575058, + 0.009520165622234344, + 0.006464474834501743, + 0.012608974240720272, + 0.011262496933341026, + 0.01178668811917305, + 0.007626086939126253, + 0.012548084370791912, + 0.011675986461341381, + 0.007665056269615889, + 0.005618928465992212, + 0.02001004107296467, + 0.011284018866717815, + 0.006558904889971018, + 0.006855407264083624, + 0.004860423505306244, + 0.003055322216823697, + 0.0034178777132183313, + 0.002913797739893198, + 0.0025466864462941885, + 0.003608959959819913, + 0.005841793958097696, + 0.00429890351369977, + 0.007300299592316151, + 0.002376551041379571, + 0.0027989153750240803, + 0.003174393204972148, + 0.0025897626765072346, + 0.004419532138854265, + 0.013830490410327911, + 0.010251669213175774, + 0.016164876520633698, + 0.009514432400465012, + 0.007006960455328226, + 0.005402977578341961, + 0.011931633576750755, + 0.013303965330123901, + 0.010110110975801945, + 0.011020752601325512, + 0.0093891816213727, + 0.005969512742012739, + 0.005892420653253794, + 0.004859955050051212, + 0.008242583833634853, + 0.006711082998663187, + 0.030848076567053795, + 0.014057750813663006, + 0.023736489936709404, + 0.016167685389518738, + 0.025741945952177048, + 0.013562050648033619, + 0.017299702391028404, + 0.03767887130379677 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 9, + 0 + ], + "selected_reward": -0.04913531243801117 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static medium shot at eye level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject a man in his thirties with neatly cropped brown hair wearing a dark blue fine-knit sweater over a white collared shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the setting appear to be a modern, minimalist home office with soft, diffuse sidelight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the man dip his right hand beneath the frame to reveal a slim smartphone with a dark glass screen and polished metal edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the man tilt the phone forward to present the screen directly to the viewer with a proud smile?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008658722043037415 + ], + "opposite_logprob_tokens": [ + -4.7586588859558105 + ], + "logprob_reward": -0.008658722043037415, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008658722043037415, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06526055932044983, + 0.3919471800327301, + 0.05458346754312515, + 0.05076395720243454, + 0.04400318115949631, + 0.034520216286182404, + 0.025864385068416595, + 0.04552869498729706, + 0.038919053971767426, + 0.24860933423042297 + ], + "credit_map": [ + 0.01348865870386362, + 0.007248167414218187, + 0.010932719334959984, + 0.007292008958756924, + 0.0067933606915175915, + 0.005361946299672127, + 0.006482102442532778, + 0.007661590352654457, + 0.1159503161907196, + 0.01705867052078247, + 0.1477482169866562, + 0.023918498307466507, + 0.021236401051282883, + 0.020965522155165672, + 0.019227126613259315, + 0.025842415168881416, + 0.012426639907062054, + 0.006527809426188469, + 0.008795332163572311, + 0.004969766363501549, + 0.00613065529614687, + 0.003500408958643675, + 0.005164782982319593, + 0.0070680673234164715, + 0.00901993177831173, + 0.006108084227889776, + 0.007853226736187935, + 0.0037123579531908035, + 0.008702834136784077, + 0.00324293808080256, + 0.005154415965080261, + 0.0069701652973890305, + 0.005396334454417229, + 0.003330885898321867, + 0.006047720089554787, + 0.0067457593977451324, + 0.0038388166576623917, + 0.0020638001151382923, + 0.011392033658921719, + 0.005187830422073603, + 0.00528752850368619, + 0.002724917372688651, + 0.00720656244084239, + 0.003074492095038295, + 0.00368759548291564, + 0.002263579983264208, + 0.00466829864308238, + 0.005607240367680788, + 0.005887201521545649, + 0.002568843774497509, + 0.004118894226849079, + 0.002471070969477296, + 0.0022366077173501253, + 0.002126287668943405, + 0.0034870635718107224, + 0.0029684153851121664, + 0.010190118104219437, + 0.0032043943647295237, + 0.00796801969408989, + 0.004411038942635059, + 0.003019867232069373, + 0.0033175116404891014, + 0.007468674331903458, + 0.005949071142822504, + 0.006064337678253651, + 0.003780535887926817, + 0.005758165381848812, + 0.004778917413204908, + 0.0029008130077272654, + 0.0027974771801382303, + 0.005411890801042318, + 0.007426920812577009, + 0.041426874697208405, + 0.02556634694337845, + 0.030364394187927246, + 0.012926257215440273, + 0.07192731648683548, + 0.021825354546308517, + 0.023797553032636642, + 0.02077522501349449 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 9, + 0 + ], + "selected_reward": -0.008658722043037415 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static medium shot at eye level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject a man in his thirties with neatly cropped brown hair wearing a dark blue fine-knit sweater over a white collared shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the setting appear to be a modern, minimalist home office with soft, diffuse sidelight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the man dip his right hand beneath the frame to reveal a slim smartphone with a dark glass screen and polished metal edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the man tilt the phone forward to present the screen directly to the viewer with a proud smile?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0052610281854867935 + ], + "opposite_logprob_tokens": [ + -5.255260944366455 + ], + "logprob_reward": -0.0052610281854867935, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.0052610281854867935, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0954645499587059, + 0.4188123941421509, + 0.03067164681851864, + 0.014238981530070305, + 0.106749027967453, + 0.012594519183039665, + 0.008534356951713562, + 0.04284925386309624, + 0.02277946099638939, + 0.24730582535266876 + ], + "credit_map": [ + 0.022210733965039253, + 0.0074213529005646706, + 0.018337151035666466, + 0.005747415591031313, + 0.01158674992620945, + 0.013092884793877602, + 0.008037334308028221, + 0.009030918590724468, + 0.1203194335103035, + 0.0172595027834177, + 0.1606525480747223, + 0.02481739968061447, + 0.022164462134242058, + 0.026308024302124977, + 0.01987706869840622, + 0.02741394191980362, + 0.0059376973658800125, + 0.0028946036472916603, + 0.005237137898802757, + 0.003180897329002619, + 0.00306446710601449, + 0.0024179245810955763, + 0.0041395979933440685, + 0.0037993185687810183, + 0.0022459602914750576, + 0.0014535320224240422, + 0.0030178911983966827, + 0.0014446943532675505, + 0.0018102374160662293, + 0.0011052328627556562, + 0.001478402060456574, + 0.0016830295789986849, + 0.015585158951580524, + 0.009071803651750088, + 0.013131155632436275, + 0.017100263386964798, + 0.008824339136481285, + 0.007284968160092831, + 0.020448293536901474, + 0.01530303992331028, + 0.002109372289851308, + 0.0013984385877847672, + 0.0019881674088537693, + 0.0009375846711918712, + 0.0012308891164138913, + 0.0011553361546248198, + 0.0017519739922136068, + 0.0020227553322911263, + 0.0014203682076185942, + 0.0008104513399302959, + 0.0013196825748309493, + 0.0008303166832774878, + 0.0006557200686074793, + 0.0010518125491216779, + 0.0007253271760419011, + 0.0017206774791702628, + 0.006489122286438942, + 0.0034850265365093946, + 0.0057159061543643475, + 0.005298512056469917, + 0.0035462290979921818, + 0.005710730329155922, + 0.005844006780534983, + 0.006759717129170895, + 0.003387234639376402, + 0.0025205828715115786, + 0.0037757137324661016, + 0.002719596726819873, + 0.002049640752375126, + 0.0013878981117159128, + 0.004075210075825453, + 0.0028635822236537933, + 0.04093312472105026, + 0.025614719837903976, + 0.029443496838212013, + 0.012712148949503899, + 0.0720917358994484, + 0.02154596894979477, + 0.023863475769758224, + 0.02110113576054573 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 9, + 4 + ], + "selected_reward": -0.0052610281854867935 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static medium shot at eye level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject a man in his thirties with neatly cropped brown hair wearing a dark blue fine-knit sweater over a white collared shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the setting appear to be a modern, minimalist home office with soft, diffuse sidelight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the man dip his right hand beneath the frame to reveal a slim smartphone with a dark glass screen and polished metal edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the man tilt the phone forward to present the screen directly to the viewer with a proud smile?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008645722642540932 + ], + "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": [ + 0.04601573944091797, + 0.13249392807483673, + 0.010524753481149673, + 0.017043087631464005, + 0.008599675260484219, + 0.008834668435156345, + 0.005991799756884575, + 0.05995158851146698, + 0.04657034948468208, + 0.6639744639396667 + ], + "credit_map": [ + 0.007376554422080517, + 0.002998689655214548, + 0.01101390179246664, + 0.005914182402193546, + 0.0033692021388560534, + 0.0029215167742222548, + 0.006409144960343838, + 0.006012544967234135, + 0.039467036724090576, + 0.005246760789304972, + 0.051140379160642624, + 0.007316669914871454, + 0.007137347012758255, + 0.007795992307364941, + 0.005711966194212437, + 0.008677772246301174, + 0.0018833675421774387, + 0.0010809830855578184, + 0.0018189470283687115, + 0.0010938069317489862, + 0.001121239853091538, + 0.000871023628860712, + 0.0012687259586527944, + 0.001386658288538456, + 0.0034599474165588617, + 0.0018418538384139538, + 0.0034090434201061726, + 0.0013184441486373544, + 0.0021879971027374268, + 0.0010502263903617859, + 0.0017429960425943136, + 0.0020325765945017338, + 0.0011343195801600814, + 0.0008912076591514051, + 0.0012674015015363693, + 0.0008568681660108268, + 0.000900006853044033, + 0.0005277515738271177, + 0.0018059649737551808, + 0.0012161543127149343, + 0.001359401154331863, + 0.0008792747394181788, + 0.0016474289586767554, + 0.0010337892454117537, + 0.000750355888158083, + 0.0008450200548395514, + 0.0009808589238673449, + 0.0013385391794145107, + 0.0009467333438806236, + 0.0007625248399563134, + 0.0013402862241491675, + 0.0005579874850809574, + 0.000647733686491847, + 0.0005474141798913479, + 0.0006142224883660674, + 0.0005748969851993024, + 0.007980754598975182, + 0.004402315244078636, + 0.006905341986566782, + 0.00732133723795414, + 0.00443114573135972, + 0.007638563867658377, + 0.0074873194098472595, + 0.013784805312752724, + 0.006627399008721113, + 0.003896416397765279, + 0.006609663832932711, + 0.0057479082606732845, + 0.004073109943419695, + 0.0031182742677628994, + 0.008739025332033634, + 0.007758548948913813, + 0.1127697229385376, + 0.06598001718521118, + 0.07798048108816147, + 0.03399452567100525, + 0.20144139230251312, + 0.05444325506687164, + 0.05943432077765465, + 0.057930707931518555 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 1, + 7 + ], + "selected_reward": -0.008645722642540932 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static medium shot at eye level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject a man in his thirties with neatly cropped brown hair wearing a dark blue fine-knit sweater over a white collared shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the setting appear to be a modern, minimalist home office with soft, diffuse sidelight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the man dip his right hand beneath the frame to reveal a slim smartphone with a dark glass screen and polished metal edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the man tilt the phone forward to present the screen directly to the viewer with a proud smile?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018193859606981277 + ], + "opposite_logprob_tokens": [ + -4.01819372177124 + ], + "logprob_reward": -0.018193859606981277, + "yes_no_margin_reward": 3.999999761581421, + "token_credit_reward": -0.018193859606981277, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08021392673254013, + 0.1737653911113739, + 0.024527370929718018, + 0.04570623114705086, + 0.020757092162966728, + 0.02309827134013176, + 0.023223675787448883, + 0.222748264670372, + 0.10900181531906128, + 0.2769579291343689 + ], + "credit_map": [ + 0.012760050594806671, + 0.005736709106713533, + 0.01714038848876953, + 0.01009716372936964, + 0.0062071532011032104, + 0.005242219194769859, + 0.010135909542441368, + 0.01289434265345335, + 0.05723825469613075, + 0.0041404590010643005, + 0.07261088490486145, + 0.006200442090630531, + 0.007332460023462772, + 0.009233453311026096, + 0.005380240734666586, + 0.011629202403128147, + 0.004701259545981884, + 0.0028032909613102674, + 0.0042075179517269135, + 0.002094089752063155, + 0.002764785662293434, + 0.0021042029839009047, + 0.0027344205882400274, + 0.0031178067438304424, + 0.009128161706030369, + 0.005144390277564526, + 0.0066000004298985004, + 0.0040288749150931835, + 0.004977950360625982, + 0.0022926179226487875, + 0.007975524291396141, + 0.005558717530220747, + 0.002932327101007104, + 0.002159043215215206, + 0.0025726419407874346, + 0.0029234362300485373, + 0.0019436654401943088, + 0.0010852221166715026, + 0.005121324211359024, + 0.002019435167312622, + 0.0043504498898983, + 0.0028076726011931896, + 0.0036577172577381134, + 0.0024311491288244724, + 0.002517594490200281, + 0.001977869775146246, + 0.0025062926579266787, + 0.0028495281003415585, + 0.004206813871860504, + 0.0033304321113973856, + 0.0034341209102422, + 0.0026679765433073044, + 0.003656501416116953, + 0.0018827818566933274, + 0.0019300447311252356, + 0.0021150074899196625, + 0.04827645793557167, + 0.020211253315210342, + 0.052255257964134216, + 0.02851942740380764, + 0.014529919251799583, + 0.010226170532405376, + 0.028295472264289856, + 0.02043432742357254, + 0.018769050016999245, + 0.01519093569368124, + 0.016317125409841537, + 0.013042116537690163, + 0.008966051042079926, + 0.007425881922245026, + 0.01214495301246643, + 0.017145706340670586, + 0.045225486159324646, + 0.036766551434993744, + 0.041849467903375626, + 0.023074327036738396, + 0.059001535177230835, + 0.017987875267863274, + 0.018876416608691216, + 0.034176282584667206 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 7, + 1 + ], + "selected_reward": -0.018193859606981277 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, medium shot, at eye level, captures a man in his thirties with neatly cropped brown hair, who initially speaks directly to the camera with an engaging and confident expression. He appears smartly casual in a dark blue fine-knit Merino wool sweater layered over a crisp white collared shirt. The setting, a modern, minimalist home office, is bathed in soft, diffuse light from an unseen sidelight, creating a gentle highlight on his features and a clean, professional aura. With a smooth, deliberate movement, his right hand dips beneath the frame before rising again, clutching a slim, modern smartphone with a dark glass screen and polished metal edges. His gaze momentarily wanders from the camera, his head tilted slightly, looking down at the device with a look of quiet satisfaction. Then, turning back to the camera, a proud smile spreading across his features, he purposefully tilts the phone forward, presenting the screen directly to the viewer in a gesture of apparent ostentation, creating an air of evocative elegance and persuasiveness.", + "seed": 8374, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000056/group_5/index_1652_seed_8374_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": -0.015019332757219672, + "vlm_logprob": -0.015019332757219672, + "vlm_yes_no_margin": 4.95, + "vlm_token_credit": -0.015019332757219672, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static medium shot at eye level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject a man in his thirties with neatly cropped brown hair wearing a dark blue fine-knit sweater over a white collared shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the setting appear to be a modern, minimalist home office with soft, diffuse sidelight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the man dip his right hand beneath the frame to reveal a slim smartphone with a dark glass screen and polished metal edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the man tilt the phone forward to present the screen directly to the viewer with a proud smile?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03837909549474716 + ], + "opposite_logprob_tokens": [ + -5.413379192352295 + ], + "logprob_reward": -0.03837909549474716, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.03837909549474716, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11288171261548996, + 0.1761825531721115, + 0.23299647867679596, + 0.02831299602985382, + 0.030294528231024742, + 0.057333704084157944, + 0.0615353100001812, + 0.08506058901548386, + 0.10829906165599823, + 0.10710308700799942 + ], + "credit_map": [ + 0.017463356256484985, + 0.01148506160825491, + 0.022141072899103165, + 0.006652274634689093, + 0.013501587323844433, + 0.015794165432453156, + 0.015965670347213745, + 0.009878512471914291, + 0.03314198553562164, + 0.020976105704903603, + 0.03200891986489296, + 0.009026296436786652, + 0.03685462847352028, + 0.024768607690930367, + 0.009050454944372177, + 0.01035554613918066, + 0.04890250787138939, + 0.01977081410586834, + 0.06534209847450256, + 0.018325285986065865, + 0.012990939430892467, + 0.03434719517827034, + 0.011285901069641113, + 0.022031724452972412, + 0.003868125146254897, + 0.003014633432030678, + 0.004776502028107643, + 0.0018729112343862653, + 0.004957077093422413, + 0.0044014109298586845, + 0.003484440268948674, + 0.0019378949655219913, + 0.005463610403239727, + 0.0028924390207976103, + 0.004659702070057392, + 0.001835871720686555, + 0.004743685945868492, + 0.004510045982897282, + 0.003616821952164173, + 0.002572346478700638, + 0.012812748551368713, + 0.0037977113388478756, + 0.014670183882117271, + 0.003442144952714443, + 0.005515597760677338, + 0.00903723482042551, + 0.003577061463147402, + 0.004481018055230379, + 0.009214084595441818, + 0.00687916437163949, + 0.012176324613392353, + 0.004157662857323885, + 0.009841353632509708, + 0.01025842409580946, + 0.004147863015532494, + 0.004860430024564266, + 0.01684807799756527, + 0.008177234791219234, + 0.013805519789457321, + 0.0072165681049227715, + 0.009407468140125275, + 0.011135169304907322, + 0.008917535655200481, + 0.00955300871282816, + 0.021102961152791977, + 0.00832816120237112, + 0.01603662222623825, + 0.008125443011522293, + 0.017082834616303444, + 0.014752973802387714, + 0.012202146463096142, + 0.010667919181287289, + 0.01719110831618309, + 0.011297735385596752, + 0.017047973349690437, + 0.01228084322065115, + 0.011510159820318222, + 0.0127993980422616, + 0.016735099256038666, + 0.00824075099080801 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 1, + 0 + ], + "selected_reward": -0.03837909549474716 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static medium shot at eye level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject a man in his thirties with neatly cropped brown hair wearing a dark blue fine-knit sweater over a white collared shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the setting appear to be a modern, minimalist home office with soft, diffuse sidelight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the man dip his right hand beneath the frame to reveal a slim smartphone with a dark glass screen and polished metal edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the man tilt the phone forward to present the screen directly to the viewer with a proud smile?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006762238685041666 + ], + "opposite_logprob_tokens": [ + -5.0067620277404785 + ], + "logprob_reward": -0.006762238685041666, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006762238685041666, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17820125818252563, + 0.15609176456928253, + 0.1892683058977127, + 0.0927617996931076, + 0.11882100999355316, + 0.050309572368860245, + 0.028981108218431473, + 0.05093620344996452, + 0.07091940194368362, + 0.06370958685874939 + ], + "credit_map": [ + 0.018041715025901794, + 0.020590173080563545, + 0.021941130980849266, + 0.01857198216021061, + 0.023687833920121193, + 0.02037881314754486, + 0.033897072076797485, + 0.02109253592789173, + 0.018148861825466156, + 0.017786715179681778, + 0.020518597215414047, + 0.008492039516568184, + 0.04510289803147316, + 0.020015254616737366, + 0.018087642267346382, + 0.007939750328660011, + 0.03910014405846596, + 0.014676891267299652, + 0.041117824614048004, + 0.012845644727349281, + 0.02346985973417759, + 0.028067007660865784, + 0.01427251286804676, + 0.01571842096745968, + 0.02016596309840679, + 0.008642971515655518, + 0.018503393977880478, + 0.007431892678141594, + 0.008512032218277454, + 0.014509457163512707, + 0.0061805774457752705, + 0.008815511129796505, + 0.021298253908753395, + 0.007663832511752844, + 0.023427646607160568, + 0.009992523118853569, + 0.01394861564040184, + 0.015466345474123955, + 0.014062255620956421, + 0.012961523607373238, + 0.005160930100828409, + 0.0037768040783703327, + 0.005161500535905361, + 0.0038026899565011263, + 0.011226080358028412, + 0.006934495642781258, + 0.009512879885733128, + 0.004734191577881575, + 0.0035859758500009775, + 0.0019097214099019766, + 0.004025962669402361, + 0.0017888913862407207, + 0.007520073093473911, + 0.00450386805459857, + 0.0027321744710206985, + 0.002914441516622901, + 0.007824825122952461, + 0.004920701961964369, + 0.006422477308660746, + 0.004200236406177282, + 0.0076280818320810795, + 0.006018526386469603, + 0.00717956805601716, + 0.00674178684130311, + 0.010267104022204876, + 0.006060193292796612, + 0.00964068528264761, + 0.004962035454809666, + 0.01282922551035881, + 0.00846316572278738, + 0.011595256626605988, + 0.007101739291101694, + 0.005512177012860775, + 0.004808207508176565, + 0.007945717312395573, + 0.0063532935455441475, + 0.009098347276449203, + 0.007896417751908302, + 0.013909426517784595, + 0.008185997605323792 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 1 + ], + "selected_reward": -0.006762238685041666 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static medium shot at eye level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject a man in his thirties with neatly cropped brown hair wearing a dark blue fine-knit sweater over a white collared shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the setting appear to be a modern, minimalist home office with soft, diffuse sidelight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the man dip his right hand beneath the frame to reveal a slim smartphone with a dark glass screen and polished metal edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the man tilt the phone forward to present the screen directly to the viewer with a proud smile?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003203262807801366 + ], + "opposite_logprob_tokens": [ + -5.753203392028809 + ], + "logprob_reward": -0.003203262807801366, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.003203262807801366, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11380825936794281, + 0.05831601098179817, + 0.3153288960456848, + 0.047041699290275574, + 0.023899763822555542, + 0.03363556042313576, + 0.01945510320365429, + 0.110793337225914, + 0.1778537482023239, + 0.09986761212348938 + ], + "credit_map": [ + 0.01813598908483982, + 0.010132473893463612, + 0.022631892934441566, + 0.008367537520825863, + 0.017583319917321205, + 0.01651551015675068, + 0.011865638196468353, + 0.00857590138912201, + 0.00893842801451683, + 0.00914167333394289, + 0.011894002556800842, + 0.003539769211784005, + 0.008035487495362759, + 0.008298179134726524, + 0.005197013262659311, + 0.003271457040682435, + 0.0638604611158371, + 0.027194635942578316, + 0.08839552104473114, + 0.026235613971948624, + 0.015322692692279816, + 0.04700825735926628, + 0.01726888306438923, + 0.03004281595349312, + 0.00651519512757659, + 0.0037881890311837196, + 0.007141122128814459, + 0.0039861793629825115, + 0.006605197675526142, + 0.005950021557509899, + 0.008268172852694988, + 0.004787622019648552, + 0.0052687483839690685, + 0.002394373994320631, + 0.005421859677881002, + 0.0014241044409573078, + 0.002342421328648925, + 0.003552032634615898, + 0.0018070423975586891, + 0.0016891809646040201, + 0.0064344750717282295, + 0.002502793213352561, + 0.008017179556190968, + 0.0018367161974310875, + 0.0038075598422437906, + 0.00505289901047945, + 0.0029926856514066458, + 0.0029912490863353014, + 0.003605222562327981, + 0.0016475433949381113, + 0.004169895779341459, + 0.001464230939745903, + 0.003463778179138899, + 0.0025277845561504364, + 0.0011739588808268309, + 0.0014026905409991741, + 0.01854582503437996, + 0.013973419554531574, + 0.026139413937926292, + 0.006806136574596167, + 0.011118295602500439, + 0.014677715487778187, + 0.00881409551948309, + 0.010718433186411858, + 0.0444982647895813, + 0.01617727056145668, + 0.033136602491140366, + 0.010547306388616562, + 0.017012396827340126, + 0.027877088636159897, + 0.015253993682563305, + 0.01335082110017538, + 0.011077003553509712, + 0.010611552745103836, + 0.01338044460862875, + 0.011845317669212818, + 0.013185372576117516, + 0.011575012467801571, + 0.01619098149240017, + 0.012001922354102135 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 8, + 0 + ], + "selected_reward": -0.003203262807801366 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static medium shot at eye level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject a man in his thirties with neatly cropped brown hair wearing a dark blue fine-knit sweater over a white collared shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the setting appear to be a modern, minimalist home office with soft, diffuse sidelight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the man dip his right hand beneath the frame to reveal a slim smartphone with a dark glass screen and polished metal edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the man tilt the phone forward to present the screen directly to the viewer with a proud smile?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012549597769975662 + ], + "opposite_logprob_tokens": [ + -4.38754940032959 + ], + "logprob_reward": -0.012549597769975662, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012549597769975662, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09961795806884766, + 0.07078451663255692, + 0.18298906087875366, + 0.03385453298687935, + 0.025637123733758926, + 0.05868132784962654, + 0.026967870071530342, + 0.07203247398138046, + 0.21684658527374268, + 0.21258848905563354 + ], + "credit_map": [ + 0.011776219122111797, + 0.007655838038772345, + 0.014429840259253979, + 0.009346989914774895, + 0.012377399951219559, + 0.010543781332671642, + 0.022410713136196136, + 0.011077187955379486, + 0.012311790138483047, + 0.007908813655376434, + 0.012463218532502651, + 0.003467721166089177, + 0.0157362911850214, + 0.009808072820305824, + 0.004420365672558546, + 0.0046682502143085, + 0.037280213087797165, + 0.014669539406895638, + 0.05005580559372902, + 0.01467687264084816, + 0.011697559617459774, + 0.026446212083101273, + 0.010682252235710621, + 0.01748059317469597, + 0.004801447968930006, + 0.002870361553505063, + 0.006474584341049194, + 0.0026387483812868595, + 0.004431987181305885, + 0.0048496113158762455, + 0.004601312801241875, + 0.0031864820048213005, + 0.004110356792807579, + 0.0021488552447408438, + 0.00397725123912096, + 0.0018728479044511914, + 0.0035919933579862118, + 0.0033571110107004642, + 0.004222503863275051, + 0.002356203505769372, + 0.012026148848235607, + 0.004631769377738237, + 0.013606703840196133, + 0.0039559006690979, + 0.004515314474701881, + 0.008905934169888496, + 0.005593139678239822, + 0.005446420516818762, + 0.0035230654757469893, + 0.0026108883321285248, + 0.0048432257026433945, + 0.0020742083434015512, + 0.004606666509062052, + 0.004026365466415882, + 0.002429110696539283, + 0.0028543395455926657, + 0.012594875879585743, + 0.00649750791490078, + 0.010805204510688782, + 0.005363954231142998, + 0.010619381442666054, + 0.009330937638878822, + 0.008416661992669106, + 0.008403952233493328, + 0.03573533892631531, + 0.020193498581647873, + 0.02843679115176201, + 0.01841280609369278, + 0.027552714571356773, + 0.02683643251657486, + 0.029780138283967972, + 0.029898887500166893, + 0.022737650200724602, + 0.021000400185585022, + 0.032486412674188614, + 0.028032204136252403, + 0.03225068002939224, + 0.027729816734790802, + 0.027261096984148026, + 0.02109021320939064 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 2 + ], + "selected_reward": -0.012549597769975662 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static medium shot at eye level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject a man in his thirties with neatly cropped brown hair wearing a dark blue fine-knit sweater over a white collared shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the setting appear to be a modern, minimalist home office with soft, diffuse sidelight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the man dip his right hand beneath the frame to reveal a slim smartphone with a dark glass screen and polished metal edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the man tilt the phone forward to present the screen directly to the viewer with a proud smile?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014202469028532505 + ], + "opposite_logprob_tokens": [ + -4.26420259475708 + ], + "logprob_reward": -0.014202469028532505, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014202469028532505, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0884939655661583, + 0.05378837510943413, + 0.16181807219982147, + 0.03813352808356285, + 0.02792135626077652, + 0.05622594803571701, + 0.02845248393714428, + 0.0645589753985405, + 0.2298671305179596, + 0.2507401704788208 + ], + "credit_map": [ + 0.008934509940445423, + 0.007362511940300465, + 0.010533705353736877, + 0.008627453818917274, + 0.01170074101537466, + 0.010194698348641396, + 0.01851896196603775, + 0.012621377594769001, + 0.010775670409202576, + 0.00424182740971446, + 0.010296979919075966, + 0.003614528104662895, + 0.0066139185801148415, + 0.007814468815922737, + 0.005097274202853441, + 0.005333708133548498, + 0.033013224601745605, + 0.012295263819396496, + 0.04361873120069504, + 0.012118589133024216, + 0.0099173653870821, + 0.023601433262228966, + 0.011438703164458275, + 0.015814773738384247, + 0.005470309406518936, + 0.003307678271085024, + 0.006631026044487953, + 0.0028415084816515446, + 0.005385990720242262, + 0.005345414392650127, + 0.005774604622274637, + 0.0033770008012652397, + 0.0054256063885986805, + 0.0022919548209756613, + 0.005697011481970549, + 0.002303004963323474, + 0.002640060381963849, + 0.003618406131863594, + 0.0034633646719157696, + 0.0024819467216730118, + 0.006292765960097313, + 0.004865117371082306, + 0.006817235145717859, + 0.003578086383640766, + 0.012942376546561718, + 0.008184950798749924, + 0.008378155529499054, + 0.005167262628674507, + 0.004511939827352762, + 0.0028657603543251753, + 0.004879665095359087, + 0.002000117674469948, + 0.0046376450918614864, + 0.0042542279697954655, + 0.0023069290909916162, + 0.0029961983673274517, + 0.007428577169775963, + 0.005280104465782642, + 0.007890922948718071, + 0.0040067946538329124, + 0.018387755379080772, + 0.01005774736404419, + 0.00632226699963212, + 0.005184811539947987, + 0.03383726254105568, + 0.02074592560529709, + 0.02766346000134945, + 0.019895052537322044, + 0.034799154847860336, + 0.029838094487786293, + 0.03662469983100891, + 0.02646348811686039, + 0.028496813029050827, + 0.024068722501397133, + 0.03229372575879097, + 0.03203408420085907, + 0.045645128935575485, + 0.03132078796625137, + 0.03277846798300743, + 0.024102449417114258 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 2 + ], + "selected_reward": -0.014202469028532505 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, medium shot, at eye level, captures a man in his thirties with neatly cropped brown hair, who initially speaks directly to the camera with an engaging and confident expression. He appears smartly casual in a dark blue fine-knit Merino wool sweater layered over a crisp white collared shirt. The setting, a modern, minimalist home office, is bathed in soft, diffuse light from an unseen sidelight, creating a gentle highlight on his features and a clean, professional aura. With a smooth, deliberate movement, his right hand dips beneath the frame before rising again, clutching a slim, modern smartphone with a dark glass screen and polished metal edges. His gaze momentarily wanders from the camera, his head tilted slightly, looking down at the device with a look of quiet satisfaction. Then, turning back to the camera, a proud smile spreading across his features, he purposefully tilts the phone forward, presenting the screen directly to the viewer in a gesture of apparent ostentation, creating an air of evocative elegance and persuasiveness.", + "seed": 8375, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000056/group_5/index_1652_seed_8375_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": -0.011383704422041774, + "vlm_logprob": -0.011383704422041774, + "vlm_yes_no_margin": 5.3, + "vlm_token_credit": -0.011383704422041774, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static medium shot at eye level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject a man in his thirties with neatly cropped brown hair wearing a dark blue fine-knit sweater over a white collared shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the setting appear to be a modern, minimalist home office with soft, diffuse sidelight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the man dip his right hand beneath the frame to reveal a slim smartphone with a dark glass screen and polished metal edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the man tilt the phone forward to present the screen directly to the viewer with a proud smile?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03957377001643181 + ], + "opposite_logprob_tokens": [ + -4.664573669433594 + ], + "logprob_reward": -0.03957377001643181, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.03957377001643181, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2809807062149048, + 0.10290905833244324, + 0.08464054018259048, + 0.06614312529563904, + 0.03451879322528839, + 0.07053478062152863, + 0.06591630727052689, + 0.0286715067923069, + 0.15276136994361877, + 0.11292380839586258 + ], + "credit_map": [ + 0.03629370406270027, + 0.025576569139957428, + 0.036507006734609604, + 0.04536260664463043, + 0.0249896589666605, + 0.03757118433713913, + 0.04565371572971344, + 0.02902623824775219, + 0.00898314081132412, + 0.011476152576506138, + 0.011916248127818108, + 0.009646361693739891, + 0.02391008287668228, + 0.014794374816119671, + 0.014648837968707085, + 0.007533852010965347, + 0.005758895538747311, + 0.012025352567434311, + 0.005297298543155193, + 0.01178665366023779, + 0.01410718634724617, + 0.014316082932054996, + 0.014149007387459278, + 0.00720005901530385, + 0.00788491778075695, + 0.00951619166880846, + 0.004212509375065565, + 0.011531420052051544, + 0.008391055278480053, + 0.011155002750456333, + 0.005606295075267553, + 0.007845727726817131, + 0.004038825631141663, + 0.0034786893520504236, + 0.0032500282395631075, + 0.005222520790994167, + 0.005729012191295624, + 0.0051559628918766975, + 0.004238789901137352, + 0.0034049598034471273, + 0.009068911895155907, + 0.006408996880054474, + 0.008527304045855999, + 0.011863759718835354, + 0.0068412101827561855, + 0.008395327255129814, + 0.008946274407207966, + 0.01048299204558134, + 0.012473090551793575, + 0.006554452236741781, + 0.009208853356540203, + 0.0083718691021204, + 0.005189381539821625, + 0.008685542270541191, + 0.007453755009919405, + 0.00797935388982296, + 0.0048664589412510395, + 0.0032769960816949606, + 0.003337893635034561, + 0.0034005718771368265, + 0.003967490512877703, + 0.004203543532639742, + 0.003105379408225417, + 0.0025131702423095703, + 0.013405749574303627, + 0.03189672529697418, + 0.022745883092284203, + 0.014614448882639408, + 0.015461091883480549, + 0.02494959719479084, + 0.01852925308048725, + 0.011158603243529797, + 0.012626084499061108, + 0.01965576596558094, + 0.01246139407157898, + 0.010377475991845131, + 0.02071884088218212, + 0.015508254989981651, + 0.01083805225789547, + 0.010737926699221134 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 9 + ], + "selected_reward": -0.03957377001643181 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static medium shot at eye level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject a man in his thirties with neatly cropped brown hair wearing a dark blue fine-knit sweater over a white collared shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the setting appear to be a modern, minimalist home office with soft, diffuse sidelight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the man dip his right hand beneath the frame to reveal a slim smartphone with a dark glass screen and polished metal edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the man tilt the phone forward to present the screen directly to the viewer with a proud smile?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003631546162068844 + ], + "opposite_logprob_tokens": [ + -5.628631591796875 + ], + "logprob_reward": -0.003631546162068844, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.003631546162068844, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15462088584899902, + 0.06812853366136551, + 0.06061674654483795, + 0.16087929904460907, + 0.033462636172771454, + 0.11024370044469833, + 0.06049858778715134, + 0.05240302532911301, + 0.21626538038253784, + 0.08288124203681946 + ], + "credit_map": [ + 0.0139345433562994, + 0.014753143303096294, + 0.015293650329113007, + 0.03138543292880058, + 0.016945568844676018, + 0.021673692390322685, + 0.02126755379140377, + 0.01936730183660984, + 0.006576383952051401, + 0.007724197581410408, + 0.005124431569129229, + 0.008214720524847507, + 0.013499027118086815, + 0.011193601414561272, + 0.009233332239091396, + 0.006562836933881044, + 0.004906334914267063, + 0.005214059259742498, + 0.004789234139025211, + 0.008056778460741043, + 0.010205947794020176, + 0.008977774530649185, + 0.01243496686220169, + 0.006031643599271774, + 0.019608160480856895, + 0.014927191659808159, + 0.013510363176465034, + 0.033417247235774994, + 0.02094583958387375, + 0.02066127024590969, + 0.01646411418914795, + 0.021345097571611404, + 0.003848185995593667, + 0.0036885261069983244, + 0.00210558925755322, + 0.005227132700383663, + 0.00673079676926136, + 0.004352305084466934, + 0.004687177948653698, + 0.002822922309860587, + 0.022340470924973488, + 0.014033660292625427, + 0.021822309121489525, + 0.013422930613160133, + 0.008043578825891018, + 0.01571236178278923, + 0.005808226298540831, + 0.009060150943696499, + 0.007380336057394743, + 0.007354955188930035, + 0.006445238366723061, + 0.0100101288408041, + 0.006585459690541029, + 0.008287226781249046, + 0.007020128425210714, + 0.00741510558873415, + 0.006527830846607685, + 0.006948752328753471, + 0.003432519966736436, + 0.009819502010941505, + 0.006226520519703627, + 0.007732221856713295, + 0.005795647390186787, + 0.005920025520026684, + 0.018842976540327072, + 0.05871403589844704, + 0.0410366952419281, + 0.021918997168540955, + 0.011792604811489582, + 0.040155891329050064, + 0.012146993540227413, + 0.011657176539301872, + 0.01078207977116108, + 0.02010229043662548, + 0.008274281397461891, + 0.0069276513531804085, + 0.01146488357335329, + 0.012817944400012493, + 0.0058293091133236885, + 0.0066827950067818165 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 3, + 0 + ], + "selected_reward": -0.003631546162068844 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static medium shot at eye level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject a man in his thirties with neatly cropped brown hair wearing a dark blue fine-knit sweater over a white collared shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the setting appear to be a modern, minimalist home office with soft, diffuse sidelight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the man dip his right hand beneath the frame to reveal a slim smartphone with a dark glass screen and polished metal edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the man tilt the phone forward to present the screen directly to the viewer with a proud smile?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005261383950710297 + ], + "opposite_logprob_tokens": [ + -5.255261421203613 + ], + "logprob_reward": -0.005261383950710297, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005261383950710297, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0882616639137268, + 0.04337802529335022, + 0.03633048012852669, + 0.08391515910625458, + 0.014967388473451138, + 0.40418538451194763, + 0.12024794518947601, + 0.08022688329219818, + 0.0756986066699028, + 0.052788447588682175 + ], + "credit_map": [ + 0.010057775303721428, + 0.010116062127053738, + 0.01399941835552454, + 0.013551359996199608, + 0.012006421573460102, + 0.01280354056507349, + 0.008263076655566692, + 0.007464007940143347, + 0.004498209338635206, + 0.0026556546799838543, + 0.0049225264228880405, + 0.00984079297631979, + 0.005338143557310104, + 0.006131477653980255, + 0.004647662863135338, + 0.005343555007129908, + 0.004251120612025261, + 0.006595764309167862, + 0.0030621776822954416, + 0.0044808960519731045, + 0.003346797311678529, + 0.008139303885400295, + 0.0032917202915996313, + 0.003162700915709138, + 0.008065922185778618, + 0.009812938049435616, + 0.008090666495263577, + 0.01672997884452343, + 0.008403419516980648, + 0.011404840275645256, + 0.007960156537592411, + 0.013447228819131851, + 0.002066822024062276, + 0.001843420322984457, + 0.0019105433020740747, + 0.0026286104694008827, + 0.0021800275426357985, + 0.0018292376771569252, + 0.0012693076860159636, + 0.0012394195655360818, + 0.09591174870729446, + 0.05183587595820427, + 0.09169639647006989, + 0.043833017349243164, + 0.02008819207549095, + 0.05589059367775917, + 0.015774384140968323, + 0.029155194759368896, + 0.02537153847515583, + 0.013715210370719433, + 0.021894296631217003, + 0.01468350924551487, + 0.009501657448709011, + 0.015884526073932648, + 0.00652635982260108, + 0.01267083827406168, + 0.018655208870768547, + 0.009389078244566917, + 0.016858404502272606, + 0.008159575052559376, + 0.006441281642764807, + 0.010449125431478024, + 0.004626461770385504, + 0.005647747311741114, + 0.00795705709606409, + 0.019972193986177444, + 0.012197267264127731, + 0.008081520907580853, + 0.004322533495724201, + 0.013583500869572163, + 0.004320986568927765, + 0.005263546481728554, + 0.0056820460595190525, + 0.012023077346384525, + 0.006162318866699934, + 0.004797755274921656, + 0.006720991805195808, + 0.008178222924470901, + 0.005080651957541704, + 0.00414338568225503 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 6, + 0 + ], + "selected_reward": -0.005261383950710297 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static medium shot at eye level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject a man in his thirties with neatly cropped brown hair wearing a dark blue fine-knit sweater over a white collared shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the setting appear to be a modern, minimalist home office with soft, diffuse sidelight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the man dip his right hand beneath the frame to reveal a slim smartphone with a dark glass screen and polished metal edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the man tilt the phone forward to present the screen directly to the viewer with a proud smile?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005257233511656523 + ], + "opposite_logprob_tokens": [ + -5.2552571296691895 + ], + "logprob_reward": -0.005257233511656523, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005257233511656523, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10040391981601715, + 0.07936368882656097, + 0.04037307947874069, + 0.046490367501974106, + 0.04727635532617569, + 0.14441703259944916, + 0.07075148820877075, + 0.038190729916095734, + 0.33701854944229126, + 0.09571484476327896 + ], + "credit_map": [ + 0.012891115620732307, + 0.015773126855492592, + 0.00938307773321867, + 0.01202705129981041, + 0.012232278473675251, + 0.021368728950619698, + 0.008970018476247787, + 0.007758524734526873, + 0.009254056960344315, + 0.013832219876348972, + 0.010913608595728874, + 0.008058784529566765, + 0.006238203961402178, + 0.01942741498351097, + 0.005561321508139372, + 0.006078071426600218, + 0.002884429646655917, + 0.006501170340925455, + 0.002984398975968361, + 0.005422016140073538, + 0.006048038601875305, + 0.007749997545033693, + 0.004614480305463076, + 0.004168549086898565, + 0.0054340544156730175, + 0.005490676034241915, + 0.005269658286124468, + 0.008258448913693428, + 0.004473647102713585, + 0.006430468056350946, + 0.004068160429596901, + 0.007065253332257271, + 0.005309389904141426, + 0.004897569306194782, + 0.0034755165688693523, + 0.010024513117969036, + 0.0061835902743041515, + 0.006441099569201469, + 0.006035348866134882, + 0.004909325856715441, + 0.03749213367700577, + 0.016785042360424995, + 0.03266749158501625, + 0.016028931364417076, + 0.007936480455100536, + 0.015230391174554825, + 0.007287926506251097, + 0.010988636873662472, + 0.010947801172733307, + 0.007466976065188646, + 0.009925980120897293, + 0.011512619443237782, + 0.0061154658906161785, + 0.011271183378994465, + 0.005332869477570057, + 0.008178587071597576, + 0.004912307020276785, + 0.0047837537713348866, + 0.0036571030505001545, + 0.005759074818342924, + 0.004593773745000362, + 0.0058661396615207195, + 0.0036228992976248264, + 0.004995678551495075, + 0.028283672407269478, + 0.09100129455327988, + 0.06469714641571045, + 0.025236554443836212, + 0.01777798868715763, + 0.0615689717233181, + 0.025078989565372467, + 0.02337392419576645, + 0.011674580164253712, + 0.014038147404789925, + 0.014245945028960705, + 0.010936467908322811, + 0.015361502766609192, + 0.011897826567292213, + 0.00823588203638792, + 0.009324494749307632 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 5, + 0 + ], + "selected_reward": -0.005257233511656523 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static medium shot at eye level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject a man in his thirties with neatly cropped brown hair wearing a dark blue fine-knit sweater over a white collared shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the setting appear to be a modern, minimalist home office with soft, diffuse sidelight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the man dip his right hand beneath the frame to reveal a slim smartphone with a dark glass screen and polished metal edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the man tilt the phone forward to present the screen directly to the viewer with a proud smile?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0031945884693413973 + ], + "opposite_logprob_tokens": [ + -5.753194808959961 + ], + "logprob_reward": -0.0031945884693413973, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.0031945884693413973, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10500143468379974, + 0.03382490947842598, + 0.01842835173010826, + 0.02906622923910618, + 0.03844767436385155, + 0.09097180515527725, + 0.03924961015582085, + 0.01946735940873623, + 0.3809995949268341, + 0.24454306066036224 + ], + "credit_map": [ + 0.01155419833958149, + 0.01250105258077383, + 0.009356489405035973, + 0.017751306295394897, + 0.010633840225636959, + 0.01643035188317299, + 0.015616475604474545, + 0.011157721281051636, + 0.0028782340232282877, + 0.005888724699616432, + 0.004164830781519413, + 0.002676707226783037, + 0.005563157144933939, + 0.007952158339321613, + 0.0022849147208034992, + 0.0024161809124052525, + 0.0019253952195867896, + 0.002145413774996996, + 0.001684662071056664, + 0.0032223595771938562, + 0.002610416617244482, + 0.0028942273929715157, + 0.0023239150177687407, + 0.0016219635726884007, + 0.0038424506783485413, + 0.0030551168601959944, + 0.002253097714856267, + 0.00626489007845521, + 0.003414508420974016, + 0.004581227898597717, + 0.002842437941581011, + 0.0028124984819442034, + 0.0053298743441700935, + 0.004004801157861948, + 0.0031279604882001877, + 0.0077148862183094025, + 0.004914180375635624, + 0.004484251141548157, + 0.003457821672782302, + 0.005413898266851902, + 0.023163218051195145, + 0.011745131574571133, + 0.0189361609518528, + 0.009675822220742702, + 0.0049379803240299225, + 0.011804837733507156, + 0.004021534696221352, + 0.006687118671834469, + 0.004749963525682688, + 0.004158373456448317, + 0.0038241969887167215, + 0.006833908148109913, + 0.004053281620144844, + 0.00566920917481184, + 0.0039051652420312166, + 0.006055515259504318, + 0.0032066828571259975, + 0.0022563044913113117, + 0.002824320225045085, + 0.0023553355131298304, + 0.0025728840846568346, + 0.0027924447786062956, + 0.0016020121984183788, + 0.0018573759589344263, + 0.02988225221633911, + 0.11464663594961166, + 0.06944917142391205, + 0.037434063851833344, + 0.016215374693274498, + 0.07215538620948792, + 0.021032053977251053, + 0.02018464170396328, + 0.02974606305360794, + 0.030414095148444176, + 0.025138897821307182, + 0.012327087111771107, + 0.06597018986940384, + 0.03839489072561264, + 0.025015814229846, + 0.01753603108227253 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 0 + ], + "selected_reward": -0.0031945884693413973 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, medium shot, at eye level, captures a man in his thirties with neatly cropped brown hair, who initially speaks directly to the camera with an engaging and confident expression. He appears smartly casual in a dark blue fine-knit Merino wool sweater layered over a crisp white collared shirt. The setting, a modern, minimalist home office, is bathed in soft, diffuse light from an unseen sidelight, creating a gentle highlight on his features and a clean, professional aura. With a smooth, deliberate movement, his right hand dips beneath the frame before rising again, clutching a slim, modern smartphone with a dark glass screen and polished metal edges. His gaze momentarily wanders from the camera, his head tilted slightly, looking down at the device with a look of quiet satisfaction. Then, turning back to the camera, a proud smile spreading across his features, he purposefully tilts the phone forward, presenting the screen directly to the viewer in a gesture of apparent ostentation, creating an air of evocative elegance and persuasiveness.", + "seed": 8376, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000056/group_5/index_1652_seed_8376_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": -0.01582157574594021, + "vlm_logprob": -0.01582157574594021, + "vlm_yes_no_margin": 4.725, + "vlm_token_credit": -0.01582157574594021, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static medium shot at eye level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject a man in his thirties with neatly cropped brown hair wearing a dark blue fine-knit sweater over a white collared shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the setting appear to be a modern, minimalist home office with soft, diffuse sidelight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the man dip his right hand beneath the frame to reveal a slim smartphone with a dark glass screen and polished metal edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the man tilt the phone forward to present the screen directly to the viewer with a proud smile?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03652440384030342 + ], + "opposite_logprob_tokens": [ + -5.036524295806885 + ], + "logprob_reward": -0.03652440384030342, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.03652440384030342, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05868912860751152, + 0.025641899555921555, + 0.054582759737968445, + 0.021730808541178703, + 0.3038482367992401, + 0.028676528483629227, + 0.037466228008270264, + 0.03939139097929001, + 0.17565374076366425, + 0.25431931018829346 + ], + "credit_map": [ + 0.008151213638484478, + 0.008941836655139923, + 0.009447956457734108, + 0.004971019923686981, + 0.010284004732966423, + 0.006158473901450634, + 0.006277811713516712, + 0.004456809256225824, + 0.0032594166696071625, + 0.0037048403173685074, + 0.003879561321809888, + 0.0023884910624474287, + 0.005450353492051363, + 0.0027523983735591173, + 0.0022704119328409433, + 0.0019364268518984318, + 0.007007260341197252, + 0.006605107337236404, + 0.006278069224208593, + 0.006617980543524027, + 0.007262720726430416, + 0.00527723366394639, + 0.010075571946799755, + 0.005458816885948181, + 0.0025272765196859837, + 0.0018843443831428885, + 0.0029318463057279587, + 0.002716518472880125, + 0.003994677681475878, + 0.002641649916768074, + 0.0033422831911593676, + 0.00169221346732229, + 0.06051192805171013, + 0.04265696182847023, + 0.04340968653559685, + 0.04262315109372139, + 0.031047511845827103, + 0.026787174865603447, + 0.0228352639824152, + 0.03397655114531517, + 0.0035810687113553286, + 0.003151385113596916, + 0.003485051915049553, + 0.0031587344128638506, + 0.005535149946808815, + 0.003909232560545206, + 0.0034006438218057156, + 0.00245526060461998, + 0.005244076251983643, + 0.007380466442555189, + 0.004958673380315304, + 0.004815271124243736, + 0.0051666018553078175, + 0.004410783760249615, + 0.0033613620325922966, + 0.002128992462530732, + 0.0040840948931872845, + 0.00461585121229291, + 0.004533312749117613, + 0.005322633311152458, + 0.007785559631884098, + 0.005239270627498627, + 0.0055072493851184845, + 0.0023034182377159595, + 0.018992234021425247, + 0.013716625049710274, + 0.016975121572613716, + 0.02159874513745308, + 0.02037947252392769, + 0.01904023252427578, + 0.04556838423013687, + 0.019382929429411888, + 0.026225676760077477, + 0.025367872789502144, + 0.02880728431046009, + 0.026604842394590378, + 0.036745935678482056, + 0.026845308020710945, + 0.06048781797289848, + 0.023234548047184944 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 9, + 8 + ], + "selected_reward": -0.03652440384030342 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static medium shot at eye level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject a man in his thirties with neatly cropped brown hair wearing a dark blue fine-knit sweater over a white collared shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the setting appear to be a modern, minimalist home office with soft, diffuse sidelight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the man dip his right hand beneath the frame to reveal a slim smartphone with a dark glass screen and polished metal edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the man tilt the phone forward to present the screen directly to the viewer with a proud smile?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00866025872528553 + ], + "opposite_logprob_tokens": [ + -4.758660316467285 + ], + "logprob_reward": -0.00866025872528553, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.00866025872528553, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10963977873325348, + 0.06703545153141022, + 0.1351432353258133, + 0.033311598002910614, + 0.1723204106092453, + 0.0522925965487957, + 0.04319276660680771, + 0.0770891010761261, + 0.09481766074895859, + 0.21515731513500214 + ], + "credit_map": [ + 0.011636273935437202, + 0.024790771305561066, + 0.011241509579122066, + 0.011832434684038162, + 0.01531983818858862, + 0.01419273391366005, + 0.014758026227355003, + 0.005868188571184874, + 0.007509640417993069, + 0.008467244915664196, + 0.008389475755393505, + 0.005578633863478899, + 0.01628459244966507, + 0.005854710936546326, + 0.010813808999955654, + 0.004137346986681223, + 0.016042659059166908, + 0.017293624579906464, + 0.018541524186730385, + 0.01587624102830887, + 0.02316967397928238, + 0.01512189581990242, + 0.017683569341897964, + 0.01141404826194048, + 0.0044653145596385, + 0.003438953310251236, + 0.004012035205960274, + 0.005378259811550379, + 0.004472979344427586, + 0.003469373332336545, + 0.005630431696772575, + 0.0024442493449896574, + 0.033035747706890106, + 0.023188045248389244, + 0.023839864879846573, + 0.023359013721346855, + 0.0181846022605896, + 0.017105743288993835, + 0.014736325480043888, + 0.018871065229177475, + 0.0047814198769629, + 0.004849130753427744, + 0.006868875119835138, + 0.005292264744639397, + 0.010936860926449299, + 0.00746181420981884, + 0.0070708757266402245, + 0.0050313533283770084, + 0.004772498272359371, + 0.005265534855425358, + 0.005238974932581186, + 0.005710140336304903, + 0.007349519990384579, + 0.005563696380704641, + 0.00519279669970274, + 0.004099603276699781, + 0.008881131187081337, + 0.008438908495008945, + 0.009591430425643921, + 0.009828979149460793, + 0.014631655998528004, + 0.010125883854925632, + 0.007304043974727392, + 0.008287073113024235, + 0.011110994033515453, + 0.007399505935609341, + 0.010782366618514061, + 0.008757125586271286, + 0.01402860227972269, + 0.008674737997353077, + 0.02383134886622429, + 0.010232990607619286, + 0.022312749177217484, + 0.019234033301472664, + 0.022444993257522583, + 0.024129699915647507, + 0.03337061032652855, + 0.021856993436813354, + 0.054357416927814484, + 0.01745082624256611 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 4, + 2 + ], + "selected_reward": -0.00866025872528553 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static medium shot at eye level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject a man in his thirties with neatly cropped brown hair wearing a dark blue fine-knit sweater over a white collared shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the setting appear to be a modern, minimalist home office with soft, diffuse sidelight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the man dip his right hand beneath the frame to reveal a slim smartphone with a dark glass screen and polished metal edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the man tilt the phone forward to present the screen directly to the viewer with a proud smile?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004645863547921181 + ], + "opposite_logprob_tokens": [ + -5.379645824432373 + ], + "logprob_reward": -0.004645863547921181, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004645863547921181, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06518889963626862, + 0.030524881556630135, + 0.021568525582551956, + 0.008958092890679836, + 0.6947107315063477, + 0.012209026142954826, + 0.016863057389855385, + 0.007938911207020283, + 0.013239121995866299, + 0.12879876792430878 + ], + "credit_map": [ + 0.008102399297058582, + 0.014349565841257572, + 0.008842537179589272, + 0.006897363346070051, + 0.009782501496374607, + 0.008649650029838085, + 0.004601957742124796, + 0.003962918650358915, + 0.0032400519121438265, + 0.005991050973534584, + 0.004315217956900597, + 0.002411888912320137, + 0.0058272103779017925, + 0.0033559410367161036, + 0.0034966222010552883, + 0.0018868966726586223, + 0.0025230429600924253, + 0.003651499282568693, + 0.0027550854720175266, + 0.0023900545202195644, + 0.002324942732229829, + 0.001858597039245069, + 0.003477472113445401, + 0.0025878273881971836, + 0.0012240050127729774, + 0.0009627247927710414, + 0.001117429230362177, + 0.001161054940894246, + 0.0018236348405480385, + 0.0009000114514492452, + 0.0011113911168649793, + 0.0006578409811481833, + 0.1347193568944931, + 0.09948694705963135, + 0.09960588812828064, + 0.09694791585206985, + 0.06825399398803711, + 0.06433895230293274, + 0.054652139544487, + 0.07670547813177109, + 0.0012821551645174623, + 0.0012189936824142933, + 0.0014533960493281484, + 0.0013526318361982703, + 0.002288184827193618, + 0.0016890757251530886, + 0.0017081636469811201, + 0.0012164243962615728, + 0.0016634170897305012, + 0.0021520480513572693, + 0.002109259134158492, + 0.0015634791925549507, + 0.0036991145461797714, + 0.002038736129179597, + 0.00247147586196661, + 0.0011655254056677222, + 0.0011232339311391115, + 0.0009464453905820847, + 0.0009969341335818172, + 0.0008380750077776611, + 0.0012433863012120128, + 0.0009902450256049633, + 0.0009693516767583787, + 0.0008312384597957134, + 0.001448148861527443, + 0.0017696080030873418, + 0.0015292152529582381, + 0.0015172972343862057, + 0.0018661586800590158, + 0.0019333945820108056, + 0.0018955011619254947, + 0.0012797965900972486, + 0.013737075962126255, + 0.009645756334066391, + 0.014453442767262459, + 0.01400933600962162, + 0.021243566647171974, + 0.011635350994765759, + 0.033149100840091705, + 0.01092512533068657 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 9, + 0 + ], + "selected_reward": -0.004645863547921181 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static medium shot at eye level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject a man in his thirties with neatly cropped brown hair wearing a dark blue fine-knit sweater over a white collared shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the setting appear to be a modern, minimalist home office with soft, diffuse sidelight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the man dip his right hand beneath the frame to reveal a slim smartphone with a dark glass screen and polished metal edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the man tilt the phone forward to present the screen directly to the viewer with a proud smile?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011087356135249138 + ], + "opposite_logprob_tokens": [ + -4.511087417602539 + ], + "logprob_reward": -0.011087356135249138, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011087356135249138, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07231062650680542, + 0.02547197788953781, + 0.015092110261321068, + 0.011415303684771061, + 0.041690219193696976, + 0.02784574218094349, + 0.022709909826517105, + 0.01086459681391716, + 0.038191117346286774, + 0.7344083786010742 + ], + "credit_map": [ + 0.010339872911572456, + 0.012845437042415142, + 0.007967639714479446, + 0.008443568833172321, + 0.009300719946622849, + 0.007609407417476177, + 0.010263398289680481, + 0.005540584214031696, + 0.0024865863379091024, + 0.003078222507610917, + 0.003081283997744322, + 0.0019190509337931871, + 0.007856028154492378, + 0.002142205834388733, + 0.003077504690736532, + 0.0018310971790924668, + 0.001952674938365817, + 0.0023101638071238995, + 0.001733465469442308, + 0.0013527406845241785, + 0.002267916454002261, + 0.0014671217650175095, + 0.002744577592238784, + 0.0012634502490982413, + 0.0013584978878498077, + 0.0014694543788209558, + 0.0014229597290977836, + 0.0017196845728904009, + 0.0020422646775841713, + 0.0012307856231927872, + 0.001352052204310894, + 0.0008196053095161915, + 0.006918671540915966, + 0.005128244869410992, + 0.00519286235794425, + 0.0060555823147296906, + 0.005039020907133818, + 0.004461313597857952, + 0.0038541483227163553, + 0.005040375981479883, + 0.002330971881747246, + 0.002792451763525605, + 0.0034898060839623213, + 0.00280128070153296, + 0.005941735580563545, + 0.004353105090558529, + 0.0035321537870913744, + 0.0026042363606393337, + 0.0028789532370865345, + 0.002899602521210909, + 0.002857996616512537, + 0.0025204040575772524, + 0.004747665487229824, + 0.0022168245632201433, + 0.003035224974155426, + 0.0015532384859398007, + 0.0011523808352649212, + 0.0013494809390977025, + 0.0013682120479643345, + 0.0013205388095229864, + 0.0020283241756260395, + 0.00129060004837811, + 0.0015635894378647208, + 0.0007914709276519716, + 0.003681082045659423, + 0.0033997080754488707, + 0.003953346982598305, + 0.004645664244890213, + 0.004690875764936209, + 0.004221074283123016, + 0.009248584508895874, + 0.0043507833033800125, + 0.07809732109308243, + 0.0625506341457367, + 0.08046866208314896, + 0.08582094311714172, + 0.10890605300664902, + 0.06332946568727493, + 0.1915711611509323, + 0.06366410851478577 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 4 + ], + "selected_reward": -0.011087356135249138 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static medium shot at eye level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject a man in his thirties with neatly cropped brown hair wearing a dark blue fine-knit sweater over a white collared shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the setting appear to be a modern, minimalist home office with soft, diffuse sidelight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the man dip his right hand beneath the frame to reveal a slim smartphone with a dark glass screen and polished metal edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the man tilt the phone forward to present the screen directly to the viewer with a proud smile?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018189996480941772 + ], + "opposite_logprob_tokens": [ + -4.018189907073975 + ], + "logprob_reward": -0.018189996480941772, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018189996480941772, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12783066928386688, + 0.04446239024400711, + 0.034237638115882874, + 0.018222017213702202, + 0.22084230184555054, + 0.03667708486318588, + 0.0425964780151844, + 0.029632356017827988, + 0.09891463816165924, + 0.34658437967300415 + ], + "credit_map": [ + 0.014036079868674278, + 0.01915050856769085, + 0.012006880715489388, + 0.010427710600197315, + 0.02523433044552803, + 0.010898477397859097, + 0.027716780081391335, + 0.008359912782907486, + 0.005230541341006756, + 0.006775349378585815, + 0.005395402666181326, + 0.0036764678079634905, + 0.010334501974284649, + 0.00439299875870347, + 0.0058129834942519665, + 0.0028441459871828556, + 0.00336446613073349, + 0.0032249540090560913, + 0.004584186710417271, + 0.0027222903445363045, + 0.007925298996269703, + 0.003427263116464019, + 0.0059237065725028515, + 0.0030654717702418566, + 0.002051097806543112, + 0.001649543410167098, + 0.0024730288423597813, + 0.003657905850559473, + 0.002776975743472576, + 0.0014454349875450134, + 0.0027670925483107567, + 0.0014009398873895407, + 0.04294171556830406, + 0.029606502503156662, + 0.030233558267354965, + 0.030607905238866806, + 0.02373536117374897, + 0.02081986330449581, + 0.017469054087996483, + 0.025428347289562225, + 0.0036720342468470335, + 0.004304096568375826, + 0.0050424267537891865, + 0.003553370712324977, + 0.007153161335736513, + 0.005344093777239323, + 0.004834610037505627, + 0.002773292362689972, + 0.0044436976313591, + 0.004659825935959816, + 0.005274798721075058, + 0.0069257719442248344, + 0.00896397139877081, + 0.004420661833137274, + 0.005041171796619892, + 0.002866581315174699, + 0.0032363145146518946, + 0.003406842239201069, + 0.003532741218805313, + 0.005037343595176935, + 0.004992263857275248, + 0.0031809958163648844, + 0.0030788714066147804, + 0.0031669819727540016, + 0.011043541133403778, + 0.007665747310966253, + 0.009169523604214191, + 0.013384127989411354, + 0.012065279297530651, + 0.010161531157791615, + 0.022244656458497047, + 0.013180237263441086, + 0.03859495744109154, + 0.03117450699210167, + 0.03906337544322014, + 0.034586355090141296, + 0.05454215034842491, + 0.04659133404493332, + 0.07166673243045807, + 0.030364936217665672 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 4, + 0 + ], + "selected_reward": -0.018189996480941772 + } + ], + "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/0000056/group_6/index_tensor([5111, 5111])_group_6_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000056/group_6/index_tensor([5111, 5111])_group_6_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..c7b50180d6277de2615c4dc79b43301cf9236187 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000056/group_6/index_tensor([5111, 5111])_group_6_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "high-speed, sports documentary-style close-up plunges the viewer directly into the heart of a fierce battle between two enormous hawkmoths in the deep night sky. The camera follows closely in a tense handheld style, capturing this aerial duel from a slightly upward angle. Backlit by the brilliant full moon, their enormous, nearly translucent emerald green wings, adorned with delicate lavender eyespots, flap with primal power, their furry white bodies tumbling through the air at high speed. They swirl and dive at each other with breathtaking speed and force, their emerald green wings clashing violently. Each collision sends countless tiny, shimmering scales scattering like stardust in the moonlight, triggering erratic camera shake and imbuing the image with the ultimate tension of action photography. This violent vibration, combined with the deep blue night sky, blurred by the high-speed motion, perfectly recreates the thrill of witnessing this aerial duel live, creating a visual spectacle both chaotic and magnificent.", + "seed": 22366, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000056/group_6/index_5111_seed_22366_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": -0.1428633530624211, + "vlm_logprob": -0.1428633530624211, + "vlm_yes_no_margin": 2.95, + "vlm_token_credit": -0.1428633530624211, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" 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 enormous hawkmoths visible engaging in a battle or aerial duel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the hawkmoths have nearly translucent emerald green wings with delicate lavender eyespots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the hawkmoths' bodies furry and white while they tumble through the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set in a deep night sky backlit by a brilliant full moon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot handheld with a slightly upward angle and erratic shake?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.22663173079490662 + ], + "opposite_logprob_tokens": [ + -1.601631760597229 + ], + "logprob_reward": -0.22663173079490662, + "yes_no_margin_reward": 1.375, + "token_credit_reward": -0.22663173079490662, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0869661197066307, + 0.07203508913516998, + 0.22531752288341522, + 0.02971670776605606, + 0.02348516322672367, + 0.3115881085395813, + 0.05332896485924721, + 0.049879975616931915, + 0.054526686668395996, + 0.09315555542707443 + ], + "credit_map": [ + 0.00971093401312828, + 0.008577593602240086, + 0.01802402175962925, + 0.01005665585398674, + 0.00759867625311017, + 0.013714718632400036, + 0.007996872067451477, + 0.011286658234894276, + 0.008840836584568024, + 0.007202555425465107, + 0.008367744274437428, + 0.01023563276976347, + 0.009708384051918983, + 0.009372764267027378, + 0.008226332254707813, + 0.010080846026539803, + 0.043160948902368546, + 0.01907459832727909, + 0.05310496315360069, + 0.024762770161032677, + 0.02687106467783451, + 0.02138754539191723, + 0.01409756951034069, + 0.02285809814929962, + 0.003909671679139137, + 0.0028301281854510307, + 0.004217332694679499, + 0.004760537762194872, + 0.0030569592490792274, + 0.003838250180706382, + 0.0023983733262866735, + 0.004705456085503101, + 0.0027710814028978348, + 0.0021409965120255947, + 0.003150187898427248, + 0.0031790060456842184, + 0.0027281278744339943, + 0.0036527886986732483, + 0.0024606299120932817, + 0.003402346046641469, + 0.1252695620059967, + 0.016301628202199936, + 0.01936759427189827, + 0.010394246317446232, + 0.011419949121773243, + 0.020941033959388733, + 0.00900479406118393, + 0.09888934344053268, + 0.004814259707927704, + 0.0062687816098332405, + 0.006406513974070549, + 0.007339576259255409, + 0.005735361482948065, + 0.0067708902060985565, + 0.007128281518816948, + 0.008865307085216045, + 0.0069059464149177074, + 0.004216648638248444, + 0.005874249618500471, + 0.004233229905366898, + 0.004860474728047848, + 0.010458501055836678, + 0.005693277809768915, + 0.00763765349984169, + 0.00416967086493969, + 0.005791401024907827, + 0.005532983224838972, + 0.004065542947500944, + 0.007973264902830124, + 0.009786293841898441, + 0.004567084368318319, + 0.012640454806387424, + 0.005686907097697258, + 0.006183246150612831, + 0.0095054367557168, + 0.006188602652400732, + 0.01252794824540615, + 0.023544183000922203, + 0.008292621932923794, + 0.02122662588953972 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 2, + 9 + ], + "selected_reward": -0.22663173079490662 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" 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 enormous hawkmoths visible engaging in a battle or aerial duel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the hawkmoths have nearly translucent emerald green wings with delicate lavender eyespots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the hawkmoths' bodies furry and white while they tumble through the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set in a deep night sky backlit by a brilliant full moon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot handheld with a slightly upward angle and erratic shake?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004645033273845911 + ], + "opposite_logprob_tokens": [ + -5.379644870758057 + ], + "logprob_reward": -0.004645033273845911, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004645033273845911, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17929749190807343, + 0.17219823598861694, + 0.09376013278961182, + 0.029378680512309074, + 0.025871360674500465, + 0.17491284012794495, + 0.07114415615797043, + 0.08646546304225922, + 0.13429518043994904, + 0.03267642855644226 + ], + "credit_map": [ + 0.017105240374803543, + 0.013170290738344193, + 0.0683322623372078, + 0.010480106808245182, + 0.006404225714504719, + 0.020071225240826607, + 0.006399999372661114, + 0.0373341366648674, + 0.021691808477044106, + 0.02195790596306324, + 0.02192077413201332, + 0.018642239272594452, + 0.02042536437511444, + 0.023754507303237915, + 0.023737190291285515, + 0.02006845362484455, + 0.01967810094356537, + 0.008593855425715446, + 0.023387053981423378, + 0.008580835536122322, + 0.010453378781676292, + 0.008615381084382534, + 0.006113256793469191, + 0.008338266052305698, + 0.0039057221729308367, + 0.004517882131040096, + 0.0036493977531790733, + 0.004812300205230713, + 0.0024049412459135056, + 0.0032905961852520704, + 0.002726851962506771, + 0.004070988856256008, + 0.0026247224304825068, + 0.005293261259794235, + 0.0029734959825873375, + 0.00371539406478405, + 0.0020851902663707733, + 0.0030158371664583683, + 0.002390749752521515, + 0.0037727116141468287, + 0.0470806285738945, + 0.011636629700660706, + 0.01861361227929592, + 0.015272093005478382, + 0.013018889352679253, + 0.022632144391536713, + 0.011845138855278492, + 0.034813690930604935, + 0.008531779050827026, + 0.0066737765446305275, + 0.007843688130378723, + 0.009073836728930473, + 0.00788798462599516, + 0.009620360098779202, + 0.008844341151416302, + 0.012668395414948463, + 0.009121892973780632, + 0.011479852721095085, + 0.009629829786717892, + 0.013410312123596668, + 0.010992543771862984, + 0.012285894714295864, + 0.007393877487629652, + 0.012151253409683704, + 0.01128888688981533, + 0.012495864182710648, + 0.014274641871452332, + 0.012167688459157944, + 0.023831291124224663, + 0.02315763756632805, + 0.012047714553773403, + 0.02503145858645439, + 0.003606747603043914, + 0.0035177033860236406, + 0.0040056174620985985, + 0.002927129389718175, + 0.005422236397862434, + 0.004492902662605047, + 0.003502593142911792, + 0.0052014985121786594 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 1 + ], + "selected_reward": -0.004645033273845911 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" 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 enormous hawkmoths visible engaging in a battle or aerial duel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the hawkmoths have nearly translucent emerald green wings with delicate lavender eyespots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the hawkmoths' bodies furry and white while they tumble through the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set in a deep night sky backlit by a brilliant full moon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot handheld with a slightly upward angle and erratic shake?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"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": [ + 0.21297040581703186, + 0.15789158642292023, + 0.2144807130098343, + 0.029982002452015877, + 0.0431167334318161, + 0.12243631482124329, + 0.07038380205631256, + 0.03887699544429779, + 0.045017749071121216, + 0.06484371423721313 + ], + "credit_map": [ + 0.020000196993350983, + 0.016235945746302605, + 0.07669993489980698, + 0.014708767645061016, + 0.008138567209243774, + 0.02484307996928692, + 0.013108590617775917, + 0.03923533111810684, + 0.017648423090577126, + 0.015692783519625664, + 0.022534819319844246, + 0.018075190484523773, + 0.019444264471530914, + 0.020830893889069557, + 0.016072748228907585, + 0.027592461556196213, + 0.05723465234041214, + 0.014995524659752846, + 0.060090843588113785, + 0.01782725751399994, + 0.021492820233106613, + 0.0160211268812418, + 0.010033845901489258, + 0.01678464189171791, + 0.0038079405203461647, + 0.005143200047314167, + 0.0037427658680826426, + 0.005850335583090782, + 0.0022583177778869867, + 0.0038714532274752855, + 0.0018443348817527294, + 0.0034636552445590496, + 0.004640594590455294, + 0.006171155720949173, + 0.00446412805467844, + 0.005426390096545219, + 0.004431380890309811, + 0.005221051629632711, + 0.004211087711155415, + 0.008550947532057762, + 0.029792403802275658, + 0.012858602218329906, + 0.014117646031081676, + 0.010819791816174984, + 0.009509564377367496, + 0.012812447734177113, + 0.007165542338043451, + 0.02536030486226082, + 0.005501282401382923, + 0.007292435970157385, + 0.008048581890761852, + 0.0066470312885940075, + 0.009260274469852448, + 0.011402135714888573, + 0.008260945789515972, + 0.013971114531159401, + 0.005044125020503998, + 0.0033756482880562544, + 0.004686092957854271, + 0.0037866586353629827, + 0.004967428278177977, + 0.006205094512552023, + 0.005197290331125259, + 0.005614657886326313, + 0.0027650543488562107, + 0.0049783699214458466, + 0.004807226825505495, + 0.0044881850481033325, + 0.009700597263872623, + 0.007418655790388584, + 0.0054009598679840565, + 0.005458700470626354, + 0.0036622255574911833, + 0.0037118312902748585, + 0.007609351072460413, + 0.004521203227341175, + 0.010048161260783672, + 0.015373962000012398, + 0.006361457519233227, + 0.013555526733398438 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 1 + ], + "selected_reward": -0.016065552830696106 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" 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 enormous hawkmoths visible engaging in a battle or aerial duel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the hawkmoths have nearly translucent emerald green wings with delicate lavender eyespots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the hawkmoths' bodies furry and white while they tumble through the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set in a deep night sky backlit by a brilliant full moon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot handheld with a slightly upward angle and erratic shake?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.038088005036115646 + ], + "opposite_logprob_tokens": [ + -3.288088083267212 + ], + "logprob_reward": -0.038088005036115646, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.038088005036115646, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0694236308336258, + 0.015204421244561672, + 0.06769869476556778, + 0.060956135392189026, + 0.03024798072874546, + 0.3279668092727661, + 0.09964415431022644, + 0.11271513998508453, + 0.19928136467933655, + 0.01686156913638115 + ], + "credit_map": [ + 0.006335464771836996, + 0.005746536888182163, + 0.02150493487715721, + 0.0062500047497451305, + 0.0046989573165774345, + 0.008183111436665058, + 0.0033978878054767847, + 0.013306736014783382, + 0.0019193501211702824, + 0.001796031603589654, + 0.0016947536496445537, + 0.0018291922751814127, + 0.0016822804464027286, + 0.002475556917488575, + 0.0018025562167167664, + 0.002004700480028987, + 0.017514511942863464, + 0.004758002236485481, + 0.015283637680113316, + 0.0048517100512981415, + 0.005432134959846735, + 0.009696145541965961, + 0.004903694614768028, + 0.005258867517113686, + 0.008361151441931725, + 0.00936525035649538, + 0.007976652123034, + 0.010270074009895325, + 0.006019589491188526, + 0.007005712948739529, + 0.004392791539430618, + 0.007564924657344818, + 0.004334953613579273, + 0.006341905333101749, + 0.004039726685732603, + 0.004813951440155506, + 0.0021881908178329468, + 0.002841915935277939, + 0.002482643583789468, + 0.003204694017767906, + 0.13489757478237152, + 0.013259698636829853, + 0.02116546779870987, + 0.009090082719922066, + 0.01291060633957386, + 0.022822773084044456, + 0.008334684185683727, + 0.10548597574234009, + 0.013278116472065449, + 0.008754153735935688, + 0.010515356436371803, + 0.011069871485233307, + 0.008051487617194653, + 0.013465609401464462, + 0.012082112953066826, + 0.02242746204137802, + 0.016911091282963753, + 0.011306952685117722, + 0.013153733685612679, + 0.01107083447277546, + 0.010865618474781513, + 0.014776911586523056, + 0.01894119195640087, + 0.015688810497522354, + 0.016479598358273506, + 0.018167881295084953, + 0.02047165483236313, + 0.018092257902026176, + 0.03407743573188782, + 0.03594624251127243, + 0.017264608293771744, + 0.038781702518463135, + 0.0012139860773459077, + 0.001249096472747624, + 0.0014926922740414739, + 0.001215440803207457, + 0.001769323367625475, + 0.003897902090102434, + 0.0018634672742336988, + 0.004159661941230297 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 8, + 7 + ], + "selected_reward": -0.038088005036115646 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" 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 enormous hawkmoths visible engaging in a battle or aerial duel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the hawkmoths have nearly translucent emerald green wings with delicate lavender eyespots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the hawkmoths' bodies furry and white while they tumble through the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set in a deep night sky backlit by a brilliant full moon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot handheld with a slightly upward angle and erratic shake?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.42888644337654114 + ], + "opposite_logprob_tokens": [ + -1.0538864135742188 + ], + "logprob_reward": -0.42888644337654114, + "yes_no_margin_reward": 0.625, + "token_credit_reward": -0.42888644337654114, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08321309089660645, + 0.033283255994319916, + 0.05253739655017853, + 0.02747996896505356, + 0.04414937645196915, + 0.44460856914520264, + 0.16486525535583496, + 0.05263197049498558, + 0.06902486085891724, + 0.02820625901222229 + ], + "credit_map": [ + 0.010065045207738876, + 0.00973539613187313, + 0.0189096350222826, + 0.008310276083648205, + 0.007672622334212065, + 0.008855882100760937, + 0.006789708510041237, + 0.012874520383775234, + 0.003492730436846614, + 0.003671824699267745, + 0.005368751939386129, + 0.005292776506394148, + 0.004426240921020508, + 0.0029615226667374372, + 0.0036474980879575014, + 0.004421907477080822, + 0.011732595972716808, + 0.004369942471385002, + 0.012684127315878868, + 0.004562030080705881, + 0.006457414943724871, + 0.004738211166113615, + 0.0029000816866755486, + 0.005092991981655359, + 0.0030845075380057096, + 0.002593590412288904, + 0.0029489388689398766, + 0.005026262253522873, + 0.0025579514913260937, + 0.003452728036791086, + 0.0031963996589183807, + 0.0046195886097848415, + 0.005799705162644386, + 0.0027450555935502052, + 0.005376703571528196, + 0.0029657501727342606, + 0.004849025513976812, + 0.00609126640483737, + 0.005067947320640087, + 0.011253922246396542, + 0.11382804065942764, + 0.035139553248882294, + 0.043296929448843, + 0.041094277054071426, + 0.033624839037656784, + 0.04834646359086037, + 0.03477361053228378, + 0.09450485557317734, + 0.020558729767799377, + 0.016610819846391678, + 0.016792403534054756, + 0.014216079376637936, + 0.02093171328306198, + 0.023833494633436203, + 0.01914767548441887, + 0.03277434781193733, + 0.007214883342385292, + 0.00518063735216856, + 0.006069434341043234, + 0.006360047496855259, + 0.004947044886648655, + 0.007004082668572664, + 0.008608797565102577, + 0.0072470358572900295, + 0.00576689001172781, + 0.005490741692483425, + 0.006280095316469669, + 0.0059792958199977875, + 0.010372845456004143, + 0.013101600110530853, + 0.006309302989393473, + 0.015724092721939087, + 0.0028903447091579437, + 0.0023845802061259747, + 0.003986194264143705, + 0.002530114958062768, + 0.005505177192389965, + 0.003994843922555447, + 0.0033871217165142298, + 0.0035278804134577513 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 6, + 0 + ], + "selected_reward": -0.42888644337654114 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "high-speed, sports documentary-style close-up plunges the viewer directly into the heart of a fierce battle between two enormous hawkmoths in the deep night sky. The camera follows closely in a tense handheld style, capturing this aerial duel from a slightly upward angle. Backlit by the brilliant full moon, their enormous, nearly translucent emerald green wings, adorned with delicate lavender eyespots, flap with primal power, their furry white bodies tumbling through the air at high speed. They swirl and dive at each other with breathtaking speed and force, their emerald green wings clashing violently. Each collision sends countless tiny, shimmering scales scattering like stardust in the moonlight, triggering erratic camera shake and imbuing the image with the ultimate tension of action photography. This violent vibration, combined with the deep blue night sky, blurred by the high-speed motion, perfectly recreates the thrill of witnessing this aerial duel live, creating a visual spectacle both chaotic and magnificent.", + "seed": 22367, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000056/group_6/index_5111_seed_22367_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": -0.2083415960893035, + "vlm_logprob": -0.2083415960893035, + "vlm_yes_no_margin": 2.95, + "vlm_token_credit": -0.2083415960893035, + "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 enormous hawkmoths visible engaging in a battle or aerial duel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the hawkmoths have nearly translucent emerald green wings with delicate lavender eyespots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the hawkmoths' bodies furry and white while they tumble through the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set in a deep night sky backlit by a brilliant full moon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot handheld with a slightly upward angle and erratic shake?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05557616427540779 + ], + "opposite_logprob_tokens": [ + -2.9305760860443115 + ], + "logprob_reward": -0.05557616427540779, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05557616427540779, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1648453176021576, + 0.07803337275981903, + 0.08464660495519638, + 0.18919169902801514, + 0.07625593990087509, + 0.050661601126194, + 0.11314347386360168, + 0.07439620792865753, + 0.09679475426673889, + 0.07203107327222824 + ], + "credit_map": [ + 0.021579131484031677, + 0.010086897760629654, + 0.031735215336084366, + 0.01614256389439106, + 0.007474562618881464, + 0.010144975036382675, + 0.03641422092914581, + 0.03126773238182068, + 0.010075253434479237, + 0.006256834138184786, + 0.011656095273792744, + 0.008064345456659794, + 0.007839693687856197, + 0.013500494882464409, + 0.007934707216918468, + 0.012705942615866661, + 0.014437162317335606, + 0.01219298504292965, + 0.008946338668465614, + 0.008638986386358738, + 0.013714759610593319, + 0.011073709465563297, + 0.005922649521380663, + 0.009720011614263058, + 0.03470326587557793, + 0.01909702643752098, + 0.02496906742453575, + 0.014674619771540165, + 0.021467356011271477, + 0.03334379568696022, + 0.014564817771315575, + 0.026371728628873825, + 0.007820437662303448, + 0.005643347278237343, + 0.010626107454299927, + 0.0110536590218544, + 0.01101023517549038, + 0.014523391611874104, + 0.0064110164530575275, + 0.009167742915451527, + 0.004199665505439043, + 0.006111806258559227, + 0.005305744707584381, + 0.003909120336174965, + 0.008307760581374168, + 0.012097887694835663, + 0.00506725674495101, + 0.005662351846694946, + 0.005271888803690672, + 0.004815620835870504, + 0.00404119398444891, + 0.003357215551659465, + 0.0041033742018043995, + 0.0441383421421051, + 0.031634848564863205, + 0.0157809816300869, + 0.006455797702074051, + 0.007572962436825037, + 0.005468475166708231, + 0.006110457703471184, + 0.017118122428655624, + 0.012436534278094769, + 0.007904047146439552, + 0.011329798959195614, + 0.011211377568542957, + 0.013268670067191124, + 0.00845849234610796, + 0.011752765625715256, + 0.01781792752444744, + 0.011589485220611095, + 0.009089931845664978, + 0.013606095686554909, + 0.006295364815741777, + 0.006437970325350761, + 0.0045678564347326756, + 0.005960436072200537, + 0.020756613463163376, + 0.015793083235621452, + 0.004320243839174509, + 0.007899504154920578 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 6 + ], + "selected_reward": -0.05557616427540779 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" 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 enormous hawkmoths visible engaging in a battle or aerial duel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the hawkmoths have nearly translucent emerald green wings with delicate lavender eyespots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the hawkmoths' bodies furry and white while they tumble through the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set in a deep night sky backlit by a brilliant full moon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot handheld with a slightly upward angle and erratic shake?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0052580637857317924 + ], + "opposite_logprob_tokens": [ + -5.255258083343506 + ], + "logprob_reward": -0.0052580637857317924, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.0052580637857317924, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13240915536880493, + 0.10478272289037704, + 0.05030694603919983, + 0.06066827103495598, + 0.055090948939323425, + 0.09524678438901901, + 0.08883073180913925, + 0.08689983189105988, + 0.27975302934646606, + 0.046011630445718765 + ], + "credit_map": [ + 0.014624254778027534, + 0.01994161494076252, + 0.019909655675292015, + 0.012195070274174213, + 0.010662227869033813, + 0.017700817435979843, + 0.015235014259815216, + 0.0221404992043972, + 0.008989902213215828, + 0.011542159132659435, + 0.02246890217065811, + 0.008046340197324753, + 0.010551342740654945, + 0.01052952278405428, + 0.006380721461027861, + 0.026273833587765694, + 0.008227912709116936, + 0.005421066656708717, + 0.006387803237885237, + 0.004402571357786655, + 0.006594713311642408, + 0.006358183454722166, + 0.005288350861519575, + 0.007626342121511698, + 0.008556649088859558, + 0.004297115374356508, + 0.008040670305490494, + 0.004434365313500166, + 0.006644600071012974, + 0.01331828348338604, + 0.007190804462879896, + 0.008185780607163906, + 0.005275431554764509, + 0.006956647150218487, + 0.006641037296503782, + 0.007186783943325281, + 0.007142872549593449, + 0.0070920586585998535, + 0.008275444619357586, + 0.006520668976008892, + 0.0072724721394479275, + 0.009535630233585835, + 0.01117478683590889, + 0.011509396135807037, + 0.01123613677918911, + 0.017585571855306625, + 0.015912866219878197, + 0.011019932106137276, + 0.005483150947839022, + 0.0070511410012841225, + 0.0030442767310887575, + 0.002725716680288315, + 0.0034845389891415834, + 0.017871933057904243, + 0.018913403153419495, + 0.030256571248173714, + 0.0075605022720992565, + 0.009570869617164135, + 0.007365584839135408, + 0.008135189302265644, + 0.017734143882989883, + 0.01484380941838026, + 0.007678666617721319, + 0.01401106920093298, + 0.05436645820736885, + 0.05386853590607643, + 0.014741113409399986, + 0.026853185147047043, + 0.023393336683511734, + 0.02241080440580845, + 0.021484140306711197, + 0.06263542920351028, + 0.004926734138280153, + 0.004921544808894396, + 0.006735090631991625, + 0.0051940265111625195, + 0.009353095665574074, + 0.006344202905893326, + 0.0038100769743323326, + 0.00472685880959034 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 1 + ], + "selected_reward": -0.0052580637857317924 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" 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 enormous hawkmoths visible engaging in a battle or aerial duel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the hawkmoths have nearly translucent emerald green wings with delicate lavender eyespots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the hawkmoths' bodies furry and white while they tumble through the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set in a deep night sky backlit by a brilliant full moon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot handheld with a slightly upward angle and erratic shake?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012537825852632523 + ], + "opposite_logprob_tokens": [ + -4.387537956237793 + ], + "logprob_reward": -0.012537825852632523, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012537825852632523, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1047598347067833, + 0.09674820303916931, + 0.0603773258626461, + 0.06728833168745041, + 0.07103107869625092, + 0.13528753817081451, + 0.10870911926031113, + 0.07693719863891602, + 0.23346389830112457, + 0.045397523790597916 + ], + "credit_map": [ + 0.013009581714868546, + 0.018796352669596672, + 0.012664098292589188, + 0.00924897100776434, + 0.011198568157851696, + 0.01607067883014679, + 0.011759118176996708, + 0.012012457475066185, + 0.009823935106396675, + 0.009002238512039185, + 0.01810121349990368, + 0.008076070807874203, + 0.012649704702198505, + 0.01148905511945486, + 0.006334947422146797, + 0.021271029487252235, + 0.00876505859196186, + 0.008800358511507511, + 0.00766155868768692, + 0.00489457743242383, + 0.009193147532641888, + 0.00766042759642005, + 0.004686224274337292, + 0.008715970441699028, + 0.009876659139990807, + 0.00516063068062067, + 0.008042397908866405, + 0.00576796242967248, + 0.006752030458301306, + 0.013594703748822212, + 0.00874896440654993, + 0.009344975464046001, + 0.008702399209141731, + 0.005250329617410898, + 0.013405309990048409, + 0.008808988146483898, + 0.007321975659579039, + 0.009880865924060345, + 0.008049937896430492, + 0.009611268527805805, + 0.011965282261371613, + 0.011788121424615383, + 0.019759153947234154, + 0.01587422750890255, + 0.01620776765048504, + 0.027200758457183838, + 0.017773235216736794, + 0.014718996360898018, + 0.006089123897254467, + 0.008149668574333191, + 0.005541590508073568, + 0.00400256272405386, + 0.004724541213363409, + 0.03180600330233574, + 0.012633147649466991, + 0.035762473940849304, + 0.006460527423769236, + 0.008612047880887985, + 0.006949168164283037, + 0.0068604424595832825, + 0.014769770205020905, + 0.013383529148995876, + 0.007357604801654816, + 0.012544097378849983, + 0.044534120708703995, + 0.04421386122703552, + 0.01307149138301611, + 0.02257038652896881, + 0.021032245829701424, + 0.018920011818408966, + 0.01877644658088684, + 0.050345320254564285, + 0.004903791937977076, + 0.0049058725126087666, + 0.005773868877440691, + 0.004819976165890694, + 0.007999624125659466, + 0.006304007489234209, + 0.003992139827460051, + 0.006698243319988251 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 5, + 6 + ], + "selected_reward": -0.012537825852632523 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two enormous hawkmoths visible engaging in a battle or aerial duel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the hawkmoths have nearly translucent emerald green wings with delicate lavender eyespots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the hawkmoths' bodies furry and white while they tumble through the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set in a deep night sky backlit by a brilliant full moon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot handheld with a slightly upward angle and erratic shake?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07000714540481567 + ], + "opposite_logprob_tokens": [ + -2.695007085800171 + ], + "logprob_reward": -0.07000714540481567, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07000714540481567, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14250782132148743, + 0.044611528515815735, + 0.04400061443448067, + 0.046641625463962555, + 0.05848070606589317, + 0.05096138268709183, + 0.08784981817007065, + 0.05863669514656067, + 0.44567182660102844, + 0.02063802443444729 + ], + "credit_map": [ + 0.013902263715863228, + 0.0253201425075531, + 0.018766969442367554, + 0.012313744984567165, + 0.014770569279789925, + 0.023270614445209503, + 0.01684732921421528, + 0.017316194251179695, + 0.004488668870180845, + 0.007189297117292881, + 0.0059342654421925545, + 0.0039745112881064415, + 0.007273797411471605, + 0.005174577236175537, + 0.0034567033872008324, + 0.007119706366211176, + 0.005448586773127317, + 0.00721672223880887, + 0.005298857111483812, + 0.00398223614320159, + 0.005184710491448641, + 0.00487359007820487, + 0.0047705513425171375, + 0.007225359324365854, + 0.006497510708868504, + 0.003165412927046418, + 0.005557690747082233, + 0.003941711969673634, + 0.0042172023095190525, + 0.010161077603697777, + 0.006495876703411341, + 0.006605144124478102, + 0.006100571248680353, + 0.005845024716109037, + 0.009312547743320465, + 0.0064282952807843685, + 0.008750570937991142, + 0.006239826790988445, + 0.010167472064495087, + 0.005636398680508137, + 0.004731995053589344, + 0.00521627813577652, + 0.005261883605271578, + 0.004154249094426632, + 0.008486701175570488, + 0.012320112437009811, + 0.004972720984369516, + 0.0058174459263682365, + 0.0064610582776367664, + 0.006844426039606333, + 0.0038698764983564615, + 0.0034143405500799417, + 0.005379512440413237, + 0.009401661343872547, + 0.029504913836717606, + 0.022974029183387756, + 0.004672430455684662, + 0.008014753460884094, + 0.006383827887475491, + 0.005019751377403736, + 0.00916871428489685, + 0.00969069916754961, + 0.005055759567767382, + 0.010630757547914982, + 0.08490744233131409, + 0.08552735298871994, + 0.023385263979434967, + 0.04459495469927788, + 0.03794367238879204, + 0.03565571829676628, + 0.031593386083841324, + 0.10206400603055954, + 0.0024237933102995157, + 0.0015864847227931023, + 0.0021497514098882675, + 0.001333417370915413, + 0.0019376191776245832, + 0.005548115819692612, + 0.0009881706209853292, + 0.004670671187341213 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 6 + ], + "selected_reward": -0.07000714540481567 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" 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 enormous hawkmoths visible engaging in a battle or aerial duel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the hawkmoths have nearly translucent emerald green wings with delicate lavender eyespots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the hawkmoths' bodies furry and white while they tumble through the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set in a deep night sky backlit by a brilliant full moon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot handheld with a slightly upward angle and erratic shake?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"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": [ + -0.8983287811279297 + ], + "opposite_logprob_tokens": [ + -0.5233287811279297 + ], + "logprob_reward": -0.8983287811279297, + "yes_no_margin_reward": -0.375, + "token_credit_reward": -0.8983287811279297, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.2457454651594162, + 0.0649140253663063, + 0.0966472253203392, + 0.11935755610466003, + 0.12518279254436493, + 0.07009023427963257, + 0.029508251696825027, + 0.028615782037377357, + 0.18992643058300018, + 0.030012233182787895 + ], + "credit_map": [ + 0.029599685221910477, + 0.018923617899417877, + 0.04629102349281311, + 0.024987101554870605, + 0.015404937788844109, + 0.01667526550590992, + 0.05416211858391762, + 0.03970171511173248, + 0.008686338551342487, + 0.005699460860341787, + 0.010363172739744186, + 0.0073104253970086575, + 0.007454474922269583, + 0.010084579698741436, + 0.007278511766344309, + 0.008037055842578411, + 0.013460999354720116, + 0.006881233770400286, + 0.014431154355406761, + 0.009931711480021477, + 0.010185343213379383, + 0.0165241751819849, + 0.008414257317781448, + 0.016818348318338394, + 0.011528012342751026, + 0.006783188320696354, + 0.013200489804148674, + 0.01016598753631115, + 0.010106218978762627, + 0.035282183438539505, + 0.0074735889211297035, + 0.024817898869514465, + 0.016054188832640648, + 0.00918214675039053, + 0.0259214099496603, + 0.0149405961856246, + 0.011446597054600716, + 0.015740158036351204, + 0.013990487903356552, + 0.01790720783174038, + 0.006477671675384045, + 0.003515085903927684, + 0.007698277942836285, + 0.005224499851465225, + 0.00995600875467062, + 0.01562768779695034, + 0.011042342521250248, + 0.010548656806349754, + 0.0058173565194010735, + 0.0026494250632822514, + 0.0022768727503716946, + 0.002316063502803445, + 0.0018166153458878398, + 0.004332576412707567, + 0.0032481218222528696, + 0.0070512196980416775, + 0.0039008010644465685, + 0.0035363149363547564, + 0.004240319132804871, + 0.004323763772845268, + 0.0032174091320484877, + 0.003004521830007434, + 0.00335480528883636, + 0.0030378468800336123, + 0.03409652039408684, + 0.037700388580560684, + 0.010252092964947224, + 0.018672261387109756, + 0.0168532133102417, + 0.017110593616962433, + 0.010569723322987556, + 0.04467164725065231, + 0.0026311019901186228, + 0.003173300065100193, + 0.0018344904528930783, + 0.0025683888234198093, + 0.006200851872563362, + 0.005952785257250071, + 0.002583067398518324, + 0.005068246275186539 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 4 + ], + "selected_reward": -0.8983287811279297 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "No" + ] + }, + { + "prompt": "high-speed, sports documentary-style close-up plunges the viewer directly into the heart of a fierce battle between two enormous hawkmoths in the deep night sky. The camera follows closely in a tense handheld style, capturing this aerial duel from a slightly upward angle. Backlit by the brilliant full moon, their enormous, nearly translucent emerald green wings, adorned with delicate lavender eyespots, flap with primal power, their furry white bodies tumbling through the air at high speed. They swirl and dive at each other with breathtaking speed and force, their emerald green wings clashing violently. Each collision sends countless tiny, shimmering scales scattering like stardust in the moonlight, triggering erratic camera shake and imbuing the image with the ultimate tension of action photography. This violent vibration, combined with the deep blue night sky, blurred by the high-speed motion, perfectly recreates the thrill of witnessing this aerial duel live, creating a visual spectacle both chaotic and magnificent.", + "seed": 22368, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000056/group_6/index_5111_seed_22368_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": -0.1400365931680426, + "vlm_logprob": -0.1400365931680426, + "vlm_yes_no_margin": 4.15, + "vlm_token_credit": -0.1400365931680426, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" 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 enormous hawkmoths visible engaging in a battle or aerial duel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the hawkmoths have nearly translucent emerald green wings with delicate lavender eyespots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the hawkmoths' bodies furry and white while they tumble through the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set in a deep night sky backlit by a brilliant full moon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot handheld with a slightly upward angle and erratic shake?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05550004169344902 + ], + "opposite_logprob_tokens": [ + -2.930500030517578 + ], + "logprob_reward": -0.05550004169344902, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05550004169344902, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2878423035144806, + 0.09281820058822632, + 0.05210987105965614, + 0.08046090602874756, + 0.031957000494003296, + 0.057744380086660385, + 0.10630625486373901, + 0.05522817745804787, + 0.11620486527681351, + 0.119328074157238 + ], + "credit_map": [ + 0.010606779716908932, + 0.02177395112812519, + 0.016003146767616272, + 0.008911271579563618, + 0.10089066624641418, + 0.09856013208627701, + 0.011420930735766888, + 0.01967542991042137, + 0.013756325468420982, + 0.01001609954982996, + 0.02021898701786995, + 0.009689079597592354, + 0.006491378881037235, + 0.01007020752876997, + 0.011579698882997036, + 0.010996419936418533, + 0.006591279525309801, + 0.0037215568590909243, + 0.008630738593637943, + 0.00472507206723094, + 0.004156390205025673, + 0.007572985719889402, + 0.004087409935891628, + 0.012624435126781464, + 0.008065947331488132, + 0.00840556900948286, + 0.010816957801580429, + 0.008193528279662132, + 0.012667596340179443, + 0.011281965300440788, + 0.011730176396667957, + 0.009299159049987793, + 0.00343089597299695, + 0.0021819944959133863, + 0.0018768531735986471, + 0.0028388865757733583, + 0.00262325257062912, + 0.008479391224682331, + 0.003335274988785386, + 0.007190452888607979, + 0.004476602654904127, + 0.008684509433805943, + 0.006617995910346508, + 0.007371419575065374, + 0.0076329829171299934, + 0.00827854871749878, + 0.006679089739918709, + 0.008003228344023228, + 0.01252462062984705, + 0.013537682592868805, + 0.012427751906216145, + 0.014777670614421368, + 0.010647900402545929, + 0.014557724818587303, + 0.015095454640686512, + 0.01273744460195303, + 0.0044539980590343475, + 0.00726191233843565, + 0.0065347496420145035, + 0.00635022297501564, + 0.007847780361771584, + 0.007473581936210394, + 0.008755608461797237, + 0.006550326012074947, + 0.013269736431539059, + 0.021483231335878372, + 0.015512369573116302, + 0.012950257398188114, + 0.011832893826067448, + 0.012730225920677185, + 0.012503004632890224, + 0.015923144295811653, + 0.014540696516633034, + 0.011205222457647324, + 0.019402606412768364, + 0.012606318108737469, + 0.015282728709280491, + 0.015109367668628693, + 0.01262175664305687, + 0.018559379503130913 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 8 + ], + "selected_reward": -0.05550004169344902 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" 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 enormous hawkmoths visible engaging in a battle or aerial duel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the hawkmoths have nearly translucent emerald green wings with delicate lavender eyespots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the hawkmoths' bodies furry and white while they tumble through the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set in a deep night sky backlit by a brilliant full moon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot handheld with a slightly upward angle and erratic shake?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002202229807153344 + ], + "opposite_logprob_tokens": [ + -6.127202033996582 + ], + "logprob_reward": -0.002202229807153344, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.002202229807153344, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.6414616107940674, + 0.039887938648462296, + 0.024808015674352646, + 0.036882489919662476, + 0.02270200103521347, + 0.022599125280976295, + 0.06271543353796005, + 0.06593602150678635, + 0.022135548293590546, + 0.060871776193380356 + ], + "credit_map": [ + 0.015149211511015892, + 0.05319635942578316, + 0.028514040634036064, + 0.013256873935461044, + 0.2451983541250229, + 0.2318296879529953, + 0.011881724931299686, + 0.04243540018796921, + 0.006475906353443861, + 0.0036654500290751457, + 0.008423002436757088, + 0.004298695828765631, + 0.002974182367324829, + 0.004337611608207226, + 0.004675653763115406, + 0.005037440452724695, + 0.002183095086365938, + 0.002921090228483081, + 0.004366246052086353, + 0.0020936436485499144, + 0.003633695188909769, + 0.002920212224125862, + 0.0027529869694262743, + 0.003937046509236097, + 0.0038192349020391703, + 0.0036756088957190514, + 0.002859955420717597, + 0.0040449476800858974, + 0.005509713664650917, + 0.0055218469351530075, + 0.0069997552782297134, + 0.004451428074389696, + 0.0022555117029696703, + 0.0018496094271540642, + 0.0014960210537537932, + 0.0014133007498458028, + 0.0028095783200114965, + 0.0054580108262598515, + 0.003579084062948823, + 0.003840887686237693, + 0.0021299608051776886, + 0.0025219444651156664, + 0.0026497014332562685, + 0.0021418300457298756, + 0.004086583852767944, + 0.002805206459015608, + 0.0029973925556987524, + 0.00326650682836771, + 0.005979685112833977, + 0.005094023421406746, + 0.006676974706351757, + 0.0077782971784472466, + 0.008608882315456867, + 0.012268131598830223, + 0.007558049634099007, + 0.008751392364501953, + 0.003798425430431962, + 0.00834302231669426, + 0.004784774035215378, + 0.0061200689524412155, + 0.010500031523406506, + 0.012397862039506435, + 0.011683296412229538, + 0.008308547548949718, + 0.002186109544709325, + 0.002618612954393029, + 0.0029827638063579798, + 0.0030649248510599136, + 0.0026763086207211018, + 0.0025149621069431305, + 0.00324888713657856, + 0.00284298206679523, + 0.00687397550791502, + 0.005816079676151276, + 0.010061272419989109, + 0.007092702202498913, + 0.006347070913761854, + 0.010415427386760712, + 0.004792116116732359, + 0.009473135694861412 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 6 + ], + "selected_reward": -0.002202229807153344 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" 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 enormous hawkmoths visible engaging in a battle or aerial duel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the hawkmoths have nearly translucent emerald green wings with delicate lavender eyespots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the hawkmoths' bodies furry and white while they tumble through the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set in a deep night sky backlit by a brilliant full moon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot handheld with a slightly upward angle and erratic shake?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008636858314275742 + ], + "opposite_logprob_tokens": [ + -4.758636951446533 + ], + "logprob_reward": -0.008636858314275742, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008636858314275742, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.7078609466552734, + 0.032986246049404144, + 0.02265419065952301, + 0.04788438230752945, + 0.007566997781395912, + 0.018433352932333946, + 0.05144773796200752, + 0.028918277472257614, + 0.02069593034684658, + 0.0615520253777504 + ], + "credit_map": [ + 0.018050899729132652, + 0.05991138145327568, + 0.03064928762614727, + 0.01489349640905857, + 0.27031609416007996, + 0.254132479429245, + 0.013056991621851921, + 0.04685027897357941, + 0.004885857924818993, + 0.003367160214111209, + 0.005897556897252798, + 0.004580986686050892, + 0.0028425799682736397, + 0.003651662031188607, + 0.003472814103588462, + 0.004287628922611475, + 0.0025928597897291183, + 0.0025208331644535065, + 0.003178358543664217, + 0.002210051752626896, + 0.0026868118438869715, + 0.0030434057116508484, + 0.0023368678521364927, + 0.0040850029326975346, + 0.004775810521095991, + 0.004853752441704273, + 0.005156876519322395, + 0.005688822828233242, + 0.006967182271182537, + 0.00709433201700449, + 0.0075058783404529095, + 0.005841725505888462, + 0.0008864825940690935, + 0.0007904625381343067, + 0.0005124256713315845, + 0.0006171806016936898, + 0.001064326148480177, + 0.0011419174261391163, + 0.0012043107999488711, + 0.0013498921180143952, + 0.0017741855699568987, + 0.0021062346640974283, + 0.002702573547139764, + 0.0018860403215512633, + 0.0025902206543833017, + 0.0023365477100014687, + 0.0025720209814608097, + 0.0024655291344970465, + 0.00429044384509325, + 0.003143470734357834, + 0.0044156587682664394, + 0.0059387716464698315, + 0.006911954376846552, + 0.012853384017944336, + 0.005232205614447594, + 0.008661848492920399, + 0.00217699003405869, + 0.004145667422562838, + 0.0027358729857951403, + 0.002883905079215765, + 0.005623499862849712, + 0.00428753811866045, + 0.0032652844674885273, + 0.0037995176389813423, + 0.0017402638914063573, + 0.001996906939893961, + 0.002528166864067316, + 0.00247835461050272, + 0.0029905347619205713, + 0.0032637284602969885, + 0.0022310034837573767, + 0.003466971917077899, + 0.006162639707326889, + 0.00477913161739707, + 0.007623943500220776, + 0.005753031000494957, + 0.005494193639606237, + 0.012857270427048206, + 0.005391428247094154, + 0.013490389101207256 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 6 + ], + "selected_reward": -0.008636858314275742 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" 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 enormous hawkmoths visible engaging in a battle or aerial duel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the hawkmoths have nearly translucent emerald green wings with delicate lavender eyespots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the hawkmoths' bodies furry and white while they tumble through the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set in a deep night sky backlit by a brilliant full moon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot handheld with a slightly upward angle and erratic shake?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0010386077919974923 + ], + "opposite_logprob_tokens": [ + -6.876038551330566 + ], + "logprob_reward": -0.0010386077919974923, + "yes_no_margin_reward": 6.875, + "token_credit_reward": -0.0010386077919974923, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.7718868255615234, + 0.023381734266877174, + 0.02093557082116604, + 0.012717628851532936, + 0.021062534302473068, + 0.023977559059858322, + 0.028594065457582474, + 0.029939254745841026, + 0.028014814481139183, + 0.0394899882376194 + ], + "credit_map": [ + 0.018474219366908073, + 0.06318912655115128, + 0.034591663628816605, + 0.016013333573937416, + 0.2944628894329071, + 0.2801068425178528, + 0.013419017195701599, + 0.05162978917360306, + 0.003903420176357031, + 0.00261111231520772, + 0.004556676838546991, + 0.0027118597645312548, + 0.0017676957650110126, + 0.0024264357052743435, + 0.002738762181252241, + 0.0026657700072973967, + 0.002875703852623701, + 0.0016945342067629099, + 0.0033711164724081755, + 0.0021348667796701193, + 0.001968582160770893, + 0.002989489585161209, + 0.0023743947967886925, + 0.003526881569996476, + 0.001196894096210599, + 0.001253457274287939, + 0.001338692964054644, + 0.0025086039677262306, + 0.0013108887942507863, + 0.0018408955074846745, + 0.0017487731529399753, + 0.0015194222796708345, + 0.0021624360233545303, + 0.0015872721560299397, + 0.0015507438220083714, + 0.001942198839969933, + 0.0021538042929023504, + 0.004298071376979351, + 0.0024618888273835182, + 0.004906116984784603, + 0.0019198713125661016, + 0.0029681394807994366, + 0.0024441173300147057, + 0.0027033232618123293, + 0.002635227283462882, + 0.003703974885866046, + 0.00258054630830884, + 0.0050223576836287975, + 0.0029313804116100073, + 0.0036061517894268036, + 0.0027082546148449183, + 0.003791817696765065, + 0.0031480586621910334, + 0.005075658671557903, + 0.0035967931617051363, + 0.0037359485868364573, + 0.0017995699308812618, + 0.002762336516752839, + 0.0026982268318533897, + 0.002874326892197132, + 0.004359651822596788, + 0.0040182555094361305, + 0.007271280512213707, + 0.004155606962740421, + 0.0032182319555431604, + 0.003970933146774769, + 0.003532037604600191, + 0.003208802081644535, + 0.003160605439916253, + 0.0030609420500695705, + 0.004051665309816599, + 0.003811595495790243, + 0.005141675006598234, + 0.004012915771454573, + 0.007978546433150768, + 0.005313266068696976, + 0.004879946820437908, + 0.004827518481761217, + 0.0033027322497218847, + 0.004033385775983334 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 7 + ], + "selected_reward": -0.0010386077919974923 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" 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 enormous hawkmoths visible engaging in a battle or aerial duel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the hawkmoths have nearly translucent emerald green wings with delicate lavender eyespots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the hawkmoths' bodies furry and white while they tumble through the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set in a deep night sky backlit by a brilliant full moon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot handheld with a slightly upward angle and erratic shake?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.6328052282333374 + ], + "opposite_logprob_tokens": [ + -0.7578052282333374 + ], + "logprob_reward": -0.6328052282333374, + "yes_no_margin_reward": 0.125, + "token_credit_reward": -0.6328052282333374, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.7112876176834106, + 0.022676007822155952, + 0.01343569252640009, + 0.016376297920942307, + 0.020122356712818146, + 0.03025614097714424, + 0.030257070437073708, + 0.041957393288612366, + 0.05368203669786453, + 0.05994939059019089 + ], + "credit_map": [ + 0.01825161837041378, + 0.05823901295661926, + 0.03283444046974182, + 0.015500888228416443, + 0.26915833353996277, + 0.25688427686691284, + 0.012916306033730507, + 0.047502703964710236, + 0.0036239998880773783, + 0.002140907570719719, + 0.003239593468606472, + 0.0031992930453270674, + 0.0019324845634400845, + 0.002697179326787591, + 0.003048412036150694, + 0.002794133499264717, + 0.0015645042294636369, + 0.0009409281192347407, + 0.0016721715219318867, + 0.0016022082418203354, + 0.0012236674083396792, + 0.0020952934864908457, + 0.0011744999792426825, + 0.003162418259307742, + 0.0019171778112649918, + 0.0021495013497769833, + 0.0021306141279637814, + 0.002010592957958579, + 0.0023724164348095655, + 0.0022574898321181536, + 0.0016037110472097993, + 0.0019347916822880507, + 0.002284270478412509, + 0.0021988607477396727, + 0.0015915523981675506, + 0.0017694543348625302, + 0.003602362237870693, + 0.003914384171366692, + 0.0021584758069366217, + 0.0026029935106635094, + 0.002645623404532671, + 0.0040115537121891975, + 0.002758615417405963, + 0.003931592218577862, + 0.004300787579268217, + 0.004530478268861771, + 0.003595795016735792, + 0.004481690935790539, + 0.002149108098819852, + 0.002167820930480957, + 0.002614818513393402, + 0.003031311556696892, + 0.004299648106098175, + 0.007247054483741522, + 0.004013087600469589, + 0.0047342171892523766, + 0.003767050802707672, + 0.004381013568490744, + 0.005041003227233887, + 0.0032795637380331755, + 0.0046080248430371284, + 0.007938682101666927, + 0.00530122360214591, + 0.007640827912837267, + 0.004493528511375189, + 0.005862112622708082, + 0.004119278397411108, + 0.005954338237643242, + 0.007006583269685507, + 0.009373227134346962, + 0.006156743969768286, + 0.010716219432651997, + 0.0059912377037107944, + 0.006391874048858881, + 0.008456959389150143, + 0.007872791029512882, + 0.01052876003086567, + 0.006167378276586533, + 0.006492780987173319, + 0.008047602139413357 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 8 + ], + "selected_reward": -0.6328052282333374 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "high-speed, sports documentary-style close-up plunges the viewer directly into the heart of a fierce battle between two enormous hawkmoths in the deep night sky. The camera follows closely in a tense handheld style, capturing this aerial duel from a slightly upward angle. Backlit by the brilliant full moon, their enormous, nearly translucent emerald green wings, adorned with delicate lavender eyespots, flap with primal power, their furry white bodies tumbling through the air at high speed. They swirl and dive at each other with breathtaking speed and force, their emerald green wings clashing violently. Each collision sends countless tiny, shimmering scales scattering like stardust in the moonlight, triggering erratic camera shake and imbuing the image with the ultimate tension of action photography. This violent vibration, combined with the deep blue night sky, blurred by the high-speed motion, perfectly recreates the thrill of witnessing this aerial duel live, creating a visual spectacle both chaotic and magnificent.", + "seed": 22369, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000056/group_6/index_5111_seed_22369_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": -0.09311057161539793, + "vlm_logprob": -0.09311057161539793, + "vlm_yes_no_margin": 3.425, + "vlm_token_credit": -0.09311057161539793, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" 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 enormous hawkmoths visible engaging in a battle or aerial duel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the hawkmoths have nearly translucent emerald green wings with delicate lavender eyespots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the hawkmoths' bodies furry and white while they tumble through the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set in a deep night sky backlit by a brilliant full moon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot handheld with a slightly upward angle and erratic shake?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1140654906630516 + ], + "opposite_logprob_tokens": [ + -2.239065408706665 + ], + "logprob_reward": -0.1140654906630516, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.1140654906630516, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.35356780886650085, + 0.06775838136672974, + 0.047039106488227844, + 0.03664064034819603, + 0.07738824188709259, + 0.026471354067325592, + 0.18495388329029083, + 0.10481560230255127, + 0.02995581552386284, + 0.07140909880399704 + ], + "credit_map": [ + 0.06880954653024673, + 0.03929220139980316, + 0.04308972880244255, + 0.035717032849788666, + 0.04032212868332863, + 0.03751102834939957, + 0.02920151874423027, + 0.059624623507261276, + 0.008388499729335308, + 0.009294245392084122, + 0.007263473700731993, + 0.006174254696816206, + 0.008824174292385578, + 0.011096484027802944, + 0.0067016868852078915, + 0.010015568695962429, + 0.005478279665112495, + 0.006623770575970411, + 0.005549498368054628, + 0.005145295523107052, + 0.007236539851874113, + 0.006366748828440905, + 0.005875757895410061, + 0.004763220902532339, + 0.006817166693508625, + 0.00495896628126502, + 0.0040591866709291935, + 0.009223873727023602, + 0.0028451462276279926, + 0.0030336356721818447, + 0.00258621945977211, + 0.0031164479441940784, + 0.012495731003582478, + 0.007325173821300268, + 0.01675248146057129, + 0.013812188059091568, + 0.006144663784652948, + 0.006589830387383699, + 0.007773840334266424, + 0.006494341883808374, + 0.0026551636401563883, + 0.003702304093167186, + 0.0035158940590918064, + 0.003495368640869856, + 0.00224063778296113, + 0.0027518162969499826, + 0.002983513055369258, + 0.005126659292727709, + 0.020937679335474968, + 0.024736428633332253, + 0.025867249816656113, + 0.017202775925397873, + 0.02527935802936554, + 0.0228351391851902, + 0.027083178982138634, + 0.0210120789706707, + 0.009210106916725636, + 0.012214197777211666, + 0.008892927318811417, + 0.005151103250682354, + 0.02467465214431286, + 0.016125274822115898, + 0.012611827813088894, + 0.015935523435473442, + 0.0022978319320827723, + 0.0036628544330596924, + 0.005509583745151758, + 0.0032199551351368427, + 0.0030652140267193317, + 0.003934214822947979, + 0.005224276799708605, + 0.0030418892856687307, + 0.006700706668198109, + 0.007633213885128498, + 0.008887751027941704, + 0.0065127224661409855, + 0.017822057008743286, + 0.008533261716365814, + 0.009758919477462769, + 0.005560469813644886 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 7 + ], + "selected_reward": -0.1140654906630516 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" 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 enormous hawkmoths visible engaging in a battle or aerial duel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the hawkmoths have nearly translucent emerald green wings with delicate lavender eyespots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the hawkmoths' bodies furry and white while they tumble through the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set in a deep night sky backlit by a brilliant full moon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot handheld with a slightly upward angle and erratic shake?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004101318307220936 + ], + "opposite_logprob_tokens": [ + -5.504101276397705 + ], + "logprob_reward": -0.004101318307220936, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004101318307220936, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.37103140354156494, + 0.10042709112167358, + 0.06476429104804993, + 0.0720611959695816, + 0.0442231260240078, + 0.023791667073965073, + 0.10582956671714783, + 0.06174621731042862, + 0.06689225882291794, + 0.08923322707414627 + ], + "credit_map": [ + 0.07698436826467514, + 0.04519875347614288, + 0.04479483142495155, + 0.035181839019060135, + 0.03593029826879501, + 0.04035823419690132, + 0.02968239039182663, + 0.06290066242218018, + 0.014485357329249382, + 0.009390158578753471, + 0.013500142842531204, + 0.017828643321990967, + 0.00996474176645279, + 0.011892592534422874, + 0.010048089548945427, + 0.013317348435521126, + 0.011082611046731472, + 0.00954578910022974, + 0.009482537396252155, + 0.0075639947317540646, + 0.007572039030492306, + 0.007024411112070084, + 0.007145681418478489, + 0.005347219295799732, + 0.010888813994824886, + 0.008468394167721272, + 0.006837738212198019, + 0.012600341811776161, + 0.007025631610304117, + 0.006828690879046917, + 0.011947864666581154, + 0.007463707588613033, + 0.0052528297528624535, + 0.004313664510846138, + 0.00813198834657669, + 0.0077138179913163185, + 0.004295724909752607, + 0.004899036139249802, + 0.0049305749125778675, + 0.0046854871325194836, + 0.003363363677635789, + 0.0024945882614701986, + 0.004542223177850246, + 0.005218869540840387, + 0.0016465806402266026, + 0.0021434053778648376, + 0.002039413433521986, + 0.002343221567571163, + 0.011552207171916962, + 0.011357818730175495, + 0.012507695704698563, + 0.013619585894048214, + 0.015739714726805687, + 0.013465192168951035, + 0.015078595839440823, + 0.012508745305240154, + 0.005516854580491781, + 0.0075455415062606335, + 0.007762385066598654, + 0.005838396493345499, + 0.00866007525473833, + 0.010399793274700642, + 0.00584018137305975, + 0.010182982310652733, + 0.005038781091570854, + 0.007224957924336195, + 0.007747113239020109, + 0.007336855866014957, + 0.007426250725984573, + 0.009969468228518963, + 0.01309403870254755, + 0.009054785594344139, + 0.005139464512467384, + 0.0070923506282269955, + 0.003720060223713517, + 0.00778593122959137, + 0.030521607026457787, + 0.0113676767796278, + 0.01809258572757244, + 0.0055135441944003105 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 1 + ], + "selected_reward": -0.004101318307220936 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" 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 enormous hawkmoths visible engaging in a battle or aerial duel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the hawkmoths have nearly translucent emerald green wings with delicate lavender eyespots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the hawkmoths' bodies furry and white while they tumble through the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set in a deep night sky backlit by a brilliant full moon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot handheld with a slightly upward angle and erratic shake?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007628239691257477 + ], + "opposite_logprob_tokens": [ + -4.882628440856934 + ], + "logprob_reward": -0.007628239691257477, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007628239691257477, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.42429879307746887, + 0.09871197491884232, + 0.1002890095114708, + 0.04567612335085869, + 0.11231781542301178, + 0.019025985151529312, + 0.0860162004828453, + 0.02452123910188675, + 0.020142795518040657, + 0.06900004297494888 + ], + "credit_map": [ + 0.08341298997402191, + 0.043946992605924606, + 0.05151443928480148, + 0.04346261918544769, + 0.04418428614735603, + 0.04753103107213974, + 0.03559042513370514, + 0.07465600967407227, + 0.010427868925035, + 0.00918962899595499, + 0.012000138871371746, + 0.02169485203921795, + 0.013247702270746231, + 0.010111605748534203, + 0.010164477862417698, + 0.011875717900693417, + 0.015888040885329247, + 0.01400065328925848, + 0.013659445568919182, + 0.011673753149807453, + 0.01313933078199625, + 0.011284307576715946, + 0.011387183330953121, + 0.009256301447749138, + 0.007373429369181395, + 0.0051432205364108086, + 0.006421415600925684, + 0.011109347455203533, + 0.0029294295236468315, + 0.00521001685410738, + 0.0036021084524691105, + 0.003887159749865532, + 0.011546860449016094, + 0.013560096733272076, + 0.01614517718553543, + 0.019532063975930214, + 0.010141870938241482, + 0.016362328082323074, + 0.01353982649743557, + 0.011489602737128735, + 0.002357789548113942, + 0.0023035143967717886, + 0.003036766778677702, + 0.004458420444279909, + 0.0016855786088854074, + 0.0009558339370414615, + 0.0014742345083504915, + 0.0027538493741303682, + 0.00954812578856945, + 0.010253500193357468, + 0.011939187534153461, + 0.008289534598588943, + 0.013195816427469254, + 0.0106351962313056, + 0.010075744241476059, + 0.01207910105586052, + 0.002379308920353651, + 0.003072191961109638, + 0.002635051729157567, + 0.0025484722573310137, + 0.0041155521757900715, + 0.0030217620078474283, + 0.0035603994037956, + 0.0031885020434856415, + 0.001999081578105688, + 0.0024615342263132334, + 0.003616953734308481, + 0.002484754426404834, + 0.0021106915082782507, + 0.0029631422366946936, + 0.0024368581362068653, + 0.0020697820000350475, + 0.004556475207209587, + 0.006158818956464529, + 0.0033543661702424288, + 0.006223619915544987, + 0.02219134010374546, + 0.009387491270899773, + 0.012068454176187515, + 0.005059477873146534 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 2 + ], + "selected_reward": -0.007628239691257477 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" 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 enormous hawkmoths visible engaging in a battle or aerial duel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the hawkmoths have nearly translucent emerald green wings with delicate lavender eyespots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the hawkmoths' bodies furry and white while they tumble through the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set in a deep night sky backlit by a brilliant full moon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot handheld with a slightly upward angle and erratic shake?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026335397735238075 + ], + "opposite_logprob_tokens": [ + -3.6513354778289795 + ], + "logprob_reward": -0.026335397735238075, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026335397735238075, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3396834135055542, + 0.02840064838528633, + 0.02866045944392681, + 0.05359701067209244, + 0.021018467843532562, + 0.05195506289601326, + 0.09123703092336655, + 0.07321631163358688, + 0.07785824686288834, + 0.23437337577342987 + ], + "credit_map": [ + 0.06923190504312515, + 0.038788098841905594, + 0.041775405406951904, + 0.03222694620490074, + 0.03653363510966301, + 0.03603598475456238, + 0.027817009016871452, + 0.05727445334196091, + 0.002974495990201831, + 0.003421642119064927, + 0.0032592499628663063, + 0.003940520342439413, + 0.0041819484904408455, + 0.003774132812395692, + 0.0032196445390582085, + 0.0036290138959884644, + 0.00323511753231287, + 0.004655174445360899, + 0.0033290379215031862, + 0.0027259215712547302, + 0.0041165826842188835, + 0.004300678614526987, + 0.00361252692528069, + 0.002685421844944358, + 0.008310834877192974, + 0.006286026444286108, + 0.004747325088828802, + 0.009946601465344429, + 0.0050654723308980465, + 0.0048335157334804535, + 0.00881230365484953, + 0.0055949329398572445, + 0.0022853389382362366, + 0.0026175258681178093, + 0.0029755982104688883, + 0.00497872568666935, + 0.0020631528459489346, + 0.0020933477208018303, + 0.0022350111976265907, + 0.001769765978679061, + 0.0069825551472604275, + 0.004106590989977121, + 0.009891909547150135, + 0.015199477784335613, + 0.004553352948278189, + 0.002586512826383114, + 0.0037367986515164375, + 0.00489786546677351, + 0.012765947729349136, + 0.010874096304178238, + 0.016717638820409775, + 0.010509210638701916, + 0.011726990342140198, + 0.00983381737023592, + 0.008199276402592659, + 0.010610055178403854, + 0.00752183236181736, + 0.0074895769357681274, + 0.01292092353105545, + 0.0118367625400424, + 0.007536537013947964, + 0.007586998865008354, + 0.007883776910603046, + 0.01043990533798933, + 0.005433289799839258, + 0.008239665068686008, + 0.006131493486464024, + 0.00774212135002017, + 0.0077383508905768394, + 0.012853289023041725, + 0.018927540630102158, + 0.010792496614158154, + 0.01736348494887352, + 0.017806287854909897, + 0.013243213295936584, + 0.01803623139858246, + 0.07990499585866928, + 0.03183072805404663, + 0.04092713072896004, + 0.01526130922138691 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 6 + ], + "selected_reward": -0.026335397735238075 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" 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 enormous hawkmoths visible engaging in a battle or aerial duel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the hawkmoths have nearly translucent emerald green wings with delicate lavender eyespots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the hawkmoths' bodies furry and white while they tumble through the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set in a deep night sky backlit by a brilliant full moon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot handheld with a slightly upward angle and erratic shake?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.31342241168022156 + ], + "opposite_logprob_tokens": [ + -1.313422441482544 + ], + "logprob_reward": -0.31342241168022156, + "yes_no_margin_reward": 1.0, + "token_credit_reward": -0.31342241168022156, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2751382887363434, + 0.20663991570472717, + 0.04748326539993286, + 0.09012506902217865, + 0.02303122542798519, + 0.027115484699606895, + 0.07020489871501923, + 0.045335572212934494, + 0.06625992804765701, + 0.14866632223129272 + ], + "credit_map": [ + 0.056096360087394714, + 0.03198365867137909, + 0.03370492905378342, + 0.028750216588377953, + 0.030796190723776817, + 0.027986325323581696, + 0.022233637049794197, + 0.04358698055148125, + 0.0273990910500288, + 0.023082202300429344, + 0.02487407624721527, + 0.03129735589027405, + 0.02547566220164299, + 0.02858945168554783, + 0.019477615132927895, + 0.026444464921951294, + 0.006836879998445511, + 0.007058465387672186, + 0.0060193189419806, + 0.005505724810063839, + 0.006016452796757221, + 0.0060251569375395775, + 0.005190637428313494, + 0.004830632358789444, + 0.017913859337568283, + 0.008596752770245075, + 0.010259445756673813, + 0.02494031935930252, + 0.004944130778312683, + 0.005929839331656694, + 0.009299392811954021, + 0.008241343311965466, + 0.0027014666702598333, + 0.002178867580369115, + 0.0039625512436032295, + 0.004103570245206356, + 0.0025798596907407045, + 0.0024684169329702854, + 0.0028833176475018263, + 0.002153177745640278, + 0.0032760442700237036, + 0.0027811285108327866, + 0.005030567292124033, + 0.006200458854436874, + 0.0025085608940571547, + 0.0029772059060633183, + 0.002223290503025055, + 0.0021182324271649122, + 0.007248356007039547, + 0.008689068257808685, + 0.009877006523311138, + 0.007563195191323757, + 0.010091833770275116, + 0.008811754174530506, + 0.008227674290537834, + 0.009696015156805515, + 0.0038934610784053802, + 0.005148072727024555, + 0.006356487050652504, + 0.0037442625034600496, + 0.008567508310079575, + 0.005621962249279022, + 0.004686697851866484, + 0.007317123934626579, + 0.007341878954321146, + 0.009494045749306679, + 0.007868890650570393, + 0.007357117719948292, + 0.007497042417526245, + 0.00936779472976923, + 0.008237008936703205, + 0.009096152149140835, + 0.008367469534277916, + 0.010605670511722565, + 0.017317138612270355, + 0.012059471569955349, + 0.04545074701309204, + 0.020095329731702805, + 0.025213688611984253, + 0.009556828066706657 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 9 + ], + "selected_reward": -0.31342241168022156 + } + ], + "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/0000056/group_7/index_tensor([5754, 5754])_group_7_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000056/group_7/index_tensor([5754, 5754])_group_7_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..1d69691ebc6d1b2be679ec9d3f3daaf4f78a50bf --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000056/group_7/index_tensor([5754, 5754])_group_7_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "dramatic low-angle close-up, rendered in a gritty, dark anime style, captures the moment a sleek electric supercar violently takes off from a standstill. The camera remains static, focusing in breakneck slow motion on the rear tires. The car's chassis is an angular marvel of shiny black carbon fiber, slicked by rain and reflecting the distorted, vibrant glow of the fuchsia and cyan neon signs emanating from the surrounding urban canyons. As the launch begins, the electric motor's immense, silent torque visibly deforms the ultra-wide, high-performance tires, their sidewalls wrinkling under the furious force as they bite into the rain-soaked asphalt. A ferocious cocktail of water is sprayed backward, and raindrops on the street and in the air are atomized by the spinning wheels, suspended in the air for a fleeting moment like a cloud of scattered, shimmering gems. The scene is a masterclass in high-contrast lighting, with deep, inky shadows sharply defining the aggressive lines of the car, while intense reflected light from the rain-soaked environment glitters off wet surfaces and suspended water droplets, creating an atmosphere of raw power and breathtakingly violent beauty.", + "seed": 25259, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000056/group_7/index_5754_seed_25259_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": -0.11919346177019179, + "vlm_logprob": -0.11919346177019179, + "vlm_yes_no_margin": 3.75, + "vlm_token_credit": -0.11919346177019179, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a gritty, dark anime style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene a low-angle close-up of a sleek electric supercar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the car's chassis made of shiny black carbon fiber and wet with rain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera remain static while focusing on the rear tires in breakneck slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the rear tires visibly deformed with wrinkling sidewalls due to the launch force?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003880234668031335 + ], + "opposite_logprob_tokens": [ + -6.253880023956299 + ], + "logprob_reward": -0.003880234668031335, + "yes_no_margin_reward": 6.25, + "token_credit_reward": -0.003880234668031335, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19202159345149994, + 0.09999216347932816, + 0.05905506759881973, + 0.15632575750350952, + 0.13650773465633392, + 0.04366101697087288, + 0.044088784605264664, + 0.029574815183877945, + 0.13024461269378662, + 0.1085284948348999 + ], + "credit_map": [ + 0.01367661077529192, + 0.017706366255879402, + 0.03508688136935234, + 0.03754201903939247, + 0.019731389358639717, + 0.01992439106106758, + 0.029907379299402237, + 0.01844656467437744, + 0.010285447351634502, + 0.01070335041731596, + 0.01884043775498867, + 0.016000263392925262, + 0.006194557063281536, + 0.011118656024336815, + 0.014702074229717255, + 0.012147379107773304, + 0.007183169946074486, + 0.0066343736834824085, + 0.005855992436408997, + 0.00984104722738266, + 0.006396998651325703, + 0.004372915718704462, + 0.011934447102248669, + 0.00683612423017621, + 0.014955275692045689, + 0.02454984374344349, + 0.022659052163362503, + 0.02151643857359886, + 0.01109832338988781, + 0.01434993464499712, + 0.021119454875588417, + 0.02607743628323078, + 0.012372217141091824, + 0.010908183641731739, + 0.0170561745762825, + 0.02016596682369709, + 0.014979892410337925, + 0.009326531551778316, + 0.03575342148542404, + 0.015945345163345337, + 0.0037713770288974047, + 0.003616942558437586, + 0.004364850465208292, + 0.005700161214917898, + 0.005383524112403393, + 0.005071588326245546, + 0.008962180465459824, + 0.006790395360440016, + 0.007124399300664663, + 0.0061253877356648445, + 0.006264660973101854, + 0.005445342045277357, + 0.005088822916150093, + 0.0029305797070264816, + 0.005674087908118963, + 0.005435502622276545, + 0.004862677771598101, + 0.0032014972530305386, + 0.002992889378219843, + 0.003300965065136552, + 0.0037412000820040703, + 0.0028856294229626656, + 0.004558461252599955, + 0.004031492862850428, + 0.015338684432208538, + 0.010526208207011223, + 0.008174024522304535, + 0.014590222388505936, + 0.008081945590674877, + 0.008524812757968903, + 0.03185272216796875, + 0.03315598890185356, + 0.02853773720562458, + 0.01802457869052887, + 0.01169181615114212, + 0.007573947310447693, + 0.007885515689849854, + 0.009596578776836395, + 0.012631283141672611, + 0.01258703600615263 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 4 + ], + "selected_reward": -0.003880234668031335 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a gritty, dark anime style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene a low-angle close-up of a sleek electric supercar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the car's chassis made of shiny black carbon fiber and wet with rain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera remain static while focusing on the rear tires in breakneck slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the rear tires visibly deformed with wrinkling sidewalls due to the launch force?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005256996490061283 + ], + "opposite_logprob_tokens": [ + -5.2552571296691895 + ], + "logprob_reward": -0.005256996490061283, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005256996490061283, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23507501184940338, + 0.07382450997829437, + 0.09081124514341354, + 0.20092087984085083, + 0.11306530982255936, + 0.07643889635801315, + 0.03921187296509743, + 0.031321391463279724, + 0.03972100839018822, + 0.0996098592877388 + ], + "credit_map": [ + 0.010900422930717468, + 0.013863318599760532, + 0.06311704218387604, + 0.04341857135295868, + 0.018807265907526016, + 0.015104287303984165, + 0.04628661274909973, + 0.023577509447932243, + 0.0076526375487446785, + 0.00851411372423172, + 0.012714830227196217, + 0.012611518613994122, + 0.005063783377408981, + 0.007622848264873028, + 0.010102439671754837, + 0.00954233855009079, + 0.01204227190464735, + 0.013924251310527325, + 0.012349864467978477, + 0.013921736739575863, + 0.007933021523058414, + 0.008883684873580933, + 0.012655417434871197, + 0.00910099409520626, + 0.01667121797800064, + 0.01838599517941475, + 0.041616495698690414, + 0.027813933789730072, + 0.009516565129160881, + 0.017261071130633354, + 0.03200898692011833, + 0.03764662146568298, + 0.015336142852902412, + 0.008246378041803837, + 0.016368402168154716, + 0.013500959612429142, + 0.011950399726629257, + 0.006702655926346779, + 0.025760119780898094, + 0.01520024798810482, + 0.009834951721131802, + 0.009406830184161663, + 0.00784580409526825, + 0.0097947483882308, + 0.009168213233351707, + 0.008051391690969467, + 0.01111235935240984, + 0.011224601417779922, + 0.004468694794923067, + 0.004336806479841471, + 0.004991243593394756, + 0.005569111555814743, + 0.003405556082725525, + 0.0035794840659946203, + 0.0063995071686804295, + 0.0064614685252308846, + 0.005630777217447758, + 0.003584644291549921, + 0.0036043436266481876, + 0.005560615565627813, + 0.002648322144523263, + 0.001700200024060905, + 0.0047571477480232716, + 0.0038353390991687775, + 0.006287258584052324, + 0.0041654035449028015, + 0.004646149463951588, + 0.004999313969165087, + 0.0028001233004033566, + 0.0028383550234138966, + 0.007273592054843903, + 0.006710812449455261, + 0.014159353449940681, + 0.016470961272716522, + 0.010958408005535603, + 0.012963002547621727, + 0.00644387723878026, + 0.009438280016183853, + 0.014185509644448757, + 0.014990455470979214 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 4 + ], + "selected_reward": -0.005256996490061283 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a gritty, dark anime style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene a low-angle close-up of a sleek electric supercar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the car's chassis made of shiny black carbon fiber and wet with rain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera remain static while focusing on the rear tires in breakneck slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the rear tires visibly deformed with wrinkling sidewalls due to the launch force?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029820306226611137 + ], + "opposite_logprob_tokens": [ + -3.529820203781128 + ], + "logprob_reward": -0.029820306226611137, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029820306226611137, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16882041096687317, + 0.1718006432056427, + 0.0747007429599762, + 0.11598115414381027, + 0.15761394798755646, + 0.06746147572994232, + 0.02825501747429371, + 0.046766456216573715, + 0.09240029007196426, + 0.07619993388652802 + ], + "credit_map": [ + 0.011461920104920864, + 0.015852274373173714, + 0.01973055489361286, + 0.030445583164691925, + 0.02264442853629589, + 0.016944393515586853, + 0.028187353163957596, + 0.02355387806892395, + 0.01536699291318655, + 0.016452740877866745, + 0.034743860363960266, + 0.028722114861011505, + 0.012431121431291103, + 0.016547152772545815, + 0.025516850873827934, + 0.022019796073436737, + 0.008053858764469624, + 0.01055354904383421, + 0.010181029327213764, + 0.011657971888780594, + 0.006367845926433802, + 0.005915536545217037, + 0.010522331111133099, + 0.011448616161942482, + 0.010379713959991932, + 0.02120465226471424, + 0.013151013292372227, + 0.01451097521930933, + 0.009911762550473213, + 0.013881263323128223, + 0.012433941476047039, + 0.020507818087935448, + 0.011343955993652344, + 0.010330366902053356, + 0.025631001219153404, + 0.020469743758440018, + 0.013018958270549774, + 0.007532939314842224, + 0.04699939861893654, + 0.022287564352154732, + 0.004580385982990265, + 0.005553522612899542, + 0.006543387658894062, + 0.009276819415390491, + 0.007913910783827305, + 0.007701316848397255, + 0.014062723144888878, + 0.011829405091702938, + 0.003700795117765665, + 0.0026054345071315765, + 0.003390600671991706, + 0.003206344787031412, + 0.0040399762801826, + 0.0024344197008758783, + 0.004782136529684067, + 0.004095308482646942, + 0.004821085836738348, + 0.004554157145321369, + 0.003716158214956522, + 0.008305019699037075, + 0.0038868961855769157, + 0.0031382227316498756, + 0.009183417074382305, + 0.009161493740975857, + 0.011953633278608322, + 0.007957622408866882, + 0.006488424260169268, + 0.010738748125731945, + 0.0056457603350281715, + 0.008660019375383854, + 0.017870664596557617, + 0.02308540791273117, + 0.010538618080317974, + 0.011127517558634281, + 0.009487427771091461, + 0.010117476806044579, + 0.004864436108618975, + 0.006873958744108677, + 0.01212386041879654, + 0.011066632345318794 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 4 + ], + "selected_reward": -0.029820306226611137 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a gritty, dark anime style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene a low-angle close-up of a sleek electric supercar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the car's chassis made of shiny black carbon fiber and wet with rain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera remain static while focusing on the rear tires in breakneck slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the rear tires visibly deformed with wrinkling sidewalls due to the launch force?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.5233203172683716 + ], + "opposite_logprob_tokens": [ + -0.8983203172683716 + ], + "logprob_reward": -0.5233203172683716, + "yes_no_margin_reward": 0.375, + "token_credit_reward": -0.5233203172683716, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18046581745147705, + 0.044040847569704056, + 0.25092706084251404, + 0.12533539533615112, + 0.10965149104595184, + 0.0426466166973114, + 0.03937956690788269, + 0.01635715179145336, + 0.12175499647855759, + 0.06944103538990021 + ], + "credit_map": [ + 0.009858459234237671, + 0.011453408747911453, + 0.04319386184215546, + 0.03767478093504906, + 0.014873651787638664, + 0.011593823321163654, + 0.032417308539152145, + 0.01940052956342697, + 0.0048400587402284145, + 0.004985468927770853, + 0.007577972952276468, + 0.008157206699252129, + 0.0027393975760787725, + 0.004188396967947483, + 0.005502817686647177, + 0.006049530114978552, + 0.02786746621131897, + 0.02814929001033306, + 0.03566761687397957, + 0.034101713448762894, + 0.03162704035639763, + 0.02298348769545555, + 0.03572290763258934, + 0.03480755537748337, + 0.019814327359199524, + 0.015530575066804886, + 0.013097003102302551, + 0.016519485041499138, + 0.008613167330622673, + 0.013223130255937576, + 0.01777701824903488, + 0.020760688930749893, + 0.012902633287012577, + 0.005964204203337431, + 0.01419169269502163, + 0.014416642487049103, + 0.01175592839717865, + 0.006157072726637125, + 0.028253963217139244, + 0.016009358689188957, + 0.006533728912472725, + 0.004970756825059652, + 0.004438817035406828, + 0.0048135812394320965, + 0.005054349545389414, + 0.005295588634908199, + 0.005971438717097044, + 0.0055683571845293045, + 0.004154703114181757, + 0.003337965812534094, + 0.004010342992842197, + 0.007569863460958004, + 0.003491811454296112, + 0.004027160350233316, + 0.006681253667920828, + 0.006106465123593807, + 0.0031462358310818672, + 0.0022423702757805586, + 0.0020308890379965305, + 0.001744366716593504, + 0.0012750428868457675, + 0.0014967939350754023, + 0.0024674399755895138, + 0.0019540118519216776, + 0.0182112455368042, + 0.00919708888977766, + 0.007941189222037792, + 0.013014431111514568, + 0.0072310916148126125, + 0.007720508147031069, + 0.029032522812485695, + 0.029406912624835968, + 0.01622857339680195, + 0.00960345845669508, + 0.007697772234678268, + 0.00872099120169878, + 0.0034642862156033516, + 0.004582451190799475, + 0.009610767476260662, + 0.009532732889056206 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 3 + ], + "selected_reward": -0.5233203172683716 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a gritty, dark anime style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene a low-angle close-up of a sleek electric supercar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the car's chassis made of shiny black carbon fiber and wet with rain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera remain static while focusing on the rear tires in breakneck slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the rear tires visibly deformed with wrinkling sidewalls due to the launch force?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.033689454197883606 + ], + "opposite_logprob_tokens": [ + -3.408689498901367 + ], + "logprob_reward": -0.033689454197883606, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.033689454197883606, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16831955313682556, + 0.0714467242360115, + 0.15257735550403595, + 0.11474865674972534, + 0.09568149596452713, + 0.11060639470815659, + 0.03935634344816208, + 0.03437080979347229, + 0.1525692492723465, + 0.06032351776957512 + ], + "credit_map": [ + 0.014004470780491829, + 0.012744780629873276, + 0.02855026163160801, + 0.026032671332359314, + 0.01812644675374031, + 0.01863396354019642, + 0.025570986792445183, + 0.024655960500240326, + 0.005777386482805014, + 0.0073702638037502766, + 0.012926307506859303, + 0.01240505464375019, + 0.005937266629189253, + 0.006665542721748352, + 0.010261890478432178, + 0.01010300312191248, + 0.018944162875413895, + 0.023551546037197113, + 0.02474374510347843, + 0.022028645500540733, + 0.013833126053214073, + 0.010794770903885365, + 0.020794743672013283, + 0.017886599525809288, + 0.009319178760051727, + 0.012908784672617912, + 0.019631722941994667, + 0.021216273307800293, + 0.007647304330021143, + 0.010515284724533558, + 0.018772447481751442, + 0.014737645164132118, + 0.009056255221366882, + 0.005355737637728453, + 0.017336217686533928, + 0.012410328723490238, + 0.006940029561519623, + 0.004858683329075575, + 0.02839617244899273, + 0.01132806297391653, + 0.010214732959866524, + 0.00855980534106493, + 0.010754970833659172, + 0.014134128578007221, + 0.013183644972741604, + 0.010346010327339172, + 0.02466377057135105, + 0.01874931901693344, + 0.004429658874869347, + 0.00349198910407722, + 0.0036111767403781414, + 0.006233884487301111, + 0.004541146103292704, + 0.003221434773877263, + 0.007995197549462318, + 0.005831850226968527, + 0.0051279934123158455, + 0.003998382482677698, + 0.004772487096488476, + 0.003786602523177862, + 0.003296938259154558, + 0.003390601370483637, + 0.004989665001630783, + 0.005008137784898281, + 0.02270391769707203, + 0.012928702868521214, + 0.011999978683888912, + 0.018842769786715508, + 0.009157347492873669, + 0.0122285857796669, + 0.029750535264611244, + 0.0349574014544487, + 0.008909708820283413, + 0.00747885275632143, + 0.004624939989298582, + 0.007011057808995247, + 0.005482789129018784, + 0.010619872249662876, + 0.007568187080323696, + 0.008628102950751781 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 8 + ], + "selected_reward": -0.033689454197883606 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "dramatic low-angle close-up, rendered in a gritty, dark anime style, captures the moment a sleek electric supercar violently takes off from a standstill. The camera remains static, focusing in breakneck slow motion on the rear tires. The car's chassis is an angular marvel of shiny black carbon fiber, slicked by rain and reflecting the distorted, vibrant glow of the fuchsia and cyan neon signs emanating from the surrounding urban canyons. As the launch begins, the electric motor's immense, silent torque visibly deforms the ultra-wide, high-performance tires, their sidewalls wrinkling under the furious force as they bite into the rain-soaked asphalt. A ferocious cocktail of water is sprayed backward, and raindrops on the street and in the air are atomized by the spinning wheels, suspended in the air for a fleeting moment like a cloud of scattered, shimmering gems. The scene is a masterclass in high-contrast lighting, with deep, inky shadows sharply defining the aggressive lines of the car, while intense reflected light from the rain-soaked environment glitters off wet surfaces and suspended water droplets, creating an atmosphere of raw power and breathtakingly violent beauty.", + "seed": 25260, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000056/group_7/index_5754_seed_25260_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": -0.14511369559913873, + "vlm_logprob": -0.14511369559913873, + "vlm_yes_no_margin": 3.225, + "vlm_token_credit": -0.14511369559913873, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a gritty, dark anime style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene a low-angle close-up of a sleek electric supercar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the car's chassis made of shiny black carbon fiber and wet with rain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera remain static while focusing on the rear tires in breakneck slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the rear tires visibly deformed with wrinkling sidewalls due to the launch force?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008772051893174648 + ], + "opposite_logprob_tokens": [ + -5.133771896362305 + ], + "logprob_reward": -0.008772051893174648, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.008772051893174648, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09281344711780548, + 0.09286760538816452, + 0.05355405434966087, + 0.08745948225259781, + 0.06628705561161041, + 0.26649346947669983, + 0.07486896961927414, + 0.0712692141532898, + 0.08586297929286957, + 0.10852374136447906 + ], + "credit_map": [ + 0.011476941406726837, + 0.008133302442729473, + 0.012712882831692696, + 0.010198500007390976, + 0.010895510204136372, + 0.011789597570896149, + 0.014249025844037533, + 0.01335768960416317, + 0.009153079241514206, + 0.012658323161303997, + 0.006395014468580484, + 0.011056936345994473, + 0.010166538879275322, + 0.021671663969755173, + 0.00872074719518423, + 0.013045302592217922, + 0.0053501734510064125, + 0.005159415304660797, + 0.00465795723721385, + 0.008008656091988087, + 0.006761655211448669, + 0.008388153277337551, + 0.007988390512764454, + 0.007239654194563627, + 0.012549271807074547, + 0.012360801920294762, + 0.009853090159595013, + 0.008711555041372776, + 0.010705908760428429, + 0.009954586625099182, + 0.012859133072197437, + 0.010465133003890514, + 0.007298248820006847, + 0.008387476205825806, + 0.007825206965208054, + 0.007269490975886583, + 0.006341865286231041, + 0.011783095076680183, + 0.008445937186479568, + 0.008935734629631042, + 0.015728075057268143, + 0.027405768632888794, + 0.0335269533097744, + 0.04166131094098091, + 0.021505169570446014, + 0.03687329217791557, + 0.050352949649095535, + 0.03943994268774986, + 0.004374448210000992, + 0.005270449910312891, + 0.0057154810056090355, + 0.006480945274233818, + 0.015112229622900486, + 0.009542323648929596, + 0.020674336701631546, + 0.007698751986026764, + 0.006452331785112619, + 0.008135409094393253, + 0.009007267653942108, + 0.013765071518719196, + 0.006135293748229742, + 0.006074028089642525, + 0.006858421489596367, + 0.014841397292912006, + 0.01069901417940855, + 0.012738727033138275, + 0.012350678443908691, + 0.013810698874294758, + 0.007194663397967815, + 0.005298548843711615, + 0.005705151706933975, + 0.018065499141812325, + 0.023181892931461334, + 0.015429512597620487, + 0.020461920648813248, + 0.018911246210336685, + 0.010871702805161476, + 0.005747886374592781, + 0.008461411111056805, + 0.005458168685436249 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 9, + 1 + ], + "selected_reward": -0.008772051893174648 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a gritty, dark anime style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene a low-angle close-up of a sleek electric supercar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the car's chassis made of shiny black carbon fiber and wet with rain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera remain static while focusing on the rear tires in breakneck slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the rear tires visibly deformed with wrinkling sidewalls due to the launch force?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01607716642320156 + ], + "opposite_logprob_tokens": [ + -4.141077041625977 + ], + "logprob_reward": -0.01607716642320156, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.01607716642320156, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08353406935930252, + 0.21511739492416382, + 0.03487251326441765, + 0.2857397496700287, + 0.05066002905368805, + 0.11498378217220306, + 0.03608736768364906, + 0.04563267156481743, + 0.09221009910106659, + 0.041162315756082535 + ], + "credit_map": [ + 0.006084998603910208, + 0.006771480198949575, + 0.010966925881803036, + 0.008080760017037392, + 0.021390490233898163, + 0.009155644103884697, + 0.010474009439349174, + 0.010609759949147701, + 0.02009294368326664, + 0.02421775832772255, + 0.028460104018449783, + 0.023561326786875725, + 0.01979154348373413, + 0.029093146324157715, + 0.03756730630993843, + 0.03233326971530914, + 0.0030658869072794914, + 0.004403890576213598, + 0.0036179947201162577, + 0.0062512923032045364, + 0.004597892053425312, + 0.0041363658383488655, + 0.004335183184593916, + 0.004464009776711464, + 0.026283929124474525, + 0.02455269545316696, + 0.02488217130303383, + 0.02367485873401165, + 0.029187554493546486, + 0.07041215896606445, + 0.0423574335873127, + 0.04438893869519234, + 0.004295946564525366, + 0.004426648840308189, + 0.007109957747161388, + 0.006661235820502043, + 0.004935049917548895, + 0.008718027733266354, + 0.00633175577968359, + 0.008181408978998661, + 0.007398105692118406, + 0.014808230102062225, + 0.01345792692154646, + 0.02274264395236969, + 0.008451992645859718, + 0.01990809477865696, + 0.012817302718758583, + 0.015399484895169735, + 0.0032903419341892004, + 0.00210502534173429, + 0.002376770367845893, + 0.003674293402582407, + 0.006942513398826122, + 0.003111388999968767, + 0.01154722273349762, + 0.0030398142989724874, + 0.0026826513931155205, + 0.004562159534543753, + 0.0030223592184484005, + 0.009205267764627934, + 0.005274809431284666, + 0.004911965690553188, + 0.0038959940429776907, + 0.012077459134161472, + 0.014756483025848866, + 0.013280926272273064, + 0.008861184120178223, + 0.009492075070738792, + 0.01203094981610775, + 0.008778185583651066, + 0.012469728477299213, + 0.012540561147034168, + 0.006571686360985041, + 0.004094723612070084, + 0.007234520744532347, + 0.006493726745247841, + 0.0053271944634616375, + 0.003266577608883381, + 0.005220420192927122, + 0.0029534664936363697 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 1, + 5 + ], + "selected_reward": -0.01607716642320156 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a gritty, dark anime style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene a low-angle close-up of a sleek electric supercar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the car's chassis made of shiny black carbon fiber and wet with rain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera remain static while focusing on the rear tires in breakneck slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the rear tires visibly deformed with wrinkling sidewalls due to the launch force?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03812656179070473 + ], + "opposite_logprob_tokens": [ + -3.2881264686584473 + ], + "logprob_reward": -0.03812656179070473, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03812656179070473, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09767711162567139, + 0.1872725933790207, + 0.06658277660608292, + 0.07929614186286926, + 0.12410176545381546, + 0.15660354495048523, + 0.06817090511322021, + 0.08143278956413269, + 0.07407538592815399, + 0.06478698551654816 + ], + "credit_map": [ + 0.010115819983184338, + 0.008058322593569756, + 0.013138754293322563, + 0.013354894705116749, + 0.013605930842459202, + 0.013111286796629429, + 0.01203963439911604, + 0.014252468943595886, + 0.020885726436972618, + 0.027806924656033516, + 0.02294653281569481, + 0.022126028314232826, + 0.02460513636469841, + 0.01963704451918602, + 0.02495771460235119, + 0.02430746890604496, + 0.007233481388539076, + 0.007258887402713299, + 0.007330260239541531, + 0.01019680593162775, + 0.006941983941942453, + 0.007931078784167767, + 0.011853541247546673, + 0.007836739532649517, + 0.00712483050301671, + 0.007765477057546377, + 0.007626617327332497, + 0.008455045521259308, + 0.00915178656578064, + 0.016857778653502464, + 0.011617543175816536, + 0.010697065852582455, + 0.010613503865897655, + 0.010859910398721695, + 0.020610937848687172, + 0.014255652204155922, + 0.014554220251739025, + 0.023726820945739746, + 0.013313228264451027, + 0.01616748981177807, + 0.010385209694504738, + 0.013783390633761883, + 0.02343936637043953, + 0.0223702322691679, + 0.014152023009955883, + 0.017466304823756218, + 0.03270449489355087, + 0.0223025344312191, + 0.006186542566865683, + 0.009717172011733055, + 0.006574202328920364, + 0.006836124695837498, + 0.008786840364336967, + 0.006657139863818884, + 0.01603979617357254, + 0.007373086642473936, + 0.004395296331495047, + 0.006811690051108599, + 0.006826123688369989, + 0.016674993559718132, + 0.008806481957435608, + 0.009818200953304768, + 0.007174920290708542, + 0.02092508226633072, + 0.010633889585733414, + 0.013242107816040516, + 0.012083741836249828, + 0.008602398447692394, + 0.007241785526275635, + 0.005192766897380352, + 0.00834967102855444, + 0.008729028515517712, + 0.012835482135415077, + 0.007982631213963032, + 0.0098795210942626, + 0.010929716750979424, + 0.007188263349235058, + 0.003234784584492445, + 0.007675237488001585, + 0.0050613535568118095 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 5, + 4 + ], + "selected_reward": -0.03812656179070473 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a gritty, dark anime style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene a low-angle close-up of a sleek electric supercar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the car's chassis made of shiny black carbon fiber and wet with rain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera remain static while focusing on the rear tires in breakneck slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the rear tires visibly deformed with wrinkling sidewalls due to the launch force?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.6328021287918091 + ], + "opposite_logprob_tokens": [ + -0.7578021287918091 + ], + "logprob_reward": -0.6328021287918091, + "yes_no_margin_reward": 0.125, + "token_credit_reward": -0.6328021287918091, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20123663544654846, + 0.05952049046754837, + 0.04701806232333183, + 0.09955260157585144, + 0.043596189469099045, + 0.20622575283050537, + 0.09842877835035324, + 0.058156680315732956, + 0.053029630333185196, + 0.13323521614074707 + ], + "credit_map": [ + 0.018700197339057922, + 0.01611349917948246, + 0.03504779189825058, + 0.019914157688617706, + 0.026736030355095863, + 0.018604328855872154, + 0.036994483321905136, + 0.029126126319169998, + 0.005567159503698349, + 0.006084983702749014, + 0.00670263497158885, + 0.005358436610549688, + 0.007736274041235447, + 0.01127262506633997, + 0.00832759402692318, + 0.008470777422189713, + 0.004665536805987358, + 0.004975341260433197, + 0.004796153865754604, + 0.007314467336982489, + 0.00580937135964632, + 0.005689461715519428, + 0.008676345460116863, + 0.005091377533972263, + 0.009576807729899883, + 0.009822379797697067, + 0.010930418968200684, + 0.009467842057347298, + 0.012806027196347713, + 0.020278222858905792, + 0.01417760644108057, + 0.012493284419178963, + 0.0031986613757908344, + 0.003571677953004837, + 0.004783142823725939, + 0.00589726073667407, + 0.004390939138829708, + 0.008465464226901531, + 0.007252809125930071, + 0.006036232225596905, + 0.015759272500872612, + 0.031214209273457527, + 0.024728599935770035, + 0.031361423432826996, + 0.022003449499607086, + 0.022711288183927536, + 0.03396449610590935, + 0.024482982233166695, + 0.008015909232199192, + 0.007402768358588219, + 0.006746101658791304, + 0.009706505574285984, + 0.02111520618200302, + 0.007466142997145653, + 0.031222717836499214, + 0.006753420457243919, + 0.0033161400351673365, + 0.00480671739205718, + 0.004171949811279774, + 0.012141508981585503, + 0.005811138544231653, + 0.007884321734309196, + 0.004289439879357815, + 0.015735462307929993, + 0.00499200914055109, + 0.005367717705667019, + 0.0062146494165062904, + 0.008248407393693924, + 0.004943301435559988, + 0.004584835842251778, + 0.00825488567352295, + 0.010423818603157997, + 0.014796189963817596, + 0.009743908420205116, + 0.01808745041489601, + 0.022489488124847412, + 0.02126249298453331, + 0.011211123317480087, + 0.02295105531811714, + 0.012693488970398903 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 9 + ], + "selected_reward": -0.6328021287918091 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a gritty, dark anime style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene a low-angle close-up of a sleek electric supercar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the car's chassis made of shiny black carbon fiber and wet with rain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera remain static while focusing on the rear tires in breakneck slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the rear tires visibly deformed with wrinkling sidewalls due to the launch force?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029790569096803665 + ], + "opposite_logprob_tokens": [ + -3.5297906398773193 + ], + "logprob_reward": -0.029790569096803665, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029790569096803665, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.30705276131629944, + 0.10493630170822144, + 0.06269565969705582, + 0.170343816280365, + 0.046378783881664276, + 0.0902397558093071, + 0.059554580599069595, + 0.06081800535321236, + 0.04993710294365883, + 0.048043303191661835 + ], + "credit_map": [ + 0.037141572684049606, + 0.03654055669903755, + 0.04851999878883362, + 0.039501823484897614, + 0.046866387128829956, + 0.03092690370976925, + 0.037499915808439255, + 0.030055569484829903, + 0.016228949651122093, + 0.01499804574996233, + 0.011797618120908737, + 0.013694697991013527, + 0.011763517744839191, + 0.009682671166956425, + 0.012622182257473469, + 0.014148611575365067, + 0.007525441236793995, + 0.00556637067347765, + 0.004828211851418018, + 0.0074973395094275475, + 0.006355757359415293, + 0.012347393669188023, + 0.009593665599822998, + 0.008981479331851006, + 0.019170017912983894, + 0.01922076940536499, + 0.018013697117567062, + 0.014980210922658443, + 0.02195614017546177, + 0.03165052458643913, + 0.02721613086760044, + 0.018136311322450638, + 0.00408201152458787, + 0.004303605295717716, + 0.007208378054201603, + 0.004389957990497351, + 0.003977042622864246, + 0.011291999369859695, + 0.00496550090610981, + 0.006160282529890537, + 0.0047100624069571495, + 0.015590287744998932, + 0.007966475561261177, + 0.017496049404144287, + 0.009275863878428936, + 0.01242617703974247, + 0.012036146596074104, + 0.01073868665844202, + 0.007034603040665388, + 0.009688046760857105, + 0.004980672616511583, + 0.006792326923459768, + 0.007561955600976944, + 0.004846279975026846, + 0.014123574830591679, + 0.0045271157287061214, + 0.00372148840688169, + 0.00442062271758914, + 0.004843297880142927, + 0.013072695583105087, + 0.0062356083653867245, + 0.006506863981485367, + 0.005042416509240866, + 0.016975004225969315, + 0.007853128015995026, + 0.005107201170176268, + 0.005925280973315239, + 0.007036383729428053, + 0.005677968263626099, + 0.004861949011683464, + 0.007146331947296858, + 0.00632885517552495, + 0.009391139261424541, + 0.004931548144668341, + 0.005924894008785486, + 0.007881912402808666, + 0.005671276710927486, + 0.003551838919520378, + 0.006485380232334137, + 0.004205312579870224 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 1 + ], + "selected_reward": -0.029790569096803665 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "dramatic low-angle close-up, rendered in a gritty, dark anime style, captures the moment a sleek electric supercar violently takes off from a standstill. The camera remains static, focusing in breakneck slow motion on the rear tires. The car's chassis is an angular marvel of shiny black carbon fiber, slicked by rain and reflecting the distorted, vibrant glow of the fuchsia and cyan neon signs emanating from the surrounding urban canyons. As the launch begins, the electric motor's immense, silent torque visibly deforms the ultra-wide, high-performance tires, their sidewalls wrinkling under the furious force as they bite into the rain-soaked asphalt. A ferocious cocktail of water is sprayed backward, and raindrops on the street and in the air are atomized by the spinning wheels, suspended in the air for a fleeting moment like a cloud of scattered, shimmering gems. The scene is a masterclass in high-contrast lighting, with deep, inky shadows sharply defining the aggressive lines of the car, while intense reflected light from the rain-soaked environment glitters off wet surfaces and suspended water droplets, creating an atmosphere of raw power and breathtakingly violent beauty.", + "seed": 25261, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000056/group_7/index_5754_seed_25261_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": -0.14325998537242413, + "vlm_logprob": -0.14325998537242413, + "vlm_yes_no_margin": 2.95, + "vlm_token_credit": -0.14325998537242413, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a gritty, dark anime style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene a low-angle close-up of a sleek electric supercar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the car's chassis made of shiny black carbon fiber and wet with rain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera remain static while focusing on the rear tires in breakneck slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the rear tires visibly deformed with wrinkling sidewalls due to the launch force?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02337891235947609 + ], + "opposite_logprob_tokens": [ + -3.898378849029541 + ], + "logprob_reward": -0.02337891235947609, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.02337891235947609, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14726710319519043, + 0.06498879194259644, + 0.06864435225725174, + 0.19630971550941467, + 0.035513754934072495, + 0.0552697516977787, + 0.21059589087963104, + 0.1063142791390419, + 0.08134361356496811, + 0.033752765506505966 + ], + "credit_map": [ + 0.010914376005530357, + 0.010499648749828339, + 0.009468134492635727, + 0.014007131569087505, + 0.03238445520401001, + 0.015376869589090347, + 0.00903777964413166, + 0.04557870700955391, + 0.00740505987778306, + 0.00819454900920391, + 0.007228091824799776, + 0.010617283172905445, + 0.008131202310323715, + 0.00894630141556263, + 0.005888489540666342, + 0.008577815257012844, + 0.0065142204985022545, + 0.005358926486223936, + 0.0062691462226212025, + 0.009673564694821835, + 0.004867506213486195, + 0.013859798200428486, + 0.009113031439483166, + 0.012988155707716942, + 0.043583329766988754, + 0.012301960028707981, + 0.03880980983376503, + 0.014476887881755829, + 0.018177516758441925, + 0.026530690491199493, + 0.014221776276826859, + 0.02820773795247078, + 0.004457917995750904, + 0.0037691376637667418, + 0.005127414129674435, + 0.007085462566465139, + 0.004345371853560209, + 0.003023649798706174, + 0.0046873497776687145, + 0.0030174534767866135, + 0.007799328770488501, + 0.007710067089647055, + 0.006555731408298016, + 0.0052841645665466785, + 0.007150168064981699, + 0.00776628265157342, + 0.006335866171866655, + 0.006668143905699253, + 0.010072792880237103, + 0.008085304871201515, + 0.07414786517620087, + 0.03298342600464821, + 0.007880615070462227, + 0.05126914009451866, + 0.015263915993273258, + 0.010892827995121479, + 0.015583883970975876, + 0.028353475034236908, + 0.008333807811141014, + 0.011451813392341137, + 0.022551655769348145, + 0.007015498820692301, + 0.007218411657959223, + 0.005805731751024723, + 0.006728798616677523, + 0.0072697363793849945, + 0.008080915547907352, + 0.00801425613462925, + 0.011691218242049217, + 0.02167460322380066, + 0.007344390731304884, + 0.010539693757891655, + 0.0033710275311022997, + 0.0035526761785149574, + 0.005087329540401697, + 0.006972658447921276, + 0.003544950159266591, + 0.0031075305305421352, + 0.0036768720019608736, + 0.004439723677933216 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 3, + 0 + ], + "selected_reward": -0.02337891235947609 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a gritty, dark anime style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene a low-angle close-up of a sleek electric supercar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the car's chassis made of shiny black carbon fiber and wet with rain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera remain static while focusing on the rear tires in breakneck slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the rear tires visibly deformed with wrinkling sidewalls due to the launch force?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008647376671433449 + ], + "opposite_logprob_tokens": [ + -4.758647441864014 + ], + "logprob_reward": -0.008647376671433449, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008647376671433449, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06941813230514526, + 0.050023145973682404, + 0.03181639313697815, + 0.23997074365615845, + 0.06845995038747787, + 0.04207922890782356, + 0.32332226634025574, + 0.05875174701213837, + 0.08772759884595871, + 0.028430813923478127 + ], + "credit_map": [ + 0.004517806228250265, + 0.0039090607315301895, + 0.0064754365012049675, + 0.0045783258974552155, + 0.01631619781255722, + 0.010009224526584148, + 0.005705324001610279, + 0.017906757071614265, + 0.008021813817322254, + 0.0069540394470095634, + 0.0046419668942689896, + 0.004455525893718004, + 0.00588558055460453, + 0.007565549109131098, + 0.0062110889703035355, + 0.006287583615630865, + 0.0038951297756284475, + 0.002450506901368499, + 0.004023273941129446, + 0.0033942984882742167, + 0.0019577518105506897, + 0.006553359795361757, + 0.004820720758289099, + 0.004721352830529213, + 0.05719838663935661, + 0.016119249165058136, + 0.025980252772569656, + 0.013482565991580486, + 0.012572516687214375, + 0.028760060667991638, + 0.019514475017786026, + 0.06634324043989182, + 0.006385196931660175, + 0.006412327289581299, + 0.00935138389468193, + 0.017986902967095375, + 0.007191105745732784, + 0.005728245247155428, + 0.01082849781960249, + 0.004576284438371658, + 0.0041048224084079266, + 0.005268205422908068, + 0.008039996027946472, + 0.0057502686977386475, + 0.00523967482149601, + 0.00332327326759696, + 0.006277897395193577, + 0.0040750871412456036, + 0.01658400148153305, + 0.010941286571323872, + 0.12442047894001007, + 0.055739693343639374, + 0.0060309735126793385, + 0.0785236582159996, + 0.016778426244854927, + 0.014303738251328468, + 0.006457386072725058, + 0.008153572678565979, + 0.005644736345857382, + 0.0070255533792078495, + 0.008742917329072952, + 0.006767073180526495, + 0.00958842784166336, + 0.006372083444148302, + 0.007709310855716467, + 0.014342443086206913, + 0.010295284911990166, + 0.011639769189059734, + 0.012710276991128922, + 0.009585968218743801, + 0.013633539900183678, + 0.007811008021235466, + 0.002439061412587762, + 0.004530822858214378, + 0.0036678940523415804, + 0.0037193335592746735, + 0.003226130036637187, + 0.0039670332334935665, + 0.003983596805483103, + 0.0028969449922442436 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 3, + 8 + ], + "selected_reward": -0.008647376671433449 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a gritty, dark anime style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene a low-angle close-up of a sleek electric supercar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the car's chassis made of shiny black carbon fiber and wet with rain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera remain static while focusing on the rear tires in breakneck slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the rear tires visibly deformed with wrinkling sidewalls due to the launch force?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0700470507144928 + ], + "opposite_logprob_tokens": [ + -2.69504714012146 + ], + "logprob_reward": -0.0700470507144928, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.0700470507144928, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12132365256547928, + 0.06725485622882843, + 0.04875689372420311, + 0.11667203158140182, + 0.08945462852716446, + 0.05991721898317337, + 0.1816098839044571, + 0.19074389338493347, + 0.06915032118558884, + 0.055116571485996246 + ], + "credit_map": [ + 0.008397008292376995, + 0.009397956542670727, + 0.012603521347045898, + 0.009623421356081963, + 0.02822587825357914, + 0.014003152959048748, + 0.006966117303818464, + 0.032106608152389526, + 0.007522250059992075, + 0.009790785610675812, + 0.007933304645121098, + 0.00949933286756277, + 0.008082677610218525, + 0.008965744636952877, + 0.007267293985933065, + 0.008193476125597954, + 0.005796320736408234, + 0.0040568336844444275, + 0.0035732996184378862, + 0.006115563213825226, + 0.0038517876528203487, + 0.012196432799100876, + 0.0032673499081283808, + 0.009899308905005455, + 0.029150232672691345, + 0.007238797377794981, + 0.025311831384897232, + 0.009301711805164814, + 0.007817558944225311, + 0.013484636321663857, + 0.007832244969904423, + 0.016535021364688873, + 0.011458280496299267, + 0.007298268377780914, + 0.015403307043015957, + 0.01914576254785061, + 0.00828070193529129, + 0.008559092879295349, + 0.01030122209340334, + 0.009007997810840607, + 0.007112923543900251, + 0.008867469616234303, + 0.009196972474455833, + 0.006670152302831411, + 0.007793740835040808, + 0.006648107431828976, + 0.007726604118943214, + 0.005901249125599861, + 0.0107488464564085, + 0.006602482404559851, + 0.06777379661798477, + 0.030089708045125008, + 0.004543694667518139, + 0.043136950582265854, + 0.010790843516588211, + 0.007923562079668045, + 0.01800217665731907, + 0.05207191780209541, + 0.01588706113398075, + 0.02519964426755905, + 0.03750620409846306, + 0.012544468976557255, + 0.02044225111603737, + 0.009090176783502102, + 0.006337632425129414, + 0.006873303558677435, + 0.006396717391908169, + 0.007036775350570679, + 0.012780473567545414, + 0.013889399357140064, + 0.007556600961834192, + 0.008279424160718918, + 0.006537517532706261, + 0.008475537411868572, + 0.006518561393022537, + 0.012248635292053223, + 0.004996058996766806, + 0.004328173119574785, + 0.007073836401104927, + 0.004938255064189434 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 6, + 0 + ], + "selected_reward": -0.0700470507144928 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a gritty, dark anime style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene a low-angle close-up of a sleek electric supercar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the car's chassis made of shiny black carbon fiber and wet with rain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera remain static while focusing on the rear tires in breakneck slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the rear tires visibly deformed with wrinkling sidewalls due to the launch force?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.5761346817016602 + ], + "opposite_logprob_tokens": [ + -0.8261346817016602 + ], + "logprob_reward": -0.5761346817016602, + "yes_no_margin_reward": 0.25, + "token_credit_reward": -0.5761346817016602, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04735761135816574, + 0.027954911813139915, + 0.06758709251880646, + 0.21002112329006195, + 0.08353132009506226, + 0.04485445097088814, + 0.18316316604614258, + 0.08679772168397903, + 0.1517588198184967, + 0.09697382152080536 + ], + "credit_map": [ + 0.0049066743813455105, + 0.004821546841412783, + 0.004998802673071623, + 0.0048463973216712475, + 0.009677167050540447, + 0.0057398672215640545, + 0.004487215541303158, + 0.007879938930273056, + 0.0054941000416874886, + 0.0036591673269867897, + 0.0026926035061478615, + 0.0031028101220726967, + 0.002774354536086321, + 0.0028515870217233896, + 0.0032074456103146076, + 0.004172844812273979, + 0.006191139109432697, + 0.0048715039156377316, + 0.005440800450742245, + 0.005697679705917835, + 0.004211792256683111, + 0.019649825990200043, + 0.007412497419863939, + 0.014111855998635292, + 0.0531061515212059, + 0.015729019418358803, + 0.025758685544133186, + 0.01283703837543726, + 0.011181910522282124, + 0.024211131036281586, + 0.01494369562715292, + 0.05225348845124245, + 0.011190768331289291, + 0.006900262087583542, + 0.0216312687844038, + 0.012519008480012417, + 0.006408995017409325, + 0.008972818031907082, + 0.009935444220900536, + 0.005972752347588539, + 0.0046067750081419945, + 0.0060690706595778465, + 0.007806659676134586, + 0.006066733971238136, + 0.005992749240249395, + 0.0035861737560480833, + 0.0065755415707826614, + 0.004150744993239641, + 0.00933349784463644, + 0.008108080364763737, + 0.0646858885884285, + 0.030076663941144943, + 0.006997533142566681, + 0.042486391961574554, + 0.01293363980948925, + 0.008541465736925602, + 0.022091498598456383, + 0.014824951067566872, + 0.007997694425284863, + 0.011164069175720215, + 0.00901658646762371, + 0.005675431806594133, + 0.010689360089600086, + 0.005338132381439209, + 0.01520899310708046, + 0.019542867317795753, + 0.018360795453190804, + 0.017637968063354492, + 0.022535746917128563, + 0.02028742991387844, + 0.02477838099002838, + 0.013406629674136639, + 0.008901393041014671, + 0.014002379961311817, + 0.011358891613781452, + 0.013871870934963226, + 0.00845723319798708, + 0.015310159884393215, + 0.013687646947801113, + 0.011384244076907635 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 6, + 8 + ], + "selected_reward": -0.5761346817016602 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a gritty, dark anime style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene a low-angle close-up of a sleek electric supercar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the car's chassis made of shiny black carbon fiber and wet with rain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera remain static while focusing on the rear tires in breakneck slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the rear tires visibly deformed with wrinkling sidewalls due to the launch force?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.038091905415058136 + ], + "opposite_logprob_tokens": [ + -3.2880918979644775 + ], + "logprob_reward": -0.038091905415058136, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.038091905415058136, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.32618212699890137, + 0.08210601657629013, + 0.039496809244155884, + 0.0873946025967598, + 0.05108810216188431, + 0.057039931416511536, + 0.09684208035469055, + 0.09885566681623459, + 0.046119995415210724, + 0.11487455666065216 + ], + "credit_map": [ + 0.016195133328437805, + 0.017328470945358276, + 0.01919800229370594, + 0.02300834283232689, + 0.0907319039106369, + 0.03395172581076622, + 0.01750423200428486, + 0.10826434940099716, + 0.013128018006682396, + 0.013739986345171928, + 0.007084098644554615, + 0.008450474590063095, + 0.01090597826987505, + 0.009807668626308441, + 0.00930647924542427, + 0.00968332588672638, + 0.004252136219292879, + 0.004053895361721516, + 0.004086805507540703, + 0.0049447836354374886, + 0.00298828212544322, + 0.009226173162460327, + 0.004139019176363945, + 0.005805717781186104, + 0.013725175522267818, + 0.0064968084916472435, + 0.013011771254241467, + 0.007391871884465218, + 0.005528145935386419, + 0.0100327767431736, + 0.00759147061035037, + 0.02361658774316311, + 0.007875779643654823, + 0.0040156166069209576, + 0.00935355480760336, + 0.005358218215405941, + 0.005798610392957926, + 0.008034257218241692, + 0.0060736509039998055, + 0.004578417167067528, + 0.004882389213889837, + 0.009570303373038769, + 0.008303537033498287, + 0.007720188237726688, + 0.009047755040228367, + 0.004035591147840023, + 0.009049605578184128, + 0.004430563189089298, + 0.006826761178672314, + 0.01012395042926073, + 0.027698777616024017, + 0.01241260301321745, + 0.0048253219574689865, + 0.023215916007757187, + 0.006921801250427961, + 0.004816957749426365, + 0.0071016401052474976, + 0.02389969676733017, + 0.008492005057632923, + 0.011671805754303932, + 0.019570551812648773, + 0.008577085100114346, + 0.012018037028610706, + 0.007524856366217136, + 0.0037218353245407343, + 0.008173622190952301, + 0.004029535222798586, + 0.005446210969239473, + 0.006787125486880541, + 0.006431013811379671, + 0.007980432361364365, + 0.003550225868821144, + 0.01362842507660389, + 0.02595353312790394, + 0.01216984074562788, + 0.013966421596705914, + 0.014835603535175323, + 0.007717054802924395, + 0.015286237001419067, + 0.011317454278469086 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 7 + ], + "selected_reward": -0.038091905415058136 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "dramatic low-angle close-up, rendered in a gritty, dark anime style, captures the moment a sleek electric supercar violently takes off from a standstill. The camera remains static, focusing in breakneck slow motion on the rear tires. The car's chassis is an angular marvel of shiny black carbon fiber, slicked by rain and reflecting the distorted, vibrant glow of the fuchsia and cyan neon signs emanating from the surrounding urban canyons. As the launch begins, the electric motor's immense, silent torque visibly deforms the ultra-wide, high-performance tires, their sidewalls wrinkling under the furious force as they bite into the rain-soaked asphalt. A ferocious cocktail of water is sprayed backward, and raindrops on the street and in the air are atomized by the spinning wheels, suspended in the air for a fleeting moment like a cloud of scattered, shimmering gems. The scene is a masterclass in high-contrast lighting, with deep, inky shadows sharply defining the aggressive lines of the car, while intense reflected light from the rain-soaked environment glitters off wet surfaces and suspended water droplets, creating an atmosphere of raw power and breathtakingly violent beauty.", + "seed": 25262, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000056/group_7/index_5754_seed_25262_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": -0.20540945269167424, + "vlm_logprob": -0.20540945269167424, + "vlm_yes_no_margin": 2.6, + "vlm_token_credit": -0.20540945269167424, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a gritty, dark anime style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene a low-angle close-up of a sleek electric supercar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the car's chassis made of shiny black carbon fiber and wet with rain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera remain static while focusing on the rear tires in breakneck slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the rear tires visibly deformed with wrinkling sidewalls due to the launch force?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.101565882563591 + ], + "opposite_logprob_tokens": [ + -2.3515658378601074 + ], + "logprob_reward": -0.101565882563591, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.101565882563591, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09537526220083237, + 0.036618854850530624, + 0.2853381633758545, + 0.028661567717790604, + 0.055485256016254425, + 0.05217546597123146, + 0.19200104475021362, + 0.07733143121004105, + 0.020308014005422592, + 0.15670493245124817 + ], + "credit_map": [ + 0.009857764467597008, + 0.009080769494175911, + 0.011722876690328121, + 0.013912754133343697, + 0.011598322540521622, + 0.010999208316206932, + 0.01462655607610941, + 0.013577009551227093, + 0.0037442068569362164, + 0.002552555175498128, + 0.00543693732470274, + 0.006929023656994104, + 0.003884667530655861, + 0.0035528778098523617, + 0.0049567678943276405, + 0.0055618188343942165, + 0.017439132556319237, + 0.01535489596426487, + 0.02797929383814335, + 0.07114149630069733, + 0.019460871815681458, + 0.026102812960743904, + 0.030916865915060043, + 0.0769428089261055, + 0.0022118070628494024, + 0.0028082875069230795, + 0.003287359606474638, + 0.0023775077424943447, + 0.0018040379509329796, + 0.003256062977015972, + 0.006721831858158112, + 0.006194672081619501, + 0.007169789168983698, + 0.005991001147776842, + 0.005124904680997133, + 0.008186507038772106, + 0.004653457552194595, + 0.005449245683848858, + 0.010347031988203526, + 0.008563317358493805, + 0.004725863691419363, + 0.0029892337042838335, + 0.009795675985515118, + 0.008534866385161877, + 0.0042055812664330006, + 0.004716960247606039, + 0.010083746165037155, + 0.0071235401555895805, + 0.021265145391225815, + 0.008347544819116592, + 0.05320831015706062, + 0.01742827333509922, + 0.01416228897869587, + 0.018238134682178497, + 0.01791507378220558, + 0.041436269879341125, + 0.01099134050309658, + 0.006437880452722311, + 0.010145737789571285, + 0.012781605124473572, + 0.006906603928655386, + 0.01117636077105999, + 0.01079561561346054, + 0.008096291683614254, + 0.002333903219550848, + 0.0020486454013735056, + 0.004177548456937075, + 0.003154124366119504, + 0.0017148094484582543, + 0.0019520065980032086, + 0.0024807469453662634, + 0.0024462295696139336, + 0.01289556547999382, + 0.01865067519247532, + 0.02811143919825554, + 0.034079957753419876, + 0.012301587499678135, + 0.01582942344248295, + 0.016124965623021126, + 0.018711313605308533 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 6, + 9 + ], + "selected_reward": -0.101565882563591 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a gritty, dark anime style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene a low-angle close-up of a sleek electric supercar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the car's chassis made of shiny black carbon fiber and wet with rain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera remain static while focusing on the rear tires in breakneck slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the rear tires visibly deformed with wrinkling sidewalls due to the launch force?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007632380351424217 + ], + "opposite_logprob_tokens": [ + -4.882632255554199 + ], + "logprob_reward": -0.007632380351424217, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007632380351424217, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1728571653366089, + 0.0625348836183548, + 0.2111266553401947, + 0.026609240099787712, + 0.07737483084201813, + 0.10011637210845947, + 0.10128351300954819, + 0.0944451317191124, + 0.04233261197805405, + 0.11131957173347473 + ], + "credit_map": [ + 0.02497323602437973, + 0.01758561097085476, + 0.018559658899903297, + 0.02503652125597, + 0.01766505092382431, + 0.014975027181208134, + 0.023371312767267227, + 0.030690746381878853, + 0.006581924390047789, + 0.004534874111413956, + 0.005625296849757433, + 0.008523602969944477, + 0.008547255769371986, + 0.008719222620129585, + 0.013162045739591122, + 0.006840656045824289, + 0.013348344713449478, + 0.008694251999258995, + 0.022822823375463486, + 0.06017905846238136, + 0.004861173685640097, + 0.013059393502771854, + 0.021090572699904442, + 0.06707103550434113, + 0.0018327765865251422, + 0.0032287135254591703, + 0.0022041904740035534, + 0.0029159020632505417, + 0.0021710784640163183, + 0.0029305238276720047, + 0.00546399038285017, + 0.005862063262611628, + 0.007263140752911568, + 0.008365518413484097, + 0.005827481392771006, + 0.015840597450733185, + 0.0037590241990983486, + 0.003691991325467825, + 0.015930812805891037, + 0.0166962631046772, + 0.01141966413706541, + 0.005949746817350388, + 0.016153695061802864, + 0.017270086333155632, + 0.006471792235970497, + 0.007465466856956482, + 0.022693509235978127, + 0.012692409567534924, + 0.012906818650662899, + 0.006233668886125088, + 0.015194499865174294, + 0.012217143550515175, + 0.0110245980322361, + 0.017093360424041748, + 0.01398899219930172, + 0.012624429538846016, + 0.013435875996947289, + 0.009808718226850033, + 0.014595973305404186, + 0.015445786528289318, + 0.009868038818240166, + 0.010071328841149807, + 0.012236207723617554, + 0.008983202278614044, + 0.004413530230522156, + 0.003585138591006398, + 0.00907021015882492, + 0.008767778053879738, + 0.0029772527050226927, + 0.004392892587929964, + 0.004492997657507658, + 0.004632808733731508, + 0.010029339231550694, + 0.01159679889678955, + 0.01944476179778576, + 0.01832437887787819, + 0.014052172191441059, + 0.012125751934945583, + 0.01255897805094719, + 0.01318738330155611 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 9 + ], + "selected_reward": -0.007632380351424217 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a gritty, dark anime style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene a low-angle close-up of a sleek electric supercar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the car's chassis made of shiny black carbon fiber and wet with rain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera remain static while focusing on the rear tires in breakneck slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the rear tires visibly deformed with wrinkling sidewalls due to the launch force?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.048679985105991364 + ], + "opposite_logprob_tokens": [ + -3.048680067062378 + ], + "logprob_reward": -0.048679985105991364, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.048679985105991364, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16459907591342926, + 0.061777520924806595, + 0.21985234320163727, + 0.05159316584467888, + 0.043009575456380844, + 0.06911187618970871, + 0.1764398217201233, + 0.06342212855815887, + 0.03465575352311134, + 0.11553874611854553 + ], + "credit_map": [ + 0.01876429095864296, + 0.013448502868413925, + 0.018713433295488358, + 0.03617402911186218, + 0.017394891008734703, + 0.01446237601339817, + 0.025459058582782745, + 0.020182499662041664, + 0.006178956478834152, + 0.004176066257059574, + 0.009904437698423862, + 0.009613760747015476, + 0.007319898344576359, + 0.005748300347477198, + 0.010088440030813217, + 0.008747664280235767, + 0.009151297621428967, + 0.009442856535315514, + 0.013951144181191921, + 0.07441527396440506, + 0.004942651838064194, + 0.01231249701231718, + 0.02092638611793518, + 0.07471023499965668, + 0.0033545629121363163, + 0.002745797159150243, + 0.00972827710211277, + 0.004060707520693541, + 0.0024213925935328007, + 0.0031151026487350464, + 0.0149129843339324, + 0.011254340410232544, + 0.0038283593021333218, + 0.0049807229079306126, + 0.003710267599672079, + 0.0073756929486989975, + 0.0046384297311306, + 0.0036955324467271566, + 0.007861741818487644, + 0.006918827537447214, + 0.0079498291015625, + 0.004232409875839949, + 0.01338203065097332, + 0.01165248267352581, + 0.005707113537937403, + 0.005974115803837776, + 0.01208682730793953, + 0.008127070032060146, + 0.020240653306245804, + 0.00829667504876852, + 0.04695315286517143, + 0.014514571987092495, + 0.013615638948976994, + 0.019240474328398705, + 0.016223039478063583, + 0.03735562041401863, + 0.006582979112863541, + 0.005914649460464716, + 0.006413345690816641, + 0.007879088632762432, + 0.006179101765155792, + 0.013111268170177937, + 0.01007047574967146, + 0.007271222770214081, + 0.0032765597570687532, + 0.004482174292206764, + 0.006579340901225805, + 0.0062019797042012215, + 0.0023694802075624466, + 0.0034521622583270073, + 0.0051445201970636845, + 0.0031495385337620974, + 0.009769137017428875, + 0.016535811126232147, + 0.02009783498942852, + 0.02953983284533024, + 0.007796537596732378, + 0.009788742288947105, + 0.01273097563534975, + 0.009279881604015827 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 6, + 0 + ], + "selected_reward": -0.048679985105991364 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a gritty, dark anime style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene a low-angle close-up of a sleek electric supercar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the car's chassis made of shiny black carbon fiber and wet with rain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera remain static while focusing on the rear tires in breakneck slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the rear tires visibly deformed with wrinkling sidewalls due to the launch force?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"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": [ + -0.8261182904243469 + ], + "opposite_logprob_tokens": [ + -0.5761182904243469 + ], + "logprob_reward": -0.8261182904243469, + "yes_no_margin_reward": -0.25, + "token_credit_reward": -0.8261182904243469, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.09286759793758392, + 0.04380052164196968, + 0.26235231757164, + 0.026242554187774658, + 0.05059739947319031, + 0.07630808651447296, + 0.135398730635643, + 0.0963093712925911, + 0.05203624814748764, + 0.16408725082874298 + ], + "credit_map": [ + 0.0149263646453619, + 0.006713517941534519, + 0.012850252911448479, + 0.013358984142541885, + 0.011588587425649166, + 0.011084916070103645, + 0.01194856595247984, + 0.01039639301598072, + 0.004581649322062731, + 0.0033376647625118494, + 0.005939417518675327, + 0.008775993250310421, + 0.0059452420100569725, + 0.005507614463567734, + 0.004355970770120621, + 0.005356960464268923, + 0.021254707127809525, + 0.01826636679470539, + 0.03862438350915909, + 0.03954653814435005, + 0.025361426174640656, + 0.03129146620631218, + 0.03968124836683273, + 0.04832614213228226, + 0.002451786072924733, + 0.005246165208518505, + 0.0026555201038718224, + 0.003544420702382922, + 0.0018667720723897219, + 0.0030499519780278206, + 0.003856556722894311, + 0.003571375971660018, + 0.0074700419791042805, + 0.0056429277174174786, + 0.004148217849433422, + 0.00684801721945405, + 0.004091434646397829, + 0.0026198732666671276, + 0.010346099734306335, + 0.00943078100681305, + 0.008412940427660942, + 0.005452493671327829, + 0.011443626135587692, + 0.01242141891270876, + 0.004998618271201849, + 0.005762824788689613, + 0.015758343040943146, + 0.012057806365191936, + 0.01482360064983368, + 0.00671157892793417, + 0.033485252410173416, + 0.012536164373159409, + 0.010437247343361378, + 0.017349950969219208, + 0.012676196172833443, + 0.02737872488796711, + 0.012061563320457935, + 0.011045328341424465, + 0.012400852516293526, + 0.015397351235151291, + 0.010245300829410553, + 0.010408176109194756, + 0.01352611929178238, + 0.011224665679037571, + 0.007794336881488562, + 0.006377018988132477, + 0.008792740292847157, + 0.008006102405488491, + 0.0037656051572412252, + 0.006277679931372404, + 0.004562503192573786, + 0.006460252683609724, + 0.01158114429563284, + 0.013716313056647778, + 0.028556138277053833, + 0.03919973224401474, + 0.018689453601837158, + 0.02088960073888302, + 0.01669357717037201, + 0.014761267229914665 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 9, + 6 + ], + "selected_reward": -0.8261182904243469 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a gritty, dark anime style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene a low-angle close-up of a sleek electric supercar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the car's chassis made of shiny black carbon fiber and wet with rain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera remain static while focusing on the rear tires in breakneck slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the rear tires visibly deformed with wrinkling sidewalls due to the launch force?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.043050725013017654 + ], + "opposite_logprob_tokens": [ + -3.168050765991211 + ], + "logprob_reward": -0.043050725013017654, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.043050725013017654, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08328871428966522, + 0.0571773424744606, + 0.22295399010181427, + 0.09377532452344894, + 0.05081072077155113, + 0.04845808818936348, + 0.13474518060684204, + 0.10787878185510635, + 0.03245481103658676, + 0.1684570461511612 + ], + "credit_map": [ + 0.011208739131689072, + 0.009413697756826878, + 0.010426036082208157, + 0.011311987414956093, + 0.008131913840770721, + 0.006593133322894573, + 0.013534924015402794, + 0.012668284587562084, + 0.006712822243571281, + 0.004399963188916445, + 0.006367271300405264, + 0.007205096539109945, + 0.007321801036596298, + 0.009726159274578094, + 0.008596895262598991, + 0.0068473308347165585, + 0.009819403290748596, + 0.00992789026349783, + 0.016008231788873672, + 0.07148156315088272, + 0.0065835691057145596, + 0.013802481815218925, + 0.0218775924295187, + 0.07345324754714966, + 0.006087968125939369, + 0.006597097963094711, + 0.016579480841755867, + 0.007282160688191652, + 0.004664773121476173, + 0.00534284021705389, + 0.027522079646587372, + 0.01969892531633377, + 0.004733334295451641, + 0.004628969356417656, + 0.003510379698127508, + 0.010564005933701992, + 0.002841348061338067, + 0.0017688279040157795, + 0.013147147372364998, + 0.00961670558899641, + 0.00460457568988204, + 0.003892544424161315, + 0.00844511017203331, + 0.007558997720479965, + 0.004760986194014549, + 0.00568145327270031, + 0.007524298503994942, + 0.005990117788314819, + 0.015147198922932148, + 0.005599033087491989, + 0.037303224205970764, + 0.012500995770096779, + 0.00976125430315733, + 0.012551063671708107, + 0.01150550041347742, + 0.030376894399523735, + 0.015663528814911842, + 0.010635774582624435, + 0.01719576120376587, + 0.017996106296777725, + 0.01066700927913189, + 0.012111826799809933, + 0.013663643971085548, + 0.009945125319063663, + 0.003917027730494738, + 0.0038938941434025764, + 0.005725306924432516, + 0.00606126431375742, + 0.0024782007094472647, + 0.002850999590009451, + 0.003763880580663681, + 0.0037642328534275293, + 0.01605977676808834, + 0.022853700444102287, + 0.028750834986567497, + 0.04016924649477005, + 0.011442984454333782, + 0.016690297052264214, + 0.017655232921242714, + 0.014834980480372906 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 9, + 6 + ], + "selected_reward": -0.043050725013017654 + } + ], + "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/0000056/group_8/index_tensor([9914, 9914])_group_8_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000056/group_8/index_tensor([9914, 9914])_group_8_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..ba5bcf8736f7d9bc8e2b32e3d813367d7fe64437 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000056/group_8/index_tensor([9914, 9914])_group_8_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "low-flying drone camera zooms in at breakneck speed, passing close to the top of a vast wheat field. The warm, setting sun of golden hour illuminates the field, a deep, brilliant amber-gold. Each ear of wheat, plump and heavy, bends with its weight, its awns gleaming with tiny shimmers in the sunlight. A gentle breeze sweeps across the vast wheat field, creating a continuous, continuous wave of golden ripples that rhythmically spread from the near side of the camera toward the distant horizon, a breathtaking sight. A few wisps of sparse white clouds punctuate the clear azure sky, creating a stark contrast to the boundless golden ocean below. The rapid camera movement creates a slight motion blur at the edges of the frame, heightening the sense of speed. The entire scene is imbued with the joy of harvest and the grandeur of nature, creating an atmosphere of freedom, openness, and a tranquil, vibrant vitality.", + "seed": 31987, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000056/group_8/index_9914_seed_31987_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": -0.15686723366379737, + "vlm_logprob": -0.15686723366379737, + "vlm_yes_no_margin": 2.925, + "vlm_token_credit": -0.15686723366379737, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 vast wheat field illuminated by a warm, setting sun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the ears of wheat depicted as plump, heavy, and bending with their weight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the field show a continuous wave of golden ripples spreading from the near side toward the distant horizon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are a few wisps of sparse white clouds visible in a clear azure sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera move rapidly with a slight motion blur at the edges 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": [ + -0.007740503177046776 + ], + "opposite_logprob_tokens": [ + -6.007740497589111 + ], + "logprob_reward": -0.007740503177046776, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.007740503177046776, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0367906391620636, + 0.8575215339660645, + 0.010254510678350925, + 0.015136546455323696, + 0.007138492073863745, + 0.007776749320328236, + 0.03379196673631668, + 0.012770812958478928, + 0.0152190076187253, + 0.0035996665246784687 + ], + "credit_map": [ + 0.002904241904616356, + 0.005130086559802294, + 0.004084075801074505, + 0.00551341287791729, + 0.003447680501267314, + 0.005728700663894415, + 0.005574270151555538, + 0.004408176057040691, + 0.1519789695739746, + 0.3088871240615845, + 0.1301434189081192, + 0.05378684401512146, + 0.10090436041355133, + 0.04809560999274254, + 0.031015669927001, + 0.03270962834358215, + 0.0006647998816333711, + 0.0005854736664332449, + 0.0019597422797232866, + 0.002306512789800763, + 0.0005600322037935257, + 0.0006809848709963262, + 0.001975876512005925, + 0.0015210898127406836, + 0.0012648238334804773, + 0.0025581989903002977, + 0.0019360160222277045, + 0.0023765203077346087, + 0.0013984362594783306, + 0.0007869576220400631, + 0.0021018905099481344, + 0.0027137042488902807, + 0.0005219504819251597, + 0.0006237017223611474, + 0.001348579302430153, + 0.0012860280694440007, + 0.0009298231452703476, + 0.000998144969344139, + 0.0008098147809505463, + 0.0006204505334608257, + 0.0005471736658364534, + 0.0008509098552167416, + 0.001285193720832467, + 0.001076146261766553, + 0.0011731797130778432, + 0.0014806835679337382, + 0.0008431742899119854, + 0.0005202885949984193, + 0.0058559877797961235, + 0.011458066292107105, + 0.005198837257921696, + 0.0022369250655174255, + 0.003920517396181822, + 0.001632525003515184, + 0.0014634478138759732, + 0.002025662688538432, + 0.0009276063065044582, + 0.0008020748500712216, + 0.0029404039960354567, + 0.0021214664448052645, + 0.0013988245045766234, + 0.0015191996935755014, + 0.0015469944337382913, + 0.0015142435440793633, + 0.0026592910289764404, + 0.004255788866430521, + 0.0008261185721494257, + 0.0011227662907913327, + 0.0021547332871705294, + 0.001998515333980322, + 0.0010436662705615163, + 0.0011581303551793098, + 0.00021931341325398535, + 0.0002679766621440649, + 0.0010191671317443252, + 0.0005457511288113892, + 0.000397370575228706, + 0.0002546399482525885, + 0.0004038297920487821, + 0.0004916179459542036 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 6 + ], + "selected_reward": -0.007740503177046776 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 vast wheat field illuminated by a warm, setting sun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the ears of wheat depicted as plump, heavy, and bending with their weight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the field show a continuous wave of golden ripples spreading from the near side toward the distant horizon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are a few wisps of sparse white clouds visible in a clear azure sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera move rapidly with a slight motion blur at the edges 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": [ + -0.42893746495246887 + ], + "opposite_logprob_tokens": [ + -1.0539374351501465 + ], + "logprob_reward": -0.42893746495246887, + "yes_no_margin_reward": 0.625, + "token_credit_reward": -0.42893746495246887, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04299228638410568, + 0.6623066663742065, + 0.04902276024222374, + 0.046422481536865234, + 0.02442186325788498, + 0.024790875613689423, + 0.06508734822273254, + 0.03986826166510582, + 0.031650565564632416, + 0.01343685481697321 + ], + "credit_map": [ + 0.002926170825958252, + 0.0033104249741882086, + 0.006363768596202135, + 0.008406106382608414, + 0.00292297569103539, + 0.003876632545143366, + 0.006981223821640015, + 0.008204986341297626, + 0.11558469384908676, + 0.23606401681900024, + 0.10361246019601822, + 0.04208666831254959, + 0.07731974869966507, + 0.03676481172442436, + 0.024863408878445625, + 0.02601085603237152, + 0.0030612063128501177, + 0.0051293401047587395, + 0.008063911460340023, + 0.01012978795915842, + 0.0022914987057447433, + 0.0016900196205824614, + 0.012764223851263523, + 0.005892775021493435, + 0.008090716786682606, + 0.015591461211442947, + 0.0035456623882055283, + 0.0036602087784558535, + 0.0053500025533139706, + 0.003133168211206794, + 0.003907870966941118, + 0.0031433908734470606, + 0.0017270916141569614, + 0.0011317962780594826, + 0.007818358950316906, + 0.005413352977484465, + 0.0018695679027587175, + 0.0020453289616853, + 0.002072439296171069, + 0.0023439275100827217, + 0.002117693657055497, + 0.003281288780272007, + 0.001962443348020315, + 0.0028657226357609034, + 0.004966864362359047, + 0.005793696269392967, + 0.0023136099334806204, + 0.0014895573258399963, + 0.014033006504178047, + 0.02613600343465805, + 0.005121273919939995, + 0.003088906640186906, + 0.008914574980735779, + 0.0036376663483679295, + 0.0016322073061019182, + 0.0025237107183784246, + 0.0024233951698988676, + 0.0027372068725526333, + 0.007993802428245544, + 0.004928576294332743, + 0.0041148304007947445, + 0.004601140506565571, + 0.007249009795486927, + 0.005820302292704582, + 0.005325556267052889, + 0.008833317086100578, + 0.002669051755219698, + 0.002637561410665512, + 0.0027244118973612785, + 0.0020423682872205973, + 0.00430002436041832, + 0.003118275199085474, + 0.0010111561277881265, + 0.0021738619543612003, + 0.0015103763435035944, + 0.0021080193109810352, + 0.0020526915322989225, + 0.0008568937773816288, + 0.00223914859816432, + 0.0014847074635326862 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 6, + 2 + ], + "selected_reward": -0.42893746495246887 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 vast wheat field illuminated by a warm, setting sun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the ears of wheat depicted as plump, heavy, and bending with their weight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the field show a continuous wave of golden ripples spreading from the near side toward the distant horizon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are a few wisps of sparse white clouds visible in a clear azure sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera move rapidly with a slight motion blur at the edges 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": [ + -0.023284347727894783 + ], + "opposite_logprob_tokens": [ + -3.773284435272217 + ], + "logprob_reward": -0.023284347727894783, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023284347727894783, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.02526986598968506, + 0.817591667175293, + 0.02796146273612976, + 0.018560733646154404, + 0.014646279625594616, + 0.006726247724145651, + 0.026519576087594032, + 0.016864154487848282, + 0.033542681485414505, + 0.012317365035414696 + ], + "credit_map": [ + 0.001686825416982174, + 0.0018494072137400508, + 0.003318856470286846, + 0.004214714281260967, + 0.0023600261192768812, + 0.0027095419354736805, + 0.00358277908526361, + 0.005547716747969389, + 0.143690824508667, + 0.2927578091621399, + 0.12539897859096527, + 0.05232909694314003, + 0.09562163800001144, + 0.04577245935797691, + 0.030275870114564896, + 0.031744956970214844, + 0.0022087336983531713, + 0.002083003520965576, + 0.004897235427051783, + 0.005073026288300753, + 0.002986753825098276, + 0.0032459991052746773, + 0.004442237317562103, + 0.0030244761146605015, + 0.0030306626576930285, + 0.0037786175962537527, + 0.001898045651614666, + 0.0025622625835239887, + 0.0015638475306332111, + 0.0009535907884128392, + 0.0027795799542218447, + 0.0019941285718232393, + 0.001036719186231494, + 0.0009127248194999993, + 0.003819948760792613, + 0.0028676579240709543, + 0.0013426347868517041, + 0.0017463715048506856, + 0.0013193817576393485, + 0.0016008417587727308, + 0.0006768150487914681, + 0.0010559101356193423, + 0.0005216025747358799, + 0.0007496392354369164, + 0.0011928348103538156, + 0.0013473773142322898, + 0.0006847920594736934, + 0.0004972770111635327, + 0.005773847457021475, + 0.010390699841082096, + 0.0013653020141646266, + 0.001598954200744629, + 0.0034618356730788946, + 0.0016074507730081677, + 0.0010010554688051343, + 0.001320432871580124, + 0.0010957400081679225, + 0.0008947129244916141, + 0.0030892372597008944, + 0.002540722955018282, + 0.0011008583242073655, + 0.0010300350841134787, + 0.003695582738146186, + 0.0034172656014561653, + 0.005549967754632235, + 0.00930075440555811, + 0.002051503164693713, + 0.0020794423762708902, + 0.0044575221836566925, + 0.004500518087297678, + 0.003184219589456916, + 0.002418756252154708, + 0.0014210263034328818, + 0.0011722065974026918, + 0.004025333095341921, + 0.0017762003699317575, + 0.0004957668134011328, + 0.0004799381422344595, + 0.0015080295270308852, + 0.0014388662530109286 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 8, + 2 + ], + "selected_reward": -0.023284347727894783 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 vast wheat field illuminated by a warm, setting sun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the ears of wheat depicted as plump, heavy, and bending with their weight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the field show a continuous wave of golden ripples spreading from the near side toward the distant horizon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are a few wisps of sparse white clouds visible in a clear azure sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera move rapidly with a slight motion blur at the edges 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": [ + -0.04306374117732048 + ], + "opposite_logprob_tokens": [ + -3.1680636405944824 + ], + "logprob_reward": -0.04306374117732048, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04306374117732048, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04860764369368553, + 0.914440929889679, + 0.0020403817761689425, + 0.007639085873961449, + 0.0027846870943903923, + 0.0035478302743285894, + 0.009229299612343311, + 0.004412392154335976, + 0.004284951835870743, + 0.003012828528881073 + ], + "credit_map": [ + 0.004134617280215025, + 0.006399665493518114, + 0.008052187971770763, + 0.006738958414644003, + 0.00319518125616014, + 0.005348998587578535, + 0.0047734398394823074, + 0.009964593686163425, + 0.16163834929466248, + 0.32859519124031067, + 0.1397855132818222, + 0.05756400153040886, + 0.10749667882919312, + 0.05126461759209633, + 0.03300289809703827, + 0.03509368747472763, + 0.0001531938323751092, + 0.00020748427778016776, + 0.0003112809790764004, + 0.00036297363112680614, + 0.00019558366329874843, + 0.00011898294906131923, + 0.0004441091150511056, + 0.0002467733283992857, + 0.000515214167535305, + 0.0007009851397015154, + 0.0007737941923551261, + 0.0012218775227665901, + 0.001065650605596602, + 0.0009083501645363867, + 0.0012923501199111342, + 0.0011608636705204844, + 0.0002743420482147485, + 0.00022154077305458486, + 0.0006678904755972326, + 0.0005274293944239616, + 0.00019727698236238211, + 0.00015065705520100892, + 0.0004193401837255806, + 0.0003262100217398256, + 0.00032073422335088253, + 0.000390856119338423, + 0.0002757792826741934, + 0.0004091172304470092, + 0.0006596396560780704, + 0.0007714141393080354, + 0.00035360929905436933, + 0.00036668061511591077, + 0.0015881656436249614, + 0.003123454051092267, + 0.0013399681774899364, + 0.0005653813132084906, + 0.0011780858039855957, + 0.0005030191969126463, + 0.0004294797545298934, + 0.0005017459625378251, + 0.00030284406966529787, + 0.0003534514980856329, + 0.0004836181178689003, + 0.00048685254296287894, + 0.0007126100827008486, + 0.000624262320343405, + 0.0007641250267624855, + 0.0006846282631158829, + 0.0007256023236550391, + 0.001191682182252407, + 0.00030967776547186077, + 0.00036559911677613854, + 0.0004488365666475147, + 0.00044998357770964503, + 0.0004984411061741412, + 0.00029512910987250507, + 0.0003264609258621931, + 0.00030888995388522744, + 0.0008824755204841495, + 0.00040299291140399873, + 0.00017245211347471923, + 0.0002148449420928955, + 0.00035855438909493387, + 0.00034615761251188815 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 6 + ], + "selected_reward": -0.04306374117732048 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 vast wheat field illuminated by a warm, setting sun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the ears of wheat depicted as plump, heavy, and bending with their weight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the field show a continuous wave of golden ripples spreading from the near side toward the distant horizon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are a few wisps of sparse white clouds visible in a clear azure sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera move rapidly with a slight motion blur at the edges 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": [ + -0.281310111284256 + ], + "opposite_logprob_tokens": [ + -1.4063100814819336 + ], + "logprob_reward": -0.281310111284256, + "yes_no_margin_reward": 1.125, + "token_credit_reward": -0.281310111284256, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15624748170375824, + 0.5222592353820801, + 0.016075437888503075, + 0.03995257243514061, + 0.011100896634161472, + 0.035587623715400696, + 0.13707564771175385, + 0.022396937012672424, + 0.050880979746580124, + 0.008423198014497757 + ], + "credit_map": [ + 0.012078146450221539, + 0.019508637487888336, + 0.025883710011839867, + 0.023277508094906807, + 0.0097678042948246, + 0.0176802147179842, + 0.016513392329216003, + 0.03153806924819946, + 0.09062038362026215, + 0.18560144305229187, + 0.08200961351394653, + 0.0336315892636776, + 0.06098246946930885, + 0.028921615332365036, + 0.019778665155172348, + 0.02071344666182995, + 0.0009614316513761878, + 0.00134363304823637, + 0.002405590610578656, + 0.0036017445381730795, + 0.0013012582203373313, + 0.0011000866070389748, + 0.0035913758911192417, + 0.001770318136550486, + 0.006126317661255598, + 0.009151194244623184, + 0.0042892019264400005, + 0.0055789640173316, + 0.0031794565729796886, + 0.0022112817969173193, + 0.005765036214143038, + 0.0036511195357888937, + 0.000853346602525562, + 0.0008305170340463519, + 0.002097165212035179, + 0.0022487863898277283, + 0.0007469863630831242, + 0.0010179163655266166, + 0.001841595396399498, + 0.0014645836781710386, + 0.002882051980122924, + 0.004281771369278431, + 0.004416235722601414, + 0.004491051193326712, + 0.00578297721222043, + 0.00760907307267189, + 0.003564530285075307, + 0.0025599333457648754, + 0.029172010719776154, + 0.054506346583366394, + 0.011691850610077381, + 0.006080377381294966, + 0.01829640567302704, + 0.00778717827051878, + 0.0038950196467339993, + 0.005646475125104189, + 0.0015162305207923055, + 0.001627685152925551, + 0.003606486599892378, + 0.0030454080551862717, + 0.0024443708825856447, + 0.0033075197134166956, + 0.0037174841854721308, + 0.003131750738248229, + 0.0073706950061023235, + 0.012082030065357685, + 0.002957433694973588, + 0.0027996201533824205, + 0.009745065122842789, + 0.007974529638886452, + 0.0038447033148258924, + 0.004106904845684767, + 0.000868680072017014, + 0.0008758401381783187, + 0.0021773476619273424, + 0.0012868904741480947, + 0.0005351954605430365, + 0.0005696489824913442, + 0.001171461888588965, + 0.0009381337440572679 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 6 + ], + "selected_reward": -0.281310111284256 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-flying drone camera zooms in at breakneck speed, passing close to the top of a vast wheat field. The warm, setting sun of golden hour illuminates the field, a deep, brilliant amber-gold. Each ear of wheat, plump and heavy, bends with its weight, its awns gleaming with tiny shimmers in the sunlight. A gentle breeze sweeps across the vast wheat field, creating a continuous, continuous wave of golden ripples that rhythmically spread from the near side of the camera toward the distant horizon, a breathtaking sight. A few wisps of sparse white clouds punctuate the clear azure sky, creating a stark contrast to the boundless golden ocean below. The rapid camera movement creates a slight motion blur at the edges of the frame, heightening the sense of speed. The entire scene is imbued with the joy of harvest and the grandeur of nature, creating an atmosphere of freedom, openness, and a tranquil, vibrant vitality.", + "seed": 31988, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000056/group_8/index_9914_seed_31988_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": -0.11217712555080653, + "vlm_logprob": -0.11217712555080653, + "vlm_yes_no_margin": 2.95, + "vlm_token_credit": -0.11217712555080653, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 vast wheat field illuminated by a warm, setting sun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the ears of wheat depicted as plump, heavy, and bending with their weight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the field show a continuous wave of golden ripples spreading from the near side toward the distant horizon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are a few wisps of sparse white clouds visible in a clear azure sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera move rapidly with a slight motion blur at the edges 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": [ + -0.008192387409508228 + ], + "opposite_logprob_tokens": [ + -5.508192539215088 + ], + "logprob_reward": -0.008192387409508228, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.008192387409508228, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2627837657928467, + 0.07364162057638168, + 0.06789523363113403, + 0.09538953751325607, + 0.06922214478254318, + 0.04372056946158409, + 0.06317788362503052, + 0.05628667399287224, + 0.09674715995788574, + 0.17113539576530457 + ], + "credit_map": [ + 0.01396333146840334, + 0.01536463387310505, + 0.09278444200754166, + 0.04082856327295303, + 0.011745037510991096, + 0.014027124270796776, + 0.03281552344560623, + 0.04125509411096573, + 0.008645325899124146, + 0.008156524039804935, + 0.007988548837602139, + 0.008705830201506615, + 0.00888940878212452, + 0.010475361719727516, + 0.011814640834927559, + 0.008965982124209404, + 0.00465521402657032, + 0.004992087371647358, + 0.01066433172672987, + 0.0111393379047513, + 0.004642321262508631, + 0.0039564515464007854, + 0.016881395131349564, + 0.010964090004563332, + 0.004189318977296352, + 0.004361353814601898, + 0.017208777368068695, + 0.016581403091549873, + 0.009437319822609425, + 0.006617933511734009, + 0.019417976960539818, + 0.017575453966856003, + 0.0036994670517742634, + 0.006337435450404882, + 0.015283905901014805, + 0.014110686257481575, + 0.008202699013054371, + 0.006214405409991741, + 0.006238680332899094, + 0.009134862571954727, + 0.0020276042632758617, + 0.002732972614467144, + 0.005621988791972399, + 0.00818499457091093, + 0.006489652208983898, + 0.003789077512919903, + 0.007472533732652664, + 0.007401746232062578, + 0.0033079213462769985, + 0.0049633728340268135, + 0.008403043262660503, + 0.0074636670760810375, + 0.012808958999812603, + 0.007534549105912447, + 0.008031263947486877, + 0.0106651084497571, + 0.0028213602490723133, + 0.0058443727903068066, + 0.010705194436013699, + 0.010350527241826057, + 0.007501161191612482, + 0.0030542598105967045, + 0.008713052608072758, + 0.007296746131032705, + 0.004506771452724934, + 0.005701569840312004, + 0.015574364922940731, + 0.016989542171359062, + 0.006195865571498871, + 0.007593982387334108, + 0.018797311931848526, + 0.021387746557593346, + 0.010383418761193752, + 0.019932091236114502, + 0.021272312849760056, + 0.03178419545292854, + 0.01825309544801712, + 0.012135069817304611, + 0.03195115551352501, + 0.025424057617783546 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 8 + ], + "selected_reward": -0.008192387409508228 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 vast wheat field illuminated by a warm, setting sun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the ears of wheat depicted as plump, heavy, and bending with their weight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the field show a continuous wave of golden ripples spreading from the near side toward the distant horizon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are a few wisps of sparse white clouds visible in a clear azure sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera move rapidly with a slight motion blur at the edges 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": [ + -0.2521343529224396 + ], + "opposite_logprob_tokens": [ + -1.5021343231201172 + ], + "logprob_reward": -0.2521343529224396, + "yes_no_margin_reward": 1.25, + "token_credit_reward": -0.2521343529224396, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07123899459838867, + 0.07736805826425552, + 0.04076677933335304, + 0.15316344797611237, + 0.10312444716691971, + 0.13004174828529358, + 0.20725910365581512, + 0.033965498208999634, + 0.11279954761266708, + 0.07027236372232437 + ], + "credit_map": [ + 0.0034229005686938763, + 0.00357432896271348, + 0.021228665485978127, + 0.012242001481354237, + 0.0052772630006074905, + 0.005297187250107527, + 0.010129587724804878, + 0.01006705779582262, + 0.0033977751154452562, + 0.0036868765018880367, + 0.011509543284773827, + 0.01376572996377945, + 0.007904179394245148, + 0.007859665900468826, + 0.015637416392564774, + 0.013606869615614414, + 0.0015265782130882144, + 0.002193654654547572, + 0.00518833426758647, + 0.006668334826827049, + 0.003631525905802846, + 0.003468381706625223, + 0.012086286209523678, + 0.00600368482992053, + 0.005822273902595043, + 0.006432622205466032, + 0.030390802770853043, + 0.03178554028272629, + 0.011822937056422234, + 0.007434032391756773, + 0.03061571903526783, + 0.028859520331025124, + 0.009985665790736675, + 0.01334460824728012, + 0.0206157099455595, + 0.02206113189458847, + 0.008825172670185566, + 0.007491051219403744, + 0.010701939463615417, + 0.010099164210259914, + 0.004985443782061338, + 0.013955075293779373, + 0.015852920711040497, + 0.02240506187081337, + 0.02186172641813755, + 0.008522886782884598, + 0.024586299434304237, + 0.01787232980132103, + 0.01308710128068924, + 0.012983899563550949, + 0.028664853423833847, + 0.03619727864861488, + 0.026645787060260773, + 0.01707935705780983, + 0.03792789205908775, + 0.03467295318841934, + 0.0017548305913805962, + 0.0029603291768580675, + 0.006597746629267931, + 0.006359400227665901, + 0.0032451178412884474, + 0.0016561257652938366, + 0.0067952703684568405, + 0.004596675280481577, + 0.00876730214804411, + 0.008092612028121948, + 0.022618381306529045, + 0.01897747628390789, + 0.011697984300553799, + 0.007632984779775143, + 0.01717381924390793, + 0.0178389810025692, + 0.0034405484329909086, + 0.007863209582865238, + 0.00984589196741581, + 0.009504922665655613, + 0.006570685654878616, + 0.0034452732652425766, + 0.021035781130194664, + 0.008566055446863174 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 3, + 5 + ], + "selected_reward": -0.2521343529224396 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 vast wheat field illuminated by a warm, setting sun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the ears of wheat depicted as plump, heavy, and bending with their weight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the field show a continuous wave of golden ripples spreading from the near side toward the distant horizon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are a few wisps of sparse white clouds visible in a clear azure sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera move rapidly with a slight motion blur at the edges 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": [ + -0.026339229196310043 + ], + "opposite_logprob_tokens": [ + -3.651339292526245 + ], + "logprob_reward": -0.026339229196310043, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026339229196310043, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21646571159362793, + 0.035964883863925934, + 0.10611783713102341, + 0.04748222231864929, + 0.09391772747039795, + 0.07237360626459122, + 0.07984185963869095, + 0.05695287883281708, + 0.059282608330249786, + 0.23160065710544586 + ], + "credit_map": [ + 0.010272138752043247, + 0.006143870297819376, + 0.0922924131155014, + 0.030273934826254845, + 0.006706967018544674, + 0.005621742457151413, + 0.016406036913394928, + 0.04874861240386963, + 0.003562080441042781, + 0.00259877135977149, + 0.005517221987247467, + 0.006092123221606016, + 0.003301201853901148, + 0.00516863539814949, + 0.005376769229769707, + 0.004348079673945904, + 0.008816489949822426, + 0.010747142136096954, + 0.01893761195242405, + 0.016394801437854767, + 0.009233388118445873, + 0.009503531269729137, + 0.018155261874198914, + 0.014329602010548115, + 0.0027746320702135563, + 0.0031577893532812595, + 0.007722157519310713, + 0.007133090402930975, + 0.005783090367913246, + 0.004068982321768999, + 0.007084188982844353, + 0.009758294560015202, + 0.006107051856815815, + 0.009365254081785679, + 0.019476132467389107, + 0.023924265056848526, + 0.006611731834709644, + 0.0060926140286028385, + 0.011654749512672424, + 0.01068592444062233, + 0.0024101079907268286, + 0.005870335269719362, + 0.008627811446785927, + 0.010109316557645798, + 0.009786940179765224, + 0.006441857665777206, + 0.01542497705668211, + 0.01370225939899683, + 0.005675435997545719, + 0.005606742110103369, + 0.01492632832378149, + 0.013919837772846222, + 0.011766904965043068, + 0.007071136496961117, + 0.009420370683073997, + 0.01145509909838438, + 0.003308794694021344, + 0.006643505766987801, + 0.007525763474404812, + 0.007887356914579868, + 0.008034233003854752, + 0.0031248037703335285, + 0.010774717666208744, + 0.009653703309595585, + 0.004406769759953022, + 0.00414803484454751, + 0.012185252271592617, + 0.010559646412730217, + 0.004259580280631781, + 0.004251091741025448, + 0.010638806037604809, + 0.008833431638777256, + 0.012314566411077976, + 0.02535429038107395, + 0.024824654683470726, + 0.04187091439962387, + 0.030696948990225792, + 0.01604992337524891, + 0.040774039924144745, + 0.03971531242132187 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 2 + ], + "selected_reward": -0.026339229196310043 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 vast wheat field illuminated by a warm, setting sun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the ears of wheat depicted as plump, heavy, and bending with their weight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the field show a continuous wave of golden ripples spreading from the near side toward the distant horizon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are a few wisps of sparse white clouds visible in a clear azure sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera move rapidly with a slight motion blur at the edges 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": [ + -0.04865546151995659 + ], + "opposite_logprob_tokens": [ + -3.0486555099487305 + ], + "logprob_reward": -0.04865546151995659, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04865546151995659, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17068912088871002, + 0.05531930550932884, + 0.18227547407150269, + 0.05740196257829666, + 0.15322370827198029, + 0.06889481097459793, + 0.099985770881176, + 0.09121610224246979, + 0.017487704753875732, + 0.10350605845451355 + ], + "credit_map": [ + 0.019406700506806374, + 0.01835455186665058, + 0.04868016764521599, + 0.021997548639774323, + 0.014857327565550804, + 0.012169119901955128, + 0.018330348655581474, + 0.016893360763788223, + 0.003432004014030099, + 0.004826878663152456, + 0.008091500028967857, + 0.009310868568718433, + 0.003367608180269599, + 0.004892341792583466, + 0.01026807352900505, + 0.011130030266940594, + 0.019367406144738197, + 0.021313974633812904, + 0.0338166318833828, + 0.028773121535778046, + 0.013578175567090511, + 0.015398357063531876, + 0.026172570884227753, + 0.023855241015553474, + 0.00308510335162282, + 0.003316713497042656, + 0.009914744645357132, + 0.011029466055333614, + 0.005828354507684708, + 0.005728664807975292, + 0.007837323471903801, + 0.0106615936383605, + 0.009029149077832699, + 0.015595739707350731, + 0.03382560610771179, + 0.049181342124938965, + 0.009372360073029995, + 0.00867337454110384, + 0.01272554136812687, + 0.014820600859820843, + 0.002673274837434292, + 0.007293889299035072, + 0.0084561537951231, + 0.013280409388244152, + 0.0133228600025177, + 0.005986550357192755, + 0.010905850678682327, + 0.0069758216850459576, + 0.004674647469073534, + 0.009263432584702969, + 0.012078550644218922, + 0.012397289276123047, + 0.024839062243700027, + 0.01034266036003828, + 0.013945850543677807, + 0.012444281950592995, + 0.004225223325192928, + 0.009519090875983238, + 0.01449424959719181, + 0.009355146437883377, + 0.026752429082989693, + 0.011481883004307747, + 0.008253411389887333, + 0.007134663872420788, + 0.0013682597782462835, + 0.0016660813707858324, + 0.003998116124421358, + 0.003196580335497856, + 0.0012495284900069237, + 0.0009580295882187784, + 0.002320405561476946, + 0.0027307034470140934, + 0.006183141842484474, + 0.011291862465441227, + 0.013639961369335651, + 0.02176651731133461, + 0.010083884000778198, + 0.00675763376057148, + 0.020178331062197685, + 0.0136047238484025 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 4 + ], + "selected_reward": -0.04865546151995659 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 vast wheat field illuminated by a warm, setting sun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the ears of wheat depicted as plump, heavy, and bending with their weight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the field show a continuous wave of golden ripples spreading from the near side toward the distant horizon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are a few wisps of sparse white clouds visible in a clear azure sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera move rapidly with a slight motion blur at the edges 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": [ + -0.22556419670581818 + ], + "opposite_logprob_tokens": [ + -1.6005642414093018 + ], + "logprob_reward": -0.22556419670581818, + "yes_no_margin_reward": 1.375, + "token_credit_reward": -0.22556419670581818, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07960060983896255, + 0.025470584630966187, + 0.0336073637008667, + 0.02860514633357525, + 0.08076157420873642, + 0.0559212863445282, + 0.023560279980301857, + 0.04669656977057457, + 0.08912163227796555, + 0.5366549491882324 + ], + "credit_map": [ + 0.0032699399162083864, + 0.004277302883565426, + 0.029138263314962387, + 0.00974030327051878, + 0.0038307509385049343, + 0.0035572119522839785, + 0.007826720364391804, + 0.017960112541913986, + 0.002173610497266054, + 0.0018056902335956693, + 0.003525134874507785, + 0.004425547085702419, + 0.003038245951756835, + 0.00405628327280283, + 0.003321755211800337, + 0.0031243166886270046, + 0.002584802685305476, + 0.0032887058332562447, + 0.006494652945548296, + 0.005673770327121019, + 0.0025193109177052975, + 0.0027543064206838608, + 0.005794366355985403, + 0.004497448913753033, + 0.001590508152730763, + 0.002778273541480303, + 0.004357627127319574, + 0.005320047028362751, + 0.0033833011984825134, + 0.003179115243256092, + 0.004485107026994228, + 0.0035111671313643456, + 0.0030459887348115444, + 0.0048631057143211365, + 0.01937405951321125, + 0.025087488815188408, + 0.005662507843226194, + 0.0047574713826179504, + 0.009990761987864971, + 0.00798018742352724, + 0.0019507523393258452, + 0.004942283965647221, + 0.007190175820142031, + 0.009396864101290703, + 0.007284136489033699, + 0.003607257269322872, + 0.0122914994135499, + 0.009258314967155457, + 0.001663863891735673, + 0.0016746489563956857, + 0.004555060062557459, + 0.004224954638630152, + 0.003342407988384366, + 0.002027330920100212, + 0.0029207810293883085, + 0.0031512323766946793, + 0.0029853368178009987, + 0.007005100138485432, + 0.004559649620205164, + 0.004208400845527649, + 0.01274101436138153, + 0.005083722062408924, + 0.005946565885096788, + 0.004166781902313232, + 0.008192325010895729, + 0.006821285467594862, + 0.019362788647413254, + 0.014115880243480206, + 0.009009825065732002, + 0.007636041380465031, + 0.01333741843700409, + 0.010646067559719086, + 0.026325218379497528, + 0.0567542128264904, + 0.05779007449746132, + 0.09663747996091843, + 0.07435119897127151, + 0.03645210713148117, + 0.105778768658638, + 0.08256584405899048 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 4 + ], + "selected_reward": -0.22556419670581818 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-flying drone camera zooms in at breakneck speed, passing close to the top of a vast wheat field. The warm, setting sun of golden hour illuminates the field, a deep, brilliant amber-gold. Each ear of wheat, plump and heavy, bends with its weight, its awns gleaming with tiny shimmers in the sunlight. A gentle breeze sweeps across the vast wheat field, creating a continuous, continuous wave of golden ripples that rhythmically spread from the near side of the camera toward the distant horizon, a breathtaking sight. A few wisps of sparse white clouds punctuate the clear azure sky, creating a stark contrast to the boundless golden ocean below. The rapid camera movement creates a slight motion blur at the edges of the frame, heightening the sense of speed. The entire scene is imbued with the joy of harvest and the grandeur of nature, creating an atmosphere of freedom, openness, and a tranquil, vibrant vitality.", + "seed": 31989, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000056/group_8/index_9914_seed_31989_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": -0.22492557298392057, + "vlm_logprob": -0.22492557298392057, + "vlm_yes_no_margin": 2.4, + "vlm_token_credit": -0.22492557298392057, + "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 vast wheat field illuminated by a warm, setting sun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the ears of wheat depicted as plump, heavy, and bending with their weight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the field show a continuous wave of golden ripples spreading from the near side toward the distant horizon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are a few wisps of sparse white clouds visible in a clear azure sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera move rapidly with a slight motion blur at the edges 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": [ + -0.009558270685374737 + ], + "opposite_logprob_tokens": [ + -5.009558200836182 + ], + "logprob_reward": -0.009558270685374737, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.009558270685374737, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04279417172074318, + 0.05705752596259117, + 0.027167564257979393, + 0.010029247030615807, + 0.0648641586303711, + 0.6336819529533386, + 0.04010371118783951, + 0.019281839951872826, + 0.09885650873184204, + 0.0061632501892745495 + ], + "credit_map": [ + 0.00525197759270668, + 0.006696846801787615, + 0.003036736510694027, + 0.003148949472233653, + 0.006720632314682007, + 0.010342302732169628, + 0.003114624647423625, + 0.0044821021147072315, + 0.0014706451911479235, + 0.024770401418209076, + 0.000715042871888727, + 0.0009080282761715353, + 0.02449532225728035, + 0.0030481938738375902, + 0.0008709185058251023, + 0.0007789698429405689, + 0.0008655943674966693, + 0.008885870687663555, + 0.0011301537742838264, + 0.001336734276264906, + 0.0095558762550354, + 0.001628093421459198, + 0.0018209312111139297, + 0.001944311079569161, + 0.0009059415315277874, + 0.0016364065231755376, + 0.0008469541207887232, + 0.0011878333752974868, + 0.002332921838387847, + 0.001588087878189981, + 0.0007797316648066044, + 0.0007513694581575692, + 0.0019081997452303767, + 0.028262265026569366, + 0.0005317845498211682, + 0.0009995565051212907, + 0.028524307534098625, + 0.0030060037970542908, + 0.0009772084886208177, + 0.0006548336241394281, + 0.018182041123509407, + 0.2802261412143707, + 0.008235607296228409, + 0.012029316276311874, + 0.26279592514038086, + 0.031562115997076035, + 0.012863085605204105, + 0.007787732407450676, + 0.0009038306307047606, + 0.0159433763474226, + 0.00046704127453267574, + 0.0006698996294289827, + 0.018156498670578003, + 0.002883200766518712, + 0.0005795779288746417, + 0.0005002900725230575, + 0.000585577916353941, + 0.007616827730089426, + 0.0003455818514339626, + 0.00035473459865897894, + 0.0081446198746562, + 0.0013224697904661298, + 0.0004508828860707581, + 0.00046114413999021053, + 0.0028236443176865578, + 0.0437241829931736, + 0.000737625639885664, + 0.0013318151468411088, + 0.04147414118051529, + 0.005786581430584192, + 0.0017887407448142767, + 0.0011897735530510545, + 0.0003597850154619664, + 0.0017968359170481563, + 0.00039478932740166783, + 0.00030368202715180814, + 0.001936942571774125, + 0.00042003486305475235, + 0.0003846603794954717, + 0.0005665199714712799 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 8, + 4 + ], + "selected_reward": -0.009558270685374737 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 vast wheat field illuminated by a warm, setting sun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the ears of wheat depicted as plump, heavy, and bending with their weight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the field show a continuous wave of golden ripples spreading from the near side toward the distant horizon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are a few wisps of sparse white clouds visible in a clear azure sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera move rapidly with a slight motion blur at the edges of the frame?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.7578499913215637 + ], + "opposite_logprob_tokens": [ + -0.6328499913215637 + ], + "logprob_reward": -0.7578499913215637, + "yes_no_margin_reward": -0.125, + "token_credit_reward": -0.7578499913215637, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.19784261286258698, + 0.05785704404115677, + 0.028251569718122482, + 0.033030278980731964, + 0.027191923931241035, + 0.3454414904117584, + 0.01639724150300026, + 0.05139220505952835, + 0.19233852624893188, + 0.050257034599781036 + ], + "credit_map": [ + 0.013474724255502224, + 0.04767655208706856, + 0.011707828380167484, + 0.013429760001599789, + 0.05018685758113861, + 0.031096432358026505, + 0.012352753430604935, + 0.01791771501302719, + 0.0026922551915049553, + 0.023974617943167686, + 0.0013266985770314932, + 0.001955897081643343, + 0.020924024283885956, + 0.0034919297322630882, + 0.002113200956955552, + 0.00137841806281358, + 0.0027546249330043793, + 0.006218058988451958, + 0.002666209591552615, + 0.0025764924939721823, + 0.0040529347024858, + 0.0037039362359791994, + 0.0029547775629907846, + 0.0033245349768549204, + 0.002678623655810952, + 0.008243466727435589, + 0.0019122279481962323, + 0.003166869515553117, + 0.009531152434647083, + 0.0025313408114016056, + 0.002134534763172269, + 0.0028320630080997944, + 0.002671591006219387, + 0.007133897393941879, + 0.00198964006267488, + 0.0017514515202492476, + 0.006117525976151228, + 0.0035286785569041967, + 0.002150073414668441, + 0.001849067397415638, + 0.009990265592932701, + 0.15246392786502838, + 0.004310667049139738, + 0.006790237035602331, + 0.14261256158351898, + 0.017621507868170738, + 0.007218168582767248, + 0.004434156697243452, + 0.0007141955429688096, + 0.0066880653612315655, + 0.00037144022644497454, + 0.00044857346802018583, + 0.006285303272306919, + 0.0010491363937035203, + 0.0004511598381213844, + 0.00038936836062930524, + 0.0014439478982239962, + 0.021788770332932472, + 0.000746803474612534, + 0.0009804946603253484, + 0.02225836180150509, + 0.002379639307036996, + 0.001021434785798192, + 0.0007727511110715568, + 0.005714836996048689, + 0.08520352840423584, + 0.001985413720831275, + 0.0025706219021230936, + 0.07863330841064453, + 0.012271850369870663, + 0.0035643125884234905, + 0.002394648501649499, + 0.0021514229010790586, + 0.016787471249699593, + 0.002387098502367735, + 0.002316058613359928, + 0.018395911902189255, + 0.0029790159314870834, + 0.002430692780762911, + 0.002809362718835473 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 8 + ], + "selected_reward": -0.7578499913215637 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 vast wheat field illuminated by a warm, setting sun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the ears of wheat depicted as plump, heavy, and bending with their weight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the field show a continuous wave of golden ripples spreading from the near side toward the distant horizon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are a few wisps of sparse white clouds visible in a clear azure sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera move rapidly with a slight motion blur at the edges 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": [ + -0.06204582005739212 + ], + "opposite_logprob_tokens": [ + -2.8120458126068115 + ], + "logprob_reward": -0.06204582005739212, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06204582005739212, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05515909567475319, + 0.034800924360752106, + 0.02970198541879654, + 0.012328899465501308, + 0.005122140981256962, + 0.6175035834312439, + 0.10202832520008087, + 0.0026186828035861254, + 0.11818542331457138, + 0.022550921887159348 + ], + "credit_map": [ + 0.004398040939122438, + 0.019184699282050133, + 0.0026153188664466143, + 0.002444826764985919, + 0.0156403798609972, + 0.003917523194104433, + 0.003531065071001649, + 0.0034272398333996534, + 0.0014016723725944757, + 0.012734913267195225, + 0.0015568441012874246, + 0.0019251107005402446, + 0.01183103583753109, + 0.001831844332627952, + 0.0020674862898886204, + 0.0014520183904096484, + 0.002509601879864931, + 0.004042804706841707, + 0.0027010838966816664, + 0.0023093661293387413, + 0.006745617836713791, + 0.004392659291625023, + 0.003680518828332424, + 0.0033203335478901863, + 0.0010650415206328034, + 0.0016114547615870833, + 0.0011337199248373508, + 0.001154569094069302, + 0.002492309780791402, + 0.002329594688490033, + 0.0011593590024858713, + 0.0013828501105308533, + 0.0005588146741501987, + 0.0008605107432231307, + 0.0004519437497947365, + 0.0003275145427323878, + 0.0011141597060486674, + 0.0008054099744185805, + 0.0005430945893749595, + 0.00046069282689131796, + 0.017809348180890083, + 0.2729056775569916, + 0.008122850209474564, + 0.011788892559707165, + 0.255687952041626, + 0.030869435518980026, + 0.012666312046349049, + 0.007653144653886557, + 0.0029048367869108915, + 0.04134520888328552, + 0.001356220105662942, + 0.001981868175789714, + 0.045198194682598114, + 0.005821833852678537, + 0.0018077784916386008, + 0.0016123855020850897, + 0.00026096487999893725, + 0.0005385702825151384, + 0.00016547898121643811, + 0.0001760784216457978, + 0.0005906158476136625, + 0.0004715894756373018, + 0.00021677964832633734, + 0.00019860515021719038, + 0.0033505228348076344, + 0.05256999284029007, + 0.0009923746110871434, + 0.0015616348246112466, + 0.04924841597676277, + 0.0070753623731434345, + 0.0020652548409998417, + 0.0013218576787039638, + 0.000809544522780925, + 0.007688835263252258, + 0.0007406551740132272, + 0.0007860501063987613, + 0.008869757875800133, + 0.001645235694013536, + 0.0009251951123587787, + 0.0010856465669348836 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 8, + 6 + ], + "selected_reward": -0.06204582005739212 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 vast wheat field illuminated by a warm, setting sun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the ears of wheat depicted as plump, heavy, and bending with their weight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the field show a continuous wave of golden ripples spreading from the near side toward the distant horizon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are a few wisps of sparse white clouds visible in a clear azure sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera move rapidly with a slight motion blur at the edges 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": [ + -0.043070364743471146 + ], + "opposite_logprob_tokens": [ + -3.1680703163146973 + ], + "logprob_reward": -0.043070364743471146, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.043070364743471146, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23102861642837524, + 0.059207405894994736, + 0.010574718937277794, + 0.010412012226879597, + 0.11439679563045502, + 0.3157135546207428, + 0.016254764050245285, + 0.03410091623663902, + 0.1998014599084854, + 0.008509722538292408 + ], + "credit_map": [ + 0.011259406805038452, + 0.07319577783346176, + 0.008460377342998981, + 0.013905530795454979, + 0.06732513755559921, + 0.034919727593660355, + 0.007714719511568546, + 0.014247956685721874, + 0.0037566900718957186, + 0.022151904180645943, + 0.0014429748989641666, + 0.0015778219094499946, + 0.021734019741415977, + 0.005346143618226051, + 0.00177007878664881, + 0.0014277738519012928, + 0.0008831529412418604, + 0.002766235498711467, + 0.0006018991116434336, + 0.0006753348279744387, + 0.0031919965986162424, + 0.001150391180999577, + 0.0007446350064128637, + 0.0005610738298855722, + 0.0011419004295021296, + 0.0024376623332500458, + 0.0005471690092235804, + 0.0008598436834290624, + 0.0030631949193775654, + 0.0012638710904866457, + 0.00045421256800182164, + 0.0006441579316742718, + 0.003231605514883995, + 0.05114179104566574, + 0.0007362514734268188, + 0.0016236029332503676, + 0.04936958849430084, + 0.005470798816531897, + 0.0017975557129830122, + 0.0010255991946905851, + 0.009149475954473019, + 0.13969896733760834, + 0.004193617030978203, + 0.0061265514232218266, + 0.12985871732234955, + 0.016204390674829483, + 0.00664327060803771, + 0.003838583594188094, + 0.0005691786063835025, + 0.006206010468304157, + 0.00030945343314670026, + 0.0003107118245679885, + 0.006843214388936758, + 0.0013293165247887373, + 0.00034471892286092043, + 0.0003421585715841502, + 0.0014217097777873278, + 0.010657858103513718, + 0.0007811106042936444, + 0.0006112883565947413, + 0.014243487268686295, + 0.00437956815585494, + 0.0008219314622692764, + 0.001183962100185454, + 0.005623821169137955, + 0.08889268338680267, + 0.0015077776042744517, + 0.0025125653482973576, + 0.08403503894805908, + 0.011677442118525505, + 0.0033378947991877794, + 0.002214247127994895, + 0.0007681067800149322, + 0.0014273080741986632, + 0.0004560181114356965, + 0.0005121613503433764, + 0.0027288913261145353, + 0.001668252283707261, + 0.0004596476792357862, + 0.0004893361474387348 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 8 + ], + "selected_reward": -0.043070364743471146 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 vast wheat field illuminated by a warm, setting sun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the ears of wheat depicted as plump, heavy, and bending with their weight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the field show a continuous wave of golden ripples spreading from the near side toward the distant horizon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are a few wisps of sparse white clouds visible in a clear azure sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera move rapidly with a slight motion blur at the edges 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": [ + -0.25210341811180115 + ], + "opposite_logprob_tokens": [ + -1.5021034479141235 + ], + "logprob_reward": -0.25210341811180115, + "yes_no_margin_reward": 1.25, + "token_credit_reward": -0.25210341811180115, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14640198647975922, + 0.026401981711387634, + 0.03190819174051285, + 0.02473285235464573, + 0.009601732715964317, + 0.4934028685092926, + 0.14163604378700256, + 0.01458543911576271, + 0.10246624797582626, + 0.008862708695232868 + ], + "credit_map": [ + 0.0064884512685239315, + 0.049763694405555725, + 0.00564704742282629, + 0.006236055865883827, + 0.042558811604976654, + 0.020478632301092148, + 0.006039580795913935, + 0.009189696982502937, + 0.001754522556439042, + 0.0034518323373049498, + 0.003567121224477887, + 0.0038006023969501257, + 0.003975487779825926, + 0.002230284968391061, + 0.0044187502935528755, + 0.0032033768948167562, + 0.0014422998065128922, + 0.009739218279719353, + 0.0017142584547400475, + 0.00158953748177737, + 0.00972583144903183, + 0.002203436568379402, + 0.0020534449722617865, + 0.003440162166953087, + 0.002111302223056555, + 0.005003283731639385, + 0.00144108384847641, + 0.0019853769335895777, + 0.007867413572967052, + 0.0031345391180366278, + 0.0016402426408603787, + 0.001549608539789915, + 0.0010477141477167606, + 0.0013969197170808911, + 0.0006709271692670882, + 0.0006741441320627928, + 0.002585822716355324, + 0.001754998229444027, + 0.0008519825059920549, + 0.0006192239816300571, + 0.014475615695118904, + 0.21739760041236877, + 0.00684879207983613, + 0.009449241682887077, + 0.2031019926071167, + 0.025082377716898918, + 0.010295030660927296, + 0.0067522102035582066, + 0.0032328013330698013, + 0.056242603808641434, + 0.0024382877163589, + 0.0030851978808641434, + 0.06244828179478645, + 0.0085701122879982, + 0.003070639446377754, + 0.0025481029879301786, + 0.0004509272985160351, + 0.00434526614844799, + 0.0004635147633962333, + 0.0004346561909187585, + 0.00580636179074645, + 0.0019166615093126893, + 0.0006110438844189048, + 0.0005570058710873127, + 0.0031208705622702837, + 0.04569222778081894, + 0.0007694031810387969, + 0.0013635017676278949, + 0.04277345538139343, + 0.005811415147036314, + 0.0017718642484396696, + 0.0011635107221081853, + 0.0004161315446253866, + 0.0028670071624219418, + 0.00039236628799699247, + 0.0004320195585023612, + 0.0031476125586777925, + 0.0007253064541146159, + 0.00045027158921584487, + 0.0004319927538745105 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 6 + ], + "selected_reward": -0.25210341811180115 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "No", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-flying drone camera zooms in at breakneck speed, passing close to the top of a vast wheat field. The warm, setting sun of golden hour illuminates the field, a deep, brilliant amber-gold. Each ear of wheat, plump and heavy, bends with its weight, its awns gleaming with tiny shimmers in the sunlight. A gentle breeze sweeps across the vast wheat field, creating a continuous, continuous wave of golden ripples that rhythmically spread from the near side of the camera toward the distant horizon, a breathtaking sight. A few wisps of sparse white clouds punctuate the clear azure sky, creating a stark contrast to the boundless golden ocean below. The rapid camera movement creates a slight motion blur at the edges of the frame, heightening the sense of speed. The entire scene is imbued with the joy of harvest and the grandeur of nature, creating an atmosphere of freedom, openness, and a tranquil, vibrant vitality.", + "seed": 31990, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000056/group_8/index_9914_seed_31990_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": -0.26090488508343695, + "vlm_logprob": -0.26090488508343695, + "vlm_yes_no_margin": 2.275, + "vlm_token_credit": -0.26090488508343695, + "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 vast wheat field illuminated by a warm, setting sun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the ears of wheat depicted as plump, heavy, and bending with their weight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the field show a continuous wave of golden ripples spreading from the near side toward the distant horizon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are a few wisps of sparse white clouds visible in a clear azure sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera move rapidly with a slight motion blur at the edges 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": [ + -0.019438251852989197 + ], + "opposite_logprob_tokens": [ + -4.26943826675415 + ], + "logprob_reward": -0.019438251852989197, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.019438251852989197, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23150084912776947, + 0.05696556717157364, + 0.040316805243492126, + 0.02105860412120819, + 0.031733665615320206, + 0.02263025753200054, + 0.04397047311067581, + 0.5111550688743591, + 0.01553034782409668, + 0.025138335302472115 + ], + "credit_map": [ + 0.01691182516515255, + 0.018280481919646263, + 0.027901889756321907, + 0.028177810832858086, + 0.0420338474214077, + 0.020342664793133736, + 0.03522811084985733, + 0.042624227702617645, + 0.006721653044223785, + 0.0028566375840455294, + 0.008519038558006287, + 0.01033172756433487, + 0.004644737113267183, + 0.0032090297900140285, + 0.012815800495445728, + 0.007866941392421722, + 0.002368978690356016, + 0.0032087182626128197, + 0.006314744241535664, + 0.006612516473978758, + 0.006084662862122059, + 0.0029702577739953995, + 0.007499076891690493, + 0.005257845390588045, + 0.0013935810420662165, + 0.0012827442260459065, + 0.0023422259837388992, + 0.004472041502594948, + 0.0023863266687840223, + 0.0010831990512087941, + 0.00447062635794282, + 0.003627859754487872, + 0.0016073104925453663, + 0.0016075813909992576, + 0.004242046736180782, + 0.007309143431484699, + 0.0028422351460903883, + 0.0022174196783453226, + 0.006437364500015974, + 0.005470563191920519, + 0.002701791003346443, + 0.0015419508563354611, + 0.004648224916309118, + 0.0029268499929457903, + 0.002502606948837638, + 0.0017443683464080095, + 0.0038110711611807346, + 0.0027533944230526686, + 0.0026429847348481417, + 0.002607041737064719, + 0.00556140998378396, + 0.00909377820789814, + 0.00668451888486743, + 0.003598707728087902, + 0.006842140108346939, + 0.006939890328794718, + 0.014995725825428963, + 0.01833947002887726, + 0.10503459721803665, + 0.14954443275928497, + 0.02393844537436962, + 0.02714821696281433, + 0.08783985674381256, + 0.08431431651115417, + 0.0013289084890857339, + 0.000841108150780201, + 0.0018727477872744203, + 0.003155552549287677, + 0.00219356594607234, + 0.0008512582862749696, + 0.0028478254098445177, + 0.002439380856230855, + 0.002715532435104251, + 0.00193820393178612, + 0.0032256469130516052, + 0.005737771280109882, + 0.002029796363785863, + 0.0014232839457690716, + 0.003961835987865925, + 0.004106265027076006 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 1 + ], + "selected_reward": -0.019438251852989197 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 vast wheat field illuminated by a warm, setting sun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the ears of wheat depicted as plump, heavy, and bending with their weight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the field show a continuous wave of golden ripples spreading from the near side toward the distant horizon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are a few wisps of sparse white clouds visible in a clear azure sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera move rapidly with a slight motion blur at the edges of the frame?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.8262233734130859 + ], + "opposite_logprob_tokens": [ + -0.5762233734130859 + ], + "logprob_reward": -0.8262233734130859, + "yes_no_margin_reward": -0.25, + "token_credit_reward": -0.8262233734130859, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.150329127907753, + 0.055260591208934784, + 0.05556248500943184, + 0.023469189181923866, + 0.06456141173839569, + 0.07684068381786346, + 0.04392458125948906, + 0.42303895950317383, + 0.03111460618674755, + 0.07589839398860931 + ], + "credit_map": [ + 0.01025941502302885, + 0.011936441995203495, + 0.022459713742136955, + 0.02206408418715, + 0.014950580894947052, + 0.0077925375662744045, + 0.027186231687664986, + 0.033680129796266556, + 0.004731436725705862, + 0.003184342524036765, + 0.007080955896526575, + 0.009950234554708004, + 0.005754446145147085, + 0.004184851422905922, + 0.011817715130746365, + 0.008556614629924297, + 0.006013307720422745, + 0.00358606968075037, + 0.007083539385348558, + 0.009496812708675861, + 0.005520435981452465, + 0.0040960065089166164, + 0.011205793358385563, + 0.008560518734157085, + 0.001785980653949082, + 0.0015461569419130683, + 0.0031053279526531696, + 0.004007779061794281, + 0.0024680010974407196, + 0.0012002333533018827, + 0.004454776179045439, + 0.004900934640318155, + 0.0035117133520543575, + 0.0036108416970819235, + 0.008565477095544338, + 0.015113581903278828, + 0.00900348648428917, + 0.0037743505090475082, + 0.013698198832571507, + 0.0072837648913264275, + 0.006422469858080149, + 0.003914669621735811, + 0.010282137431204319, + 0.011254391632974148, + 0.005970936734229326, + 0.003518741810694337, + 0.025672487914562225, + 0.009804843924939632, + 0.0038792318664491177, + 0.002718812320381403, + 0.007443776819854975, + 0.008716816082596779, + 0.0032948425505310297, + 0.0021486200857907534, + 0.007930899038910866, + 0.007791582960635424, + 0.010894760489463806, + 0.015865309163928032, + 0.0858093798160553, + 0.12096598744392395, + 0.01726246252655983, + 0.022977575659751892, + 0.0772024467587471, + 0.07206104695796967, + 0.0026978496462106705, + 0.002743592718616128, + 0.004298302810639143, + 0.004269871395081282, + 0.004070066846907139, + 0.0018818769603967667, + 0.0060636610724031925, + 0.0050893849693238735, + 0.003436356782913208, + 0.004051096737384796, + 0.010496138595044613, + 0.014996824786067009, + 0.007638442795723677, + 0.004863867536187172, + 0.019239718094468117, + 0.01117594912648201 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 5 + ], + "selected_reward": -0.8262233734130859 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 vast wheat field illuminated by a warm, setting sun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the ears of wheat depicted as plump, heavy, and bending with their weight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the field show a continuous wave of golden ripples spreading from the near side toward the distant horizon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are a few wisps of sparse white clouds visible in a clear azure sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera move rapidly with a slight motion blur at the edges 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": [ + -0.038097646087408066 + ], + "opposite_logprob_tokens": [ + -3.288097620010376 + ], + "logprob_reward": -0.038097646087408066, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.038097646087408066, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10285644233226776, + 0.06271404027938843, + 0.023274077102541924, + 0.019998058676719666, + 0.04680227115750313, + 0.040240924805402756, + 0.026329491287469864, + 0.6245442032814026, + 0.02419002540409565, + 0.029050510376691818 + ], + "credit_map": [ + 0.006866260897368193, + 0.011237159371376038, + 0.011365436017513275, + 0.014263338409364223, + 0.018198220059275627, + 0.0060009039007127285, + 0.014101075939834118, + 0.020824050530791283, + 0.004510880913585424, + 0.002659603487700224, + 0.009334947913885117, + 0.013829096220433712, + 0.004056181758642197, + 0.0028145299293100834, + 0.014526855200529099, + 0.010981947183609009, + 0.001529097673483193, + 0.0016565585974603891, + 0.002960935467854142, + 0.004970632493495941, + 0.0031278759706765413, + 0.0017588109476491809, + 0.0037998417392373085, + 0.0034703253768384457, + 0.0011544341687113047, + 0.0009423570591025054, + 0.0029606225434690714, + 0.0042453547939658165, + 0.0016424241475760937, + 0.0011153838131576777, + 0.0038415640592575073, + 0.004095918498933315, + 0.002195782260969281, + 0.0018619729671627283, + 0.007647349499166012, + 0.010974232107400894, + 0.003222699509933591, + 0.0026582791469991207, + 0.009620935656130314, + 0.008621021173894405, + 0.003997928462922573, + 0.002250140532851219, + 0.009664378128945827, + 0.004821902606636286, + 0.0030022526625543833, + 0.002485261531546712, + 0.007021310273557901, + 0.006997749675065279, + 0.0017079438548535109, + 0.0017327051609754562, + 0.003583603072911501, + 0.005873570218682289, + 0.0030387977603822947, + 0.0018073826795443892, + 0.003193051088601351, + 0.005392438732087612, + 0.018414616584777832, + 0.02227484993636608, + 0.12576860189437866, + 0.17750968039035797, + 0.028004266321659088, + 0.032469917088747025, + 0.11897854506969452, + 0.10112372785806656, + 0.001218851888552308, + 0.001018201932311058, + 0.0037893166299909353, + 0.003691192716360092, + 0.0024891027715057135, + 0.0012373261852189898, + 0.006171592976897955, + 0.004574439022690058, + 0.0016042367788031697, + 0.0016874343855306506, + 0.003922321368008852, + 0.006612548604607582, + 0.0029578960966318846, + 0.001787803485058248, + 0.006563923787325621, + 0.00391434645280242 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 1 + ], + "selected_reward": -0.038097646087408066 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 vast wheat field illuminated by a warm, setting sun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the ears of wheat depicted as plump, heavy, and bending with their weight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the field show a continuous wave of golden ripples spreading from the near side toward the distant horizon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are a few wisps of sparse white clouds visible in a clear azure sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera move rapidly with a slight motion blur at the edges 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": [ + -0.03370582312345505 + ], + "opposite_logprob_tokens": [ + -3.408705711364746 + ], + "logprob_reward": -0.03370582312345505, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03370582312345505, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16917218267917633, + 0.13261789083480835, + 0.1495729386806488, + 0.030600737780332565, + 0.06411734968423843, + 0.08792515099048615, + 0.012763590551912785, + 0.28556931018829346, + 0.028925444930791855, + 0.03873540833592415 + ], + "credit_map": [ + 0.021670224145054817, + 0.024194154888391495, + 0.01641509309411049, + 0.01729246787726879, + 0.04319367557764053, + 0.013488579541444778, + 0.01629881002008915, + 0.016619188711047173, + 0.007930020801723003, + 0.007260128855705261, + 0.020016243681311607, + 0.03247666731476784, + 0.009004556573927402, + 0.00764349102973938, + 0.018930427730083466, + 0.029356366023421288, + 0.012346794828772545, + 0.018048550933599472, + 0.01591319963335991, + 0.018550250679254532, + 0.037257786840200424, + 0.014737032353878021, + 0.014741218648850918, + 0.017978113144636154, + 0.0030402028933167458, + 0.002478279173374176, + 0.0033987811766564846, + 0.00368435587733984, + 0.005386441946029663, + 0.0027618808671832085, + 0.006123609375208616, + 0.0037271869368851185, + 0.003555167466402054, + 0.0038541853427886963, + 0.007696656044572592, + 0.012270123697817326, + 0.008088583126664162, + 0.004610979463905096, + 0.014307132922112942, + 0.009734530933201313, + 0.0038488940335810184, + 0.007395899388939142, + 0.013441904447972775, + 0.01118426863104105, + 0.018403274938464165, + 0.007798822596669197, + 0.015035588294267654, + 0.010816499590873718, + 0.0011025257408618927, + 0.0009772282792255282, + 0.001754693454131484, + 0.0018338701920583844, + 0.0012610771227627993, + 0.000859825755469501, + 0.0026260651648044586, + 0.0023483065888285637, + 0.010704686865210533, + 0.012758292257785797, + 0.05500224232673645, + 0.0805833637714386, + 0.02362104132771492, + 0.01713498868048191, + 0.04229550063610077, + 0.043469224125146866, + 0.0028037100564688444, + 0.0015403834404423833, + 0.0033963865134865046, + 0.005374233704060316, + 0.003758771577849984, + 0.0017132095526903868, + 0.0057806153781712055, + 0.00455813854932785, + 0.004846479743719101, + 0.0027129300870001316, + 0.004694180563092232, + 0.008228272199630737, + 0.0031523611396551132, + 0.0025677168741822243, + 0.00706693809479475, + 0.0054665300995111465 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 2 + ], + "selected_reward": -0.03370582312345505 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 vast wheat field illuminated by a warm, setting sun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the ears of wheat depicted as plump, heavy, and bending with their weight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the field show a continuous wave of golden ripples spreading from the near side toward the distant horizon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are a few wisps of sparse white clouds visible in a clear azure sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera move rapidly with a slight motion blur at the edges 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": [ + -0.3870593309402466 + ], + "opposite_logprob_tokens": [ + -1.1370593309402466 + ], + "logprob_reward": -0.3870593309402466, + "yes_no_margin_reward": 0.75, + "token_credit_reward": -0.3870593309402466, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10046496242284775, + 0.12855875492095947, + 0.057542189955711365, + 0.06672122329473495, + 0.1512848287820816, + 0.1656978279352188, + 0.0398329496383667, + 0.1648608148097992, + 0.07922230660915375, + 0.04581416770815849 + ], + "credit_map": [ + 0.007538514211773872, + 0.006757467519491911, + 0.012957841157913208, + 0.017859317362308502, + 0.011378343217074871, + 0.005012771114706993, + 0.018481221050024033, + 0.020479490980505943, + 0.012796257622539997, + 0.007137992884963751, + 0.020764417946338654, + 0.01995568536221981, + 0.00978335551917553, + 0.006717160809785128, + 0.02873150445520878, + 0.022672384977340698, + 0.00449809106066823, + 0.0028694872744381428, + 0.009410101920366287, + 0.01080384012311697, + 0.004229514393955469, + 0.003834127215668559, + 0.013234980404376984, + 0.008662047795951366, + 0.004762591328471899, + 0.004870994482189417, + 0.007997941225767136, + 0.010678963735699654, + 0.011050957255065441, + 0.004181762225925922, + 0.013658636249601841, + 0.009519374929368496, + 0.00756411999464035, + 0.007565724663436413, + 0.019640449434518814, + 0.027041828259825706, + 0.01466305647045374, + 0.010923055931925774, + 0.037658799439668655, + 0.026227787137031555, + 0.007255236618220806, + 0.006619545165449381, + 0.031974900513887405, + 0.025261428207159042, + 0.01219309400767088, + 0.008201641030609608, + 0.04829002171754837, + 0.025901971384882927, + 0.0024619540199637413, + 0.0018493349198251963, + 0.006051290314644575, + 0.008793001994490623, + 0.0035595661029219627, + 0.0026189128402620554, + 0.005876225419342518, + 0.008622664026916027, + 0.009617026895284653, + 0.006112140137702227, + 0.028646430000662804, + 0.04052865505218506, + 0.007525008171796799, + 0.008297569118440151, + 0.030939318239688873, + 0.03319466859102249, + 0.004863070324063301, + 0.00395932188257575, + 0.013110039755702019, + 0.01333855465054512, + 0.009180361405014992, + 0.006761896423995495, + 0.015237816609442234, + 0.012771239504218102, + 0.005856030620634556, + 0.0031309514306485653, + 0.005937524605542421, + 0.00888492539525032, + 0.003241981379687786, + 0.0027369745075702667, + 0.008031855337321758, + 0.00799392256885767 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 7, + 4 + ], + "selected_reward": -0.3870593309402466 + } + ], + "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/0000056/group_9/index_tensor([9769, 9769])_group_9_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000056/group_9/index_tensor([9769, 9769])_group_9_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..408b0b347c55e1b434c532d131a2440a12307812 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000056/group_9/index_tensor([9769, 9769])_group_9_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "slow, cinematic push-in begins in medium close-up, capturing a meticulously detailed, mechanical version of Pomu seated on a futuristic, sun-drenched stage. Her glowing green fiber-optic hair is styled in a high ponytail, with individual pulses of light falling on her face. Her limbs are polished chrome, their surfaces capturing and distorting the vibrant violet and cyan lighting of the environment, revealing the intricate gears and servos at her joints. She leans forward in rapt concentration, her expressive LED eyes, a brilliant emerald green, reflecting the complex digital battlefield—a series of strategic maps, data streams, and miniature explosions. Her chrome hands blur over the translucent three-dimensional holographic interface floating before her; her fingertips, equipped with glowing sensors, move swiftly and decisively, tapping the glowing icons with preternatural speed. The background is a vast stadium, filled with blurred silhouettes and massive holographic screens displaying match footage. The floor beneath her feet is a dark, polished material, creating a rain-smooth reflection of the entire scene, amplifying the cyberpunk aesthetic. The atmosphere is thrilling and tense, with the cool glow of the hologram contrasting with the warm neon lights, all filmed in stunning 8K HD, showcasing every detail of her cybernetic form and the high-stakes esports competition she dominates.", + "seed": 15890, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000056/group_9/index_9769_seed_15890_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": -0.0468664139509201, + "vlm_logprob": -0.0468664139509201, + "vlm_yes_no_margin": 3.475, + "vlm_token_credit": -0.0468664139509201, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a mechanical version of a character named Pomu with polished chrome limbs and glowing green fiber-optic hair in a high ponytail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character seated on a futuristic stage with a dark, polished floor that reflects the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character leaning forward in concentration while interacting with a translucent three-dimensional holographic interface using glowing sensor-equipped hands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a vast stadium filled with blurred silhouettes and massive holographic screens displaying match footage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video begin with a slow, cinematic push-in shot starting from a 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": [ + -0.055880580097436905 + ], + "opposite_logprob_tokens": [ + -2.930880546569824 + ], + "logprob_reward": -0.055880580097436905, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.055880580097436905, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2087385207414627, + 0.06716331839561462, + 0.10035207122564316, + 0.11609488725662231, + 0.056025903671979904, + 0.1551174521446228, + 0.05124308541417122, + 0.05004777014255524, + 0.16690628230571747, + 0.028310788795351982 + ], + "credit_map": [ + 0.02639911323785782, + 0.031075624749064445, + 0.023442210629582405, + 0.023954737931489944, + 0.03299002721905708, + 0.021938970312476158, + 0.02396613173186779, + 0.024971691891551018, + 0.007755144964903593, + 0.00860065408051014, + 0.010391994379460812, + 0.010078446939587593, + 0.008559410460293293, + 0.006913343910127878, + 0.009386616759002209, + 0.0054776971228420734, + 0.009809763170778751, + 0.010943838395178318, + 0.011639527045190334, + 0.02521657571196556, + 0.009481051005423069, + 0.008752694353461266, + 0.009085633791983128, + 0.015422981232404709, + 0.025592273101210594, + 0.011352621018886566, + 0.034908026456832886, + 0.011807135306298733, + 0.008955754339694977, + 0.005928436294198036, + 0.008898996748030186, + 0.008651630021631718, + 0.007793523371219635, + 0.007372025400400162, + 0.005433971062302589, + 0.005950699094682932, + 0.005695650819689035, + 0.005272401496767998, + 0.007608472835272551, + 0.010899154469370842, + 0.0180327408015728, + 0.023279890418052673, + 0.010725251398980618, + 0.023300157859921455, + 0.025712236762046814, + 0.016910498961806297, + 0.01182188093662262, + 0.02533476985991001, + 0.008184131234884262, + 0.0051873899064958096, + 0.007030990440398455, + 0.007154679391533136, + 0.0064483629539608955, + 0.005747244693338871, + 0.005771829281002283, + 0.00571844819933176, + 0.00845311488956213, + 0.0045251925475895405, + 0.005007119849324226, + 0.004874098114669323, + 0.007576046511530876, + 0.006136177573353052, + 0.004688380751758814, + 0.008787636645138264, + 0.052618175745010376, + 0.008723239414393902, + 0.06362639367580414, + 0.01368845533579588, + 0.004248025827109814, + 0.007917394861578941, + 0.006045485381036997, + 0.010039097629487514, + 0.005402006674557924, + 0.003759281476959586, + 0.0026761945337057114, + 0.0025096454191952944, + 0.0035175897646695375, + 0.0033688030671328306, + 0.002745033474639058, + 0.004332232289016247 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 5 + ], + "selected_reward": -0.055880580097436905 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a mechanical version of a character named Pomu with polished chrome limbs and glowing green fiber-optic hair in a high ponytail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character seated on a futuristic stage with a dark, polished floor that reflects the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character leaning forward in concentration while interacting with a translucent three-dimensional holographic interface using glowing sensor-equipped hands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a vast stadium filled with blurred silhouettes and massive holographic screens displaying match footage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video begin with a slow, cinematic push-in shot starting from a 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": [ + -0.005253438837826252 + ], + "opposite_logprob_tokens": [ + -5.255253314971924 + ], + "logprob_reward": -0.005253438837826252, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005253438837826252, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08938289433717728, + 0.08943942934274673, + 0.08297896385192871, + 0.10905236750841141, + 0.2896004915237427, + 0.03674287721514702, + 0.016854694113135338, + 0.03519127145409584, + 0.2342073619365692, + 0.016549663618206978 + ], + "credit_map": [ + 0.012263967655599117, + 0.011138489469885826, + 0.008466167375445366, + 0.015690051019191742, + 0.011778374202549458, + 0.00764223700389266, + 0.0111037977039814, + 0.011299812234938145, + 0.009307307191193104, + 0.015537027269601822, + 0.008550681173801422, + 0.020489875227212906, + 0.007325856480747461, + 0.007115114480257034, + 0.011771421879529953, + 0.009342138655483723, + 0.006908762734383345, + 0.006615606602281332, + 0.006848354358226061, + 0.02978232316672802, + 0.003688785247504711, + 0.004496569279581308, + 0.005620951298624277, + 0.01901761256158352, + 0.03135718032717705, + 0.007780659012496471, + 0.042718466371297836, + 0.0068998029455542564, + 0.003987258765846491, + 0.0033397299703210592, + 0.004407481290400028, + 0.008561793714761734, + 0.06612002849578857, + 0.02207818441092968, + 0.01459551602602005, + 0.021291086450219154, + 0.013272952288389206, + 0.011341678909957409, + 0.028187820687890053, + 0.11271321028470993, + 0.0038680541329085827, + 0.0035987559240311384, + 0.002779996022582054, + 0.007675583474338055, + 0.00697186728939414, + 0.0030631080735474825, + 0.0038876670878380537, + 0.004897845443338156, + 0.003443810623139143, + 0.0016697051469236612, + 0.002072053262963891, + 0.0023310924880206585, + 0.0015372294001281261, + 0.0020183396991342306, + 0.0017173157539218664, + 0.0020651486702263355, + 0.006291531957685947, + 0.0021244490053504705, + 0.002180685754865408, + 0.002801246242597699, + 0.0034566933754831553, + 0.0030766823329031467, + 0.0032940583769232035, + 0.011965923011302948, + 0.07015757262706757, + 0.011562894098460674, + 0.09056196361780167, + 0.01892290636897087, + 0.006612933706492186, + 0.013895940966904163, + 0.010026006028056145, + 0.012467136606574059, + 0.0021178878378123045, + 0.0030386175494641066, + 0.001349282800219953, + 0.0017232855316251516, + 0.003186300862580538, + 0.001989044714719057, + 0.0017257730942219496, + 0.0014194714603945613 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 8, + 3 + ], + "selected_reward": -0.005253438837826252 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a mechanical version of a character named Pomu with polished chrome limbs and glowing green fiber-optic hair in a high ponytail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character seated on a futuristic stage with a dark, polished floor that reflects the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character leaning forward in concentration while interacting with a translucent three-dimensional holographic interface using glowing sensor-equipped hands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a vast stadium filled with blurred silhouettes and massive holographic screens displaying match footage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video begin with a slow, cinematic push-in shot starting from a 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": [ + -0.014195299707353115 + ], + "opposite_logprob_tokens": [ + -4.264195442199707 + ], + "logprob_reward": -0.014195299707353115, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014195299707353115, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1993199735879898, + 0.1064257025718689, + 0.07659640908241272, + 0.1937793493270874, + 0.026227839291095734, + 0.16161805391311646, + 0.023890089243650436, + 0.06003952771425247, + 0.11263053119182587, + 0.03947252035140991 + ], + "credit_map": [ + 0.023926112800836563, + 0.019786978140473366, + 0.02402423694729805, + 0.02114732377231121, + 0.031237486749887466, + 0.014706187881529331, + 0.037767086178064346, + 0.02672455459833145, + 0.009946073405444622, + 0.009288677014410496, + 0.009758279658854008, + 0.015439743176102638, + 0.021450050175189972, + 0.008546355180442333, + 0.022468606010079384, + 0.009527914226055145, + 0.01203263271600008, + 0.012174406088888645, + 0.010822615586221218, + 0.006693505682051182, + 0.009845087304711342, + 0.008645154535770416, + 0.007270314730703831, + 0.009112689644098282, + 0.06445705145597458, + 0.012938115745782852, + 0.07376700639724731, + 0.013200132176280022, + 0.008806892670691013, + 0.00552115635946393, + 0.008654779754579067, + 0.006434215698391199, + 0.003411358455196023, + 0.0036277000326663256, + 0.002299191430211067, + 0.003096197033300996, + 0.003455956932157278, + 0.002758184215053916, + 0.002839165972545743, + 0.004740085452795029, + 0.02474450133740902, + 0.02429133653640747, + 0.011524487286806107, + 0.030861280858516693, + 0.022096766158938408, + 0.014706428162753582, + 0.01654319278895855, + 0.016850074753165245, + 0.003287392435595393, + 0.0025546113029122353, + 0.003046335419639945, + 0.002951564732939005, + 0.0031123661901801825, + 0.0028873018454760313, + 0.002606942318379879, + 0.0034435726702213287, + 0.010193831287324429, + 0.003627446945756674, + 0.00415643909946084, + 0.004798968322575092, + 0.005755494814366102, + 0.004611110780388117, + 0.005509505048394203, + 0.021386733278632164, + 0.028871333226561546, + 0.008816076442599297, + 0.036968689411878586, + 0.011645922437310219, + 0.00708741694688797, + 0.007783744018524885, + 0.004735881928354502, + 0.006721457000821829, + 0.005337477661669254, + 0.0050635309889912605, + 0.0033293338492512703, + 0.004316091537475586, + 0.008239283226430416, + 0.004801352508366108, + 0.004013987258076668, + 0.0043714637868106365 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 5 + ], + "selected_reward": -0.014195299707353115 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a mechanical version of a character named Pomu with polished chrome limbs and glowing green fiber-optic hair in a high ponytail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character seated on a futuristic stage with a dark, polished floor that reflects the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character leaning forward in concentration while interacting with a translucent three-dimensional holographic interface using glowing sensor-equipped hands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a vast stadium filled with blurred silhouettes and massive holographic screens displaying match footage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video begin with a slow, cinematic push-in shot starting from a 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": [ + -0.08901983499526978 + ], + "opposite_logprob_tokens": [ + -2.464019775390625 + ], + "logprob_reward": -0.08901983499526978, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08901983499526978, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20196017622947693, + 0.1270277202129364, + 0.09975483268499374, + 0.059016838669776917, + 0.24019676446914673, + 0.048570048063993454, + 0.021414034068584442, + 0.05212066322565079, + 0.12118270248174667, + 0.028756285086274147 + ], + "credit_map": [ + 0.021997585892677307, + 0.022450223565101624, + 0.01483065728098154, + 0.025835607200860977, + 0.03266588971018791, + 0.016376035287976265, + 0.039133284240961075, + 0.028670847415924072, + 0.01637829840183258, + 0.019849952310323715, + 0.012192636728286743, + 0.02295623905956745, + 0.011762849986553192, + 0.010999992489814758, + 0.022191623225808144, + 0.010696116834878922, + 0.018263360485434532, + 0.006866234354674816, + 0.0053771850652992725, + 0.01043019536882639, + 0.004073761403560638, + 0.007836227305233479, + 0.012099692597985268, + 0.034808166325092316, + 0.011653834022581577, + 0.006303892005234957, + 0.019110066816210747, + 0.004221122711896896, + 0.003866806160658598, + 0.0035619998816400766, + 0.004472465720027685, + 0.0058266413398087025, + 0.04899444431066513, + 0.01881958171725273, + 0.0127253457903862, + 0.018186941742897034, + 0.013152522034943104, + 0.012650086544454098, + 0.025487734004855156, + 0.09018007665872574, + 0.007199546787887812, + 0.006424478720873594, + 0.003578132251277566, + 0.006236010696738958, + 0.007756936829537153, + 0.005354469176381826, + 0.004881200846284628, + 0.007139264605939388, + 0.0035680802538990974, + 0.0030564952176064253, + 0.002045032335445285, + 0.0024144425988197327, + 0.002120067598298192, + 0.0023680157028138638, + 0.0021937538404017687, + 0.0036481437273323536, + 0.009917099960148335, + 0.0028106123208999634, + 0.002587081165984273, + 0.0038027337286621332, + 0.003978982102125883, + 0.003421907778829336, + 0.004701454192399979, + 0.02090078592300415, + 0.03669785335659981, + 0.005980611313134432, + 0.04623812437057495, + 0.009671532548964024, + 0.003803045954555273, + 0.0071121565997600555, + 0.004339616745710373, + 0.007339733187109232, + 0.003638794645667076, + 0.003979857079684734, + 0.002693126443773508, + 0.003756157588213682, + 0.004564473871141672, + 0.0033924623858183622, + 0.0034864437766373158, + 0.0032449644058942795 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 1 + ], + "selected_reward": -0.08901983499526978 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a mechanical version of a character named Pomu with polished chrome limbs and glowing green fiber-optic hair in a high ponytail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character seated on a futuristic stage with a dark, polished floor that reflects the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character leaning forward in concentration while interacting with a translucent three-dimensional holographic interface using glowing sensor-equipped hands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a vast stadium filled with blurred silhouettes and massive holographic screens displaying match footage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video begin with a slow, cinematic push-in shot starting from a 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": [ + -0.06998291611671448 + ], + "opposite_logprob_tokens": [ + -2.6949830055236816 + ], + "logprob_reward": -0.06998291611671448, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.06998291611671448, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14368587732315063, + 0.15785355865955353, + 0.1701219528913498, + 0.07774627953767776, + 0.03519057109951973, + 0.2839062809944153, + 0.023878188803792, + 0.04142414778470993, + 0.05000649020075798, + 0.016186606138944626 + ], + "credit_map": [ + 0.012398408725857735, + 0.008686034008860588, + 0.01193715538829565, + 0.03939678519964218, + 0.015927262604236603, + 0.009631918743252754, + 0.020543117076158524, + 0.025165215134620667, + 0.013222864829003811, + 0.01501486636698246, + 0.011352541856467724, + 0.02618827298283577, + 0.03379027172923088, + 0.014478222467005253, + 0.027830855920910835, + 0.01597568579018116, + 0.013743642717599869, + 0.012939282692968845, + 0.01114706415683031, + 0.04760560020804405, + 0.00966532714664936, + 0.01164416503161192, + 0.012250576168298721, + 0.05112631991505623, + 0.024057870730757713, + 0.006956914439797401, + 0.023743359372019768, + 0.005219213664531708, + 0.0050123161636292934, + 0.0036036414094269276, + 0.00510329520329833, + 0.004049677401781082, + 0.0037881210446357727, + 0.003970354795455933, + 0.0037928044330328703, + 0.003755051176995039, + 0.003915548790246248, + 0.004604564514011145, + 0.00530412420630455, + 0.0060600051656365395, + 0.04359409958124161, + 0.04548954963684082, + 0.018216663971543312, + 0.04240208491683006, + 0.045883435755968094, + 0.02760292775928974, + 0.022384481504559517, + 0.03833306208252907, + 0.0032380204647779465, + 0.0034871185198426247, + 0.002259235130622983, + 0.002702343976125121, + 0.004680091515183449, + 0.002150473417714238, + 0.002528330311179161, + 0.0028325768653303385, + 0.006163690704852343, + 0.0027741380035877228, + 0.0033315334003418684, + 0.003867262974381447, + 0.003248703433200717, + 0.005104245617985725, + 0.0034827941562980413, + 0.01345178671181202, + 0.015308400616049767, + 0.002675382886081934, + 0.01912475936114788, + 0.0038250908255577087, + 0.001336000277660787, + 0.002755951602011919, + 0.001822607940994203, + 0.0031583039090037346, + 0.002307737013325095, + 0.0021327268332242966, + 0.0015905278269201517, + 0.0015663529047742486, + 0.002543842885643244, + 0.0019536048639565706, + 0.0015061757294461131, + 0.0025856380816549063 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 2, + 1 + ], + "selected_reward": -0.06998291611671448 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "slow, cinematic push-in begins in medium close-up, capturing a meticulously detailed, mechanical version of Pomu seated on a futuristic, sun-drenched stage. Her glowing green fiber-optic hair is styled in a high ponytail, with individual pulses of light falling on her face. Her limbs are polished chrome, their surfaces capturing and distorting the vibrant violet and cyan lighting of the environment, revealing the intricate gears and servos at her joints. She leans forward in rapt concentration, her expressive LED eyes, a brilliant emerald green, reflecting the complex digital battlefield—a series of strategic maps, data streams, and miniature explosions. Her chrome hands blur over the translucent three-dimensional holographic interface floating before her; her fingertips, equipped with glowing sensors, move swiftly and decisively, tapping the glowing icons with preternatural speed. The background is a vast stadium, filled with blurred silhouettes and massive holographic screens displaying match footage. The floor beneath her feet is a dark, polished material, creating a rain-smooth reflection of the entire scene, amplifying the cyberpunk aesthetic. The atmosphere is thrilling and tense, with the cool glow of the hologram contrasting with the warm neon lights, all filmed in stunning 8K HD, showcasing every detail of her cybernetic form and the high-stakes esports competition she dominates.", + "seed": 15891, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000056/group_9/index_9769_seed_15891_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": -0.0394119662232697, + "vlm_logprob": -0.0394119662232697, + "vlm_yes_no_margin": 3.575, + "vlm_token_credit": -0.0394119662232697, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a mechanical version of a character named Pomu with polished chrome limbs and glowing green fiber-optic hair in a high ponytail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character seated on a futuristic stage with a dark, polished floor that reflects the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character leaning forward in concentration while interacting with a translucent three-dimensional holographic interface using glowing sensor-equipped hands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a vast stadium filled with blurred silhouettes and massive holographic screens displaying match footage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video begin with a slow, cinematic push-in shot starting from a 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": [ + -0.050188545137643814 + ], + "opposite_logprob_tokens": [ + -3.0501885414123535 + ], + "logprob_reward": -0.050188545137643814, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.050188545137643814, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16672295331954956, + 0.19449898600578308, + 0.06412036716938019, + 0.05362476035952568, + 0.0532730370759964, + 0.03668925538659096, + 0.04003963991999626, + 0.1201644018292427, + 0.0559026375412941, + 0.21496398746967316 + ], + "credit_map": [ + 0.01422934141010046, + 0.020020920783281326, + 0.012705102562904358, + 0.020687924697995186, + 0.019854336977005005, + 0.029760459437966347, + 0.03516775369644165, + 0.014297098852694035, + 0.025510024279356003, + 0.01810174062848091, + 0.020766420289874077, + 0.029589418321847916, + 0.02119378373026848, + 0.013845445588231087, + 0.04078727215528488, + 0.02470485493540764, + 0.007986625656485558, + 0.005407834891229868, + 0.005505344830453396, + 0.006016631610691547, + 0.013016846962273121, + 0.007083497941493988, + 0.010676407255232334, + 0.008427174761891365, + 0.005427783355116844, + 0.00652249064296484, + 0.006227519828826189, + 0.005885924678295851, + 0.011268416419625282, + 0.005815085489302874, + 0.006690352689474821, + 0.005787180736660957, + 0.007853595539927483, + 0.005903950426727533, + 0.006335655692964792, + 0.005202032625675201, + 0.009405470453202724, + 0.005590849090367556, + 0.005637744907289743, + 0.0073437318205833435, + 0.0029407937545329332, + 0.006051457021385431, + 0.003907722886651754, + 0.00387816084548831, + 0.007261153310537338, + 0.004094818141311407, + 0.0042196642607450485, + 0.004335481207817793, + 0.0034597930498421192, + 0.0054382383823394775, + 0.0035365966614335775, + 0.005397971253842115, + 0.005833763163536787, + 0.005096751265227795, + 0.005293147638440132, + 0.005983375012874603, + 0.014220322482287884, + 0.013502465561032295, + 0.009812632575631142, + 0.010143160820007324, + 0.016032909974455833, + 0.024425415322184563, + 0.016750585287809372, + 0.015276903286576271, + 0.010784653015434742, + 0.007618858013302088, + 0.004968507215380669, + 0.007273336872458458, + 0.005411538295447826, + 0.006800059229135513, + 0.007350137922912836, + 0.00569553766399622, + 0.026929115876555443, + 0.037808194756507874, + 0.018063466995954514, + 0.025686394423246384, + 0.0234315637499094, + 0.019497498869895935, + 0.03645913675427437, + 0.02708858996629715 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 1, + 0 + ], + "selected_reward": -0.050188545137643814 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a mechanical version of a character named Pomu with polished chrome limbs and glowing green fiber-optic hair in a high ponytail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character seated on a futuristic stage with a dark, polished floor that reflects the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character leaning forward in concentration while interacting with a translucent three-dimensional holographic interface using glowing sensor-equipped hands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a vast stadium filled with blurred silhouettes and massive holographic screens displaying match footage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video begin with a slow, cinematic push-in shot starting from a 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": [ + -0.006739267613738775 + ], + "opposite_logprob_tokens": [ + -5.006739139556885 + ], + "logprob_reward": -0.006739267613738775, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006739267613738775, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15462268888950348, + 0.087044358253479, + 0.13759872317314148, + 0.08463997393846512, + 0.036588944494724274, + 0.06089227646589279, + 0.04521765559911728, + 0.1487058401107788, + 0.12691046297550201, + 0.11777906864881516 + ], + "credit_map": [ + 0.017725002020597458, + 0.022979291155934334, + 0.01348907221108675, + 0.019571499899029732, + 0.022319594398140907, + 0.0201566144824028, + 0.021196605637669563, + 0.017184998840093613, + 0.009297697804868221, + 0.008107451722025871, + 0.011038637720048428, + 0.009944142773747444, + 0.008684954605996609, + 0.008403434418141842, + 0.012474313378334045, + 0.019093727692961693, + 0.013737885281443596, + 0.013581112958490849, + 0.01622951030731201, + 0.015618943609297276, + 0.015078668482601643, + 0.017535986378788948, + 0.025253593921661377, + 0.020563025027513504, + 0.010424436070024967, + 0.009311969392001629, + 0.011765025556087494, + 0.009027332067489624, + 0.012988670729100704, + 0.010743740946054459, + 0.010192284360527992, + 0.010186512023210526, + 0.005544761661440134, + 0.0036350255832076073, + 0.004106046166270971, + 0.005019739270210266, + 0.003733371151611209, + 0.0036336907651275396, + 0.0040075527504086494, + 0.006908758543431759, + 0.00796855241060257, + 0.012415069155395031, + 0.00783934723585844, + 0.006410379894077778, + 0.006836385000497103, + 0.0069437213242053986, + 0.006800089031457901, + 0.0056787291541695595, + 0.005431871861219406, + 0.00764675484970212, + 0.005205407273024321, + 0.004762616474181414, + 0.004534419625997543, + 0.007056588772684336, + 0.00535478675737977, + 0.005225207190960646, + 0.01294654794037342, + 0.018929312005639076, + 0.009364555589854717, + 0.01079106517136097, + 0.02208860218524933, + 0.03422975912690163, + 0.014704099856317043, + 0.025651903823018074, + 0.021251413971185684, + 0.015648838132619858, + 0.009881499223411083, + 0.018709739670157433, + 0.011663679033517838, + 0.019200030714273453, + 0.015024803578853607, + 0.015530463308095932, + 0.019285086542367935, + 0.010831446386873722, + 0.008942062966525555, + 0.014191940426826477, + 0.015113003551959991, + 0.012777778320014477, + 0.016734464094042778, + 0.01990327052772045 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 2 + ], + "selected_reward": -0.006739267613738775 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a mechanical version of a character named Pomu with polished chrome limbs and glowing green fiber-optic hair in a high ponytail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character seated on a futuristic stage with a dark, polished floor that reflects the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character leaning forward in concentration while interacting with a translucent three-dimensional holographic interface using glowing sensor-equipped hands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a vast stadium filled with blurred silhouettes and massive holographic screens displaying match footage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video begin with a slow, cinematic push-in shot starting from a 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": [ + -0.01253888476639986 + ], + "opposite_logprob_tokens": [ + -4.387538909912109 + ], + "logprob_reward": -0.01253888476639986, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.01253888476639986, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2321833372116089, + 0.3302386701107025, + 0.11692594736814499, + 0.046872761100530624, + 0.03786027431488037, + 0.02274717576801777, + 0.044080592691898346, + 0.057820048183202744, + 0.03890877962112427, + 0.07236247509717941 + ], + "credit_map": [ + 0.024608520790934563, + 0.036787111312150955, + 0.019988540560007095, + 0.033359769731760025, + 0.0348835252225399, + 0.022143855690956116, + 0.03254683315753937, + 0.02786516584455967, + 0.0373319648206234, + 0.02576509676873684, + 0.0369243286550045, + 0.04484236240386963, + 0.03861192613840103, + 0.025287318974733353, + 0.05953330174088478, + 0.06194233521819115, + 0.011670983396470547, + 0.011336671188473701, + 0.0122586190700531, + 0.015356576070189476, + 0.015159384347498417, + 0.01762997917830944, + 0.016733666881918907, + 0.016780074685811996, + 0.004189877305179834, + 0.004282356705516577, + 0.00583528820425272, + 0.006430156994611025, + 0.006666598375886679, + 0.0056682550348341465, + 0.007011183071881533, + 0.006789050064980984, + 0.005591127090156078, + 0.004167195875197649, + 0.005049566738307476, + 0.0037538642063736916, + 0.005748780909925699, + 0.0037287455052137375, + 0.00461007934063673, + 0.0052109151147305965, + 0.0020794805604964495, + 0.004114710260182619, + 0.0030635541770607233, + 0.0022055585868656635, + 0.004063169471919537, + 0.002644142834469676, + 0.0019457254093140364, + 0.002630834002047777, + 0.006081198807805777, + 0.0050797597505152225, + 0.0053210132755339146, + 0.005916982423514128, + 0.004545027390122414, + 0.004838827531784773, + 0.005577806383371353, + 0.006719977129250765, + 0.009400499984622002, + 0.007010267581790686, + 0.005613229237496853, + 0.004510163329541683, + 0.00811500009149313, + 0.00919169932603836, + 0.008093634620308876, + 0.0058855535462498665, + 0.004711391404271126, + 0.00712950062006712, + 0.0032874024473130703, + 0.004272674210369587, + 0.006575550884008408, + 0.005428860429674387, + 0.003431623103097081, + 0.00407177722081542, + 0.009754827246069908, + 0.009625611826777458, + 0.005520693026483059, + 0.011495782062411308, + 0.013258285820484161, + 0.008949264883995056, + 0.007710135076195002, + 0.006047872360795736 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 2 + ], + "selected_reward": -0.01253888476639986 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a mechanical version of a character named Pomu with polished chrome limbs and glowing green fiber-optic hair in a high ponytail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character seated on a futuristic stage with a dark, polished floor that reflects the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character leaning forward in concentration while interacting with a translucent three-dimensional holographic interface using glowing sensor-equipped hands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a vast stadium filled with blurred silhouettes and massive holographic screens displaying match footage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video begin with a slow, cinematic push-in shot starting from a 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": [ + -0.07896696776151657 + ], + "opposite_logprob_tokens": [ + -2.5789668560028076 + ], + "logprob_reward": -0.07896696776151657, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07896696776151657, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2855210304260254, + 0.17702414095401764, + 0.0760844349861145, + 0.12691843509674072, + 0.05636969208717346, + 0.08282748609781265, + 0.043150946497917175, + 0.06315752863883972, + 0.04854311794042587, + 0.04040319472551346 + ], + "credit_map": [ + 0.013131160289049149, + 0.02357511594891548, + 0.01132096815854311, + 0.034970369189977646, + 0.02475358545780182, + 0.05675887316465378, + 0.0897526815533638, + 0.031258273869752884, + 0.018597571179270744, + 0.01713329367339611, + 0.027118531987071037, + 0.017292728647589684, + 0.024770762771368027, + 0.015456059016287327, + 0.02716146595776081, + 0.02949371375143528, + 0.010030468925833702, + 0.007980803959071636, + 0.009445974603295326, + 0.009378289803862572, + 0.00852633360773325, + 0.006798582151532173, + 0.010846024379134178, + 0.013077959418296814, + 0.01825677789747715, + 0.015947530046105385, + 0.013617528602480888, + 0.013823851011693478, + 0.020655149593949318, + 0.01526243519037962, + 0.015294366516172886, + 0.014060795307159424, + 0.009041676297783852, + 0.005320667289197445, + 0.006221840623766184, + 0.00617563258856535, + 0.005458270199596882, + 0.0069812810979783535, + 0.006293507292866707, + 0.01087681669741869, + 0.012535433284938335, + 0.016486752778291702, + 0.007420648820698261, + 0.010089443996548653, + 0.009447995573282242, + 0.008658980950713158, + 0.009534638375043869, + 0.008653596043586731, + 0.007144875358790159, + 0.0046427007764577866, + 0.005035784095525742, + 0.0065595610067248344, + 0.004373971838504076, + 0.006008986383676529, + 0.004311985336244106, + 0.005073079839348793, + 0.008647570386528969, + 0.0067694950848817825, + 0.0037180804647505283, + 0.00443605799227953, + 0.009277586825191975, + 0.013795565813779831, + 0.006189933978021145, + 0.010323243215680122, + 0.005844823084771633, + 0.007820701226592064, + 0.004695781972259283, + 0.007265069056302309, + 0.004842902068048716, + 0.006602253299206495, + 0.006101947743445635, + 0.005369642749428749, + 0.005463059991598129, + 0.006184360943734646, + 0.0031705021392554045, + 0.0035964036360383034, + 0.008617999032139778, + 0.004499154631048441, + 0.003949835430830717, + 0.004921876359730959 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 3 + ], + "selected_reward": -0.07896696776151657 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a mechanical version of a character named Pomu with polished chrome limbs and glowing green fiber-optic hair in a high ponytail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character seated on a futuristic stage with a dark, polished floor that reflects the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character leaning forward in concentration while interacting with a translucent three-dimensional holographic interface using glowing sensor-equipped hands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a vast stadium filled with blurred silhouettes and massive holographic screens displaying match footage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video begin with a slow, cinematic push-in shot starting from a 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": [ + -0.048626165837049484 + ], + "opposite_logprob_tokens": [ + -3.048626184463501 + ], + "logprob_reward": -0.048626165837049484, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.048626165837049484, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1421421617269516, + 0.21484553813934326, + 0.10238707810640335, + 0.060561370104551315, + 0.0696236863732338, + 0.06680447608232498, + 0.04081954434514046, + 0.0913388803601265, + 0.1542699635028839, + 0.05720736086368561 + ], + "credit_map": [ + 0.011421565897762775, + 0.01515624113380909, + 0.011393705382943153, + 0.01716047152876854, + 0.015757983550429344, + 0.016847604885697365, + 0.03528640419244766, + 0.0191181767731905, + 0.02112443372607231, + 0.02700144611299038, + 0.02364037185907364, + 0.027549689635634422, + 0.03049449622631073, + 0.016307517886161804, + 0.03025103732943535, + 0.038476523011922836, + 0.014301994815468788, + 0.013883830979466438, + 0.00852152332663536, + 0.011396010406315327, + 0.022134054452180862, + 0.010012800805270672, + 0.01202553603798151, + 0.01011131051927805, + 0.00920015387237072, + 0.008002934977412224, + 0.007586524821817875, + 0.006966008804738522, + 0.00829300843179226, + 0.006568627897650003, + 0.007173178251832724, + 0.00677092419937253, + 0.011049335822463036, + 0.0058412072248756886, + 0.008636167272925377, + 0.009409911930561066, + 0.00787866860628128, + 0.006316560786217451, + 0.009493568912148476, + 0.010998248122632504, + 0.007109162863343954, + 0.018690329045057297, + 0.005785059649497271, + 0.007319488096982241, + 0.006739749573171139, + 0.006433561444282532, + 0.007430471479892731, + 0.007296646013855934, + 0.005288053769618273, + 0.005895840469747782, + 0.0038802719209343195, + 0.005023558158427477, + 0.004682098515331745, + 0.005672165658324957, + 0.005652119405567646, + 0.004725432489067316, + 0.00981005560606718, + 0.01391675416380167, + 0.005174099467694759, + 0.007890325970947742, + 0.0144107760861516, + 0.015737639740109444, + 0.009187694638967514, + 0.015211518853902817, + 0.018788428977131844, + 0.023341314867138863, + 0.012860661372542381, + 0.02084347978234291, + 0.01421571709215641, + 0.031116966158151627, + 0.013361979275941849, + 0.01974140666425228, + 0.008140516467392445, + 0.011341278441250324, + 0.004264743532985449, + 0.006896715145558119, + 0.009862108156085014, + 0.007483837194740772, + 0.004857433959841728, + 0.004360721446573734 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 8, + 0 + ], + "selected_reward": -0.048626165837049484 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "slow, cinematic push-in begins in medium close-up, capturing a meticulously detailed, mechanical version of Pomu seated on a futuristic, sun-drenched stage. Her glowing green fiber-optic hair is styled in a high ponytail, with individual pulses of light falling on her face. Her limbs are polished chrome, their surfaces capturing and distorting the vibrant violet and cyan lighting of the environment, revealing the intricate gears and servos at her joints. She leans forward in rapt concentration, her expressive LED eyes, a brilliant emerald green, reflecting the complex digital battlefield—a series of strategic maps, data streams, and miniature explosions. Her chrome hands blur over the translucent three-dimensional holographic interface floating before her; her fingertips, equipped with glowing sensors, move swiftly and decisively, tapping the glowing icons with preternatural speed. The background is a vast stadium, filled with blurred silhouettes and massive holographic screens displaying match footage. The floor beneath her feet is a dark, polished material, creating a rain-smooth reflection of the entire scene, amplifying the cyberpunk aesthetic. The atmosphere is thrilling and tense, with the cool glow of the hologram contrasting with the warm neon lights, all filmed in stunning 8K HD, showcasing every detail of her cybernetic form and the high-stakes esports competition she dominates.", + "seed": 15892, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000056/group_9/index_9769_seed_15892_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": -0.04629840729758143, + "vlm_logprob": -0.04629840729758143, + "vlm_yes_no_margin": 3.475, + "vlm_token_credit": -0.04629840729758143, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a mechanical version of a character named Pomu with polished chrome limbs and glowing green fiber-optic hair in a high ponytail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character seated on a futuristic stage with a dark, polished floor that reflects the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character leaning forward in concentration while interacting with a translucent three-dimensional holographic interface using glowing sensor-equipped hands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a vast stadium filled with blurred silhouettes and massive holographic screens displaying match footage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video begin with a slow, cinematic push-in shot starting from a 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": [ + -0.05577562376856804 + ], + "opposite_logprob_tokens": [ + -2.9307756423950195 + ], + "logprob_reward": -0.05577562376856804, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05577562376856804, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15401466190814972, + 0.037950750440359116, + 0.035840537399053574, + 0.039684001356363297, + 0.032945506274700165, + 0.04028836637735367, + 0.47209373116493225, + 0.07752563059329987, + 0.049750205129384995, + 0.059906598180532455 + ], + "credit_map": [ + 0.024767877534031868, + 0.015770334750413895, + 0.025648348033428192, + 0.02223728783428669, + 0.007701833266764879, + 0.030197132378816605, + 0.011400006711483002, + 0.016291849315166473, + 0.0048413584008812904, + 0.004295787774026394, + 0.005523231811821461, + 0.005566283129155636, + 0.0038658538833260536, + 0.0051331063732504845, + 0.004952568095177412, + 0.0037725588772445917, + 0.00686806533485651, + 0.003714409191161394, + 0.004739346448332071, + 0.003868170315399766, + 0.004202163312584162, + 0.004629071336239576, + 0.003926451783627272, + 0.0038928568828850985, + 0.008400795049965382, + 0.003981725312769413, + 0.0067888121120631695, + 0.004291478544473648, + 0.0034836370032280684, + 0.005270606838166714, + 0.00421487120911479, + 0.003252074820920825, + 0.002732476918026805, + 0.004478164948523045, + 0.004982536192983389, + 0.0032035394106060266, + 0.004340538289397955, + 0.005283297970890999, + 0.0036108247004449368, + 0.004314127843827009, + 0.005000748205929995, + 0.00427762046456337, + 0.007765446323901415, + 0.004922440275549889, + 0.004716860596090555, + 0.005674367770552635, + 0.004479997791349888, + 0.0034508865792304277, + 0.05260717123746872, + 0.06503094732761383, + 0.06619080901145935, + 0.044530220329761505, + 0.05551068112254143, + 0.058769065886735916, + 0.045566849410533905, + 0.08388801664113998, + 0.011346892453730106, + 0.007970317266881466, + 0.01095204520970583, + 0.012713324278593063, + 0.006223752163350582, + 0.009486782364547253, + 0.006522795185446739, + 0.012309727258980274, + 0.004506308119744062, + 0.008856096304953098, + 0.005295949522405863, + 0.006383635103702545, + 0.007469578180462122, + 0.004332444630563259, + 0.004110714420676231, + 0.008795475587248802, + 0.00599098764359951, + 0.008738376200199127, + 0.014540834352374077, + 0.004426876083016396, + 0.004512292332947254, + 0.007168291136622429, + 0.0033900428097695112, + 0.011138894595205784 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 7 + ], + "selected_reward": -0.05577562376856804 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a mechanical version of a character named Pomu with polished chrome limbs and glowing green fiber-optic hair in a high ponytail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character seated on a futuristic stage with a dark, polished floor that reflects the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character leaning forward in concentration while interacting with a translucent three-dimensional holographic interface using glowing sensor-equipped hands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a vast stadium filled with blurred silhouettes and massive holographic screens displaying match footage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video begin with a slow, cinematic push-in shot starting from a 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": [ + -0.006737255025655031 + ], + "opposite_logprob_tokens": [ + -5.006737232208252 + ], + "logprob_reward": -0.006737255025655031, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006737255025655031, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17364345490932465, + 0.06112196296453476, + 0.035102762281894684, + 0.3150516450405121, + 0.06854831427335739, + 0.04316230118274689, + 0.0984862670302391, + 0.06259918957948685, + 0.060228198766708374, + 0.0820559561252594 + ], + "credit_map": [ + 0.023651141673326492, + 0.01549383345991373, + 0.026469510048627853, + 0.02426452562212944, + 0.013042059727013111, + 0.026128573343157768, + 0.022753795608878136, + 0.021840017288923264, + 0.007352240849286318, + 0.007005353923887014, + 0.009104675613343716, + 0.010458423756062984, + 0.00626605749130249, + 0.007677155081182718, + 0.006505441851913929, + 0.006752611603587866, + 0.005687872413545847, + 0.004380472004413605, + 0.004975603893399239, + 0.0048198578879237175, + 0.004644469358026981, + 0.0046659051440656185, + 0.0029724452178925276, + 0.002956136129796505, + 0.05812782794237137, + 0.029781775549054146, + 0.06007908657193184, + 0.03643026947975159, + 0.021909257397055626, + 0.042869534343481064, + 0.03488890081644058, + 0.030964991077780724, + 0.006097914185374975, + 0.010500662960112095, + 0.01051575317978859, + 0.007573300041258335, + 0.006809237413108349, + 0.008998775854706764, + 0.009424472227692604, + 0.00862819142639637, + 0.005371490027755499, + 0.004675974138081074, + 0.005186208989471197, + 0.007133910432457924, + 0.004083186853677034, + 0.004801999777555466, + 0.006327578332275152, + 0.005581950768828392, + 0.01278800517320633, + 0.01153239980340004, + 0.012725964188575745, + 0.01222844235599041, + 0.009044365957379341, + 0.01264365017414093, + 0.01127915270626545, + 0.016244281083345413, + 0.00655363081023097, + 0.009152424521744251, + 0.00985674373805523, + 0.007912537083029747, + 0.00724116200581193, + 0.006716229021549225, + 0.006472567096352577, + 0.008693891577422619, + 0.006548081990331411, + 0.010920876637101173, + 0.0063318293541669846, + 0.005895600654184818, + 0.010676971636712551, + 0.006722753401845694, + 0.006481912452727556, + 0.006650176830589771, + 0.010772381909191608, + 0.011063982732594013, + 0.016886258497834206, + 0.006682306528091431, + 0.008315404877066612, + 0.007931341417133808, + 0.0049973465502262115, + 0.015406937338411808 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 6 + ], + "selected_reward": -0.006737255025655031 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a mechanical version of a character named Pomu with polished chrome limbs and glowing green fiber-optic hair in a high ponytail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character seated on a futuristic stage with a dark, polished floor that reflects the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character leaning forward in concentration while interacting with a translucent three-dimensional holographic interface using glowing sensor-equipped hands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a vast stadium filled with blurred silhouettes and massive holographic screens displaying match footage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video begin with a slow, cinematic push-in shot starting from a 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": [ + -0.011074741370975971 + ], + "opposite_logprob_tokens": [ + -4.511074542999268 + ], + "logprob_reward": -0.011074741370975971, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011074741370975971, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18535283207893372, + 0.09799879044294357, + 0.08390717953443527, + 0.04673098400235176, + 0.06636787205934525, + 0.06573904305696487, + 0.11693662405014038, + 0.0724906399846077, + 0.11046478152275085, + 0.15401127934455872 + ], + "credit_map": [ + 0.027136333286762238, + 0.018159683793783188, + 0.01952063851058483, + 0.019033903256058693, + 0.009940609335899353, + 0.05075011029839516, + 0.02048383466899395, + 0.020327717065811157, + 0.017349375411868095, + 0.012559824623167515, + 0.011923584155738354, + 0.016218028962612152, + 0.011263278312981129, + 0.013529385440051556, + 0.007612557616084814, + 0.00754275918006897, + 0.015429341234266758, + 0.009650910273194313, + 0.017779992893338203, + 0.008287993259727955, + 0.008267269469797611, + 0.009794707410037518, + 0.00746506592258811, + 0.007231898605823517, + 0.0061622196808457375, + 0.005935947876423597, + 0.005745589733123779, + 0.005420947913080454, + 0.006339679472148418, + 0.00641667190939188, + 0.005777499172836542, + 0.004932431038469076, + 0.006517306435853243, + 0.009443936869502068, + 0.00939114112406969, + 0.006457449868321419, + 0.008459790609776974, + 0.009610828012228012, + 0.007917899638414383, + 0.008569514378905296, + 0.006407696288079023, + 0.006619398947805166, + 0.010181818157434464, + 0.00577216362580657, + 0.008583614602684975, + 0.009862647391855717, + 0.008272483944892883, + 0.010039214976131916, + 0.012158485129475594, + 0.015470334328711033, + 0.01399420015513897, + 0.016536200419068336, + 0.011090900748968124, + 0.012697767466306686, + 0.013270747847855091, + 0.021717986091971397, + 0.013601493090391159, + 0.008888731710612774, + 0.007749305106699467, + 0.0068056355230510235, + 0.00657636858522892, + 0.00901386234909296, + 0.005843067541718483, + 0.014012176543474197, + 0.010069307871162891, + 0.025323176756501198, + 0.007196863181889057, + 0.009168676100671291, + 0.010590353980660439, + 0.00916325394064188, + 0.012533425353467464, + 0.02641972340643406, + 0.018920082598924637, + 0.023246126249432564, + 0.029855892062187195, + 0.013455867767333984, + 0.01733114756643772, + 0.014658837579190731, + 0.01062077283859253, + 0.025922562927007675 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 6 + ], + "selected_reward": -0.011074741370975971 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a mechanical version of a character named Pomu with polished chrome limbs and glowing green fiber-optic hair in a high ponytail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character seated on a futuristic stage with a dark, polished floor that reflects the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character leaning forward in concentration while interacting with a translucent three-dimensional holographic interface using glowing sensor-equipped hands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a vast stadium filled with blurred silhouettes and massive holographic screens displaying match footage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video begin with a slow, cinematic push-in shot starting from a 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": [ + -0.07896620035171509 + ], + "opposite_logprob_tokens": [ + -2.5789661407470703 + ], + "logprob_reward": -0.07896620035171509, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07896620035171509, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4245702624320984, + 0.0770341008901596, + 0.07620017975568771, + 0.08778238296508789, + 0.058538224548101425, + 0.0678570494055748, + 0.05149192735552788, + 0.029179243370890617, + 0.05487943813204765, + 0.07246719300746918 + ], + "credit_map": [ + 0.07545403391122818, + 0.041915636509656906, + 0.05006629601120949, + 0.055638138204813004, + 0.02915572002530098, + 0.06803864240646362, + 0.05240471661090851, + 0.0518970750272274, + 0.013595659285783768, + 0.007997657172381878, + 0.013892958872020245, + 0.013035045005381107, + 0.004203722346574068, + 0.010583808645606041, + 0.007380530703812838, + 0.006344716064631939, + 0.02166593074798584, + 0.007166504859924316, + 0.011499043554067612, + 0.009528621099889278, + 0.005124165676534176, + 0.010873581282794476, + 0.005094605963677168, + 0.005247730761766434, + 0.01622612774372101, + 0.00883684866130352, + 0.015166320838034153, + 0.009423534385859966, + 0.0072265309281647205, + 0.010850189253687859, + 0.010258126072585583, + 0.009794708341360092, + 0.005592463072389364, + 0.00828554853796959, + 0.00816244725137949, + 0.0060523031279444695, + 0.006222318857908249, + 0.011019334197044373, + 0.0056812698021531105, + 0.007522539235651493, + 0.008944417349994183, + 0.004536070860922337, + 0.016811996698379517, + 0.006047498434782028, + 0.004520447924733162, + 0.013620235957205296, + 0.006677720230072737, + 0.006698664277791977, + 0.004832021426409483, + 0.00573293212801218, + 0.00697300722822547, + 0.00502148037776351, + 0.006875248625874519, + 0.0058830976486206055, + 0.0055498359724879265, + 0.01062430627644062, + 0.004724834579974413, + 0.0029668554197996855, + 0.004912190139293671, + 0.0037254251074045897, + 0.0021502883173525333, + 0.003222380531951785, + 0.002691596280783415, + 0.004785674624145031, + 0.0048801228404045105, + 0.012600059621036053, + 0.004134161397814751, + 0.0038360573817044497, + 0.010505866259336472, + 0.004869299940764904, + 0.0036839423701167107, + 0.01036993507295847, + 0.010690310969948769, + 0.011910267174243927, + 0.011862645857036114, + 0.007547458168119192, + 0.008933346718549728, + 0.0063921804539859295, + 0.00456561055034399, + 0.010565373115241528 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 1 + ], + "selected_reward": -0.07896620035171509 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a mechanical version of a character named Pomu with polished chrome limbs and glowing green fiber-optic hair in a high ponytail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character seated on a futuristic stage with a dark, polished floor that reflects the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character leaning forward in concentration while interacting with a translucent three-dimensional holographic interface using glowing sensor-equipped hands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a vast stadium filled with blurred silhouettes and massive holographic screens displaying match footage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video begin with a slow, cinematic push-in shot starting from a 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": [ + -0.07893821597099304 + ], + "opposite_logprob_tokens": [ + -2.5789382457733154 + ], + "logprob_reward": -0.07893821597099304, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07893821597099304, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20685403048992157, + 0.0764075443148613, + 0.10184325277805328, + 0.07820438593626022, + 0.07569180428981781, + 0.09246060252189636, + 0.1898277997970581, + 0.03812314197421074, + 0.029492763802409172, + 0.11109471321105957 + ], + "credit_map": [ + 0.035553596913814545, + 0.020060159265995026, + 0.026712439954280853, + 0.027589788660407066, + 0.018133388832211494, + 0.03061545267701149, + 0.01948123797774315, + 0.0287079568952322, + 0.02223929762840271, + 0.010458448901772499, + 0.009403458796441555, + 0.007126702927052975, + 0.005243297666311264, + 0.00921275932341814, + 0.006591654382646084, + 0.006131917238235474, + 0.024493025615811348, + 0.009271116927266121, + 0.017660781741142273, + 0.012446515262126923, + 0.00866504106670618, + 0.014520856551826, + 0.007406546268612146, + 0.007379363756626844, + 0.013516416773200035, + 0.00807338859885931, + 0.014221577905118465, + 0.01054380927234888, + 0.007434737868607044, + 0.009273895062506199, + 0.00801010150462389, + 0.007130454294383526, + 0.00628513703122735, + 0.009395976550877094, + 0.010305152274668217, + 0.007948490791022778, + 0.008692449890077114, + 0.011044288985431194, + 0.01118392962962389, + 0.010836372151970863, + 0.011622920632362366, + 0.006009718868881464, + 0.022010035812854767, + 0.009409486316144466, + 0.006975775118917227, + 0.01740904524922371, + 0.010286788456141949, + 0.008736835792660713, + 0.016383115202188492, + 0.024746062234044075, + 0.027322297915816307, + 0.01755819283425808, + 0.03149722144007683, + 0.0252949520945549, + 0.018533362075686455, + 0.028492584824562073, + 0.00708795664831996, + 0.004409603774547577, + 0.0052168588154017925, + 0.003910197876393795, + 0.004388207104057074, + 0.004026344977319241, + 0.0026061376556754112, + 0.006477830465883017, + 0.0029351776465773582, + 0.004356573801487684, + 0.0031963346991688013, + 0.004525850526988506, + 0.003724131267517805, + 0.0029623208101838827, + 0.003398050321266055, + 0.0043943230994045734, + 0.01307523064315319, + 0.018347548320889473, + 0.020142219960689545, + 0.010476537048816681, + 0.014189410023391247, + 0.011487671174108982, + 0.008305427618324757, + 0.015070662833750248 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 9 + ], + "selected_reward": -0.07893821597099304 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "slow, cinematic push-in begins in medium close-up, capturing a meticulously detailed, mechanical version of Pomu seated on a futuristic, sun-drenched stage. Her glowing green fiber-optic hair is styled in a high ponytail, with individual pulses of light falling on her face. Her limbs are polished chrome, their surfaces capturing and distorting the vibrant violet and cyan lighting of the environment, revealing the intricate gears and servos at her joints. She leans forward in rapt concentration, her expressive LED eyes, a brilliant emerald green, reflecting the complex digital battlefield—a series of strategic maps, data streams, and miniature explosions. Her chrome hands blur over the translucent three-dimensional holographic interface floating before her; her fingertips, equipped with glowing sensors, move swiftly and decisively, tapping the glowing icons with preternatural speed. The background is a vast stadium, filled with blurred silhouettes and massive holographic screens displaying match footage. The floor beneath her feet is a dark, polished material, creating a rain-smooth reflection of the entire scene, amplifying the cyberpunk aesthetic. The atmosphere is thrilling and tense, with the cool glow of the hologram contrasting with the warm neon lights, all filmed in stunning 8K HD, showcasing every detail of her cybernetic form and the high-stakes esports competition she dominates.", + "seed": 15893, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000056/group_9/index_9769_seed_15893_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": -0.06711066663265228, + "vlm_logprob": -0.06711066663265228, + "vlm_yes_no_margin": 3.15, + "vlm_token_credit": -0.06711066663265228, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a mechanical version of a character named Pomu with polished chrome limbs and glowing green fiber-optic hair in a high ponytail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character seated on a futuristic stage with a dark, polished floor that reflects the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character leaning forward in concentration while interacting with a translucent three-dimensional holographic interface using glowing sensor-equipped hands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a vast stadium filled with blurred silhouettes and massive holographic screens displaying match footage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video begin with a slow, cinematic push-in shot starting from a 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": [ + -0.07989362627267838 + ], + "opposite_logprob_tokens": [ + -2.5798935890197754 + ], + "logprob_reward": -0.07989362627267838, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07989362627267838, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.27358806133270264, + 0.04093923047184944, + 0.05321188643574715, + 0.04414735734462738, + 0.04194246232509613, + 0.23521997034549713, + 0.14425024390220642, + 0.0528571754693985, + 0.07934416085481644, + 0.0344993993639946 + ], + "credit_map": [ + 0.03218352049589157, + 0.041043996810913086, + 0.039829906076192856, + 0.02696792408823967, + 0.03101312182843685, + 0.031455326825380325, + 0.041837166994810104, + 0.02925710566341877, + 0.005694726016372442, + 0.0040933797135949135, + 0.003653182415291667, + 0.004829739220440388, + 0.0054277703166007996, + 0.005723973270505667, + 0.005619542673230171, + 0.005896919872611761, + 0.006032573524862528, + 0.00543385511264205, + 0.005542709492146969, + 0.00450865039601922, + 0.0056396424770355225, + 0.012345821596682072, + 0.005907838232815266, + 0.007800798397511244, + 0.0054796002805233, + 0.004425054416060448, + 0.009113321080803871, + 0.004210142884403467, + 0.00611727312207222, + 0.0054174624383449554, + 0.004113437607884407, + 0.005271067377179861, + 0.0050208717584609985, + 0.006310354918241501, + 0.005809814669191837, + 0.0034910673275589943, + 0.009272045455873013, + 0.003792396280914545, + 0.004986558109521866, + 0.0032593603245913982, + 0.03495152294635773, + 0.0437595508992672, + 0.034789856523275375, + 0.04104156419634819, + 0.033171821385622025, + 0.0138587336987257, + 0.019641760736703873, + 0.014005179516971111, + 0.015025229193270206, + 0.01710525155067444, + 0.017292410135269165, + 0.020549966022372246, + 0.020664576441049576, + 0.019320763647556305, + 0.01692914590239525, + 0.017362909391522408, + 0.006770974025130272, + 0.006747066508978605, + 0.0060310959815979, + 0.00493664899840951, + 0.004364448133856058, + 0.012405375018715858, + 0.005068799015134573, + 0.006532774772495031, + 0.010664565488696098, + 0.019686952233314514, + 0.009642611257731915, + 0.008416625671088696, + 0.009670164436101913, + 0.007876135408878326, + 0.00776281300932169, + 0.005624296609312296, + 0.004324824083596468, + 0.005218528676778078, + 0.0032822999637573957, + 0.0026677006389945745, + 0.006087163928896189, + 0.006331492215394974, + 0.0036285356618463993, + 0.0029588553588837385 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 6 + ], + "selected_reward": -0.07989362627267838 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a mechanical version of a character named Pomu with polished chrome limbs and glowing green fiber-optic hair in a high ponytail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character seated on a futuristic stage with a dark, polished floor that reflects the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character leaning forward in concentration while interacting with a translucent three-dimensional holographic interface using glowing sensor-equipped hands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a vast stadium filled with blurred silhouettes and massive holographic screens displaying match footage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video begin with a slow, cinematic push-in shot starting from a 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": [ + -0.01107768900692463 + ], + "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": [ + 0.2874757647514343, + 0.15474113821983337, + 0.10027319192886353, + 0.05762845277786255, + 0.06051301956176758, + 0.058066967874765396, + 0.05045901611447334, + 0.035087957978248596, + 0.11196186393499374, + 0.08379270881414413 + ], + "credit_map": [ + 0.03706042468547821, + 0.03432653844356537, + 0.03301256522536278, + 0.023640260100364685, + 0.04670238122344017, + 0.03811702877283096, + 0.040135275572538376, + 0.034481268376111984, + 0.014638934284448624, + 0.016215207055211067, + 0.01597101427614689, + 0.022216061130166054, + 0.01671641692519188, + 0.02415233850479126, + 0.02497723512351513, + 0.019853917881846428, + 0.013237033039331436, + 0.014903672970831394, + 0.01057733315974474, + 0.009229599498212337, + 0.012339708395302296, + 0.013182606548070908, + 0.014218643307685852, + 0.012584584765136242, + 0.0073735774494707584, + 0.003999706357717514, + 0.013275123201310635, + 0.004242641851305962, + 0.004218680318444967, + 0.006614815443754196, + 0.005294489674270153, + 0.012609411031007767, + 0.009849774651229382, + 0.0054720803163945675, + 0.00756407342851162, + 0.005743907764554024, + 0.011614429764449596, + 0.005265248008072376, + 0.008890816010534763, + 0.006112684495747089, + 0.004320717882364988, + 0.006842716597020626, + 0.006908580660820007, + 0.012454905547201633, + 0.006625466980040073, + 0.007864291779696941, + 0.006877873092889786, + 0.006172409281134605, + 0.006071938201785088, + 0.004069955553859472, + 0.007001703139394522, + 0.008230420760810375, + 0.006425173487514257, + 0.007602699566632509, + 0.005492371506989002, + 0.0055647497065365314, + 0.00440873671323061, + 0.0050139352679252625, + 0.00553562818095088, + 0.003860766999423504, + 0.003544105216860771, + 0.00658060796558857, + 0.003255834337323904, + 0.002888338640332222, + 0.01775536872446537, + 0.01447378657758236, + 0.01667756401002407, + 0.012897097505629063, + 0.008655215613543987, + 0.01683666557073593, + 0.011626321822404861, + 0.01303983572870493, + 0.015784302726387978, + 0.006107848137617111, + 0.01104669738560915, + 0.00928688608109951, + 0.005804781801998615, + 0.01523410715162754, + 0.011288569308817387, + 0.009239507839083672 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 8 + ], + "selected_reward": -0.01107768900692463 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a mechanical version of a character named Pomu with polished chrome limbs and glowing green fiber-optic hair in a high ponytail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character seated on a futuristic stage with a dark, polished floor that reflects the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character leaning forward in concentration while interacting with a translucent three-dimensional holographic interface using glowing sensor-equipped hands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a vast stadium filled with blurred silhouettes and massive holographic screens displaying match footage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video begin with a slow, cinematic push-in shot starting from a 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": [ + -0.01606954075396061 + ], + "opposite_logprob_tokens": [ + -4.141069412231445 + ], + "logprob_reward": -0.01606954075396061, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.01606954075396061, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2500351369380951, + 0.07961158454418182, + 0.06629301607608795, + 0.03854776918888092, + 0.03841787949204445, + 0.17758984863758087, + 0.16217055916786194, + 0.05972269922494888, + 0.06798490136861801, + 0.05962666869163513 + ], + "credit_map": [ + 0.033841636031866074, + 0.023560665547847748, + 0.03383547440171242, + 0.02208680287003517, + 0.03012116253376007, + 0.03236938640475273, + 0.040108583867549896, + 0.0341113843023777, + 0.008688459172844887, + 0.007374773267656565, + 0.006283903028815985, + 0.009059558622539043, + 0.009382284246385098, + 0.017650900408625603, + 0.012228633277118206, + 0.008943064138293266, + 0.008368045091629028, + 0.006553663406521082, + 0.005768679082393646, + 0.0044139716774225235, + 0.008225882425904274, + 0.013852201402187347, + 0.009813597425818443, + 0.009296968579292297, + 0.004349790513515472, + 0.0032047952990978956, + 0.005798844620585442, + 0.002832228085026145, + 0.0038215601816773415, + 0.007511436007916927, + 0.004128076136112213, + 0.006901038344949484, + 0.005025473423302174, + 0.004764426499605179, + 0.002945943037047982, + 0.004429014865309, + 0.007283568847924471, + 0.00468671228736639, + 0.00546093937009573, + 0.003821797901764512, + 0.026682235300540924, + 0.03343610838055611, + 0.028713874518871307, + 0.023025719448924065, + 0.02620294690132141, + 0.010721142403781414, + 0.017587032169103622, + 0.011220786720514297, + 0.016208099201321602, + 0.02675263211131096, + 0.018582431599497795, + 0.02776273526251316, + 0.021108532324433327, + 0.019344821572303772, + 0.017120089381933212, + 0.01529120746999979, + 0.007460311520844698, + 0.010512256994843483, + 0.008038349449634552, + 0.005076718516647816, + 0.007125618401914835, + 0.006185157690197229, + 0.007547078188508749, + 0.00777720520272851, + 0.011870117858052254, + 0.010118997655808926, + 0.008932581171393394, + 0.009313385933637619, + 0.006483312230557203, + 0.0078499224036932, + 0.007279898971319199, + 0.006136675830930471, + 0.006612788420170546, + 0.01046720240265131, + 0.010938409715890884, + 0.005305727943778038, + 0.0062453486025333405, + 0.006704892963171005, + 0.006992499344050884, + 0.0063597941771149635 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 6 + ], + "selected_reward": -0.01606954075396061 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a mechanical version of a character named Pomu with polished chrome limbs and glowing green fiber-optic hair in a high ponytail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character seated on a futuristic stage with a dark, polished floor that reflects the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character leaning forward in concentration while interacting with a translucent three-dimensional holographic interface using glowing sensor-equipped hands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a vast stadium filled with blurred silhouettes and massive holographic screens displaying match footage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video begin with a slow, cinematic push-in shot starting from a 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": [ + -0.17988517880439758 + ], + "opposite_logprob_tokens": [ + -1.8048851490020752 + ], + "logprob_reward": -0.17988517880439758, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.17988517880439758, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1850634515285492, + 0.054813411086797714, + 0.3209981322288513, + 0.08193229138851166, + 0.030688287690281868, + 0.06528257578611374, + 0.08360006660223007, + 0.04166338965296745, + 0.07374098151922226, + 0.062217358499765396 + ], + "credit_map": [ + 0.029151078313589096, + 0.02534591220319271, + 0.02141457423567772, + 0.014593418687582016, + 0.02913486585021019, + 0.025323469191789627, + 0.024335086345672607, + 0.01576506905257702, + 0.00616934010758996, + 0.00651237927377224, + 0.007691000588238239, + 0.00578716117888689, + 0.007875901646912098, + 0.008063522167503834, + 0.005741553381085396, + 0.006972558796405792, + 0.05005315691232681, + 0.02977644093334675, + 0.028949793428182602, + 0.02945861406624317, + 0.03277239948511124, + 0.06707358360290527, + 0.032218579202890396, + 0.05069559067487717, + 0.006023927591741085, + 0.005843299441039562, + 0.019186032935976982, + 0.005224681459367275, + 0.006946698296815157, + 0.016305236145853996, + 0.0064602927304804325, + 0.015942130237817764, + 0.005665984470397234, + 0.0032813060097396374, + 0.0057326629757881165, + 0.002377904951572418, + 0.0033088698983192444, + 0.002089638728648424, + 0.004561935551464558, + 0.0036699874326586723, + 0.008077731356024742, + 0.0119924983009696, + 0.011478228494524956, + 0.005736805964261293, + 0.009335500188171864, + 0.006074467208236456, + 0.006038379855453968, + 0.006548972334712744, + 0.009591813199222088, + 0.008691465482115746, + 0.009518499486148357, + 0.01153553370386362, + 0.010352677665650845, + 0.015111292712390423, + 0.01055104099214077, + 0.008247751742601395, + 0.004870086442679167, + 0.0035117310471832752, + 0.00464383652433753, + 0.004006719216704369, + 0.002651888644322753, + 0.0074804783798754215, + 0.005073826294392347, + 0.009424827061593533, + 0.011267362162470818, + 0.005274375434964895, + 0.010840166360139847, + 0.00907311774790287, + 0.00542081892490387, + 0.01293863169848919, + 0.008338522166013718, + 0.010587993077933788, + 0.008488638326525688, + 0.004371596500277519, + 0.0104557229205966, + 0.007182599510997534, + 0.004927746951580048, + 0.01109407376497984, + 0.00880296528339386, + 0.006894020829349756 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 6 + ], + "selected_reward": -0.17988517880439758 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a mechanical version of a character named Pomu with polished chrome limbs and glowing green fiber-optic hair in a high ponytail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character seated on a futuristic stage with a dark, polished floor that reflects the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character leaning forward in concentration while interacting with a translucent three-dimensional holographic interface using glowing sensor-equipped hands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a vast stadium filled with blurred silhouettes and massive holographic screens displaying match footage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video begin with a slow, cinematic push-in shot starting from a 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": [ + -0.04862729832530022 + ], + "opposite_logprob_tokens": [ + -3.0486273765563965 + ], + "logprob_reward": -0.04862729832530022, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04862729832530022, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1691354215145111, + 0.08913788199424744, + 0.2157231569290161, + 0.15247474610805511, + 0.04560108855366707, + 0.11842210590839386, + 0.04512428864836693, + 0.07126358896493912, + 0.057499293237924576, + 0.03561846911907196 + ], + "credit_map": [ + 0.025451313704252243, + 0.0269574336707592, + 0.01855791173875332, + 0.01545382384210825, + 0.027044834569096565, + 0.023194311186671257, + 0.019869433715939522, + 0.012606356292963028, + 0.015246626920998096, + 0.008738083764910698, + 0.007753495126962662, + 0.0098745571449399, + 0.01044786162674427, + 0.011206671595573425, + 0.012516869232058525, + 0.013353715650737286, + 0.0320342481136322, + 0.03029739111661911, + 0.022938305512070656, + 0.016718488186597824, + 0.029469314962625504, + 0.032765619456768036, + 0.028133446350693703, + 0.023366348817944527, + 0.015059174969792366, + 0.016740791499614716, + 0.02874334156513214, + 0.011853349395096302, + 0.022022970020771027, + 0.027613794431090355, + 0.01206937525421381, + 0.01837194710969925, + 0.009823966771364212, + 0.008198360912501812, + 0.0046326699666678905, + 0.0035596955567598343, + 0.009324771352112293, + 0.0028752190992236137, + 0.004136012401431799, + 0.0030503938905894756, + 0.018012922257184982, + 0.012379183433949947, + 0.009194912388920784, + 0.04415295273065567, + 0.00912485271692276, + 0.005888595711439848, + 0.012642346322536469, + 0.007026342209428549, + 0.004979768767952919, + 0.005060647614300251, + 0.004645244218409061, + 0.0070801605470478535, + 0.00530667370185256, + 0.008798649534583092, + 0.004689635708928108, + 0.004563512280583382, + 0.009552385658025742, + 0.007008287589997053, + 0.011630605906248093, + 0.0050057293847203255, + 0.005567293148487806, + 0.014318340457975864, + 0.006002089008688927, + 0.012178858742117882, + 0.008785785175859928, + 0.00801268219947815, + 0.006878932937979698, + 0.007562397047877312, + 0.0057627116329967976, + 0.0071131945587694645, + 0.00629901047796011, + 0.007084579672664404, + 0.004139149561524391, + 0.005836188327521086, + 0.004084787331521511, + 0.005181086715310812, + 0.004736101720482111, + 0.003561020363122225, + 0.004325995221734047, + 0.0037541401106864214 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 3 + ], + "selected_reward": -0.04862729832530022 + } + ], + "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/0000121/group_0/index_tensor([8306, 8306])_group_0_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000121/group_0/index_tensor([8306, 8306])_group_0_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..2b83e87acf02ff9aeb9834cdead5103336108170 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000121/group_0/index_tensor([8306, 8306])_group_0_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "slowly advancing medium shot, shot from a level angle, brings us into a warm, dimly lit living room. At the center of the frame is an antique wooden radio, its dark brown walnut case gleaming with a warm sheen in the dim light. Next to the finely textured speaker grille, a round glass tuning dial steadily emits a warm, soft amber glow. A family gathered intimately around a radio. A middle-aged father, his neatly combed dark hair streaked with silver, wore a beige turtleneck sweater and leaned forward slightly, listening attentively with a gentle, focused expression on his face. Beside him sat a mother with shoulder-length, curly brown hair, wearing a crimson wool cardigan. Her head rested gently on her husband's shoulder, her eyes slightly closed, a contented smile playing on her lips, completely absorbed in the sound. A boy, perhaps seven or eight, sat cross-legged on the thick carpet before them, his chin resting on his hands, his head tilted back, his large eyes gleaming with curiosity and fascination. The radio's sole, dim light was the primary source of light in the scene, gently illuminating each face, outlining its soft contours and casting long, almost motionless shadows on the wall behind them. The entire scene, bathed in warm tones, was permeated with a timeless tranquility, nostalgia, and familial warmth.", + "seed": 11615, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000121/group_0/index_8306_seed_11615_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": -0.041643927060067656, + "vlm_logprob": -0.041643927060067656, + "vlm_yes_no_margin": 3.724999952316284, + "vlm_token_credit": -0.041643927060067656, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 warm, dimly lit living room?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is an antique wooden radio with a dark brown walnut case visible at the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a round glass tuning dial on the radio emitting a warm, soft amber glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are three family members present: a father in a beige turtleneck, a mother in a crimson wool cardigan, and a young boy sitting cross-legged on the carpet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated primarily by the dim light of the radio, casting long shadows on the wall?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03230872377753258 + ], + "opposite_logprob_tokens": [ + -4.032308578491211 + ], + "logprob_reward": -0.03230872377753258, + "yes_no_margin_reward": 3.999999761581421, + "token_credit_reward": -0.03230872377753258, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2407413274049759, + 0.06782769411802292, + 0.03851810470223427, + 0.0335397906601429, + 0.04783210530877113, + 0.13030003011226654, + 0.19136282801628113, + 0.10253613442182541, + 0.0550653338432312, + 0.0922766849398613 + ], + "credit_map": [ + 0.018117109313607216, + 0.022161409258842468, + 0.02960384637117386, + 0.01867080107331276, + 0.028000231832265854, + 0.04421175643801689, + 0.02221749722957611, + 0.05775866657495499, + 0.005864176899194717, + 0.00859674159437418, + 0.008515174500644207, + 0.0072148824110627174, + 0.008431586436927319, + 0.008175675757229328, + 0.010865075513720512, + 0.010164371691644192, + 0.0030452809296548367, + 0.004968114662915468, + 0.005331230815500021, + 0.003982632886618376, + 0.006254532374441624, + 0.004508380312472582, + 0.0056928363628685474, + 0.0047350977547466755, + 0.002339553087949753, + 0.0035546645522117615, + 0.00406354945152998, + 0.0036976311821490526, + 0.004373465199023485, + 0.004283669404685497, + 0.006521517410874367, + 0.004705740604549646, + 0.0035848210100084543, + 0.0067276982590556145, + 0.005878557451069355, + 0.003734149504452944, + 0.006271139718592167, + 0.006585101597011089, + 0.006195876747369766, + 0.008854759857058525, + 0.008277052082121372, + 0.013180679641664028, + 0.016123926267027855, + 0.01184649858623743, + 0.013113658875226974, + 0.022842245176434517, + 0.012952163815498352, + 0.03196379914879799, + 0.012697196565568447, + 0.01587303727865219, + 0.021630223840475082, + 0.016767587512731552, + 0.02150687761604786, + 0.029589639976620674, + 0.0355931855738163, + 0.037705086171627045, + 0.008012142032384872, + 0.010563117451965809, + 0.01597689650952816, + 0.009831885807216167, + 0.01084914244711399, + 0.013286017812788486, + 0.01626836508512497, + 0.017748571932315826, + 0.002607818925753236, + 0.003610266838222742, + 0.00454824510961771, + 0.0032712710089981556, + 0.0033241920173168182, + 0.017966507002711296, + 0.003504495369270444, + 0.016232535243034363, + 0.00584139209240675, + 0.010877645574510098, + 0.00957967434078455, + 0.007759410887956619, + 0.010993736796081066, + 0.023712288588285446, + 0.009232506155967712, + 0.014280027709901333 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 5 + ], + "selected_reward": -0.03230872377753258 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 warm, dimly lit living room?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is an antique wooden radio with a dark brown walnut case visible at the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a round glass tuning dial on the radio emitting a warm, soft amber glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are three family members present: a father in a beige turtleneck, a mother in a crimson wool cardigan, and a young boy sitting cross-legged on the carpet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated primarily by the dim light of the radio, casting long shadows on the wall?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023303447291254997 + ], + "opposite_logprob_tokens": [ + -3.773303508758545 + ], + "logprob_reward": -0.023303447291254997, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023303447291254997, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3660331666469574, + 0.12907539308071136, + 0.06589873135089874, + 0.08313656598329544, + 0.04318396747112274, + 0.04164518043398857, + 0.05884590744972229, + 0.05339295044541359, + 0.059752486646175385, + 0.0990355983376503 + ], + "credit_map": [ + 0.023491082713007927, + 0.03685298189520836, + 0.04064299538731575, + 0.034654878079891205, + 0.04499358311295509, + 0.06057620421051979, + 0.054818201810121536, + 0.07000324875116348, + 0.008228764869272709, + 0.012910251505672932, + 0.014967872761189938, + 0.014497638680040836, + 0.015528973191976547, + 0.016180966049432755, + 0.02697245590388775, + 0.01978846825659275, + 0.00475622434169054, + 0.0062844459898769855, + 0.007345458026975393, + 0.00792381726205349, + 0.007976165041327477, + 0.008060810156166553, + 0.013272631913423538, + 0.01027918141335249, + 0.004857576917856932, + 0.007312139496207237, + 0.007435486186295748, + 0.009128483012318611, + 0.008605459704995155, + 0.011631551198661327, + 0.017069373279809952, + 0.017096498981118202, + 0.003120721084997058, + 0.0034674559719860554, + 0.004910156596451998, + 0.0036811314057558775, + 0.0038061002269387245, + 0.008740561082959175, + 0.0051431371830403805, + 0.010314702056348324, + 0.002653110073879361, + 0.003577145515009761, + 0.0044981446117162704, + 0.003835894400253892, + 0.00529909785836935, + 0.00629069097340107, + 0.0073576378636062145, + 0.00813345704227686, + 0.004454619251191616, + 0.007096611894667149, + 0.007462128531187773, + 0.0067098382860422134, + 0.007628070190548897, + 0.0093142781406641, + 0.00891268439590931, + 0.007267675828188658, + 0.00519985007122159, + 0.005204824730753899, + 0.006378730293363333, + 0.005768206436187029, + 0.005944857373833656, + 0.007489196956157684, + 0.011000566184520721, + 0.006406719330698252, + 0.005172929260879755, + 0.00796420592814684, + 0.01077363733202219, + 0.005069159436970949, + 0.006123627070337534, + 0.00738383736461401, + 0.0061491006053984165, + 0.011115984059870243, + 0.010533117689192295, + 0.012660977430641651, + 0.01390468142926693, + 0.008633165620267391, + 0.012130766175687313, + 0.013636967167258263, + 0.010935519821941853, + 0.01660040020942688 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 9 + ], + "selected_reward": -0.023303447291254997 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 warm, dimly lit living room?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is an antique wooden radio with a dark brown walnut case visible at the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a round glass tuning dial on the radio emitting a warm, soft amber glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are three family members present: a father in a beige turtleneck, a mother in a crimson wool cardigan, and a young boy sitting cross-legged on the carpet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated primarily by the dim light of the radio, casting long shadows on the wall?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003619787283241749 + ], + "opposite_logprob_tokens": [ + -5.62861967086792 + ], + "logprob_reward": -0.003619787283241749, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.003619787283241749, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2533550560474396, + 0.1070585623383522, + 0.09955629706382751, + 0.07193989306688309, + 0.05518164858222008, + 0.06162051483988762, + 0.06124653294682503, + 0.037813566625118256, + 0.03905561566352844, + 0.2131723314523697 + ], + "credit_map": [ + 0.019245309755206108, + 0.02591806836426258, + 0.02856381982564926, + 0.029117032885551453, + 0.03233583644032478, + 0.0321049727499485, + 0.05259498581290245, + 0.033475033938884735, + 0.007514888420701027, + 0.010634932667016983, + 0.011710482649505138, + 0.010968904942274094, + 0.013173045590519905, + 0.01639159768819809, + 0.014846994541585445, + 0.021817706525325775, + 0.007648609112948179, + 0.00847373716533184, + 0.011334391310811043, + 0.010420954786241055, + 0.014919091947376728, + 0.011154497042298317, + 0.02206016331911087, + 0.013544846326112747, + 0.0051747155375778675, + 0.006590760312974453, + 0.008416474796831608, + 0.0077861277386546135, + 0.010586380027234554, + 0.009325414896011353, + 0.016037466004490852, + 0.008022553287446499, + 0.00466250441968441, + 0.005471743643283844, + 0.005949187558144331, + 0.005183620378375053, + 0.0068683139979839325, + 0.007038095500320196, + 0.01114944089204073, + 0.008858746849000454, + 0.004694279748946428, + 0.004999008495360613, + 0.007020018994808197, + 0.005644700024276972, + 0.008053532801568508, + 0.01049085147678852, + 0.011217032559216022, + 0.00950109027326107, + 0.004813948646187782, + 0.006264377851039171, + 0.008181494660675526, + 0.006977350451052189, + 0.006348690949380398, + 0.006670221220701933, + 0.013447030447423458, + 0.008543416857719421, + 0.003015053691342473, + 0.003599877003580332, + 0.004269518423825502, + 0.00419717188924551, + 0.004827749915421009, + 0.004086689557880163, + 0.009210830554366112, + 0.004606671165674925, + 0.002900240244343877, + 0.0045986357145011425, + 0.005381220486015081, + 0.003367421915754676, + 0.0034880712628364563, + 0.009582212194800377, + 0.004143558908253908, + 0.005594255402684212, + 0.010301467962563038, + 0.011224261485040188, + 0.02153993397951126, + 0.009800990112125874, + 0.01235986314713955, + 0.07543929666280746, + 0.015567298978567123, + 0.056939225643873215 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 1 + ], + "selected_reward": -0.003619787283241749 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 warm, dimly lit living room?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is an antique wooden radio with a dark brown walnut case visible at the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a round glass tuning dial on the radio emitting a warm, soft amber glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are three family members present: a father in a beige turtleneck, a mother in a crimson wool cardigan, and a young boy sitting cross-legged on the carpet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated primarily by the dim light of the radio, casting long shadows on the wall?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04866408929228783 + ], + "opposite_logprob_tokens": [ + -3.048664093017578 + ], + "logprob_reward": -0.04866408929228783, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04866408929228783, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3040740489959717, + 0.1381475180387497, + 0.09652779251337051, + 0.04333781823515892, + 0.061852697283029556, + 0.044241782277822495, + 0.06515294313430786, + 0.06587810814380646, + 0.06314364820718765, + 0.11764362454414368 + ], + "credit_map": [ + 0.022737564519047737, + 0.04133956879377365, + 0.042190976440906525, + 0.04054488614201546, + 0.047154974192380905, + 0.03749231621623039, + 0.03171228617429733, + 0.040901511907577515, + 0.007418135181069374, + 0.012245266698300838, + 0.014950969256460667, + 0.013230615295469761, + 0.01364944875240326, + 0.031514883041381836, + 0.012641001492738724, + 0.032497208565473557, + 0.00597990769892931, + 0.009509114548563957, + 0.011310025118291378, + 0.01126704178750515, + 0.010208596475422382, + 0.014587469398975372, + 0.018874259665608406, + 0.014791376888751984, + 0.0029795051086694, + 0.005265753250569105, + 0.0045798891223967075, + 0.00526438420638442, + 0.005496826022863388, + 0.004993858747184277, + 0.0083934236317873, + 0.0063641807064414024, + 0.0037098790053278208, + 0.005185799207538366, + 0.006882920395582914, + 0.005386210512369871, + 0.007310396991670132, + 0.011877837590873241, + 0.006916578393429518, + 0.01458308007568121, + 0.002981588477268815, + 0.003804219188168645, + 0.00523108197376132, + 0.004058422986418009, + 0.0056247771717607975, + 0.00637503806501627, + 0.007739606779068708, + 0.008427050895988941, + 0.004658964462578297, + 0.008298024535179138, + 0.008647495880723, + 0.006698370445519686, + 0.008557752706110477, + 0.00879801157861948, + 0.010114148259162903, + 0.009380181320011616, + 0.005600583739578724, + 0.007382780313491821, + 0.00772357452660799, + 0.006357845384627581, + 0.006261073052883148, + 0.009478067979216576, + 0.010007564909756184, + 0.013066619634628296, + 0.004735168535262346, + 0.0065206182189285755, + 0.010712996125221252, + 0.005912454333156347, + 0.006289207376539707, + 0.007136001251637936, + 0.00873575359582901, + 0.013101451098918915, + 0.006199853029102087, + 0.007121863774955273, + 0.012217243202030659, + 0.006776651367545128, + 0.007067359983921051, + 0.04043405130505562, + 0.010273161344230175, + 0.02755345217883587 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 9 + ], + "selected_reward": -0.04866408929228783 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 warm, dimly lit living room?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is an antique wooden radio with a dark brown walnut case visible at the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a round glass tuning dial on the radio emitting a warm, soft amber glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are three family members present: a father in a beige turtleneck, a mother in a crimson wool cardigan, and a young boy sitting cross-legged on the carpet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated primarily by the dim light of the radio, casting long shadows on the wall?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10032358765602112 + ], + "opposite_logprob_tokens": [ + -2.3503236770629883 + ], + "logprob_reward": -0.10032358765602112, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10032358765602112, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2796514928340912, + 0.11034233123064041, + 0.060145244002342224, + 0.060463324189186096, + 0.09802927821874619, + 0.04744350165128708, + 0.05146714299917221, + 0.04190871864557266, + 0.06154971197247505, + 0.18899919092655182 + ], + "credit_map": [ + 0.025938155129551888, + 0.034131016582250595, + 0.029986100271344185, + 0.02737846039235592, + 0.05194340646266937, + 0.041088785976171494, + 0.03137915953993797, + 0.03780645877122879, + 0.010485725477337837, + 0.012900454923510551, + 0.010104148648679256, + 0.009226029738783836, + 0.02166799269616604, + 0.01836629956960678, + 0.009824266657233238, + 0.01776742935180664, + 0.004786175675690174, + 0.006122425198554993, + 0.00684878695756197, + 0.006982376333326101, + 0.011170078068971634, + 0.008257817476987839, + 0.009208924137055874, + 0.006768667604774237, + 0.0050402297638356686, + 0.005864844191819429, + 0.005799765232950449, + 0.0073308516293764114, + 0.008391090668737888, + 0.007790491450577974, + 0.011677419766783714, + 0.008568641729652882, + 0.00854543037712574, + 0.011959025636315346, + 0.011405609548091888, + 0.008448141627013683, + 0.014364541508257389, + 0.014304537326097488, + 0.015194537118077278, + 0.013807471841573715, + 0.004340049345046282, + 0.004877005238085985, + 0.005247155670076609, + 0.0052798292599618435, + 0.006280723959207535, + 0.006683853920549154, + 0.007358554285019636, + 0.007376336958259344, + 0.004022030159831047, + 0.00594810675829649, + 0.007559188175946474, + 0.005065247416496277, + 0.004941833671182394, + 0.0061226896941661835, + 0.007777543272823095, + 0.010030509904026985, + 0.003604155033826828, + 0.004284055437892675, + 0.004585203714668751, + 0.003751305630430579, + 0.004302452318370342, + 0.005876054055988789, + 0.007777557708323002, + 0.007727938238531351, + 0.004374596290290356, + 0.005607324186712503, + 0.006737447343766689, + 0.0038616855163127184, + 0.0051993574015796185, + 0.015725255012512207, + 0.00400423351675272, + 0.016039816662669182, + 0.02403314597904682, + 0.02160046063363552, + 0.026852257549762726, + 0.01732317917048931, + 0.019419291988015175, + 0.04307136684656143, + 0.014369793236255646, + 0.02232970856130123 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 1 + ], + "selected_reward": -0.10032358765602112 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "slowly advancing medium shot, shot from a level angle, brings us into a warm, dimly lit living room. At the center of the frame is an antique wooden radio, its dark brown walnut case gleaming with a warm sheen in the dim light. Next to the finely textured speaker grille, a round glass tuning dial steadily emits a warm, soft amber glow. A family gathered intimately around a radio. A middle-aged father, his neatly combed dark hair streaked with silver, wore a beige turtleneck sweater and leaned forward slightly, listening attentively with a gentle, focused expression on his face. Beside him sat a mother with shoulder-length, curly brown hair, wearing a crimson wool cardigan. Her head rested gently on her husband's shoulder, her eyes slightly closed, a contented smile playing on her lips, completely absorbed in the sound. A boy, perhaps seven or eight, sat cross-legged on the thick carpet before them, his chin resting on his hands, his head tilted back, his large eyes gleaming with curiosity and fascination. The radio's sole, dim light was the primary source of light in the scene, gently illuminating each face, outlining its soft contours and casting long, almost motionless shadows on the wall behind them. The entire scene, bathed in warm tones, was permeated with a timeless tranquility, nostalgia, and familial warmth.", + "seed": 11616, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000121/group_0/index_8306_seed_11616_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": -0.06777299651876093, + "vlm_logprob": -0.06777299651876093, + "vlm_yes_no_margin": 3.45, + "vlm_token_credit": -0.06777299651876093, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 warm, dimly lit living room?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is an antique wooden radio with a dark brown walnut case visible at the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a round glass tuning dial on the radio emitting a warm, soft amber glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are three family members present: a father in a beige turtleneck, a mother in a crimson wool cardigan, and a young boy sitting cross-legged on the carpet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated primarily by the dim light of the radio, casting long shadows on the wall?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.021195685490965843 + ], + "opposite_logprob_tokens": [ + -4.396195888519287 + ], + "logprob_reward": -0.021195685490965843, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.021195685490965843, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2764935791492462, + 0.09212519973516464, + 0.040241532027721405, + 0.044603049755096436, + 0.08977967500686646, + 0.20530664920806885, + 0.05843861773610115, + 0.041385021060705185, + 0.10206212103366852, + 0.04956452548503876 + ], + "credit_map": [ + 0.03178098797798157, + 0.027230408042669296, + 0.04366346076130867, + 0.038387976586818695, + 0.024757346138358116, + 0.03147151693701744, + 0.03729579597711563, + 0.041906118392944336, + 0.008115006610751152, + 0.008389188908040524, + 0.017284095287322998, + 0.014829742722213268, + 0.012897004373371601, + 0.009802207350730896, + 0.013278421014547348, + 0.007529542315751314, + 0.0042547970078885555, + 0.004419170320034027, + 0.006744364742189646, + 0.005077153444290161, + 0.004156418610364199, + 0.004745954181998968, + 0.005361226852983236, + 0.005482451058924198, + 0.0041683209128677845, + 0.004074073862284422, + 0.00585919851437211, + 0.004027993883937597, + 0.007056327536702156, + 0.0062582711689174175, + 0.006089182570576668, + 0.007069686893373728, + 0.006693490780889988, + 0.008244662545621395, + 0.011500678025186062, + 0.0060771615244448185, + 0.010268141515552998, + 0.01735336147248745, + 0.013094916008412838, + 0.016547275707125664, + 0.01832488738000393, + 0.013471025042235851, + 0.033962447196245193, + 0.013461003080010414, + 0.01841866783797741, + 0.040135059505701065, + 0.030498208478093147, + 0.03703538328409195, + 0.005768896080553532, + 0.00826133694499731, + 0.007845010608434677, + 0.006578220520168543, + 0.00840760674327612, + 0.00602188752964139, + 0.008605299517512321, + 0.006950365845113993, + 0.003783817170187831, + 0.005036402028053999, + 0.0059353625401854515, + 0.0040294099599123, + 0.005684426054358482, + 0.004402007441967726, + 0.006167009472846985, + 0.0063465917482972145, + 0.008161093108355999, + 0.007510871160775423, + 0.01339771505445242, + 0.011062043718993664, + 0.01212509348988533, + 0.016482438892126083, + 0.014602256938815117, + 0.01872061751782894, + 0.004299862310290337, + 0.0039031433407217264, + 0.006933409720659256, + 0.00684988871216774, + 0.007789915893226862, + 0.005755069199949503, + 0.007201786153018475, + 0.006831453647464514 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 8 + ], + "selected_reward": -0.021195685490965843 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 warm, dimly lit living room?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is an antique wooden radio with a dark brown walnut case visible at the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a round glass tuning dial on the radio emitting a warm, soft amber glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are three family members present: a father in a beige turtleneck, a mother in a crimson wool cardigan, and a young boy sitting cross-legged on the carpet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated primarily by the dim light of the radio, casting long shadows on the wall?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.038118645548820496 + ], + "opposite_logprob_tokens": [ + -3.288118600845337 + ], + "logprob_reward": -0.038118645548820496, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.038118645548820496, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18526321649551392, + 0.17702794075012207, + 0.12601134181022644, + 0.0697745755314827, + 0.19016975164413452, + 0.057761289179325104, + 0.03393649682402611, + 0.04086991399526596, + 0.06725221127271652, + 0.05193323642015457 + ], + "credit_map": [ + 0.028129031881690025, + 0.01635568030178547, + 0.03596381098031998, + 0.015957854688167572, + 0.023016568273305893, + 0.021842362359166145, + 0.025721684098243713, + 0.018276218324899673, + 0.018753478303551674, + 0.015105430968105793, + 0.025484897196292877, + 0.01733170449733734, + 0.015425747260451317, + 0.02186240814626217, + 0.02254444546997547, + 0.040519844740629196, + 0.01603611931204796, + 0.011520533822476864, + 0.029797149822115898, + 0.011871332302689552, + 0.012779044918715954, + 0.01584586501121521, + 0.018306223675608635, + 0.009855085983872414, + 0.005128942895680666, + 0.005705597810447216, + 0.010790984146296978, + 0.0069009470753371716, + 0.011202740482985973, + 0.012373574078083038, + 0.010900775901973248, + 0.006771019194275141, + 0.01478175912052393, + 0.014463103376328945, + 0.03504345193505287, + 0.012269441969692707, + 0.025193369016051292, + 0.03383522108197212, + 0.029298007488250732, + 0.025285402312874794, + 0.008710544556379318, + 0.006750164553523064, + 0.008497439324855804, + 0.004903651308268309, + 0.006910291034728289, + 0.007844881154596806, + 0.007886765524744987, + 0.006257552653551102, + 0.004103252198547125, + 0.0032870229333639145, + 0.0041043926030397415, + 0.0026543273124843836, + 0.005266581661999226, + 0.005644878372550011, + 0.00503530353307724, + 0.003840738907456398, + 0.006582483649253845, + 0.003945811185985804, + 0.005771327763795853, + 0.0036880795378237963, + 0.006967096123844385, + 0.005224125925451517, + 0.005056244786828756, + 0.003634748049080372, + 0.009112725034356117, + 0.007157362997531891, + 0.010080906562507153, + 0.005657339934259653, + 0.007221468258649111, + 0.007287828717380762, + 0.00995691679418087, + 0.010777665302157402, + 0.005302507430315018, + 0.004199305549263954, + 0.006094248965382576, + 0.007162374909967184, + 0.010408943518996239, + 0.006249383557587862, + 0.007236723322421312, + 0.005279750097543001 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 1 + ], + "selected_reward": -0.038118645548820496 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 warm, dimly lit living room?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is an antique wooden radio with a dark brown walnut case visible at the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a round glass tuning dial on the radio emitting a warm, soft amber glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are three family members present: a father in a beige turtleneck, a mother in a crimson wool cardigan, and a young boy sitting cross-legged on the carpet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated primarily by the dim light of the radio, casting long shadows on the wall?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0052567594684660435 + ], + "opposite_logprob_tokens": [ + -5.255256652832031 + ], + "logprob_reward": -0.0052567594684660435, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.0052567594684660435, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.36585840582847595, + 0.18232078850269318, + 0.05627085268497467, + 0.045878566801548004, + 0.06835294514894485, + 0.09045381098985672, + 0.03628682717680931, + 0.049165163189172745, + 0.05891970545053482, + 0.04649295657873154 + ], + "credit_map": [ + 0.051908254623413086, + 0.02499675564467907, + 0.04547221586108208, + 0.027749935165047646, + 0.040960200130939484, + 0.04780774563550949, + 0.05136356130242348, + 0.07559975236654282, + 0.015363240614533424, + 0.0135853486135602, + 0.024578817188739777, + 0.02154211699962616, + 0.02593967132270336, + 0.02384200133383274, + 0.025208402425050735, + 0.0322611965239048, + 0.005421923473477364, + 0.0052211228758096695, + 0.010773186571896076, + 0.005859958007931709, + 0.00849066860973835, + 0.005999756045639515, + 0.007762342691421509, + 0.00674189580604434, + 0.0030392634216696024, + 0.004598353989422321, + 0.005559496581554413, + 0.003589226631447673, + 0.0066938032396137714, + 0.00808352418243885, + 0.005295991897583008, + 0.009018911980092525, + 0.006494298577308655, + 0.005474064499139786, + 0.0077784499153494835, + 0.006169924046844244, + 0.007110644597560167, + 0.012992484495043755, + 0.008670123293995857, + 0.013662957586348057, + 0.006884000264108181, + 0.007630827371031046, + 0.010823906399309635, + 0.005886375904083252, + 0.009984869509935379, + 0.018706314265727997, + 0.013226393610239029, + 0.017311127856373787, + 0.004352155141532421, + 0.0028375054243952036, + 0.006644459906965494, + 0.0027097188867628574, + 0.005274548195302486, + 0.004482637159526348, + 0.006040416657924652, + 0.003945390228182077, + 0.008069307543337345, + 0.004915742669254541, + 0.007947249338030815, + 0.006017481442540884, + 0.006769149098545313, + 0.0034600638318806887, + 0.006888009142130613, + 0.005098163615912199, + 0.005790642462670803, + 0.006299609784036875, + 0.0075484164990484715, + 0.00464674411341548, + 0.005964101757854223, + 0.007441544905304909, + 0.007987570017576218, + 0.01324107963591814, + 0.006052887532860041, + 0.005102195776998997, + 0.007755057420581579, + 0.005018881056457758, + 0.005613121669739485, + 0.005106967408210039, + 0.007120748516172171, + 0.004723099060356617 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 5 + ], + "selected_reward": -0.0052567594684660435 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 warm, dimly lit living room?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is an antique wooden radio with a dark brown walnut case visible at the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a round glass tuning dial on the radio emitting a warm, soft amber glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are three family members present: a father in a beige turtleneck, a mother in a crimson wool cardigan, and a young boy sitting cross-legged on the carpet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated primarily by the dim light of the radio, casting long shadows on the wall?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.048663295805454254 + ], + "opposite_logprob_tokens": [ + -3.048663377761841 + ], + "logprob_reward": -0.048663295805454254, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.048663295805454254, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3370437026023865, + 0.14718055725097656, + 0.07341433316469193, + 0.029886417090892792, + 0.06545943766832352, + 0.0946919247508049, + 0.08329009264707565, + 0.08269493281841278, + 0.0473211370408535, + 0.03901747986674309 + ], + "credit_map": [ + 0.032329875975847244, + 0.0352226197719574, + 0.04316010698676109, + 0.04015492647886276, + 0.049080487340688705, + 0.043226711452007294, + 0.05049712955951691, + 0.043371859937906265, + 0.014842471107840538, + 0.01933843269944191, + 0.022505123168230057, + 0.02058251015841961, + 0.02107958309352398, + 0.01284965593367815, + 0.021520929411053658, + 0.014461850747466087, + 0.008474339731037617, + 0.007969730533659458, + 0.01097807940095663, + 0.00910206325352192, + 0.009602990932762623, + 0.009863773360848427, + 0.009922580793499947, + 0.007500773295760155, + 0.0029493619222193956, + 0.004398796241730452, + 0.004523574840277433, + 0.0032816175371408463, + 0.003091952297836542, + 0.004657563287764788, + 0.0038195617962628603, + 0.003163987537845969, + 0.0059569282457232475, + 0.007519579958170652, + 0.010777143761515617, + 0.007469938602298498, + 0.007700345944613218, + 0.008652451448142529, + 0.010133148171007633, + 0.007249897811561823, + 0.01551990956068039, + 0.011257395148277283, + 0.01094811037182808, + 0.009775999002158642, + 0.01014222763478756, + 0.01010938361287117, + 0.013180259615182877, + 0.013758636079728603, + 0.011075847782194614, + 0.011204004287719727, + 0.012612955644726753, + 0.008266228251159191, + 0.010649465024471283, + 0.008106636814773083, + 0.012481600977480412, + 0.008893348276615143, + 0.010342301800847054, + 0.00745510496199131, + 0.01449175551533699, + 0.007090229541063309, + 0.010763989761471748, + 0.008325034752488136, + 0.0120638906955719, + 0.01216263510286808, + 0.004708166234195232, + 0.007564912084490061, + 0.007223647553473711, + 0.004064077977091074, + 0.007457304280251265, + 0.00413453159853816, + 0.006573792546987534, + 0.005594702437520027, + 0.005659932270646095, + 0.00409869896247983, + 0.006042109336704016, + 0.006376610603183508, + 0.0037041865289211273, + 0.0027266577817499638, + 0.005919100251048803, + 0.0044901869259774685 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 5 + ], + "selected_reward": -0.048663295805454254 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 warm, dimly lit living room?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is an antique wooden radio with a dark brown walnut case visible at the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a round glass tuning dial on the radio emitting a warm, soft amber glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are three family members present: a father in a beige turtleneck, a mother in a crimson wool cardigan, and a young boy sitting cross-legged on the carpet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated primarily by the dim light of the radio, casting long shadows on the wall?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.22563059628009796 + ], + "opposite_logprob_tokens": [ + -1.6006306409835815 + ], + "logprob_reward": -0.22563059628009796, + "yes_no_margin_reward": 1.375, + "token_credit_reward": -0.22563059628009796, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.267643541097641, + 0.1512490063905716, + 0.0573025718331337, + 0.0337567962706089, + 0.02712238021194935, + 0.28434520959854126, + 0.05292949080467224, + 0.03330662474036217, + 0.051579251885414124, + 0.040765129029750824 + ], + "credit_map": [ + 0.02740846946835518, + 0.025460729375481606, + 0.03634529933333397, + 0.028048448264598846, + 0.03943301737308502, + 0.03204222023487091, + 0.038202203810214996, + 0.04070315510034561, + 0.012301236391067505, + 0.010302357375621796, + 0.03005349077284336, + 0.011969617567956448, + 0.018497804179787636, + 0.014959934167563915, + 0.02169431746006012, + 0.03147026523947716, + 0.00546224694699049, + 0.006237156223505735, + 0.0070680854842066765, + 0.00596303166821599, + 0.009053574874997139, + 0.008368500508368015, + 0.007852667942643166, + 0.007297309581190348, + 0.002138894284144044, + 0.0024040844291448593, + 0.004806101322174072, + 0.002596046542748809, + 0.005559294018894434, + 0.005343504250049591, + 0.004757972434163094, + 0.0061509027145802975, + 0.002421194687485695, + 0.0027655556332319975, + 0.0035785322543233633, + 0.002322871470823884, + 0.0031737072858959436, + 0.004900437314063311, + 0.0038540810346603394, + 0.004106004256755114, + 0.026469437405467033, + 0.0212444756180048, + 0.033429767936468124, + 0.019174393266439438, + 0.02729221247136593, + 0.05564672872424126, + 0.043717965483665466, + 0.057370223104953766, + 0.0044417791068553925, + 0.005240579601377249, + 0.00696946308016777, + 0.004616101738065481, + 0.009267299436032772, + 0.006371514406055212, + 0.007708590477705002, + 0.008314169012010098, + 0.003794904565438628, + 0.0038979332894086838, + 0.004076855257153511, + 0.003732306882739067, + 0.003934475593268871, + 0.0037880034651607275, + 0.004972804803401232, + 0.005109346006065607, + 0.005601265002042055, + 0.0061174542643129826, + 0.005052102729678154, + 0.003754907753318548, + 0.006453824695199728, + 0.0069029624573886395, + 0.007358313072472811, + 0.010338426567614079, + 0.0033030654303729534, + 0.0036017356906086206, + 0.006126205902546644, + 0.00307058310136199, + 0.006887214723974466, + 0.006451249122619629, + 0.005774955730885267, + 0.00555012421682477 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 1 + ], + "selected_reward": -0.22563059628009796 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "slowly advancing medium shot, shot from a level angle, brings us into a warm, dimly lit living room. At the center of the frame is an antique wooden radio, its dark brown walnut case gleaming with a warm sheen in the dim light. Next to the finely textured speaker grille, a round glass tuning dial steadily emits a warm, soft amber glow. A family gathered intimately around a radio. A middle-aged father, his neatly combed dark hair streaked with silver, wore a beige turtleneck sweater and leaned forward slightly, listening attentively with a gentle, focused expression on his face. Beside him sat a mother with shoulder-length, curly brown hair, wearing a crimson wool cardigan. Her head rested gently on her husband's shoulder, her eyes slightly closed, a contented smile playing on her lips, completely absorbed in the sound. A boy, perhaps seven or eight, sat cross-legged on the thick carpet before them, his chin resting on his hands, his head tilted back, his large eyes gleaming with curiosity and fascination. The radio's sole, dim light was the primary source of light in the scene, gently illuminating each face, outlining its soft contours and casting long, almost motionless shadows on the wall behind them. The entire scene, bathed in warm tones, was permeated with a timeless tranquility, nostalgia, and familial warmth.", + "seed": 11617, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000121/group_0/index_8306_seed_11617_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": -0.02863248309586197, + "vlm_logprob": -0.02863248309586197, + "vlm_yes_no_margin": 4.225, + "vlm_token_credit": -0.02863248309586197, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 warm, dimly lit living room?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is an antique wooden radio with a dark brown walnut case visible at the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a round glass tuning dial on the radio emitting a warm, soft amber glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are three family members present: a father in a beige turtleneck, a mother in a crimson wool cardigan, and a young boy sitting cross-legged on the carpet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated primarily by the dim light of the radio, casting long shadows on the wall?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01883596181869507 + ], + "opposite_logprob_tokens": [ + -4.26883602142334 + ], + "logprob_reward": -0.01883596181869507, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.01883596181869507, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24153639376163483, + 0.12778615951538086, + 0.10611497610807419, + 0.0659046396613121, + 0.07414406538009644, + 0.05024366453289986, + 0.11785895377397537, + 0.029985159635543823, + 0.14097823202610016, + 0.04544776305556297 + ], + "credit_map": [ + 0.028412770479917526, + 0.03127209469676018, + 0.03861142322421074, + 0.03241566941142082, + 0.027978911995887756, + 0.0321638397872448, + 0.027950923889875412, + 0.02273075468838215, + 0.018593499436974525, + 0.016078919172286987, + 0.018086683005094528, + 0.016814684495329857, + 0.016223475337028503, + 0.015066164545714855, + 0.013478787615895271, + 0.01344394963234663, + 0.010520847514271736, + 0.011264875531196594, + 0.017374126240611076, + 0.0154562471434474, + 0.015451336279511452, + 0.014821527525782585, + 0.012004580348730087, + 0.00922143179923296, + 0.009216598235070705, + 0.007240342907607555, + 0.009202120825648308, + 0.008235766552388668, + 0.010444614104926586, + 0.009717167355120182, + 0.006377825513482094, + 0.0054702055640518665, + 0.008279840461909771, + 0.010969002731144428, + 0.009694774635136127, + 0.010327798314392567, + 0.01081128604710102, + 0.009216256439685822, + 0.008495993912220001, + 0.006349114701151848, + 0.00452541233971715, + 0.006865581497550011, + 0.0069244978949427605, + 0.00533087644726038, + 0.006171069573611021, + 0.008815468288958073, + 0.004596142563968897, + 0.00701461685821414, + 0.011140669696033001, + 0.013963941484689713, + 0.017601193860173225, + 0.017397942021489143, + 0.014804577454924583, + 0.013997445814311504, + 0.016891740262508392, + 0.012061444111168385, + 0.0035491124726831913, + 0.004358185455203056, + 0.0038084823172539473, + 0.0033408510498702526, + 0.003521825885400176, + 0.0036461511626839638, + 0.0035486870910972357, + 0.004211864899843931, + 0.017528001219034195, + 0.03578021749854088, + 0.01929546147584915, + 0.01690998300909996, + 0.0172052513808012, + 0.014515322633087635, + 0.008924468420445919, + 0.010819517076015472, + 0.004287043120712042, + 0.005660547409206629, + 0.00644475594162941, + 0.005380650516599417, + 0.0062181660905480385, + 0.00549223693087697, + 0.007333814632147551, + 0.004630547482520342 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 1 + ], + "selected_reward": -0.01883596181869507 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 warm, dimly lit living room?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is an antique wooden radio with a dark brown walnut case visible at the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a round glass tuning dial on the radio emitting a warm, soft amber glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are three family members present: a father in a beige turtleneck, a mother in a crimson wool cardigan, and a young boy sitting cross-legged on the carpet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated primarily by the dim light of the radio, casting long shadows on the wall?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014209285378456116 + ], + "opposite_logprob_tokens": [ + -4.264209270477295 + ], + "logprob_reward": -0.014209285378456116, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014209285378456116, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12762495875358582, + 0.10746181756258011, + 0.05899043753743172, + 0.05479303002357483, + 0.11271744966506958, + 0.13669395446777344, + 0.09546078741550446, + 0.07942278683185577, + 0.1364685297012329, + 0.09036625176668167 + ], + "credit_map": [ + 0.013534597121179104, + 0.013926287181675434, + 0.018913771957159042, + 0.01881341263651848, + 0.015769019722938538, + 0.017075054347515106, + 0.013927574269473553, + 0.015665248036384583, + 0.01640324667096138, + 0.010758405551314354, + 0.016034698113799095, + 0.017888886854052544, + 0.01202284637838602, + 0.013356954790651798, + 0.009346343576908112, + 0.01165043842047453, + 0.007219234947115183, + 0.006878830958157778, + 0.008201070129871368, + 0.0075319926254451275, + 0.008487517945468426, + 0.006590822711586952, + 0.008983682841062546, + 0.005097285844385624, + 0.007668413687497377, + 0.005597129464149475, + 0.008161908015608788, + 0.008646652102470398, + 0.00620215805247426, + 0.008200865238904953, + 0.0049546584486961365, + 0.005361242685467005, + 0.013866385444998741, + 0.017568806186318398, + 0.014776220545172691, + 0.015620294027030468, + 0.013499674387276173, + 0.013129673898220062, + 0.01314668357372284, + 0.01110970601439476, + 0.01228768564760685, + 0.021392878144979477, + 0.017744336277246475, + 0.018364276736974716, + 0.01620902493596077, + 0.0160139761865139, + 0.017965126782655716, + 0.016716646030545235, + 0.009783067740499973, + 0.010716820135712624, + 0.012966630980372429, + 0.015043112449347973, + 0.012416575103998184, + 0.010230548679828644, + 0.014466873370110989, + 0.00983716081827879, + 0.008051218464970589, + 0.008118282072246075, + 0.010772133246064186, + 0.009593389928340912, + 0.008089229464530945, + 0.007163789123296738, + 0.017645757645368576, + 0.00998898409307003, + 0.013054565526545048, + 0.02340312860906124, + 0.01540609635412693, + 0.020377250388264656, + 0.016743028536438942, + 0.014137910678982735, + 0.021252034232020378, + 0.012094524689018726, + 0.008942140266299248, + 0.010729481466114521, + 0.012851797044277191, + 0.013544614426791668, + 0.015421702526509762, + 0.01047071348875761, + 0.010112984105944633, + 0.00829282682389021 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 8, + 0 + ], + "selected_reward": -0.014209285378456116 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 warm, dimly lit living room?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is an antique wooden radio with a dark brown walnut case visible at the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a round glass tuning dial on the radio emitting a warm, soft amber glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are three family members present: a father in a beige turtleneck, a mother in a crimson wool cardigan, and a young boy sitting cross-legged on the carpet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated primarily by the dim light of the radio, casting long shadows on the wall?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0013388964580371976 + ], + "opposite_logprob_tokens": [ + -6.626338958740234 + ], + "logprob_reward": -0.0013388964580371976, + "yes_no_margin_reward": 6.625, + "token_credit_reward": -0.0013388964580371976, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09890776872634888, + 0.0639772117137909, + 0.02911379188299179, + 0.1146552786231041, + 0.11203359812498093, + 0.15042082965373993, + 0.15242813527584076, + 0.10408876836299896, + 0.1143706738948822, + 0.06000387668609619 + ], + "credit_map": [ + 0.010342820547521114, + 0.011304348707199097, + 0.01572228968143463, + 0.01445621345192194, + 0.011631130240857601, + 0.013592110946774483, + 0.010244053788483143, + 0.011614813469350338, + 0.010054551064968109, + 0.007876037620007992, + 0.008199897594749928, + 0.008038663305342197, + 0.008336719125509262, + 0.007918126881122589, + 0.00704935472458601, + 0.00650386419147253, + 0.002612043172121048, + 0.003961489070206881, + 0.004720224998891354, + 0.004382312763482332, + 0.0033131379168480635, + 0.004697245545685291, + 0.0030423440039157867, + 0.002384996972978115, + 0.01347806490957737, + 0.014342320151627064, + 0.015408296138048172, + 0.014396050944924355, + 0.019855236634612083, + 0.016443058848381042, + 0.009843247011303902, + 0.010889008641242981, + 0.011334671638906002, + 0.021523458883166313, + 0.01537568960338831, + 0.01305696927011013, + 0.017011158168315887, + 0.011006425134837627, + 0.01565798930823803, + 0.0070672486908733845, + 0.011200737208127975, + 0.019451087340712547, + 0.020220911130309105, + 0.019554419443011284, + 0.01859758049249649, + 0.017410172149538994, + 0.027349812909960747, + 0.016636118292808533, + 0.015009147115051746, + 0.02039177156984806, + 0.022012844681739807, + 0.023273201659321785, + 0.017553478479385376, + 0.017703980207443237, + 0.020823245868086815, + 0.01566047966480255, + 0.008651194162666798, + 0.01244115736335516, + 0.012223209254443645, + 0.014044392853975296, + 0.01076226495206356, + 0.010553836822509766, + 0.02324257232248783, + 0.012170142494142056, + 0.010320311412215233, + 0.01978992484509945, + 0.012973901815712452, + 0.01889648474752903, + 0.010645696893334389, + 0.012627003714442253, + 0.019690565764904022, + 0.009426801465451717, + 0.006230936851352453, + 0.009124615229666233, + 0.007945488207042217, + 0.007074940018355846, + 0.007891707122325897, + 0.007116927765309811, + 0.008425172418355942, + 0.006194096989929676 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 5, + 3 + ], + "selected_reward": -0.0013388964580371976 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 warm, dimly lit living room?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is an antique wooden radio with a dark brown walnut case visible at the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a round glass tuning dial on the radio emitting a warm, soft amber glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are three family members present: a father in a beige turtleneck, a mother in a crimson wool cardigan, and a young boy sitting cross-legged on the carpet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated primarily by the dim light of the radio, casting long shadows on the wall?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07898029685020447 + ], + "opposite_logprob_tokens": [ + -2.5789802074432373 + ], + "logprob_reward": -0.07898029685020447, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07898029685020447, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16268804669380188, + 0.07534930855035782, + 0.0767248347401619, + 0.12667542695999146, + 0.17969758808612823, + 0.0780017301440239, + 0.0843103900551796, + 0.11527542024850845, + 0.055062707513570786, + 0.04621452838182449 + ], + "credit_map": [ + 0.022297246381640434, + 0.01884881965816021, + 0.021773388609290123, + 0.023671358823776245, + 0.019021105021238327, + 0.025469470769166946, + 0.013717211782932281, + 0.01788945309817791, + 0.011851095594465733, + 0.007562041282653809, + 0.010824001394212246, + 0.010230518877506256, + 0.007843059487640858, + 0.010886003263294697, + 0.008046029135584831, + 0.008106554858386517, + 0.007991672493517399, + 0.0073927403427660465, + 0.01035811472684145, + 0.012182943522930145, + 0.011111944913864136, + 0.010482476092875004, + 0.007541879080235958, + 0.009663069620728493, + 0.015129008330404758, + 0.012948485091328621, + 0.016338177025318146, + 0.01561072375625372, + 0.01909305900335312, + 0.020586345344781876, + 0.012843513861298561, + 0.014126124791800976, + 0.019802318885922432, + 0.028749169781804085, + 0.022205816581845284, + 0.02209298498928547, + 0.02424137108027935, + 0.02245987392961979, + 0.0246026199311018, + 0.015543442219495773, + 0.007976077497005463, + 0.010874642990529537, + 0.010667343623936176, + 0.010066991671919823, + 0.009608913213014603, + 0.011100783944129944, + 0.007557297591120005, + 0.01014967542141676, + 0.007601103745400906, + 0.009160827845335007, + 0.012010606937110424, + 0.01186634786427021, + 0.007449197117239237, + 0.012003184296190739, + 0.014805137179791927, + 0.009413981810212135, + 0.009389051236212254, + 0.02338140271604061, + 0.01249401643872261, + 0.014896210283041, + 0.012641177512705326, + 0.011938911862671375, + 0.01452670618891716, + 0.01600794866681099, + 0.00496229063719511, + 0.006606760434806347, + 0.00636577932164073, + 0.0093458853662014, + 0.006892009172588587, + 0.006794867105782032, + 0.007967737503349781, + 0.00612737936899066, + 0.00432016933336854, + 0.006932008545845747, + 0.005184597801417112, + 0.006554204970598221, + 0.006283042952418327, + 0.004348639398813248, + 0.007321495097130537, + 0.005270369816571474 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 3 + ], + "selected_reward": -0.07898029685020447 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 warm, dimly lit living room?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is an antique wooden radio with a dark brown walnut case visible at the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a round glass tuning dial on the radio emitting a warm, soft amber glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are three family members present: a father in a beige turtleneck, a mother in a crimson wool cardigan, and a young boy sitting cross-legged on the carpet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated primarily by the dim light of the radio, casting long shadows on the wall?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029797974973917007 + ], + "opposite_logprob_tokens": [ + -3.5297980308532715 + ], + "logprob_reward": -0.029797974973917007, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029797974973917007, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18868336081504822, + 0.08197104930877686, + 0.04985464736819267, + 0.08331286907196045, + 0.1777782291173935, + 0.07347772270441055, + 0.07207144051790237, + 0.11485171318054199, + 0.04809211567044258, + 0.10990681499242783 + ], + "credit_map": [ + 0.01956746354699135, + 0.021970100700855255, + 0.030965980142354965, + 0.039464548230171204, + 0.01725144125521183, + 0.02248743176460266, + 0.018601328134536743, + 0.01837507076561451, + 0.009988420642912388, + 0.00901113823056221, + 0.0108639569953084, + 0.013095344416797161, + 0.010599225759506226, + 0.013212371617555618, + 0.0073560988530516624, + 0.007844493724405766, + 0.0051170638762414455, + 0.004898826591670513, + 0.006560726556926966, + 0.009416414424777031, + 0.004730733577162027, + 0.007942873053252697, + 0.00614768173545599, + 0.005040331277996302, + 0.01062415074557066, + 0.00887078046798706, + 0.010960153304040432, + 0.012803476303815842, + 0.013278535567224026, + 0.010852211154997349, + 0.007489535491913557, + 0.008434038609266281, + 0.018714047968387604, + 0.032796118408441544, + 0.023163385689258575, + 0.022144433110952377, + 0.02739824168384075, + 0.019463712349534035, + 0.02301677130162716, + 0.011081530712544918, + 0.005242913030087948, + 0.00882711261510849, + 0.009059425443410873, + 0.010020801797509193, + 0.007289811968803406, + 0.00821088906377554, + 0.01715262606739998, + 0.007674152497202158, + 0.0067177750170230865, + 0.007585960440337658, + 0.00952887162566185, + 0.011057977564632893, + 0.006167566403746605, + 0.007923096418380737, + 0.013307844288647175, + 0.009782355278730392, + 0.007664037868380547, + 0.013550697825849056, + 0.013830471783876419, + 0.011899412609636784, + 0.015276232734322548, + 0.01264534518122673, + 0.015058991499245167, + 0.024926528334617615, + 0.005380789749324322, + 0.00655633769929409, + 0.006172368302941322, + 0.0068054115399718285, + 0.004889983683824539, + 0.005853726994246244, + 0.005084263626486063, + 0.007349237333983183, + 0.010810301639139652, + 0.015375988557934761, + 0.01507533062249422, + 0.015564560890197754, + 0.01360250823199749, + 0.014754651114344597, + 0.01708207279443741, + 0.007641409989446402 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 7 + ], + "selected_reward": -0.029797974973917007 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "slowly advancing medium shot, shot from a level angle, brings us into a warm, dimly lit living room. At the center of the frame is an antique wooden radio, its dark brown walnut case gleaming with a warm sheen in the dim light. Next to the finely textured speaker grille, a round glass tuning dial steadily emits a warm, soft amber glow. A family gathered intimately around a radio. A middle-aged father, his neatly combed dark hair streaked with silver, wore a beige turtleneck sweater and leaned forward slightly, listening attentively with a gentle, focused expression on his face. Beside him sat a mother with shoulder-length, curly brown hair, wearing a crimson wool cardigan. Her head rested gently on her husband's shoulder, her eyes slightly closed, a contented smile playing on her lips, completely absorbed in the sound. A boy, perhaps seven or eight, sat cross-legged on the thick carpet before them, his chin resting on his hands, his head tilted back, his large eyes gleaming with curiosity and fascination. The radio's sole, dim light was the primary source of light in the scene, gently illuminating each face, outlining its soft contours and casting long, almost motionless shadows on the wall behind them. The entire scene, bathed in warm tones, was permeated with a timeless tranquility, nostalgia, and familial warmth.", + "seed": 11618, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000121/group_0/index_8306_seed_11618_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": -0.06837776713073254, + "vlm_logprob": -0.06837776713073254, + "vlm_yes_no_margin": 2.9, + "vlm_token_credit": -0.06837776713073254, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 warm, dimly lit living room?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is an antique wooden radio with a dark brown walnut case visible at the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a round glass tuning dial on the radio emitting a warm, soft amber glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are three family members present: a father in a beige turtleneck, a mother in a crimson wool cardigan, and a young boy sitting cross-legged on the carpet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated primarily by the dim light of the radio, casting long shadows on the wall?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.055314965546131134 + ], + "opposite_logprob_tokens": [ + -3.0553150177001953 + ], + "logprob_reward": -0.055314965546131134, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.055314965546131134, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.29632359743118286, + 0.15218882262706757, + 0.06033776327967644, + 0.02943754754960537, + 0.029340630397200584, + 0.09635050594806671, + 0.10638698935508728, + 0.05476686730980873, + 0.15018145740032196, + 0.024685783311724663 + ], + "credit_map": [ + 0.036050789058208466, + 0.024622909724712372, + 0.027810905128717422, + 0.02829761616885662, + 0.050609465688467026, + 0.05057740584015846, + 0.039409004151821136, + 0.0389455109834671, + 0.020760463550686836, + 0.01685490272939205, + 0.013736991211771965, + 0.014720713719725609, + 0.02179667539894581, + 0.025862516835331917, + 0.02125523053109646, + 0.017201319336891174, + 0.007000582758337259, + 0.006279362365603447, + 0.006623927503824234, + 0.006396378856152296, + 0.009981431998312473, + 0.009061441756784916, + 0.008099138736724854, + 0.00689550070092082, + 0.0027458968106657267, + 0.002942788414657116, + 0.004040713422000408, + 0.0032225127797573805, + 0.004060497507452965, + 0.0055305068381130695, + 0.004154979716986418, + 0.0027396520599722862, + 0.002965135034173727, + 0.00302413129247725, + 0.004430667497217655, + 0.0034530849661678076, + 0.004682573024183512, + 0.004660891368985176, + 0.0038847150281071663, + 0.0022394326515495777, + 0.005682326853275299, + 0.007339785806834698, + 0.009123140014708042, + 0.008780870586633682, + 0.012450110167264938, + 0.021372569724917412, + 0.023115525022149086, + 0.008486175909638405, + 0.01193774864077568, + 0.014963643625378609, + 0.011652953922748566, + 0.012105671688914299, + 0.013873652555048466, + 0.01378533337265253, + 0.01310821995139122, + 0.014959760941565037, + 0.005961439106613398, + 0.01109875738620758, + 0.006249644327908754, + 0.005232254974544048, + 0.005267908331006765, + 0.009528960101306438, + 0.008447281084954739, + 0.002980620600283146, + 0.016637418419122696, + 0.016672194004058838, + 0.016538476571440697, + 0.013878264464437962, + 0.019035013392567635, + 0.030046265572309494, + 0.02436518296599388, + 0.01300864852964878, + 0.003026651218533516, + 0.002784276381134987, + 0.002276870422065258, + 0.0023493501357734203, + 0.0033106578048318624, + 0.004657639656215906, + 0.004014002624899149, + 0.0022663359995931387 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 8 + ], + "selected_reward": -0.055314965546131134 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 warm, dimly lit living room?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is an antique wooden radio with a dark brown walnut case visible at the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a round glass tuning dial on the radio emitting a warm, soft amber glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are three family members present: a father in a beige turtleneck, a mother in a crimson wool cardigan, and a young boy sitting cross-legged on the carpet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated primarily by the dim light of the radio, casting long shadows on the wall?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04867067560553551 + ], + "opposite_logprob_tokens": [ + -3.048670768737793 + ], + "logprob_reward": -0.04867067560553551, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04867067560553551, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.43583258986473083, + 0.23529942333698273, + 0.08575118333101273, + 0.05586790293455124, + 0.0367036871612072, + 0.03965574502944946, + 0.027343200519680977, + 0.03802267462015152, + 0.019914956763386726, + 0.025608627125620842 + ], + "credit_map": [ + 0.04736718907952309, + 0.0773017629981041, + 0.04446173831820488, + 0.052804894745349884, + 0.06198001280426979, + 0.0614338219165802, + 0.04701881855726242, + 0.04346438869833946, + 0.031420618295669556, + 0.036895591765642166, + 0.025005685165524483, + 0.025116214528679848, + 0.030654218047857285, + 0.038091376423835754, + 0.028461255133152008, + 0.01965448446571827, + 0.012681292369961739, + 0.009339561685919762, + 0.009619811549782753, + 0.010575304739177227, + 0.012529677711427212, + 0.012842722237110138, + 0.010844566859304905, + 0.007318257819861174, + 0.007204393856227398, + 0.006191266234964132, + 0.008483241312205791, + 0.007001106161624193, + 0.008438664488494396, + 0.007119806483387947, + 0.006960778031498194, + 0.0044686500914394855, + 0.0038490535225719213, + 0.00355891534127295, + 0.003563067177310586, + 0.0047377184964716434, + 0.005433492362499237, + 0.0050008282996714115, + 0.006124515552073717, + 0.004436099901795387, + 0.0033723630476742983, + 0.002671030582860112, + 0.0028252380434423685, + 0.003533622482791543, + 0.0038461547810584307, + 0.008591894060373306, + 0.010013220831751823, + 0.00480222562327981, + 0.0024752262979745865, + 0.002794107189401984, + 0.002894122153520584, + 0.002999026793986559, + 0.004857896361500025, + 0.004007881507277489, + 0.0036279456689953804, + 0.003687000134959817, + 0.004136068280786276, + 0.007152612321078777, + 0.004393011797219515, + 0.004387387540191412, + 0.003513087285682559, + 0.005959152244031429, + 0.005727591458708048, + 0.0027537692803889513, + 0.002189483493566513, + 0.002013386692851782, + 0.0027765529230237007, + 0.002624684479087591, + 0.0022910977713763714, + 0.0034554956946521997, + 0.002844548085704446, + 0.0017197099514305592, + 0.0028156971093267202, + 0.0027302755042910576, + 0.0027882459107786417, + 0.0023404948879033327, + 0.0035031368024647236, + 0.004340396728366613, + 0.0032211679499596357, + 0.0038692152593284845 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.04867067560553551 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 warm, dimly lit living room?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is an antique wooden radio with a dark brown walnut case visible at the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a round glass tuning dial on the radio emitting a warm, soft amber glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are three family members present: a father in a beige turtleneck, a mother in a crimson wool cardigan, and a young boy sitting cross-legged on the carpet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated primarily by the dim light of the radio, casting long shadows on the wall?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016072357073426247 + ], + "opposite_logprob_tokens": [ + -4.1410722732543945 + ], + "logprob_reward": -0.016072357073426247, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016072357073426247, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3733616769313812, + 0.18609736859798431, + 0.09553559869527817, + 0.05248059332370758, + 0.06436450034379959, + 0.05731271952390671, + 0.03987724706530571, + 0.04833215847611427, + 0.04270927980542183, + 0.03992881625890732 + ], + "credit_map": [ + 0.0392066054046154, + 0.06834516674280167, + 0.03246108815073967, + 0.042434174567461014, + 0.0493670254945755, + 0.059929583221673965, + 0.045921795070171356, + 0.03569624945521355, + 0.02272016927599907, + 0.024454953148961067, + 0.01612107828259468, + 0.017673494294285774, + 0.02347555197775364, + 0.034999702125787735, + 0.024413246661424637, + 0.022239167243242264, + 0.012013793922960758, + 0.009607565589249134, + 0.01016814261674881, + 0.01073442678898573, + 0.012874148786067963, + 0.01515643484890461, + 0.012511995621025562, + 0.012469088658690453, + 0.0059183258563280106, + 0.005740009248256683, + 0.006078312639147043, + 0.0057226866483688354, + 0.008881979621946812, + 0.008305211551487446, + 0.006377368699759245, + 0.005456701852381229, + 0.005920690484344959, + 0.007176731713116169, + 0.007229197304695845, + 0.007769846357405186, + 0.008059870451688766, + 0.01019847858697176, + 0.013056725263595581, + 0.0049529592506587505, + 0.005122146103531122, + 0.0034822309389710426, + 0.004297259729355574, + 0.004529980942606926, + 0.0068825483322143555, + 0.013058178126811981, + 0.015848860144615173, + 0.0040915170684456825, + 0.004734339192509651, + 0.003477860940620303, + 0.0028301130514591932, + 0.00321289524435997, + 0.004370705224573612, + 0.008330998942255974, + 0.009655658155679703, + 0.0032646763138473034, + 0.005608567502349615, + 0.008246066980063915, + 0.005436870269477367, + 0.004577397368848324, + 0.005262594670057297, + 0.008718647994101048, + 0.005465760361403227, + 0.005016255658119917, + 0.004975166637450457, + 0.0068211629986763, + 0.006357770878821611, + 0.004688564222306013, + 0.004492618143558502, + 0.006866509094834328, + 0.004563338588923216, + 0.003944146912544966, + 0.005112664774060249, + 0.0059167128056287766, + 0.004203883931040764, + 0.0032314632553607225, + 0.0049129086546599865, + 0.0067719160579144955, + 0.0042207022197544575, + 0.0055585638619959354 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.016072357073426247 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a warm, dimly lit living room?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is an antique wooden radio with a dark brown walnut case visible at the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a round glass tuning dial on the radio emitting a warm, soft amber glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are three family members present: a father in a beige turtleneck, a mother in a crimson wool cardigan, and a young boy sitting cross-legged on the carpet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated primarily by the dim light of the radio, casting long shadows on the wall?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07899969071149826 + ], + "opposite_logprob_tokens": [ + -2.5789997577667236 + ], + "logprob_reward": -0.07899969071149826, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07899969071149826, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3705328106880188, + 0.1314544379711151, + 0.06799735128879547, + 0.06311182677745819, + 0.08277576416730881, + 0.08928423374891281, + 0.07904589176177979, + 0.05100065469741821, + 0.036615774035453796, + 0.028181269764900208 + ], + "credit_map": [ + 0.040286898612976074, + 0.05586862936615944, + 0.05493783950805664, + 0.031320761889219284, + 0.04931362718343735, + 0.06189104542136192, + 0.04190768674015999, + 0.03500628098845482, + 0.017591465264558792, + 0.014145812019705772, + 0.01437663845717907, + 0.01219326350837946, + 0.017095373943448067, + 0.02375316061079502, + 0.018358904868364334, + 0.013939820230007172, + 0.009228848852217197, + 0.00737689808011055, + 0.00818528514355421, + 0.00768240774050355, + 0.010286341421306133, + 0.008731378242373466, + 0.010547684505581856, + 0.005958513356745243, + 0.006809819024056196, + 0.008608462288975716, + 0.007727408315986395, + 0.007522663101553917, + 0.012366394512355328, + 0.006361711770296097, + 0.00778674753382802, + 0.005928618833422661, + 0.007461012806743383, + 0.0077181654050946236, + 0.010121033526957035, + 0.00823797658085823, + 0.009595468640327454, + 0.01494177058339119, + 0.017811648547649384, + 0.0068886918015778065, + 0.00791751965880394, + 0.008866703137755394, + 0.011571965180337429, + 0.007861414924263954, + 0.012070270255208015, + 0.01723708026111126, + 0.017061656340956688, + 0.006697626784443855, + 0.00985676795244217, + 0.008754904381930828, + 0.011312794871628284, + 0.009440616704523563, + 0.013916394673287868, + 0.011560915969312191, + 0.00858634989708662, + 0.005617152433842421, + 0.006188328843563795, + 0.01043263915926218, + 0.006082319654524326, + 0.004746326245367527, + 0.0049038296565413475, + 0.008419308811426163, + 0.006871214602142572, + 0.003356687491759658, + 0.0037293590139597654, + 0.0039008085150271654, + 0.005129532422870398, + 0.0047194743528962135, + 0.0057321880012750626, + 0.006377882324159145, + 0.0037948093377053738, + 0.0032317182049155235, + 0.0029235694091767073, + 0.002872888231649995, + 0.0034001527819782495, + 0.003030400723218918, + 0.0040567186661064625, + 0.004473784007132053, + 0.0031650562305003405, + 0.004258699249476194 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 5 + ], + "selected_reward": -0.07899969071149826 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 warm, dimly lit living room?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is an antique wooden radio with a dark brown walnut case visible at the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a round glass tuning dial on the radio emitting a warm, soft amber glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are three family members present: a father in a beige turtleneck, a mother in a crimson wool cardigan, and a young boy sitting cross-legged on the carpet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated primarily by the dim light of the radio, casting long shadows on the wall?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.14283114671707153 + ], + "opposite_logprob_tokens": [ + -2.0178310871124268 + ], + "logprob_reward": -0.14283114671707153, + "yes_no_margin_reward": 1.875, + "token_credit_reward": -0.14283114671707153, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3776940107345581, + 0.1865544319152832, + 0.06528425961732864, + 0.04523853585124016, + 0.06602390855550766, + 0.13974395394325256, + 0.030373921617865562, + 0.04125955328345299, + 0.02133246324956417, + 0.02649497054517269 + ], + "credit_map": [ + 0.04015238955616951, + 0.04727817699313164, + 0.03794363886117935, + 0.03145137056708336, + 0.07044976949691772, + 0.05177738890051842, + 0.04860614612698555, + 0.05003512278199196, + 0.02162843570113182, + 0.02166898176074028, + 0.020835019648075104, + 0.017049020156264305, + 0.028761345893144608, + 0.02974414825439453, + 0.024985404685139656, + 0.0218820758163929, + 0.0065010120160877705, + 0.005999046843498945, + 0.005398437846451998, + 0.005754522047936916, + 0.012917800806462765, + 0.010284163989126682, + 0.01027597114443779, + 0.008153305388987064, + 0.004829529672861099, + 0.005343809723854065, + 0.0051334756426513195, + 0.004663324914872646, + 0.005727220792323351, + 0.008443235419690609, + 0.005418508313596249, + 0.0056794313713908195, + 0.006180428899824619, + 0.007391391787678003, + 0.008666599169373512, + 0.0060225059278309345, + 0.009427946992218494, + 0.01044673752039671, + 0.01201714389026165, + 0.005871147848665714, + 0.010135223157703876, + 0.014248461462557316, + 0.012830386869609356, + 0.013287454843521118, + 0.016153153032064438, + 0.023916516453027725, + 0.03405352309346199, + 0.015119249001145363, + 0.003058280097320676, + 0.0033630249090492725, + 0.003251422429457307, + 0.0030595508869737387, + 0.0062255882658064365, + 0.00495603121817112, + 0.0035758463200181723, + 0.0028841770254075527, + 0.0039009954780340195, + 0.005809498019516468, + 0.005360539071261883, + 0.004091909620910883, + 0.004960633814334869, + 0.008058306761085987, + 0.005828164517879486, + 0.0032495048362761736, + 0.002640416147187352, + 0.0023137556854635477, + 0.002733263885602355, + 0.0024380425456911325, + 0.0030752092134207487, + 0.0037168420385569334, + 0.0027582747861742973, + 0.0016566591802984476, + 0.0032442237716168165, + 0.0028043859638273716, + 0.00357125885784626, + 0.0027865241281688213, + 0.0035793050192296505, + 0.004277860280126333, + 0.003148395102471113, + 0.0030830155592411757 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 5 + ], + "selected_reward": -0.14283114671707153 + } + ], + "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/0000121/group_1/index_tensor([2191, 2191])_group_1_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000121/group_1/index_tensor([2191, 2191])_group_1_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..dd686d361264d69ac6021615f3d5ce3c164d60c0 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000121/group_1/index_tensor([2191, 2191])_group_1_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "sweeping low-angle, long shot follows the camera as it slowly pans right, tracking a lone camel across a vast expanse of desert bathed in the golden glow of a setting sun. The dromedary's entire body is silhouetted, rendered in a deep black, its outline outlined only by a thin, fiery halo of light from the setting sun below the horizon, clearly delineating its signature hump and gracefully curved neck. It advances with a leisurely, rhythmic gait, its powerful backlight casting a vast, elongated, almost grotesque shadow across the undulating sand dunes behind it. This shadow twists and flows across the dunes' undulating forms with every step. The dune crests, lit by the last rays of the setting sun, glow with a warm orange-gold hue, while the hollows between them are bathed in a deepening shade of deep purple, creating a stark contrast between light and dark. The sky in the distance is a magnificent painting of fiery orange, deep red and soft purple. The whole scene is filled with an epic sense of tranquility and loneliness, full of cinematic visual impact.", + "seed": 20091, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000121/group_1/index_2191_seed_20091_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": -0.03544348627328873, + "vlm_logprob": -0.03544348627328873, + "vlm_yes_no_margin": 4.3, + "vlm_token_credit": -0.03544348627328873, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 long shot that pans slowly to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a lone dromedary camel visible, silhouetted in deep black with a thin fiery halo outlining its hump and neck?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camel advancing with a leisurely, rhythmic gait while casting a vast, elongated, twisting shadow across the sand dunes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the sand dune crests glow with warm orange-gold light while the hollows are bathed in deep purple?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the sky in the distance painted with fiery orange, deep red, and soft purple colors?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.14252996444702148 + ], + "opposite_logprob_tokens": [ + -2.1425299644470215 + ], + "logprob_reward": -0.14252996444702148, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.14252996444702148, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2601628005504608, + 0.06849843263626099, + 0.051720939576625824, + 0.15757018327713013, + 0.03379400447010994, + 0.13109475374221802, + 0.046573005616664886, + 0.11127771437168121, + 0.06867953389883041, + 0.0706285759806633 + ], + "credit_map": [ + 0.054216351360082626, + 0.02234441414475441, + 0.026593029499053955, + 0.023454802110791206, + 0.017658699303865433, + 0.023281237110495567, + 0.02084578573703766, + 0.07176848500967026, + 0.004931080620735884, + 0.006065023131668568, + 0.010160272940993309, + 0.009152468293905258, + 0.0064203087240457535, + 0.009015576913952827, + 0.01259537786245346, + 0.010158325545489788, + 0.005514734890311956, + 0.004908408038318157, + 0.007023754995316267, + 0.008248362690210342, + 0.004527626093477011, + 0.0043817381374537945, + 0.010557787492871284, + 0.006558527704328299, + 0.01743641495704651, + 0.013874020427465439, + 0.012442623265087605, + 0.023316429927945137, + 0.024439407512545586, + 0.023513812571763992, + 0.02228427119553089, + 0.02026320993900299, + 0.0029033299069851637, + 0.003489861963316798, + 0.004620823077857494, + 0.0035467694979161024, + 0.0033772101160138845, + 0.003819122677668929, + 0.006978525314480066, + 0.005058364011347294, + 0.027151361107826233, + 0.004971765913069248, + 0.013118596747517586, + 0.006028309464454651, + 0.005891642067581415, + 0.009858865290880203, + 0.01918076165020466, + 0.044893451035022736, + 0.0036241067573428154, + 0.0044534713961184025, + 0.007324793841689825, + 0.006375598255544901, + 0.004559731110930443, + 0.005976399406790733, + 0.006673886440694332, + 0.007585019338876009, + 0.006456318777054548, + 0.009207510389387608, + 0.016065528616309166, + 0.018311405554413795, + 0.010099373757839203, + 0.011566022410988808, + 0.021969400346279144, + 0.017602160573005676, + 0.003644713666290045, + 0.005724738817662001, + 0.009589708410203457, + 0.012753254733979702, + 0.004774503875523806, + 0.00823716539889574, + 0.013528320007026196, + 0.010427129454910755, + 0.006098642013967037, + 0.005601709708571434, + 0.012316207401454449, + 0.008900626562535763, + 0.008390191942453384, + 0.005195070058107376, + 0.011173444800078869, + 0.012952686287462711 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 5 + ], + "selected_reward": -0.14252996444702148 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 long shot that pans slowly to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a lone dromedary camel visible, silhouetted in deep black with a thin fiery halo outlining its hump and neck?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camel advancing with a leisurely, rhythmic gait while casting a vast, elongated, twisting shadow across the sand dunes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the sand dune crests glow with warm orange-gold light while the hollows are bathed in deep purple?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the sky in the distance painted with fiery orange, deep red, and soft purple colors?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008660731837153435 + ], + "opposite_logprob_tokens": [ + -4.758660793304443 + ], + "logprob_reward": -0.008660731837153435, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008660731837153435, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.34355032444000244, + 0.056506361812353134, + 0.036634355783462524, + 0.04388079047203064, + 0.02981610596179962, + 0.03405791148543358, + 0.055559977889060974, + 0.02273336611688137, + 0.06597676873207092, + 0.31128400564193726 + ], + "credit_map": [ + 0.08326014131307602, + 0.023291684687137604, + 0.03825481981039047, + 0.01429029367864132, + 0.018873829394578934, + 0.034567687660455704, + 0.02180781215429306, + 0.10920407623052597, + 0.005946403834968805, + 0.005199255887418985, + 0.009510056115686893, + 0.006522345822304487, + 0.004397979471832514, + 0.0056907800026237965, + 0.007288325112313032, + 0.011951214633882046, + 0.0031181524973362684, + 0.005297945812344551, + 0.005803550593554974, + 0.004014334175735712, + 0.003497587749734521, + 0.004480556584894657, + 0.005736195016652346, + 0.004686033818870783, + 0.003837401047348976, + 0.004745580721646547, + 0.0046745529398322105, + 0.005405969452112913, + 0.0067366003058850765, + 0.005833221133798361, + 0.0071515110321342945, + 0.00549595570191741, + 0.0036599193699657917, + 0.0035294867120683193, + 0.004934978671371937, + 0.0028477488085627556, + 0.003468411276116967, + 0.002715677721425891, + 0.004139028023928404, + 0.004520857241004705, + 0.0032390502747148275, + 0.002477742964401841, + 0.004464501049369574, + 0.0041634379886090755, + 0.0019764136523008347, + 0.0028798992279917, + 0.006190589629113674, + 0.00866627972573042, + 0.0030436855740845203, + 0.003589747939258814, + 0.007232354953885078, + 0.007581544108688831, + 0.00707979453727603, + 0.0074567170813679695, + 0.01003055926412344, + 0.00954557117074728, + 0.0016338654095306993, + 0.002083219587802887, + 0.0042172325775027275, + 0.0028375221882015467, + 0.0025465034414082766, + 0.0025452300906181335, + 0.0034749063197523355, + 0.0033948852214962244, + 0.004561855457723141, + 0.006653772667050362, + 0.009559785947203636, + 0.00869064126163721, + 0.009981600567698479, + 0.006575798615813255, + 0.010304119437932968, + 0.009649199433624744, + 0.024004047736525536, + 0.029307685792446136, + 0.04474930092692375, + 0.03155684098601341, + 0.05859030783176422, + 0.03470873460173607, + 0.04244372248649597, + 0.0459233783185482 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 8 + ], + "selected_reward": -0.008660731837153435 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 long shot that pans slowly to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a lone dromedary camel visible, silhouetted in deep black with a thin fiery halo outlining its hump and neck?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camel advancing with a leisurely, rhythmic gait while casting a vast, elongated, twisting shadow across the sand dunes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the sand dune crests glow with warm orange-gold light while the hollows are bathed in deep purple?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the sky in the distance painted with fiery orange, deep red, and soft purple colors?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014195770025253296 + ], + "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": [ + 0.37080755829811096, + 0.07417454570531845, + 0.08079904317855835, + 0.12122372537851334, + 0.02982526458799839, + 0.1215708777308464, + 0.043514180928468704, + 0.05943462625145912, + 0.07042758166790009, + 0.0282225850969553 + ], + "credit_map": [ + 0.08685723692178726, + 0.022455181926488876, + 0.03512384742498398, + 0.022864244878292084, + 0.03154478967189789, + 0.03176381438970566, + 0.02648097090423107, + 0.1137174665927887, + 0.007709738798439503, + 0.006611413788050413, + 0.00863924901932478, + 0.007333108689635992, + 0.008430649526417255, + 0.010414390824735165, + 0.011298766359686852, + 0.01373722217977047, + 0.007582142483443022, + 0.009662322700023651, + 0.01341500785201788, + 0.01170528493821621, + 0.007297057658433914, + 0.006298406980931759, + 0.014506482519209385, + 0.010332337580621243, + 0.01452853437513113, + 0.009300744161009789, + 0.024042917415499687, + 0.01593204215168953, + 0.011252532713115215, + 0.016262134537100792, + 0.013562996871769428, + 0.016341829672455788, + 0.0024824279826134443, + 0.0027123787440359592, + 0.004307691939175129, + 0.0033779204823076725, + 0.0033672009594738483, + 0.0038167082238942385, + 0.005226543173193932, + 0.004534393548965454, + 0.02621065266430378, + 0.004806395620107651, + 0.011770900338888168, + 0.005912340711802244, + 0.007242234889417887, + 0.010888874530792236, + 0.01679067499935627, + 0.03794880211353302, + 0.0029293035622686148, + 0.004031319171190262, + 0.006650567054748535, + 0.005091068800538778, + 0.00575126800686121, + 0.0056265853345394135, + 0.006491420790553093, + 0.0069426498375833035, + 0.003933163825422525, + 0.004698086064308882, + 0.008631477132439613, + 0.009144101291894913, + 0.0059437877498567104, + 0.005589418578892946, + 0.013200880959630013, + 0.008293713442981243, + 0.003445293754339218, + 0.0056316745467484, + 0.011620260775089264, + 0.013461404480040073, + 0.005832246970385313, + 0.006575029343366623, + 0.012024189345538616, + 0.011837480589747429, + 0.0024438442196696997, + 0.0024115147534757853, + 0.004042220767587423, + 0.0031685514841228724, + 0.004329085815697908, + 0.002675925847142935, + 0.00419216463342309, + 0.004959278739988804 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 3 + ], + "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 camera shot a low-angle long shot that pans slowly to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a lone dromedary camel visible, silhouetted in deep black with a thin fiery halo outlining its hump and neck?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camel advancing with a leisurely, rhythmic gait while casting a vast, elongated, twisting shadow across the sand dunes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the sand dune crests glow with warm orange-gold light while the hollows are bathed in deep purple?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the sky in the distance painted with fiery orange, deep red, and soft purple colors?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008638040162622929 + ], + "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": [ + 0.15800534188747406, + 0.13632239401340485, + 0.09521957486867905, + 0.1650581806898117, + 0.17434462904930115, + 0.06352776288986206, + 0.07269700616598129, + 0.03979102522134781, + 0.04896078631281853, + 0.04607335478067398 + ], + "credit_map": [ + 0.02150038816034794, + 0.01800154149532318, + 0.022502442821860313, + 0.017981072887778282, + 0.023530365899205208, + 0.017065133899450302, + 0.01781465671956539, + 0.019609728828072548, + 0.016456278041005135, + 0.009095069020986557, + 0.03209817409515381, + 0.014012693427503109, + 0.0072845956310629845, + 0.013460103422403336, + 0.012674996629357338, + 0.031240476295351982, + 0.008011002093553543, + 0.004934013821184635, + 0.03368336707353592, + 0.010873459279537201, + 0.005461175460368395, + 0.00467204162850976, + 0.00914512388408184, + 0.018439389765262604, + 0.01938684657216072, + 0.01206327322870493, + 0.034074217081069946, + 0.022727414965629578, + 0.01875615306198597, + 0.016942694783210754, + 0.019610479474067688, + 0.021497083827853203, + 0.030026983469724655, + 0.01119694672524929, + 0.05064021050930023, + 0.011532900854945183, + 0.010577908717095852, + 0.009901395998895168, + 0.015692507848143578, + 0.03477576747536659, + 0.012382936663925648, + 0.002985844388604164, + 0.00929830502718687, + 0.005174674559384584, + 0.0037363707087934017, + 0.004601278807967901, + 0.0077043562196195126, + 0.017643995583057404, + 0.004512183368206024, + 0.0051729422993958, + 0.008508226834237576, + 0.010879218578338623, + 0.009294804185628891, + 0.009840191341936588, + 0.014663556590676308, + 0.009825878776609898, + 0.003463100641965866, + 0.002791563281789422, + 0.008772880770266056, + 0.004302399232983589, + 0.004165348596870899, + 0.004961847327649593, + 0.0053391410037875175, + 0.005994740407913923, + 0.0030302535742521286, + 0.004357402678579092, + 0.007954010739922523, + 0.006327684503048658, + 0.005363524425774813, + 0.005874644964933395, + 0.008813747204840183, + 0.007239514496177435, + 0.004251799080520868, + 0.004287827294319868, + 0.007820354774594307, + 0.005303849466145039, + 0.007616438437253237, + 0.005524247419089079, + 0.005330732092261314, + 0.0059381029568612576 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 3, + 0 + ], + "selected_reward": -0.008638040162622929 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 long shot that pans slowly to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a lone dromedary camel visible, silhouetted in deep black with a thin fiery halo outlining its hump and neck?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camel advancing with a leisurely, rhythmic gait while casting a vast, elongated, twisting shadow across the sand dunes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the sand dune crests glow with warm orange-gold light while the hollows are bathed in deep purple?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the sky in the distance painted with fiery orange, deep red, and soft purple colors?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0031929248943924904 + ], + "opposite_logprob_tokens": [ + -5.753192901611328 + ], + "logprob_reward": -0.0031929248943924904, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.0031929248943924904, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2857549488544464, + 0.12390666455030441, + 0.09579929709434509, + 0.08198673278093338, + 0.06883928924798965, + 0.06439689546823502, + 0.09947223961353302, + 0.047051817178726196, + 0.052924804389476776, + 0.07986723631620407 + ], + "credit_map": [ + 0.06318975239992142, + 0.02022976614534855, + 0.022710775956511497, + 0.021693997085094452, + 0.04348336532711983, + 0.03411116823554039, + 0.020855437964200974, + 0.05948072299361229, + 0.016889795660972595, + 0.011840621940791607, + 0.020959269255399704, + 0.011011569760739803, + 0.0156266950070858, + 0.01591668091714382, + 0.012574217282235622, + 0.01908782869577408, + 0.012499052099883556, + 0.006881460547447205, + 0.025994744151830673, + 0.009380190633237362, + 0.007012617774307728, + 0.007038270588964224, + 0.00949923787266016, + 0.017493726685643196, + 0.00801070872694254, + 0.007708001881837845, + 0.011739015579223633, + 0.011676589027047157, + 0.008955365046858788, + 0.01089354045689106, + 0.010704339481890202, + 0.012299182824790478, + 0.010793600231409073, + 0.005491344258189201, + 0.019755447283387184, + 0.004603404086083174, + 0.004486145451664925, + 0.005712638143450022, + 0.005629177670925856, + 0.012367533519864082, + 0.009415240958333015, + 0.0033837410155683756, + 0.008447516709566116, + 0.005565920379012823, + 0.0033240451011806726, + 0.004780028481036425, + 0.012959354557096958, + 0.01652105525135994, + 0.006781782489269972, + 0.006619434803724289, + 0.01304345577955246, + 0.014534339308738708, + 0.012093235738575459, + 0.013407810591161251, + 0.019392283633351326, + 0.013599907048046589, + 0.0042558517307043076, + 0.003881508717313409, + 0.010188085027039051, + 0.004900843370705843, + 0.004861592315137386, + 0.00604231096804142, + 0.006011767778545618, + 0.006909861229360104, + 0.004204736091196537, + 0.005106052849441767, + 0.0070217326283454895, + 0.00860869511961937, + 0.005705434363335371, + 0.005963952746242285, + 0.008512889966368675, + 0.00780131621286273, + 0.009821904823184013, + 0.00618628179654479, + 0.013247289694845676, + 0.007948809303343296, + 0.009842212311923504, + 0.006847955286502838, + 0.010816080495715141, + 0.015156706795096397 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 6 + ], + "selected_reward": -0.0031929248943924904 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "sweeping low-angle, long shot follows the camera as it slowly pans right, tracking a lone camel across a vast expanse of desert bathed in the golden glow of a setting sun. The dromedary's entire body is silhouetted, rendered in a deep black, its outline outlined only by a thin, fiery halo of light from the setting sun below the horizon, clearly delineating its signature hump and gracefully curved neck. It advances with a leisurely, rhythmic gait, its powerful backlight casting a vast, elongated, almost grotesque shadow across the undulating sand dunes behind it. This shadow twists and flows across the dunes' undulating forms with every step. The dune crests, lit by the last rays of the setting sun, glow with a warm orange-gold hue, while the hollows between them are bathed in a deepening shade of deep purple, creating a stark contrast between light and dark. The sky in the distance is a magnificent painting of fiery orange, deep red and soft purple. The whole scene is filled with an epic sense of tranquility and loneliness, full of cinematic visual impact.", + "seed": 20092, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000121/group_1/index_2191_seed_20092_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": -0.029386894404888154, + "vlm_logprob": -0.029386894404888154, + "vlm_yes_no_margin": 4.2, + "vlm_token_credit": -0.029386894404888154, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 long shot that pans slowly to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a lone dromedary camel visible, silhouetted in deep black with a thin fiery halo outlining its hump and neck?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camel advancing with a leisurely, rhythmic gait while casting a vast, elongated, twisting shadow across the sand dunes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the sand dune crests glow with warm orange-gold light while the hollows are bathed in deep purple?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the sky in the distance painted with fiery orange, deep red, and soft purple colors?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10118316113948822 + ], + "opposite_logprob_tokens": [ + -2.4761831760406494 + ], + "logprob_reward": -0.10118316113948822, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.10118316113948822, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5579482316970825, + 0.0493270680308342, + 0.029623333364725113, + 0.05463987588882446, + 0.048703499138355255, + 0.027064360678195953, + 0.07492154091596603, + 0.11983271688222885, + 0.024738438427448273, + 0.013200961984694004 + ], + "credit_map": [ + 0.06247447803616524, + 0.06508694589138031, + 0.1836565136909485, + 0.04090462252497673, + 0.010555149056017399, + 0.01677464134991169, + 0.025641953572630882, + 0.15285390615463257, + 0.00511727761477232, + 0.004303750582039356, + 0.008197257295250893, + 0.008080737665295601, + 0.005117159336805344, + 0.005032401531934738, + 0.007644742261618376, + 0.005833740346133709, + 0.002850034972652793, + 0.002584320493042469, + 0.005011955741792917, + 0.004534248728305101, + 0.0024815138895064592, + 0.0027926296461373568, + 0.0054472568444907665, + 0.003921372815966606, + 0.004469683393836021, + 0.004353847820311785, + 0.006938553415238857, + 0.010449158027768135, + 0.007024938240647316, + 0.004995803348720074, + 0.010166766121983528, + 0.006241127848625183, + 0.0030990629456937313, + 0.0035381182096898556, + 0.006347572430968285, + 0.005564887076616287, + 0.0046770209446549416, + 0.007989158853888512, + 0.0116952545940876, + 0.005792425014078617, + 0.0020505290012806654, + 0.0026206085458397865, + 0.003711553756147623, + 0.005183078348636627, + 0.002267820993438363, + 0.0028273263014853, + 0.004269863478839397, + 0.004133581183850765, + 0.007016380783170462, + 0.00745185324922204, + 0.008295108564198017, + 0.00878063403069973, + 0.008636167272925377, + 0.00898592546582222, + 0.01505024079233408, + 0.010705235414206982, + 0.00796031579375267, + 0.00792154110968113, + 0.013123338110744953, + 0.009681078605353832, + 0.016818994656205177, + 0.015800204128026962, + 0.03150200471282005, + 0.017025241628289223, + 0.002094860887154937, + 0.0021974320989102125, + 0.0034455512650310993, + 0.002258735243231058, + 0.0017636360134929419, + 0.0025296409148722887, + 0.006113689858466387, + 0.004334893077611923, + 0.0013676739763468504, + 0.0014547440223395824, + 0.0014065077994018793, + 0.002040946390479803, + 0.0010029147379100323, + 0.001434003934264183, + 0.002585178939625621, + 0.0019089924171566963 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 6 + ], + "selected_reward": -0.10118316113948822 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 long shot that pans slowly to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a lone dromedary camel visible, silhouetted in deep black with a thin fiery halo outlining its hump and neck?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camel advancing with a leisurely, rhythmic gait while casting a vast, elongated, twisting shadow across the sand dunes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the sand dune crests glow with warm orange-gold light while the hollows are bathed in deep purple?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the sky in the distance painted with fiery orange, deep red, and soft purple colors?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006758213043212891 + ], + "opposite_logprob_tokens": [ + -5.006758213043213 + ], + "logprob_reward": -0.006758213043212891, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006758213043212891, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.44780391454696655, + 0.12114086002111435, + 0.034679535776376724, + 0.05452651157975197, + 0.06457947939634323, + 0.02097892016172409, + 0.03328034281730652, + 0.029603131115436554, + 0.10673762857913971, + 0.08666966110467911 + ], + "credit_map": [ + 0.07772420346736908, + 0.11452625691890717, + 0.10160887241363525, + 0.027154643088579178, + 0.016099808737635612, + 0.01768154464662075, + 0.02044510468840599, + 0.07256345450878143, + 0.01827814243733883, + 0.016946859657764435, + 0.01792517490684986, + 0.01714266836643219, + 0.008936027064919472, + 0.008398734964430332, + 0.013930173590779305, + 0.019583074375987053, + 0.00364617258310318, + 0.0027320203371345997, + 0.006090146489441395, + 0.006693764589726925, + 0.0023526104632765055, + 0.0026085604913532734, + 0.006542988587170839, + 0.004013272933661938, + 0.007484108209609985, + 0.00632181391119957, + 0.0078769875690341, + 0.0071480534970760345, + 0.004113045521080494, + 0.00393945537507534, + 0.007186646573245525, + 0.010456404648721218, + 0.00804926361888647, + 0.005784037057310343, + 0.007525877095758915, + 0.007565270643681288, + 0.00409574992954731, + 0.0052347127348184586, + 0.016437344253063202, + 0.009887227788567543, + 0.003268231637775898, + 0.0022514224983751774, + 0.0021478328853845596, + 0.0023192407097667456, + 0.001714574173092842, + 0.0025446193758398294, + 0.003327277023345232, + 0.0034057237207889557, + 0.0026998466346412897, + 0.0024429578334093094, + 0.0037758355028927326, + 0.003514052601531148, + 0.0030636785086244345, + 0.0034030722454190254, + 0.00957455299794674, + 0.004806348122656345, + 0.0026399886701256037, + 0.0020668746437877417, + 0.004673895891755819, + 0.0032882767263799906, + 0.0031195534393191338, + 0.002865493530407548, + 0.006007843650877476, + 0.004941205959767103, + 0.006827703211456537, + 0.007753909565508366, + 0.02680063061416149, + 0.01457968819886446, + 0.006817201618105173, + 0.007004349026829004, + 0.019262049347162247, + 0.017692100256681442, + 0.011513685807585716, + 0.008728018030524254, + 0.009463249705731869, + 0.015134497545659542, + 0.0058110859245061874, + 0.006135303061455488, + 0.015603462234139442, + 0.01428036205470562 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 8 + ], + "selected_reward": -0.006758213043212891 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 long shot that pans slowly to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a lone dromedary camel visible, silhouetted in deep black with a thin fiery halo outlining its hump and neck?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camel advancing with a leisurely, rhythmic gait while casting a vast, elongated, twisting shadow across the sand dunes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the sand dune crests glow with warm orange-gold light while the hollows are bathed in deep purple?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the sky in the distance painted with fiery orange, deep red, and soft purple colors?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01107556652277708 + ], + "opposite_logprob_tokens": [ + -4.511075496673584 + ], + "logprob_reward": -0.01107556652277708, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.01107556652277708, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3797215521335602, + 0.08423378318548203, + 0.04774875566363335, + 0.10015936940908432, + 0.03785513713955879, + 0.043253738433122635, + 0.13446299731731415, + 0.03171224147081375, + 0.08736053854227066, + 0.05349196121096611 + ], + "credit_map": [ + 0.041261959820985794, + 0.05586214363574982, + 0.10329670459032059, + 0.03037647344172001, + 0.010573460720479488, + 0.014907518401741982, + 0.02423839084804058, + 0.09920487552881241, + 0.006109973415732384, + 0.007248532027006149, + 0.015763623639941216, + 0.014325215481221676, + 0.009051562286913395, + 0.008189517073333263, + 0.013010100461542606, + 0.010535256005823612, + 0.0051797074265778065, + 0.004745309241116047, + 0.007746165618300438, + 0.006857261061668396, + 0.003952815663069487, + 0.005270299501717091, + 0.00834983866661787, + 0.005647353362292051, + 0.00884474441409111, + 0.008105544373393059, + 0.017451070249080658, + 0.019779590889811516, + 0.00850381888449192, + 0.006483145989477634, + 0.017248453572392464, + 0.013742990791797638, + 0.00579327205196023, + 0.0035413915757089853, + 0.0042637730948626995, + 0.004497942049056292, + 0.003279386553913355, + 0.0030338261276483536, + 0.008747433312237263, + 0.004698109347373247, + 0.004631386138498783, + 0.004557062406092882, + 0.006540524307638407, + 0.005207932088524103, + 0.0050444635562598705, + 0.0041257780976593494, + 0.006389830261468887, + 0.006756763439625502, + 0.006702861282974482, + 0.008786764927208424, + 0.014359460212290287, + 0.017297852784395218, + 0.012124814093112946, + 0.014201819896697998, + 0.04044453427195549, + 0.020544875413179398, + 0.003376387059688568, + 0.0030710750725120306, + 0.004175622947514057, + 0.0033626214135438204, + 0.0038089845329523087, + 0.0032010118011385202, + 0.005938949529081583, + 0.004777586553245783, + 0.006572316400706768, + 0.008118235506117344, + 0.020460152998566628, + 0.0107073113322258, + 0.006073619704693556, + 0.0045686145313084126, + 0.016637343913316727, + 0.01422293670475483, + 0.005120732355862856, + 0.004231273662298918, + 0.00693961326032877, + 0.012299664318561554, + 0.00368108949624002, + 0.0030823962297290564, + 0.009420198388397694, + 0.008716988377273083 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 3 + ], + "selected_reward": -0.01107556652277708 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 long shot that pans slowly to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a lone dromedary camel visible, silhouetted in deep black with a thin fiery halo outlining its hump and neck?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camel advancing with a leisurely, rhythmic gait while casting a vast, elongated, twisting shadow across the sand dunes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the sand dune crests glow with warm orange-gold light while the hollows are bathed in deep purple?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the sky in the distance painted with fiery orange, deep red, and soft purple colors?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02328341454267502 + ], + "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": [ + 0.3604211211204529, + 0.11775006353855133, + 0.019225507974624634, + 0.02789471484720707, + 0.02224539779126644, + 0.028210239484906197, + 0.04571656882762909, + 0.041222378611564636, + 0.16907742619514465, + 0.168236643075943 + ], + "credit_map": [ + 0.11128946393728256, + 0.1545964628458023, + 0.016712676733732224, + 0.01872575841844082, + 0.016828948631882668, + 0.019817929714918137, + 0.009604580700397491, + 0.012845281511545181, + 0.018535520881414413, + 0.01641731522977352, + 0.021096551790833473, + 0.014998241327702999, + 0.006757560186088085, + 0.005447597708553076, + 0.008983259089291096, + 0.025514010339975357, + 0.002157400129362941, + 0.0019357213750481606, + 0.0031031607650220394, + 0.0039587258361279964, + 0.001320944633334875, + 0.0018234460148960352, + 0.00229895138181746, + 0.002627156674861908, + 0.002345156390219927, + 0.0022807379718869925, + 0.00537214195355773, + 0.006882581394165754, + 0.0014865922275930643, + 0.0014691282995045185, + 0.0033800839446485043, + 0.004678293596953154, + 0.0019323391607031226, + 0.002627616049721837, + 0.004171372391283512, + 0.002626082394272089, + 0.0011292427079752088, + 0.0013704634038731456, + 0.0034150017891079187, + 0.004973278846591711, + 0.002697430085390806, + 0.002415603958070278, + 0.004785253666341305, + 0.0041159819811582565, + 0.002790206577628851, + 0.0023895050399005413, + 0.004539185669273138, + 0.004477072972804308, + 0.007357874885201454, + 0.00555756688117981, + 0.005768040660768747, + 0.005159197840839624, + 0.0027005162555724382, + 0.0033852842170745134, + 0.007472186349332333, + 0.008315898478031158, + 0.0034329877234995365, + 0.00305211846716702, + 0.005528791341930628, + 0.00389450928196311, + 0.0035809676628559828, + 0.005206302274018526, + 0.00999950896948576, + 0.006527193356305361, + 0.018338829278945923, + 0.017617585137486458, + 0.04763689264655113, + 0.02263960801064968, + 0.006374023854732513, + 0.0072537860833108425, + 0.020114274695515633, + 0.029102414846420288, + 0.018332192674279213, + 0.014886189252138138, + 0.019667088985443115, + 0.029822876676917076, + 0.01047732587903738, + 0.008520424365997314, + 0.029603568837046623, + 0.03692697361111641 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 9 + ], + "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 camera shot a low-angle long shot that pans slowly to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a lone dromedary camel visible, silhouetted in deep black with a thin fiery halo outlining its hump and neck?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camel advancing with a leisurely, rhythmic gait while casting a vast, elongated, twisting shadow across the sand dunes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the sand dune crests glow with warm orange-gold light while the hollows are bathed in deep purple?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the sky in the distance painted with fiery orange, deep red, and soft purple colors?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004634116776287556 + ], + "opposite_logprob_tokens": [ + -5.379633903503418 + ], + "logprob_reward": -0.004634116776287556, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004634116776287556, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3031999170780182, + 0.2646295428276062, + 0.03683415427803993, + 0.025455426424741745, + 0.03009966015815735, + 0.07039475440979004, + 0.045316144824028015, + 0.03481390327215195, + 0.09657695144414902, + 0.09267959743738174 + ], + "credit_map": [ + 0.06207913160324097, + 0.07378555834293365, + 0.06461621820926666, + 0.01822376623749733, + 0.00905540119856596, + 0.011620180681347847, + 0.013760642148554325, + 0.0500589981675148, + 0.03568271920084953, + 0.029974212870001793, + 0.0543406717479229, + 0.030105052515864372, + 0.014331256039440632, + 0.010223542340099812, + 0.02377278357744217, + 0.06619929522275925, + 0.004800873342901468, + 0.004450060427188873, + 0.005366231314837933, + 0.0061956848949193954, + 0.0022393714170902967, + 0.0028489951509982347, + 0.0059702396392822266, + 0.004962698090821505, + 0.002770096529275179, + 0.00250828149728477, + 0.003858615644276142, + 0.005466552916914225, + 0.0017725109355524182, + 0.0020741906482726336, + 0.003690339857712388, + 0.003314840141683817, + 0.0032769031822681427, + 0.0032685953192412853, + 0.004370903596282005, + 0.004537141416221857, + 0.0019790371879935265, + 0.002284771529957652, + 0.005435022059828043, + 0.004947285167872906, + 0.005134804639965296, + 0.006332497112452984, + 0.008916452527046204, + 0.010322212241590023, + 0.008139881305396557, + 0.00892830640077591, + 0.01367831602692604, + 0.00894228182733059, + 0.007454447913914919, + 0.005877654068171978, + 0.006105550564825535, + 0.004164187703281641, + 0.0032658386044204235, + 0.0033970822114497423, + 0.006272548343986273, + 0.0087788300588727, + 0.003111200872808695, + 0.0031891365069895983, + 0.004382725805044174, + 0.0035351316910237074, + 0.003913937136530876, + 0.002005366375669837, + 0.009670541621744633, + 0.005005864426493645, + 0.009470394812524319, + 0.008731240406632423, + 0.027663247659802437, + 0.012163737788796425, + 0.004050000105053186, + 0.004631063435226679, + 0.012841860763728619, + 0.017025403678417206, + 0.010385618545114994, + 0.008305166848003864, + 0.011352951638400555, + 0.018063481897115707, + 0.005763534922152758, + 0.004684555809944868, + 0.016407527029514313, + 0.017716756090521812 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 8 + ], + "selected_reward": -0.004634116776287556 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "sweeping low-angle, long shot follows the camera as it slowly pans right, tracking a lone camel across a vast expanse of desert bathed in the golden glow of a setting sun. The dromedary's entire body is silhouetted, rendered in a deep black, its outline outlined only by a thin, fiery halo of light from the setting sun below the horizon, clearly delineating its signature hump and gracefully curved neck. It advances with a leisurely, rhythmic gait, its powerful backlight casting a vast, elongated, almost grotesque shadow across the undulating sand dunes behind it. This shadow twists and flows across the dunes' undulating forms with every step. The dune crests, lit by the last rays of the setting sun, glow with a warm orange-gold hue, while the hollows between them are bathed in a deepening shade of deep purple, creating a stark contrast between light and dark. The sky in the distance is a magnificent painting of fiery orange, deep red and soft purple. The whole scene is filled with an epic sense of tranquility and loneliness, full of cinematic visual impact.", + "seed": 20093, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000121/group_1/index_2191_seed_20093_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": -0.02979777967557311, + "vlm_logprob": -0.02979777967557311, + "vlm_yes_no_margin": 4.475, + "vlm_token_credit": -0.02979777967557311, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 long shot that pans slowly to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a lone dromedary camel visible, silhouetted in deep black with a thin fiery halo outlining its hump and neck?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camel advancing with a leisurely, rhythmic gait while casting a vast, elongated, twisting shadow across the sand dunes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the sand dune crests glow with warm orange-gold light while the hollows are bathed in deep purple?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the sky in the distance painted with fiery orange, deep red, and soft purple colors?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11106274276971817 + ], + "opposite_logprob_tokens": [ + -2.36106276512146 + ], + "logprob_reward": -0.11106274276971817, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.11106274276971817, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04338986054062843, + 0.013487905263900757, + 0.01656135357916355, + 0.0295160673558712, + 0.3090602457523346, + 0.03363926336169243, + 0.033888716250658035, + 0.021055223420262337, + 0.48469895124435425, + 0.01470231544226408 + ], + "credit_map": [ + 0.00476286094635725, + 0.005398068577051163, + 0.005825020372867584, + 0.008710380643606186, + 0.003169912612065673, + 0.004040285013616085, + 0.00694520166143775, + 0.004538131412118673, + 0.0012487734202295542, + 0.0013796703424304724, + 0.0020231769885867834, + 0.002010033233091235, + 0.001343296840786934, + 0.0014498181408271194, + 0.0020249374210834503, + 0.00200819899328053, + 0.0014744683867320418, + 0.0018158911261707544, + 0.0028343426529318094, + 0.002888898830860853, + 0.0012965475907549262, + 0.0012976231519132853, + 0.0022563922684639692, + 0.00269719073548913, + 0.005487146787345409, + 0.002086907858029008, + 0.005222423002123833, + 0.003584388643503189, + 0.0023410574067384005, + 0.004314683377742767, + 0.00318126636557281, + 0.0032981927506625652, + 0.13102373480796814, + 0.008889865130186081, + 0.007308587897568941, + 0.003649420803412795, + 0.006041632499545813, + 0.13281957805156708, + 0.010771560482680798, + 0.008555872365832329, + 0.008814138360321522, + 0.0013429997488856316, + 0.003416789695620537, + 0.0020578568801283836, + 0.0019767116755247116, + 0.009622400626540184, + 0.003180295694619417, + 0.003228069981560111, + 0.004097479861229658, + 0.0029670714866369963, + 0.0065278164111077785, + 0.004126533400267363, + 0.002713033463805914, + 0.004236817825585604, + 0.004334412515163422, + 0.0048855505883693695, + 0.0030611788388341665, + 0.002546201692894101, + 0.002196735702455044, + 0.002488011959940195, + 0.002038508653640747, + 0.003393014660105109, + 0.00286369607783854, + 0.00246787560172379, + 0.1819618046283722, + 0.012929406017065048, + 0.02287207916378975, + 0.013464758172631264, + 0.01198819000273943, + 0.20597989857196808, + 0.012225794605910778, + 0.02327703684568405, + 0.0016171310562640429, + 0.001366175478324294, + 0.003097084816545248, + 0.0018168588867411017, + 0.0011802022345364094, + 0.0018115483690053225, + 0.001266947714611888, + 0.0025463667698204517 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 4, + 0 + ], + "selected_reward": -0.11106274276971817 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 long shot that pans slowly to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a lone dromedary camel visible, silhouetted in deep black with a thin fiery halo outlining its hump and neck?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camel advancing with a leisurely, rhythmic gait while casting a vast, elongated, twisting shadow across the sand dunes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the sand dune crests glow with warm orange-gold light while the hollows are bathed in deep purple?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the sky in the distance painted with fiery orange, deep red, and soft purple colors?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005965523421764374 + ], + "opposite_logprob_tokens": [ + -5.130965709686279 + ], + "logprob_reward": -0.005965523421764374, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005965523421764374, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.048603180795907974, + 0.0444425605237484, + 0.0244984719902277, + 0.04521646723151207, + 0.4660727083683014, + 0.09561949968338013, + 0.06903640925884247, + 0.043964359909296036, + 0.15085025131702423, + 0.011696110479533672 + ], + "credit_map": [ + 0.004324755631387234, + 0.0043101985938847065, + 0.005275642033666372, + 0.006580254528671503, + 0.0037730748299509287, + 0.004389953799545765, + 0.006732698995620012, + 0.013216602616012096, + 0.004012920428067446, + 0.005015344358980656, + 0.0038288014475256205, + 0.005531006958335638, + 0.005364325828850269, + 0.005315156187862158, + 0.008150516077876091, + 0.007224489469081163, + 0.003162496257573366, + 0.0030297988560050726, + 0.0034607204142957926, + 0.003964607138186693, + 0.0020965018775314093, + 0.0016312947263941169, + 0.0032430340070277452, + 0.003910018131136894, + 0.0052222199738025665, + 0.0038626529276371002, + 0.009518911130726337, + 0.0054044099524617195, + 0.004546074662357569, + 0.0058004725724458694, + 0.00459836283698678, + 0.00626336270943284, + 0.19514039158821106, + 0.013197984546422958, + 0.013124869205057621, + 0.006964736618101597, + 0.009430993348360062, + 0.19757285714149475, + 0.015924492850899696, + 0.014716378413140774, + 0.009738683700561523, + 0.006899947766214609, + 0.020970690995454788, + 0.008426900021731853, + 0.008968528360128403, + 0.012499586679041386, + 0.012056994251906872, + 0.016058163717389107, + 0.009739878587424755, + 0.004884956404566765, + 0.014534644782543182, + 0.006984003819525242, + 0.004912635777145624, + 0.009762097150087357, + 0.007813110947608948, + 0.010405084118247032, + 0.00523442542180419, + 0.005240392405539751, + 0.00491966912522912, + 0.006140774115920067, + 0.004610645584762096, + 0.0061296140775084496, + 0.006392433308064938, + 0.0052964044734835625, + 0.05741655454039574, + 0.004031643271446228, + 0.00720761576667428, + 0.0036297219339758158, + 0.0037088755052536726, + 0.06436308473348618, + 0.00374283897690475, + 0.006749916821718216, + 0.0010900357738137245, + 0.0010564690455794334, + 0.0023122630082070827, + 0.002104700542986393, + 0.0008974569500423968, + 0.0012851988431066275, + 0.001278999843634665, + 0.0016709869960322976 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 8, + 5 + ], + "selected_reward": -0.005965523421764374 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 long shot that pans slowly to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a lone dromedary camel visible, silhouetted in deep black with a thin fiery halo outlining its hump and neck?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camel advancing with a leisurely, rhythmic gait while casting a vast, elongated, twisting shadow across the sand dunes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the sand dune crests glow with warm orange-gold light while the hollows are bathed in deep purple?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the sky in the distance painted with fiery orange, deep red, and soft purple colors?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023287607356905937 + ], + "opposite_logprob_tokens": [ + -3.773287534713745 + ], + "logprob_reward": -0.023287607356905937, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023287607356905937, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08803975582122803, + 0.06386630237102509, + 0.03681120648980141, + 0.11730802804231644, + 0.35820141434669495, + 0.06659188121557236, + 0.09821321815252304, + 0.061048876494169235, + 0.06958679109811783, + 0.040332503616809845 + ], + "credit_map": [ + 0.009883665479719639, + 0.010927927680313587, + 0.011430754326283932, + 0.0157574862241745, + 0.00868700910359621, + 0.008758483454585075, + 0.011725937947630882, + 0.010868490673601627, + 0.005541328806430101, + 0.0060820551589131355, + 0.007225442212074995, + 0.008903962559998035, + 0.006715766154229641, + 0.006647094618529081, + 0.011779138818383217, + 0.010971513576805592, + 0.0033500944264233112, + 0.004015762358903885, + 0.005464576184749603, + 0.005631617270410061, + 0.0032869791612029076, + 0.003225713036954403, + 0.006005448289215565, + 0.005831015296280384, + 0.018190408125519753, + 0.009051486849784851, + 0.024567559361457825, + 0.014683142304420471, + 0.008851968683302402, + 0.016191251575946808, + 0.01277886237949133, + 0.012993356212973595, + 0.146955668926239, + 0.010921085253357887, + 0.011384619399905205, + 0.00777776725590229, + 0.008234229870140553, + 0.1471175104379654, + 0.01410236582159996, + 0.011708158068358898, + 0.010823138989508152, + 0.004860464483499527, + 0.011113897897303104, + 0.005266459193080664, + 0.005383842624723911, + 0.011730539612472057, + 0.007165292277932167, + 0.010248249396681786, + 0.013043089769780636, + 0.00660680141299963, + 0.021475043147802353, + 0.010081465356051922, + 0.007456835824996233, + 0.014398782514035702, + 0.010301066562533379, + 0.014850139617919922, + 0.011221961118280888, + 0.006193460896611214, + 0.008633630350232124, + 0.00561937503516674, + 0.00603076396510005, + 0.010060240514576435, + 0.006255084183067083, + 0.007034358102828264, + 0.010920180939137936, + 0.0033698962070047855, + 0.008170882239937782, + 0.007290369365364313, + 0.0034586265683174133, + 0.015032181516289711, + 0.00983963068574667, + 0.011505017057061195, + 0.004748417530208826, + 0.0027730115689337254, + 0.005734910257160664, + 0.005786889232695103, + 0.002643297193571925, + 0.0055311936885118484, + 0.004614369012415409, + 0.008500414900481701 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 3, + 6 + ], + "selected_reward": -0.023287607356905937 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 long shot that pans slowly to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a lone dromedary camel visible, silhouetted in deep black with a thin fiery halo outlining its hump and neck?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camel advancing with a leisurely, rhythmic gait while casting a vast, elongated, twisting shadow across the sand dunes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the sand dune crests glow with warm orange-gold light while the hollows are bathed in deep purple?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the sky in the distance painted with fiery orange, deep red, and soft purple colors?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0067337024956941605 + ], + "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": [ + 0.1098427101969719, + 0.03699817508459091, + 0.06947827339172363, + 0.04989837482571602, + 0.5449733138084412, + 0.040087927132844925, + 0.025728456676006317, + 0.04358779639005661, + 0.06110655888915062, + 0.018298406153917313 + ], + "credit_map": [ + 0.009446767158806324, + 0.00927545316517353, + 0.020492350682616234, + 0.024268154054880142, + 0.005129246972501278, + 0.006876811850816011, + 0.013097750023007393, + 0.02125616744160652, + 0.003956187050789595, + 0.004552451428025961, + 0.004441350698471069, + 0.004644966684281826, + 0.0033115013502538204, + 0.0034485552459955215, + 0.004663093946874142, + 0.007980065420269966, + 0.00878195557743311, + 0.00858751218765974, + 0.011735411360859871, + 0.013175420463085175, + 0.004790647886693478, + 0.004383708816021681, + 0.006333013065159321, + 0.011690590530633926, + 0.010512569919228554, + 0.00386842992156744, + 0.009738373570144176, + 0.004193808883428574, + 0.002428740728646517, + 0.009580970741808414, + 0.003302596742287278, + 0.006272879429161549, + 0.23087862133979797, + 0.015464621596038342, + 0.012945903465151787, + 0.0069915130734443665, + 0.01078468281775713, + 0.2337835431098938, + 0.01872975379228592, + 0.015394681133329868, + 0.0079522505402565, + 0.002745912177488208, + 0.006526322569698095, + 0.0027941628359258175, + 0.0027653523720800877, + 0.007708800490945578, + 0.0042137689888477325, + 0.005381355062127113, + 0.003487631445750594, + 0.0018080431036651134, + 0.005151734687387943, + 0.002648301189765334, + 0.0018641743808984756, + 0.0025751162320375443, + 0.003124989103525877, + 0.005068464670330286, + 0.005801618564873934, + 0.004326520022004843, + 0.008084752596914768, + 0.0043443250469863415, + 0.0034044187050312757, + 0.007047485094517469, + 0.004352985881268978, + 0.006225686985999346, + 0.012140347622334957, + 0.003158179111778736, + 0.008004232309758663, + 0.0046607693657279015, + 0.0033201139885932207, + 0.014605876989662647, + 0.007250511087477207, + 0.007966525852680206, + 0.0030360191594809294, + 0.0013103893725201488, + 0.0026293869595974684, + 0.0015989610692486167, + 0.0014736898010596633, + 0.003300921292975545, + 0.0015909936046227813, + 0.003358043497428298 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 2 + ], + "selected_reward": -0.0067337024956941605 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a low-angle long shot that pans slowly to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a lone dromedary camel visible, silhouetted in deep black with a thin fiery halo outlining its hump and neck?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camel advancing with a leisurely, rhythmic gait while casting a vast, elongated, twisting shadow across the sand dunes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the sand dune crests glow with warm orange-gold light while the hollows are bathed in deep purple?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the sky in the distance painted with fiery orange, deep red, and soft purple colors?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0019393223337829113 + ], + "opposite_logprob_tokens": [ + -6.251939296722412 + ], + "logprob_reward": -0.0019393223337829113, + "yes_no_margin_reward": 6.25, + "token_credit_reward": -0.0019393223337829113, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1867797076702118, + 0.05840035900473595, + 0.059665679931640625, + 0.07804583013057709, + 0.29962342977523804, + 0.1433701366186142, + 0.04398660734295845, + 0.05648388713598251, + 0.03938312083482742, + 0.034261204302310944 + ], + "credit_map": [ + 0.01620992086827755, + 0.01598190888762474, + 0.0314888097345829, + 0.03940804675221443, + 0.010485509410500526, + 0.014056936837732792, + 0.020937640219926834, + 0.03821094334125519, + 0.006201567128300667, + 0.006483003031462431, + 0.006868104916065931, + 0.011622187681496143, + 0.005064823664724827, + 0.004832729697227478, + 0.008234676904976368, + 0.009093266911804676, + 0.006994672119617462, + 0.006655571516603231, + 0.00900830514729023, + 0.010842474177479744, + 0.004803847521543503, + 0.004166120663285255, + 0.00775544811040163, + 0.009439242072403431, + 0.00962216779589653, + 0.007786944042891264, + 0.01731341890990734, + 0.008658056147396564, + 0.00679037906229496, + 0.009162730537354946, + 0.009944350458681583, + 0.008767787367105484, + 0.11559891700744629, + 0.009917087852954865, + 0.013330689631402493, + 0.008341887965798378, + 0.010360396467149258, + 0.1183113306760788, + 0.011378355324268341, + 0.012384800240397453, + 0.03967548906803131, + 0.010132987052202225, + 0.015799812972545624, + 0.007492066826671362, + 0.0075117116793990135, + 0.04222679138183594, + 0.010457132011651993, + 0.010074137710034847, + 0.004579049069434404, + 0.005010731052607298, + 0.011230725795030594, + 0.0036900024861097336, + 0.0030510162468999624, + 0.004645207896828651, + 0.00474221957847476, + 0.00703765545040369, + 0.00935538299381733, + 0.004824284464120865, + 0.006572700571268797, + 0.005842751357704401, + 0.004009837284684181, + 0.00993225909769535, + 0.006492982152849436, + 0.009453696198761463, + 0.004227411467581987, + 0.0034716345835477114, + 0.0068211788311600685, + 0.004986026789993048, + 0.003174418583512306, + 0.005807242821902037, + 0.004304961767047644, + 0.006590249948203564, + 0.006634208373725414, + 0.002503360854461789, + 0.004680668469518423, + 0.0035620187409222126, + 0.0026098729576915503, + 0.006412675604224205, + 0.0035424812231212854, + 0.0043159183114767075 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 5 + ], + "selected_reward": -0.0019393223337829113 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "sweeping low-angle, long shot follows the camera as it slowly pans right, tracking a lone camel across a vast expanse of desert bathed in the golden glow of a setting sun. The dromedary's entire body is silhouetted, rendered in a deep black, its outline outlined only by a thin, fiery halo of light from the setting sun below the horizon, clearly delineating its signature hump and gracefully curved neck. It advances with a leisurely, rhythmic gait, its powerful backlight casting a vast, elongated, almost grotesque shadow across the undulating sand dunes behind it. This shadow twists and flows across the dunes' undulating forms with every step. The dune crests, lit by the last rays of the setting sun, glow with a warm orange-gold hue, while the hollows between them are bathed in a deepening shade of deep purple, creating a stark contrast between light and dark. The sky in the distance is a magnificent painting of fiery orange, deep red and soft purple. The whole scene is filled with an epic sense of tranquility and loneliness, full of cinematic visual impact.", + "seed": 20094, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000121/group_1/index_2191_seed_20094_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": -0.02237014186102897, + "vlm_logprob": -0.02237014186102897, + "vlm_yes_no_margin": 4.9, + "vlm_token_credit": -0.02237014186102897, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 long shot that pans slowly to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a lone dromedary camel visible, silhouetted in deep black with a thin fiery halo outlining its hump and neck?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camel advancing with a leisurely, rhythmic gait while casting a vast, elongated, twisting shadow across the sand dunes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the sand dune crests glow with warm orange-gold light while the hollows are bathed in deep purple?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the sky in the distance painted with fiery orange, deep red, and soft purple colors?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.09276434034109116 + ], + "opposite_logprob_tokens": [ + -2.592764377593994 + ], + "logprob_reward": -0.09276434034109116, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.09276434034109116, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18448103964328766, + 0.10237347334623337, + 0.10571970790624619, + 0.1763714849948883, + 0.09266624599695206, + 0.11331018060445786, + 0.052988678216934204, + 0.07871604710817337, + 0.06888704746961594, + 0.024486131966114044 + ], + "credit_map": [ + 0.010299536399543285, + 0.03721623122692108, + 0.014833416789770126, + 0.04235873371362686, + 0.024711577221751213, + 0.020373189821839333, + 0.024343902245163918, + 0.01034445408731699, + 0.008586482144892216, + 0.011900557205080986, + 0.020289581269025803, + 0.018576184287667274, + 0.008314598351716995, + 0.008572805672883987, + 0.015526587143540382, + 0.010606671683490276, + 0.012410897761583328, + 0.009725324809551239, + 0.015853917226195335, + 0.03450128436088562, + 0.006033327896147966, + 0.0037095423322170973, + 0.010833609849214554, + 0.012651798315346241, + 0.016319068148732185, + 0.022578870877623558, + 0.01940697617828846, + 0.02885964885354042, + 0.020957371219992638, + 0.023829806596040726, + 0.024702012538909912, + 0.01971772499382496, + 0.008301591500639915, + 0.008039431646466255, + 0.012099221348762512, + 0.010077793151140213, + 0.013168508186936378, + 0.011033935472369194, + 0.012648828327655792, + 0.01729693077504635, + 0.010949728079140186, + 0.01150261890143156, + 0.016685349866747856, + 0.018414635211229324, + 0.01122713927179575, + 0.012524190358817577, + 0.015443428419530392, + 0.016563091427087784, + 0.005267784930765629, + 0.003958815708756447, + 0.00828919094055891, + 0.008388123475015163, + 0.005040598567575216, + 0.005343612749129534, + 0.008307719603180885, + 0.008392831310629845, + 0.015498919412493706, + 0.006507247220724821, + 0.014032482169568539, + 0.008362122811377048, + 0.00922109093517065, + 0.00582865672186017, + 0.009216710925102234, + 0.010048811323940754, + 0.005122572183609009, + 0.0074001094326376915, + 0.004071923904120922, + 0.012330800294876099, + 0.011913971975445747, + 0.004969799425452948, + 0.013868432492017746, + 0.00920942798256874, + 0.003396872663870454, + 0.0026763123460114002, + 0.0023751449771225452, + 0.0032355443108826876, + 0.003423241199925542, + 0.0023699430748820305, + 0.0036385238636285067, + 0.0033705478999763727 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 5 + ], + "selected_reward": -0.09276434034109116 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 long shot that pans slowly to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a lone dromedary camel visible, silhouetted in deep black with a thin fiery halo outlining its hump and neck?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camel advancing with a leisurely, rhythmic gait while casting a vast, elongated, twisting shadow across the sand dunes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the sand dune crests glow with warm orange-gold light while the hollows are bathed in deep purple?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the sky in the distance painted with fiery orange, deep red, and soft purple colors?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004107492044568062 + ], + "opposite_logprob_tokens": [ + -5.504107475280762 + ], + "logprob_reward": -0.004107492044568062, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004107492044568062, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12168090045452118, + 0.19161704182624817, + 0.0718645378947258, + 0.07908904552459717, + 0.07614118605852127, + 0.041931334882974625, + 0.2134438008069992, + 0.09862303733825684, + 0.06766827404499054, + 0.037940863519907 + ], + "credit_map": [ + 0.008350515738129616, + 0.01713954657316208, + 0.009145939722657204, + 0.03322141245007515, + 0.014046851545572281, + 0.009676270186901093, + 0.018424849957227707, + 0.011675510555505753, + 0.01339618768543005, + 0.020771978422999382, + 0.04955647885799408, + 0.025203896686434746, + 0.01774037815630436, + 0.017737409099936485, + 0.021165765821933746, + 0.026044927537441254, + 0.005922171752899885, + 0.014020465314388275, + 0.012193062342703342, + 0.010219988413155079, + 0.008931346237659454, + 0.005871812347322702, + 0.0068993899039924145, + 0.0078062983229756355, + 0.00719060841947794, + 0.013621922582387924, + 0.010003190487623215, + 0.011760166846215725, + 0.009110364131629467, + 0.010387121699750423, + 0.008277456276118755, + 0.008738215081393719, + 0.006826456636190414, + 0.009925060905516148, + 0.011162512935698032, + 0.008603091351687908, + 0.009112677536904812, + 0.007012497633695602, + 0.013887090608477592, + 0.009611803106963634, + 0.003207223955541849, + 0.00462990440428257, + 0.005910159088671207, + 0.005767894443124533, + 0.0050715492106974125, + 0.0039937179535627365, + 0.006856719963252544, + 0.006494165398180485, + 0.031281404197216034, + 0.012717383913695812, + 0.06287889182567596, + 0.020393196493387222, + 0.01467581745237112, + 0.022155648097395897, + 0.016951078549027443, + 0.032390378415584564, + 0.013709096238017082, + 0.01099978107959032, + 0.0109662264585495, + 0.012812399305403233, + 0.015085299499332905, + 0.007169763557612896, + 0.014644487760961056, + 0.013235983438789845, + 0.005737998057156801, + 0.0091861542314291, + 0.005033519584685564, + 0.01007932610809803, + 0.013479997403919697, + 0.005464005284011364, + 0.010008692741394043, + 0.008678574115037918, + 0.004876235034316778, + 0.004220571368932724, + 0.002618520287796855, + 0.004452580586075783, + 0.00814778357744217, + 0.0023565213195979595, + 0.005791281815618277, + 0.0054773688316345215 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 1, + 0 + ], + "selected_reward": -0.004107492044568062 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 long shot that pans slowly to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a lone dromedary camel visible, silhouetted in deep black with a thin fiery halo outlining its hump and neck?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camel advancing with a leisurely, rhythmic gait while casting a vast, elongated, twisting shadow across the sand dunes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the sand dune crests glow with warm orange-gold light while the hollows are bathed in deep purple?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the sky in the distance painted with fiery orange, deep red, and soft purple colors?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00863425899296999 + ], + "opposite_logprob_tokens": [ + -4.758634090423584 + ], + "logprob_reward": -0.00863425899296999, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.00863425899296999, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.055902205407619476, + 0.06492871791124344, + 0.1483021229505539, + 0.1895219087600708, + 0.21010737121105194, + 0.09991169720888138, + 0.050562527030706406, + 0.07376940548419952, + 0.06990964710712433, + 0.03708445280790329 + ], + "credit_map": [ + 0.003472184529528022, + 0.009269844740629196, + 0.00415849220007658, + 0.013621008023619652, + 0.006340850610285997, + 0.004362578969448805, + 0.009017121978104115, + 0.005660124588757753, + 0.009437479078769684, + 0.006253585685044527, + 0.016751376911997795, + 0.007882334291934967, + 0.004172494634985924, + 0.0030427926685661077, + 0.01084929145872593, + 0.006539362482726574, + 0.012234268710017204, + 0.01008335780352354, + 0.031025046482682228, + 0.05799049884080887, + 0.006580700632184744, + 0.0044097607024014, + 0.014818750321865082, + 0.011159753426909447, + 0.026729067787528038, + 0.019575145095586777, + 0.02262837067246437, + 0.0252228956669569, + 0.01897408813238144, + 0.036061111837625504, + 0.018609512597322464, + 0.02172173000872135, + 0.028028856962919235, + 0.01681881956756115, + 0.02939593233168125, + 0.027388300746679306, + 0.020305953919887543, + 0.038920942693948746, + 0.02109331078827381, + 0.0281552504748106, + 0.009312243200838566, + 0.008838997222483158, + 0.015985792502760887, + 0.015947571024298668, + 0.008783359080553055, + 0.010927499271929264, + 0.011767820455133915, + 0.018348420038819313, + 0.006747599225491285, + 0.004541889298707247, + 0.009453271515667439, + 0.0071221934631466866, + 0.0060108499601483345, + 0.0050543975085020065, + 0.004975807853043079, + 0.006656517740339041, + 0.008070175535976887, + 0.006689051166176796, + 0.014768002554774284, + 0.006741048768162727, + 0.00839286483824253, + 0.009959946386516094, + 0.009326048195362091, + 0.009822270832955837, + 0.005675183609127998, + 0.01034544501453638, + 0.005834055133163929, + 0.009120019152760506, + 0.014513736590743065, + 0.005942601710557938, + 0.009208295494318008, + 0.009270308539271355, + 0.007398258429020643, + 0.0034455882851034403, + 0.0060096909292042255, + 0.0043141464702785015, + 0.0031299374531954527, + 0.002276310231536627, + 0.00481518404558301, + 0.005695335566997528 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 3, + 2 + ], + "selected_reward": -0.00863425899296999 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 long shot that pans slowly to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a lone dromedary camel visible, silhouetted in deep black with a thin fiery halo outlining its hump and neck?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camel advancing with a leisurely, rhythmic gait while casting a vast, elongated, twisting shadow across the sand dunes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the sand dune crests glow with warm orange-gold light while the hollows are bathed in deep purple?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the sky in the distance painted with fiery orange, deep red, and soft purple colors?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004633998032659292 + ], + "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": [ + 0.12293603271245956, + 0.11917800456285477, + 0.05760110914707184, + 0.1596277505159378, + 0.1658499389886856, + 0.09041839092969894, + 0.0877315029501915, + 0.12921586632728577, + 0.046164967119693756, + 0.02127647213637829 + ], + "credit_map": [ + 0.007749936543405056, + 0.018785512074828148, + 0.012060320004820824, + 0.033964723348617554, + 0.012799043208360672, + 0.009162544272840023, + 0.016847532242536545, + 0.011566417291760445, + 0.0076157404109835625, + 0.008505221456289291, + 0.02779253199696541, + 0.01724950410425663, + 0.006844487506896257, + 0.0069845994003117085, + 0.03537285700440407, + 0.008813063614070415, + 0.005799326580017805, + 0.006283474620431662, + 0.01157007459551096, + 0.014387266710400581, + 0.004646709654480219, + 0.003722674446180463, + 0.007125246804207563, + 0.0040663303807377815, + 0.019630100578069687, + 0.019136054441332817, + 0.018948940560221672, + 0.023599958047270775, + 0.01752256229519844, + 0.025357859209179878, + 0.01939309574663639, + 0.01603916846215725, + 0.023063568398356438, + 0.013271329924464226, + 0.022961938753724098, + 0.021809171885252, + 0.0173899345099926, + 0.02729656733572483, + 0.01905229315161705, + 0.021005135029554367, + 0.009971555322408676, + 0.007917926646769047, + 0.014074266888201237, + 0.011346645653247833, + 0.008281860500574112, + 0.009693748317658901, + 0.010587138123810291, + 0.018545251339673996, + 0.011392251588404179, + 0.005335292778909206, + 0.028169982135295868, + 0.0076937139965593815, + 0.0035280198790133, + 0.006982059217989445, + 0.009414056316018105, + 0.015216127038002014, + 0.02897423505783081, + 0.00825233943760395, + 0.03176194801926613, + 0.011225991882383823, + 0.009685010649263859, + 0.012554147280752659, + 0.00919201597571373, + 0.01757018454372883, + 0.0036015408113598824, + 0.0065889074467122555, + 0.003833414986729622, + 0.007352413143962622, + 0.008947115391492844, + 0.003117949003353715, + 0.006730761844664812, + 0.005992859601974487, + 0.0031835800036787987, + 0.0019507819088175893, + 0.0023999859113246202, + 0.0028758863918483257, + 0.0020866619888693094, + 0.0016103979432955384, + 0.0036703497171401978, + 0.0034988296683877707 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 3, + 7 + ], + "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 the camera shot a low-angle long shot that pans slowly to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a lone dromedary camel visible, silhouetted in deep black with a thin fiery halo outlining its hump and neck?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camel advancing with a leisurely, rhythmic gait while casting a vast, elongated, twisting shadow across the sand dunes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the sand dune crests glow with warm orange-gold light while the hollows are bathed in deep purple?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the sky in the distance painted with fiery orange, deep red, and soft purple colors?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0017106198938563466 + ], + "opposite_logprob_tokens": [ + -6.376710414886475 + ], + "logprob_reward": -0.0017106198938563466, + "yes_no_margin_reward": 6.375, + "token_credit_reward": -0.0017106198938563466, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12496306002140045, + 0.19194358587265015, + 0.07936214655637741, + 0.163863405585289, + 0.1648787558078766, + 0.03656957671046257, + 0.051530998200178146, + 0.11822038143873215, + 0.027432812377810478, + 0.04123526066541672 + ], + "credit_map": [ + 0.010924217291176319, + 0.02394096553325653, + 0.011765100061893463, + 0.024214446544647217, + 0.014238756150007248, + 0.012031727470457554, + 0.015917176380753517, + 0.011930673383176327, + 0.015865786001086235, + 0.007958161644637585, + 0.07572582364082336, + 0.027024731040000916, + 0.006960913073271513, + 0.008885964751243591, + 0.02919894829392433, + 0.0203232504427433, + 0.013471392914652824, + 0.011093837209045887, + 0.011530839838087559, + 0.010169130750000477, + 0.007748087402433157, + 0.006851439829915762, + 0.007521587423980236, + 0.01097582746297121, + 0.02175094000995159, + 0.019732128828763962, + 0.01945570670068264, + 0.022893555462360382, + 0.01758749596774578, + 0.026436712592840195, + 0.018796725198626518, + 0.017210152000188828, + 0.022619381546974182, + 0.013554535806179047, + 0.02255801297724247, + 0.021125245839357376, + 0.018189063295722008, + 0.02670109272003174, + 0.019919676706194878, + 0.020211759954690933, + 0.0040295859798789024, + 0.003176039084792137, + 0.005419176071882248, + 0.0057182046584784985, + 0.0042969598434865475, + 0.005291844252496958, + 0.003976650070399046, + 0.004661118611693382, + 0.006914098747074604, + 0.003982193302363157, + 0.009556733071804047, + 0.0065495711751282215, + 0.005363258998841047, + 0.005107739008963108, + 0.0073018912225961685, + 0.006755512673407793, + 0.019335372373461723, + 0.007730337791144848, + 0.027920391410589218, + 0.012422402389347553, + 0.010415778495371342, + 0.010497739538550377, + 0.013703066855669022, + 0.016195299103856087, + 0.00281985173933208, + 0.0029623962473124266, + 0.0031168542336672544, + 0.0040984065271914005, + 0.0034450681414455175, + 0.0022482771892100573, + 0.004630107432603836, + 0.004111852031201124, + 0.005420365370810032, + 0.004063445143401623, + 0.005741218104958534, + 0.0036278555635362864, + 0.005888238083571196, + 0.004914518445730209, + 0.005496462807059288, + 0.006083157379180193 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 4, + 3 + ], + "selected_reward": -0.0017106198938563466 + } + ], + "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/0000121/group_10/index_tensor([6300, 6300])_group_10_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000121/group_10/index_tensor([6300, 6300])_group_10_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..a6bdbfb96a093e1e09c4cbbd89c7c2c02552ff46 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000121/group_10/index_tensor([6300, 6300])_group_10_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "stunning close-up shot, shot from an extremely low angle, almost touching the ground, focuses on a massive rocket engine nozzle. Its complex metal structure, thermal insulation tiles, and dense network of pipes create an overwhelming industrial aesthetic, with parts of the surface still blackened by soot. In an instant, a blinding white spark erupts from the nozzle's center. Immediately, a dense column of fire, composed of a blazing white core and swirling orange-yellow flames, erupts violently. This unstoppable torrent of energy rapidly expands, crashing down and completely engulfing the concrete launch platform below. The flames violently impact the rough-hewn guide ducts, sending sparks flying through the sky, while billowing white steam and gray smoke billow out from all sides, rapidly filling the frame. The dazzling light of this raw power illuminates every metallic detail of the engine, casting violently dancing shadows. The entire frame vibrates violently with the immense energy release, creating an awe-inspiring atmosphere of raw power and technological wonder.", + "seed": 27430, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000121/group_10/index_6300_seed_27430_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": -0.016259132884442806, + "vlm_logprob": -0.016259132884442806, + "vlm_yes_no_margin": 5.4, + "vlm_token_credit": -0.016259132884442806, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 massive rocket engine nozzle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle extremely low, almost touching the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the nozzle show complex metal structure, thermal insulation tiles, and soot-blackened surfaces?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a blinding white spark erupt from the nozzle center followed by a dense column of fire with a white core and orange-yellow flames?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do billowing white steam and gray smoke rapidly fill the frame while the frame vibrates violently?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007669171784073114 + ], + "opposite_logprob_tokens": [ + -6.382668972015381 + ], + "logprob_reward": -0.007669171784073114, + "yes_no_margin_reward": 6.375, + "token_credit_reward": -0.007669171784073114, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08371245115995407, + 0.06644511222839355, + 0.08103914558887482, + 0.06722946465015411, + 0.06026645004749298, + 0.06334536522626877, + 0.08327209949493408, + 0.07944463938474655, + 0.0640902891755104, + 0.35115504264831543 + ], + "credit_map": [ + 0.009469245560467243, + 0.008349834941327572, + 0.008265413343906403, + 0.011173965409398079, + 0.018589701503515244, + 0.0063058165833354, + 0.012288742698729038, + 0.009269729256629944, + 0.009963327087461948, + 0.007755551487207413, + 0.005554129835218191, + 0.009066242724657059, + 0.011060402728617191, + 0.006338108796626329, + 0.00967140682041645, + 0.007035931572318077, + 0.012845460325479507, + 0.005801150109618902, + 0.00673385476693511, + 0.012894497253000736, + 0.008670276962220669, + 0.005916063208132982, + 0.015870144590735435, + 0.012307687662541866, + 0.0052417004480957985, + 0.004030696116387844, + 0.0033788818400353193, + 0.02034003846347332, + 0.004853931255638599, + 0.0033847452141344547, + 0.0213328767567873, + 0.004666589200496674, + 0.0078084575943648815, + 0.006741655524820089, + 0.004883752670139074, + 0.009737493470311165, + 0.007237820886075497, + 0.008167505264282227, + 0.008981872349977493, + 0.006707888562232256, + 0.006629212759435177, + 0.015809396281838417, + 0.004478424321860075, + 0.008000163361430168, + 0.011040209792554379, + 0.005446034017950296, + 0.00494571216404438, + 0.0069962069392204285, + 0.013981270603835583, + 0.010978697799146175, + 0.008486870676279068, + 0.01117795705795288, + 0.013439937494695187, + 0.007640555500984192, + 0.01106367725878954, + 0.006503122393041849, + 0.007348996587097645, + 0.010668400675058365, + 0.011595400981605053, + 0.009382902644574642, + 0.012126736342906952, + 0.006550775375217199, + 0.013235711492598057, + 0.00853570457547903, + 0.005008258391171694, + 0.00995495356619358, + 0.00952167995274067, + 0.0067557403817772865, + 0.008376841433346272, + 0.010432427749037743, + 0.0055022467859089375, + 0.008538134396076202, + 0.03369754180312157, + 0.04645717889070511, + 0.05827679857611656, + 0.048953305929899216, + 0.03924163803458214, + 0.04907902702689171, + 0.041736919432878494, + 0.033712588250637054 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 6 + ], + "selected_reward": -0.007669171784073114 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 massive rocket engine nozzle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle extremely low, almost touching the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the nozzle show complex metal structure, thermal insulation tiles, and soot-blackened surfaces?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a blinding white spark erupt from the nozzle center followed by a dense column of fire with a white core and orange-yellow flames?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do billowing white steam and gray smoke rapidly fill the frame while the frame vibrates violently?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06207215040922165 + ], + "opposite_logprob_tokens": [ + -2.8120720386505127 + ], + "logprob_reward": -0.06207215040922165, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06207215040922165, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13603031635284424, + 0.057613275945186615, + 0.051743004471063614, + 0.02178732305765152, + 0.0723656639456749, + 0.11905894428491592, + 0.17839409410953522, + 0.15180879831314087, + 0.10161877423524857, + 0.10957985371351242 + ], + "credit_map": [ + 0.02314302884042263, + 0.013991030864417553, + 0.014849337749183178, + 0.021511811763048172, + 0.021944919601082802, + 0.01071279589086771, + 0.017576392740011215, + 0.012300986796617508, + 0.006110170390456915, + 0.006974054966121912, + 0.0045018610544502735, + 0.007766938302665949, + 0.015209701843559742, + 0.005217096768319607, + 0.006825563497841358, + 0.0050078872591257095, + 0.0045573473908007145, + 0.005095379427075386, + 0.0054924385622143745, + 0.00848978478461504, + 0.007317666430026293, + 0.005176512524485588, + 0.01004274282604456, + 0.005571130197495222, + 0.002013159915804863, + 0.0024363393895328045, + 0.0027155571151524782, + 0.004292156547307968, + 0.0027474071830511093, + 0.001734389690682292, + 0.0035014061722904444, + 0.0023469047155231237, + 0.010660247877240181, + 0.009875896386802197, + 0.006954419892281294, + 0.010475757531821728, + 0.008310189470648766, + 0.007790809031575918, + 0.010482634417712688, + 0.00781569816172123, + 0.010719395242631435, + 0.02211783267557621, + 0.013495469465851784, + 0.017956290394067764, + 0.01628762297332287, + 0.013812471181154251, + 0.014836205169558525, + 0.009833656251430511, + 0.01741415075957775, + 0.022702913731336594, + 0.020394759252667427, + 0.03241884335875511, + 0.03014531172811985, + 0.014696110039949417, + 0.025934187695384026, + 0.014687810093164444, + 0.017080729827284813, + 0.016972586512565613, + 0.02665402926504612, + 0.026623040437698364, + 0.021576952189207077, + 0.011167224496603012, + 0.016834309324622154, + 0.0148999048396945, + 0.006762078031897545, + 0.011080733500421047, + 0.01646655611693859, + 0.012422388419508934, + 0.011243641376495361, + 0.01610102504491806, + 0.012598163448274136, + 0.014944178983569145, + 0.01224795263260603, + 0.01306504663079977, + 0.016718050464987755, + 0.01948940008878708, + 0.014144317246973515, + 0.012203417718410492, + 0.011886888183653355, + 0.009824777022004128 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 7, + 0 + ], + "selected_reward": -0.06207215040922165 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 massive rocket engine nozzle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle extremely low, almost touching the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the nozzle show complex metal structure, thermal insulation tiles, and soot-blackened surfaces?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a blinding white spark erupt from the nozzle center followed by a dense column of fire with a white core and orange-yellow flames?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do billowing white steam and gray smoke rapidly fill the frame while the frame vibrates violently?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003203381784260273 + ], + "opposite_logprob_tokens": [ + -5.753203392028809 + ], + "logprob_reward": -0.003203381784260273, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.003203381784260273, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08059237152338028, + 0.11105939000844955, + 0.05854594334959984, + 0.25191399455070496, + 0.047235071659088135, + 0.05542515218257904, + 0.041893038898706436, + 0.14293313026428223, + 0.052187707275152206, + 0.1582142412662506 + ], + "credit_map": [ + 0.008599773049354553, + 0.00914801750332117, + 0.0064958129078149796, + 0.014690394513309002, + 0.01438630186021328, + 0.006842738017439842, + 0.012184699065983295, + 0.008244636468589306, + 0.011918751522898674, + 0.010732474736869335, + 0.009461735375225544, + 0.020830368623137474, + 0.015443289652466774, + 0.010253035463392735, + 0.01408364251255989, + 0.018336093053221703, + 0.0060615031979978085, + 0.004539776593446732, + 0.005373412277549505, + 0.015080936253070831, + 0.0053090848959982395, + 0.0032602157443761826, + 0.010204781778156757, + 0.008716234937310219, + 0.017588747665286064, + 0.013821400701999664, + 0.008261761628091335, + 0.08017010241746902, + 0.01717505045235157, + 0.009165645577013493, + 0.08839412033557892, + 0.01733718439936638, + 0.0058538345620036125, + 0.00658201752230525, + 0.003890935331583023, + 0.006381272338330746, + 0.005858142860233784, + 0.004726871848106384, + 0.0090012913569808, + 0.004940707702189684, + 0.006736147217452526, + 0.009254439733922482, + 0.005086617544293404, + 0.008587535470724106, + 0.007826922461390495, + 0.003265964798629284, + 0.007857215590775013, + 0.006810312625020742, + 0.0032342011108994484, + 0.005621972493827343, + 0.003953850362449884, + 0.006089889910072088, + 0.006874831393361092, + 0.0037149470299482346, + 0.008143185637891293, + 0.0042601595632731915, + 0.012631814926862717, + 0.017982928082346916, + 0.027286989614367485, + 0.027922548353672028, + 0.021296102553606033, + 0.008517823182046413, + 0.01667899079620838, + 0.010615929961204529, + 0.004503137897700071, + 0.00866638869047165, + 0.005359625443816185, + 0.005511557217687368, + 0.0063821556977927685, + 0.010934248566627502, + 0.005047840531915426, + 0.005782753694802523, + 0.015542196109890938, + 0.02305947057902813, + 0.022893264889717102, + 0.025850238278508186, + 0.01857806369662285, + 0.020185044035315514, + 0.017975151538848877, + 0.014130805619060993 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 9, + 7 + ], + "selected_reward": -0.003203381784260273 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 massive rocket engine nozzle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle extremely low, almost touching the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the nozzle show complex metal structure, thermal insulation tiles, and soot-blackened surfaces?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a blinding white spark erupt from the nozzle center followed by a dense column of fire with a white core and orange-yellow flames?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do billowing white steam and gray smoke rapidly fill the frame while the frame vibrates violently?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0007195267826318741 + ], + "opposite_logprob_tokens": [ + -7.2507195472717285 + ], + "logprob_reward": -0.0007195267826318741, + "yes_no_margin_reward": 7.25, + "token_credit_reward": -0.0007195267826318741, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.062160976231098175, + 0.0702134221792221, + 0.06557301431894302, + 0.1753568798303604, + 0.0653112381696701, + 0.03361129388213158, + 0.0585615336894989, + 0.05883535370230675, + 0.1752638965845108, + 0.23511242866516113 + ], + "credit_map": [ + 0.007494049146771431, + 0.006211637053638697, + 0.005276137962937355, + 0.010924668051302433, + 0.009571981616318226, + 0.00519480649381876, + 0.010406924411654472, + 0.0070807733573019505, + 0.00826912559568882, + 0.0063409567810595036, + 0.004312200006097555, + 0.012608325108885765, + 0.012567148543894291, + 0.005763264372944832, + 0.009800419211387634, + 0.010551981627941132, + 0.007899879477918148, + 0.00801827572286129, + 0.0055411141365766525, + 0.012659058906137943, + 0.008489929139614105, + 0.0049413410015404224, + 0.011830094270408154, + 0.006193322129547596, + 0.012595743872225285, + 0.00991043634712696, + 0.006017609033733606, + 0.05430067703127861, + 0.01263171061873436, + 0.006312154233455658, + 0.06133162975311279, + 0.012256931513547897, + 0.00626470847055316, + 0.008405103348195553, + 0.006457718554884195, + 0.012830846011638641, + 0.009170493111014366, + 0.007829174399375916, + 0.008517857640981674, + 0.005835336167365313, + 0.003089851699769497, + 0.006859130226075649, + 0.0035880571231245995, + 0.004276954103261232, + 0.0047699566930532455, + 0.0034428902436047792, + 0.005117847118526697, + 0.002466607838869095, + 0.007922609336674213, + 0.00695195235311985, + 0.00799696147441864, + 0.008482035249471664, + 0.00984808336943388, + 0.004781484138220549, + 0.007536215707659721, + 0.005042190197855234, + 0.0057228426449000835, + 0.005500334315001965, + 0.007416270207613707, + 0.008397131226956844, + 0.010442206636071205, + 0.004694342613220215, + 0.012655265629291534, + 0.004006957169622183, + 0.01199119072407484, + 0.018040930852293968, + 0.018851254135370255, + 0.017281150445342064, + 0.02178335189819336, + 0.043815676122903824, + 0.01898006536066532, + 0.024520261213183403, + 0.02828332968056202, + 0.019625475630164146, + 0.05118288844823837, + 0.03492957353591919, + 0.02925320342183113, + 0.021631231531500816, + 0.027063248679041862, + 0.023143479600548744 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 3, + 8 + ], + "selected_reward": -0.0007195267826318741 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 massive rocket engine nozzle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle extremely low, almost touching the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the nozzle show complex metal structure, thermal insulation tiles, and soot-blackened surfaces?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a blinding white spark erupt from the nozzle center followed by a dense column of fire with a white core and orange-yellow flames?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do billowing white steam and gray smoke rapidly fill the frame while the frame vibrates violently?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"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": [ + 0.0549807995557785, + 0.03289501741528511, + 0.023028934374451637, + 0.12565553188323975, + 0.07453981786966324, + 0.03475262597203255, + 0.037835508584976196, + 0.06813627481460571, + 0.04187421500682831, + 0.5063012838363647 + ], + "credit_map": [ + 0.00651745218783617, + 0.005490132141858339, + 0.004447543993592262, + 0.010194207541644573, + 0.011233491823077202, + 0.0038712285459041595, + 0.008306723088026047, + 0.004920014645904303, + 0.0031066357623785734, + 0.004039961379021406, + 0.0032377480529248714, + 0.0032299859449267387, + 0.010242759250104427, + 0.002623665612190962, + 0.004050776828080416, + 0.0023634813260287046, + 0.002022598870098591, + 0.002400008961558342, + 0.0017469485756009817, + 0.0032085541170090437, + 0.00422128988429904, + 0.0019847857765853405, + 0.004016059450805187, + 0.0034286861773580313, + 0.008188948966562748, + 0.008060435764491558, + 0.0046045733615756035, + 0.040360935032367706, + 0.007915202528238297, + 0.004367116838693619, + 0.04319235309958458, + 0.008965957909822464, + 0.010956920683383942, + 0.007651483174413443, + 0.006215957459062338, + 0.01241535134613514, + 0.008188609965145588, + 0.007453022059053183, + 0.010105292312800884, + 0.011553174816071987, + 0.003223904874175787, + 0.004576500970870256, + 0.004846697673201561, + 0.0058271498419344425, + 0.0038650408387184143, + 0.0030895338859409094, + 0.006550594232976437, + 0.002773202722892165, + 0.0030223808716982603, + 0.0054659005254507065, + 0.003972661681473255, + 0.005602619610726833, + 0.006680817808955908, + 0.003457980463281274, + 0.0049267299473285675, + 0.004706413019448519, + 0.004564329516142607, + 0.006461633834987879, + 0.010745728388428688, + 0.012814881280064583, + 0.01190793514251709, + 0.004279097076505423, + 0.013603202998638153, + 0.0037594593595713377, + 0.002573557198047638, + 0.004930056631565094, + 0.00460920250043273, + 0.004560170695185661, + 0.005609298590570688, + 0.010623847134411335, + 0.0037318465765565634, + 0.005236232187598944, + 0.06477846205234528, + 0.06459300965070724, + 0.09485934674739838, + 0.06457993388175964, + 0.060111746191978455, + 0.05362953618168831, + 0.05666033923625946, + 0.04708889126777649 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 3, + 4 + ], + "selected_reward": -0.007631433662027121 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "stunning close-up shot, shot from an extremely low angle, almost touching the ground, focuses on a massive rocket engine nozzle. Its complex metal structure, thermal insulation tiles, and dense network of pipes create an overwhelming industrial aesthetic, with parts of the surface still blackened by soot. In an instant, a blinding white spark erupts from the nozzle's center. Immediately, a dense column of fire, composed of a blazing white core and swirling orange-yellow flames, erupts violently. This unstoppable torrent of energy rapidly expands, crashing down and completely engulfing the concrete launch platform below. The flames violently impact the rough-hewn guide ducts, sending sparks flying through the sky, while billowing white steam and gray smoke billow out from all sides, rapidly filling the frame. The dazzling light of this raw power illuminates every metallic detail of the engine, casting violently dancing shadows. The entire frame vibrates violently with the immense energy release, creating an awe-inspiring atmosphere of raw power and technological wonder.", + "seed": 27431, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000121/group_10/index_6300_seed_27431_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": -0.030025455972645432, + "vlm_logprob": -0.030025455972645432, + "vlm_yes_no_margin": 5.125, + "vlm_token_credit": -0.030025455972645432, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 massive rocket engine nozzle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle extremely low, almost touching the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the nozzle show complex metal structure, thermal insulation tiles, and soot-blackened surfaces?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a blinding white spark erupt from the nozzle center followed by a dense column of fire with a white core and orange-yellow flames?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do billowing white steam and gray smoke rapidly fill the frame while the frame vibrates violently?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01627504453063011 + ], + "opposite_logprob_tokens": [ + -4.891274929046631 + ], + "logprob_reward": -0.01627504453063011, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.01627504453063011, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5579338669776917, + 0.043399877846241, + 0.033047497272491455, + 0.020841676741838455, + 0.03911921754479408, + 0.03799824416637421, + 0.04309391230344772, + 0.061807528138160706, + 0.049137700349092484, + 0.11362054198980331 + ], + "credit_map": [ + 0.09166549146175385, + 0.014164150692522526, + 0.24480867385864258, + 0.07645826786756516, + 0.023383740335702896, + 0.06384678930044174, + 0.022743580862879753, + 0.020863126963377, + 0.004092508926987648, + 0.0035376469604671, + 0.0063427104614675045, + 0.006906058639287949, + 0.004231754224747419, + 0.004777718335390091, + 0.010202625766396523, + 0.003308858722448349, + 0.00412187073379755, + 0.004131803289055824, + 0.004448421765118837, + 0.005204205866903067, + 0.003225731197744608, + 0.0045854318886995316, + 0.004700785037130117, + 0.0026292474940419197, + 0.0018371673068031669, + 0.002232588594779372, + 0.0027173750568181276, + 0.0031280294060707092, + 0.0024875220842659473, + 0.003724333830177784, + 0.0025781416334211826, + 0.002136518945917487, + 0.003316848073154688, + 0.0047096433117985725, + 0.0041045863181352615, + 0.006785710342228413, + 0.004633745644241571, + 0.007897536270320415, + 0.004113197326660156, + 0.0035579486284404993, + 0.005316019058227539, + 0.004405312240123749, + 0.004724413156509399, + 0.00537682743743062, + 0.0030639723408967257, + 0.006003714632242918, + 0.005135898012667894, + 0.003972088452428579, + 0.0042774248868227005, + 0.005752573721110821, + 0.005681205075234175, + 0.005494000855833292, + 0.006087092682719231, + 0.0052111390978097916, + 0.006997440010309219, + 0.00359303317964077, + 0.004040812607854605, + 0.008428707718849182, + 0.013088488951325417, + 0.011941078118979931, + 0.004378883633762598, + 0.005860006436705589, + 0.0091708367690444, + 0.004898711107671261, + 0.0035857982002198696, + 0.005874757654964924, + 0.007887900806963444, + 0.007892937399446964, + 0.0055281175300478935, + 0.0054194144904613495, + 0.008311047218739986, + 0.004637726582586765, + 0.008310887031257153, + 0.01265873946249485, + 0.015264047309756279, + 0.01861126720905304, + 0.019785303622484207, + 0.013185199350118637, + 0.01786218024790287, + 0.007942908443510532 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 7 + ], + "selected_reward": -0.01627504453063011 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 massive rocket engine nozzle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle extremely low, almost touching the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the nozzle show complex metal structure, thermal insulation tiles, and soot-blackened surfaces?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a blinding white spark erupt from the nozzle center followed by a dense column of fire with a white core and orange-yellow flames?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do billowing white steam and gray smoke rapidly fill the frame while the frame vibrates violently?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1271010786294937 + ], + "opposite_logprob_tokens": [ + -2.127101182937622 + ], + "logprob_reward": -0.1271010786294937, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.1271010786294937, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17548856139183044, + 0.13504110276699066, + 0.0563247911632061, + 0.05795285478234291, + 0.09987303614616394, + 0.07288771122694016, + 0.057042196393013, + 0.06035592034459114, + 0.16991128027439117, + 0.11512253433465958 + ], + "credit_map": [ + 0.026924841105937958, + 0.009296488016843796, + 0.06346169114112854, + 0.023701339960098267, + 0.014030680060386658, + 0.020382443442940712, + 0.009334775619208813, + 0.008356294594705105, + 0.011219356209039688, + 0.012576160952448845, + 0.01700790412724018, + 0.018855571746826172, + 0.015414344146847725, + 0.013797006569802761, + 0.03577534109354019, + 0.010395425371825695, + 0.00490144919604063, + 0.004397685639560223, + 0.00862123817205429, + 0.00870763324201107, + 0.00485521974042058, + 0.013475239276885986, + 0.006868536584079266, + 0.004497789777815342, + 0.00687984237447381, + 0.006852435879409313, + 0.006916760932654142, + 0.00595775805413723, + 0.005413268692791462, + 0.011051871813833714, + 0.009093417786061764, + 0.005787503905594349, + 0.00893347803503275, + 0.013948999345302582, + 0.014783148653805256, + 0.025836853310465813, + 0.008501486852765083, + 0.01060399692505598, + 0.00902305357158184, + 0.008242017589509487, + 0.006254227366298437, + 0.008878229185938835, + 0.00996960699558258, + 0.01252133771777153, + 0.006438812706619501, + 0.010127130895853043, + 0.008001117967069149, + 0.010697243735194206, + 0.0057691242545843124, + 0.007169018033891916, + 0.00735314330086112, + 0.007721804082393646, + 0.008259441703557968, + 0.007029525004327297, + 0.008328549563884735, + 0.005411592777818441, + 0.004490363411605358, + 0.0095243314281106, + 0.010804271325469017, + 0.010097727179527283, + 0.004951623734086752, + 0.007385427597910166, + 0.007785862311720848, + 0.005316316150128841, + 0.014704314060509205, + 0.02332264930009842, + 0.03702099993824959, + 0.023296061903238297, + 0.013717645779252052, + 0.02171231433749199, + 0.02009616419672966, + 0.016041120514273643, + 0.0081802261993289, + 0.011992412619292736, + 0.014335746876895428, + 0.016916297376155853, + 0.02200663462281227, + 0.013091080822050571, + 0.018218744546175003, + 0.01038138847798109 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 1 + ], + "selected_reward": -0.1271010786294937 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 massive rocket engine nozzle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle extremely low, almost touching the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the nozzle show complex metal structure, thermal insulation tiles, and soot-blackened surfaces?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a blinding white spark erupt from the nozzle center followed by a dense column of fire with a white core and orange-yellow flames?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do billowing white steam and gray smoke rapidly fill the frame while the frame vibrates violently?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"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": [ + 0.11969070136547089, + 0.1870691329240799, + 0.06830095499753952, + 0.08989440649747849, + 0.08544111251831055, + 0.053570546209812164, + 0.09312127530574799, + 0.08965639024972916, + 0.06771786510944366, + 0.1455376297235489 + ], + "credit_map": [ + 0.01039151567965746, + 0.011485610157251358, + 0.017723307013511658, + 0.023854102939367294, + 0.012780958786606789, + 0.020838914439082146, + 0.009711411781609058, + 0.012904862873256207, + 0.0164395272731781, + 0.01730002462863922, + 0.024484381079673767, + 0.025543643161654472, + 0.022472383454442024, + 0.02051077038049698, + 0.045904986560344696, + 0.014413404278457165, + 0.004091350361704826, + 0.009219706058502197, + 0.007455969229340553, + 0.01294603943824768, + 0.0065611256286501884, + 0.011903365142643452, + 0.011169343255460262, + 0.004954046104103327, + 0.010678951628506184, + 0.009909636341035366, + 0.009161578491330147, + 0.00808935146778822, + 0.01822483167052269, + 0.015772942453622818, + 0.01024698093533516, + 0.007810128852725029, + 0.008219853974878788, + 0.010756745003163815, + 0.012091116048395634, + 0.017269492149353027, + 0.007496314123272896, + 0.011427901685237885, + 0.009981730952858925, + 0.008197952061891556, + 0.006800901610404253, + 0.006887287832796574, + 0.007724888157099485, + 0.00874693226069212, + 0.004354363773018122, + 0.00614556111395359, + 0.0068135010078549385, + 0.006097105797380209, + 0.008621522225439548, + 0.01120068784803152, + 0.015251081436872482, + 0.012505874037742615, + 0.011420561000704765, + 0.00880663376301527, + 0.013418730348348618, + 0.011896169744431973, + 0.005610907915979624, + 0.010900994762778282, + 0.022050289437174797, + 0.01972280628979206, + 0.007480846717953682, + 0.006843487732112408, + 0.009976070374250412, + 0.007070975378155708, + 0.006118879187852144, + 0.006975458934903145, + 0.011091865599155426, + 0.008974962867796421, + 0.005845323670655489, + 0.013438447378575802, + 0.008287668228149414, + 0.006985253654420376, + 0.00965784303843975, + 0.017340227961540222, + 0.020367467775940895, + 0.023720109835267067, + 0.024922331795096397, + 0.01623072661459446, + 0.022405166178941727, + 0.010893743485212326 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 9, + 0 + ], + "selected_reward": -0.0036292895674705505 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 massive rocket engine nozzle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle extremely low, almost touching the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the nozzle show complex metal structure, thermal insulation tiles, and soot-blackened surfaces?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a blinding white spark erupt from the nozzle center followed by a dense column of fire with a white core and orange-yellow flames?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do billowing white steam and gray smoke rapidly fill the frame while the frame vibrates violently?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0009218971827067435 + ], + "opposite_logprob_tokens": [ + -7.000921726226807 + ], + "logprob_reward": -0.0009218971827067435, + "yes_no_margin_reward": 7.0, + "token_credit_reward": -0.0009218971827067435, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.36957797408103943, + 0.0838451236486435, + 0.03431107476353645, + 0.03849976509809494, + 0.024486279115080833, + 0.053309109061956406, + 0.10862556099891663, + 0.09132619947195053, + 0.09418770670890808, + 0.10183116048574448 + ], + "credit_map": [ + 0.061896052211523056, + 0.011568848975002766, + 0.1457577645778656, + 0.049135565757751465, + 0.02478138357400894, + 0.04324677586555481, + 0.016418611630797386, + 0.016772989183664322, + 0.007006000727415085, + 0.0075074126943945885, + 0.01086015347391367, + 0.011147946119308472, + 0.008731108158826828, + 0.008975517936050892, + 0.023227283731102943, + 0.006389697548002005, + 0.004394769202917814, + 0.003861876903101802, + 0.0047607384622097015, + 0.005788065493106842, + 0.0036118989810347557, + 0.005152852740138769, + 0.0036457341630011797, + 0.003095141611993313, + 0.003761107102036476, + 0.004049714654684067, + 0.004558647517114878, + 0.005136952269822359, + 0.004185813944786787, + 0.007403071038424969, + 0.005971353501081467, + 0.0034331073984503746, + 0.002924140775576234, + 0.0023665884509682655, + 0.0027183908969163895, + 0.003846961073577404, + 0.0039204191416502, + 0.004339063074439764, + 0.0027231168933212757, + 0.0016475992742925882, + 0.007295457646250725, + 0.006304455921053886, + 0.00730182183906436, + 0.009527596645057201, + 0.004645275883376598, + 0.006059589330106974, + 0.007520534563809633, + 0.00465437863022089, + 0.009226524271070957, + 0.011750648729503155, + 0.011941834352910519, + 0.017542900517582893, + 0.022094417363405228, + 0.0123137217015028, + 0.017838750034570694, + 0.005916764959692955, + 0.007534143049269915, + 0.013476546853780746, + 0.013310789130628109, + 0.014858328737318516, + 0.010079426690936089, + 0.010605383664369583, + 0.013372418470680714, + 0.008089163340628147, + 0.008660891093313694, + 0.013827630318701267, + 0.016489267349243164, + 0.011960936710238457, + 0.010126599110662937, + 0.01275542937219143, + 0.012276988476514816, + 0.008089971728622913, + 0.007826360873878002, + 0.012773826718330383, + 0.008079331368207932, + 0.016240380704402924, + 0.025076689198613167, + 0.012057862244546413, + 0.014267714694142342, + 0.005508992355316877 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 9 + ], + "selected_reward": -0.0009218971827067435 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 massive rocket engine nozzle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle extremely low, almost touching the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the nozzle show complex metal structure, thermal insulation tiles, and soot-blackened surfaces?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a blinding white spark erupt from the nozzle center followed by a dense column of fire with a white core and orange-yellow flames?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do billowing white steam and gray smoke rapidly fill the frame while the frame vibrates violently?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0021999699529260397 + ], + "opposite_logprob_tokens": [ + -6.127200126647949 + ], + "logprob_reward": -0.0021999699529260397, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.0021999699529260397, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5443463921546936, + 0.047683876007795334, + 0.01267636101692915, + 0.01911981776356697, + 0.014769721776247025, + 0.025616927072405815, + 0.0398702472448349, + 0.06732650101184845, + 0.11318260431289673, + 0.11540745943784714 + ], + "credit_map": [ + 0.0896063894033432, + 0.013225642964243889, + 0.2398410588502884, + 0.0737866684794426, + 0.023193025961518288, + 0.06226653233170509, + 0.021493656560778618, + 0.020933497697114944, + 0.004087603650987148, + 0.003971266560256481, + 0.005571683868765831, + 0.006765673402696848, + 0.005563235841691494, + 0.004973548464477062, + 0.013295142911374569, + 0.003455725032836199, + 0.0009498285362496972, + 0.0010422412306070328, + 0.0016811622772365808, + 0.0015847096219658852, + 0.0010021142661571503, + 0.0035790693946182728, + 0.0018273015739396214, + 0.0010099356295540929, + 0.001585223712027073, + 0.002364012412726879, + 0.00319573818705976, + 0.0027516467962414026, + 0.0017779695335775614, + 0.0033930460922420025, + 0.002022730652242899, + 0.002029451774433255, + 0.0016551206354051828, + 0.001869910629466176, + 0.0016867053927853703, + 0.002222630428150296, + 0.0017684834310784936, + 0.0027307174168527126, + 0.0017768675461411476, + 0.0010592879261821508, + 0.0028412532992661, + 0.0030336191412061453, + 0.004166985861957073, + 0.005654988344758749, + 0.002094600349664688, + 0.0036457066889852285, + 0.002329831011593342, + 0.0018499448196962476, + 0.0031639414373785257, + 0.004369988571852446, + 0.006489422172307968, + 0.004799814894795418, + 0.004958844278007746, + 0.003956800326704979, + 0.007505740970373154, + 0.004625699017196894, + 0.005628060083836317, + 0.010767229832708836, + 0.00990984495729208, + 0.010006514377892017, + 0.006598897743970156, + 0.007446934003382921, + 0.010257684625685215, + 0.006711340043693781, + 0.00885954312980175, + 0.013344510458409786, + 0.030255122110247612, + 0.014790822751820087, + 0.011424214579164982, + 0.011058276519179344, + 0.012293335981667042, + 0.01115679182112217, + 0.008185292594134808, + 0.012418297119438648, + 0.020822230726480484, + 0.018962129950523376, + 0.015675030648708344, + 0.01245870627462864, + 0.017486736178398132, + 0.0093990508466959 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 8 + ], + "selected_reward": -0.0021999699529260397 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "stunning close-up shot, shot from an extremely low angle, almost touching the ground, focuses on a massive rocket engine nozzle. Its complex metal structure, thermal insulation tiles, and dense network of pipes create an overwhelming industrial aesthetic, with parts of the surface still blackened by soot. In an instant, a blinding white spark erupts from the nozzle's center. Immediately, a dense column of fire, composed of a blazing white core and swirling orange-yellow flames, erupts violently. This unstoppable torrent of energy rapidly expands, crashing down and completely engulfing the concrete launch platform below. The flames violently impact the rough-hewn guide ducts, sending sparks flying through the sky, while billowing white steam and gray smoke billow out from all sides, rapidly filling the frame. The dazzling light of this raw power illuminates every metallic detail of the engine, casting violently dancing shadows. The entire frame vibrates violently with the immense energy release, creating an awe-inspiring atmosphere of raw power and technological wonder.", + "seed": 27432, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000121/group_10/index_6300_seed_27432_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": -0.01966525145689957, + "vlm_logprob": -0.01966525145689957, + "vlm_yes_no_margin": 5.975, + "vlm_token_credit": -0.01966525145689957, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 massive rocket engine nozzle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle extremely low, almost touching the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the nozzle show complex metal structure, thermal insulation tiles, and soot-blackened surfaces?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a blinding white spark erupt from the nozzle center followed by a dense column of fire with a white core and orange-yellow flames?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do billowing white steam and gray smoke rapidly fill the frame while the frame vibrates violently?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006293478421866894 + ], + "opposite_logprob_tokens": [ + -6.131293296813965 + ], + "logprob_reward": -0.006293478421866894, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.006293478421866894, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06895149499177933, + 0.05932175740599632, + 0.04963715374469757, + 0.045716773718595505, + 0.0503845252096653, + 0.1600053608417511, + 0.08104079216718674, + 0.0687074214220047, + 0.21153897047042847, + 0.2046957165002823 + ], + "credit_map": [ + 0.010115601122379303, + 0.0056711179204285145, + 0.008186781778931618, + 0.006136878859251738, + 0.008141431957483292, + 0.009998541325330734, + 0.008383887819945812, + 0.01231725700199604, + 0.005399387329816818, + 0.006533467210829258, + 0.008647145703434944, + 0.006303821224719286, + 0.009716133587062359, + 0.0055847568437457085, + 0.006486308295279741, + 0.010650737211108208, + 0.004129315260797739, + 0.0037116538733243942, + 0.016341980546712875, + 0.006277629639953375, + 0.004057357553392649, + 0.00439134007319808, + 0.003637992776930332, + 0.0070898812264204025, + 0.005719298962503672, + 0.008749272674322128, + 0.0072129410691559315, + 0.005540796089917421, + 0.005745016969740391, + 0.004158602096140385, + 0.0042414539493620396, + 0.004349390510469675, + 0.005512828938663006, + 0.0050697592087090015, + 0.007402250077575445, + 0.006876186467707157, + 0.008115647360682487, + 0.006180896889418364, + 0.005692183971405029, + 0.005534774158149958, + 0.011592711322009563, + 0.025672730058431625, + 0.013091866858303547, + 0.04150009527802467, + 0.021974334493279457, + 0.009900382719933987, + 0.02310170792043209, + 0.013171529397368431, + 0.011099913157522678, + 0.011354351416230202, + 0.01023135520517826, + 0.01364162378013134, + 0.008299940265715122, + 0.00752947386354208, + 0.009714794345200062, + 0.00916934385895729, + 0.009912628680467606, + 0.007074568420648575, + 0.013364867307245731, + 0.010442662984132767, + 0.005785737652331591, + 0.00673576770350337, + 0.007736052852123976, + 0.007655136287212372, + 0.02140260860323906, + 0.08430498093366623, + 0.048459891229867935, + 0.01956919953227043, + 0.010995442979037762, + 0.010362150147557259, + 0.006675136275589466, + 0.00976954773068428, + 0.01655121147632599, + 0.013142136856913567, + 0.051339197903871536, + 0.016957558691501617, + 0.0117412731051445, + 0.017725756391882896, + 0.018213095143437386, + 0.059025492519140244 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 5 + ], + "selected_reward": -0.006293478421866894 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 massive rocket engine nozzle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle extremely low, almost touching the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the nozzle show complex metal structure, thermal insulation tiles, and soot-blackened surfaces?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a blinding white spark erupt from the nozzle center followed by a dense column of fire with a white core and orange-yellow flames?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do billowing white steam and gray smoke rapidly fill the frame while the frame vibrates violently?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08904938399791718 + ], + "opposite_logprob_tokens": [ + -2.4640493392944336 + ], + "logprob_reward": -0.08904938399791718, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08904938399791718, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0482405386865139, + 0.044618912041187286, + 0.04128298535943031, + 0.048156026750802994, + 0.08540859818458557, + 0.2892901599407196, + 0.1376054286956787, + 0.07330327481031418, + 0.12433768808841705, + 0.10775639116764069 + ], + "credit_map": [ + 0.00461594806984067, + 0.005207419395446777, + 0.006981520913541317, + 0.006791942752897739, + 0.005486614536494017, + 0.006004112306982279, + 0.0055684372782707214, + 0.007584541104733944, + 0.004321891814470291, + 0.005535488482564688, + 0.006634901277720928, + 0.00282870139926672, + 0.009034761227667332, + 0.004768102429807186, + 0.004687846172600985, + 0.00680721877142787, + 0.0048121982254087925, + 0.003219964914023876, + 0.0059668924659490585, + 0.0042235590517520905, + 0.003742966568097472, + 0.007238429039716721, + 0.004888661205768585, + 0.007190316915512085, + 0.00682465685531497, + 0.006437825970351696, + 0.007779600098729134, + 0.006729775108397007, + 0.005336528178304434, + 0.004977640230208635, + 0.004436503164470196, + 0.005633499938994646, + 0.010787226259708405, + 0.010995235294103622, + 0.010954179801046848, + 0.010100163519382477, + 0.010270504280924797, + 0.009489458054304123, + 0.009935352019965649, + 0.012876475229859352, + 0.023615039885044098, + 0.050615884363651276, + 0.02509608305990696, + 0.08194562792778015, + 0.03302853927016258, + 0.019870616495609283, + 0.02929244376718998, + 0.025825943797826767, + 0.020656168460845947, + 0.020004015415906906, + 0.016417311504483223, + 0.022713102400302887, + 0.014477400109171867, + 0.013297058641910553, + 0.012969166971743107, + 0.017071207985281944, + 0.00967799685895443, + 0.00783652625977993, + 0.018223337829113007, + 0.008172456175088882, + 0.009119098074734211, + 0.006739137694239616, + 0.00752425380051136, + 0.0060104685835540295, + 0.008365045301616192, + 0.02616235427558422, + 0.05105849355459213, + 0.009305204264819622, + 0.004171811975538731, + 0.004169961903244257, + 0.006218604743480682, + 0.01488620974123478, + 0.011692902073264122, + 0.009435418993234634, + 0.01302225049585104, + 0.015869103372097015, + 0.008252255618572235, + 0.015245407819747925, + 0.011589485220611095, + 0.02264956757426262 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 6, + 8 + ], + "selected_reward": -0.08904938399791718 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 massive rocket engine nozzle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle extremely low, almost touching the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the nozzle show complex metal structure, thermal insulation tiles, and soot-blackened surfaces?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a blinding white spark erupt from the nozzle center followed by a dense column of fire with a white core and orange-yellow flames?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do billowing white steam and gray smoke rapidly fill the frame while the frame vibrates violently?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0017204972682520747 + ], + "opposite_logprob_tokens": [ + -6.376720428466797 + ], + "logprob_reward": -0.0017204972682520747, + "yes_no_margin_reward": 6.375, + "token_credit_reward": -0.0017204972682520747, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08004169911146164, + 0.14877931773662567, + 0.03621252626180649, + 0.047002796083688736, + 0.06514675170183182, + 0.12060565501451492, + 0.09177184104919434, + 0.05794178694486618, + 0.15735793113708496, + 0.1951398253440857 + ], + "credit_map": [ + 0.006111168768256903, + 0.011378196068108082, + 0.015333118848502636, + 0.009465131908655167, + 0.008038392290472984, + 0.006266243755817413, + 0.013567888177931309, + 0.009881545789539814, + 0.016350140795111656, + 0.01789047196507454, + 0.022796211764216423, + 0.011929665692150593, + 0.019759414717555046, + 0.018751615658402443, + 0.014986908063292503, + 0.02631487138569355, + 0.004696802236139774, + 0.003288337029516697, + 0.00776261230930686, + 0.0035560934338718653, + 0.0031779143027961254, + 0.005395460873842239, + 0.0036309550050646067, + 0.00470434594899416, + 0.008320544846355915, + 0.005847894586622715, + 0.007927059195935726, + 0.004064796958118677, + 0.005943481344729662, + 0.006093244068324566, + 0.003587949089705944, + 0.005217821337282658, + 0.0068557169288396835, + 0.007553779054433107, + 0.00868072547018528, + 0.01112302578985691, + 0.0075714425183832645, + 0.00498192198574543, + 0.009917262941598892, + 0.008462871424853802, + 0.009744917973876, + 0.020543592050671577, + 0.009418954141438007, + 0.03319677338004112, + 0.011820780113339424, + 0.008757353760302067, + 0.01643061824142933, + 0.010692652314901352, + 0.008161566220223904, + 0.013788902200758457, + 0.012013043276965618, + 0.01672576367855072, + 0.011349792592227459, + 0.007857579737901688, + 0.012164304964244366, + 0.009710880927741528, + 0.007982763461768627, + 0.006207660306245089, + 0.015287429094314575, + 0.007543638814240694, + 0.0037484103813767433, + 0.006358643993735313, + 0.00604722835123539, + 0.0047660027630627155, + 0.01258273795247078, + 0.05916764214634895, + 0.04890361428260803, + 0.010110422968864441, + 0.006795130670070648, + 0.006171177141368389, + 0.004407869186252356, + 0.00921932328492403, + 0.018070463091135025, + 0.007023544982075691, + 0.04753127694129944, + 0.01354300882667303, + 0.010925373993813992, + 0.018975934013724327, + 0.018944550305604935, + 0.06012565270066261 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 1 + ], + "selected_reward": -0.0017204972682520747 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 massive rocket engine nozzle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle extremely low, almost touching the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the nozzle show complex metal structure, thermal insulation tiles, and soot-blackened surfaces?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a blinding white spark erupt from the nozzle center followed by a dense column of fire with a white core and orange-yellow flames?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do billowing white steam and gray smoke rapidly fill the frame while the frame vibrates violently?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00034231049357913435 + ], + "opposite_logprob_tokens": [ + -8.00034236907959 + ], + "logprob_reward": -0.00034231049357913435, + "yes_no_margin_reward": 8.0, + "token_credit_reward": -0.00034231049357913435, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.032010890543460846, + 0.035757876932621, + 0.029949503019452095, + 0.05249419063329697, + 0.07810131460428238, + 0.061829909682273865, + 0.1172681376338005, + 0.07955460250377655, + 0.37194621562957764, + 0.1410873830318451 + ], + "credit_map": [ + 0.0034031763207167387, + 0.003655404318124056, + 0.004369473084807396, + 0.003636443056166172, + 0.003994864411652088, + 0.0030515328980982304, + 0.006120750214904547, + 0.003779242280870676, + 0.004087358713150024, + 0.003818363882601261, + 0.004180897958576679, + 0.0033899424597620964, + 0.005019045434892178, + 0.004809814505279064, + 0.003545805811882019, + 0.006906646769493818, + 0.0033330617006868124, + 0.0029927431605756283, + 0.004189170431345701, + 0.00287896697409451, + 0.003202955238521099, + 0.005403647664934397, + 0.003158118575811386, + 0.004790839739143848, + 0.006971308495849371, + 0.0071779475547373295, + 0.010242772288620472, + 0.006884499453008175, + 0.005851742345839739, + 0.004595465492457151, + 0.006142863538116217, + 0.004627590999007225, + 0.00919136218726635, + 0.01092247199267149, + 0.014496992342174053, + 0.010116090066730976, + 0.009606316685676575, + 0.0078735938295722, + 0.006464780308306217, + 0.009429708123207092, + 0.00768033042550087, + 0.009395060129463673, + 0.005873882677406073, + 0.01180078461766243, + 0.007043689955025911, + 0.006829401012510061, + 0.00745873199775815, + 0.005748030263930559, + 0.01316610723733902, + 0.019756970927119255, + 0.010076712816953659, + 0.022117402404546738, + 0.015472550876438618, + 0.011508152820169926, + 0.016470326110720634, + 0.008699921891093254, + 0.010734779760241508, + 0.008907984010875225, + 0.021933654323220253, + 0.012183722108602524, + 0.006395548582077026, + 0.007528515066951513, + 0.007003827020525932, + 0.004866572562605143, + 0.03439643606543541, + 0.14052516222000122, + 0.10411074012517929, + 0.0281257051974535, + 0.019609160721302032, + 0.01312147919088602, + 0.011724844574928284, + 0.020332707092165947, + 0.012232075445353985, + 0.00619625486433506, + 0.03148410841822624, + 0.011973021551966667, + 0.008213644847273827, + 0.0153122553601861, + 0.014068576507270336, + 0.04160744696855545 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 6 + ], + "selected_reward": -0.00034231049357913435 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 massive rocket engine nozzle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle extremely low, almost touching the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the nozzle show complex metal structure, thermal insulation tiles, and soot-blackened surfaces?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a blinding white spark erupt from the nozzle center followed by a dense column of fire with a white core and orange-yellow flames?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do billowing white steam and gray smoke rapidly fill the frame while the frame vibrates violently?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0009205871028825641 + ], + "opposite_logprob_tokens": [ + -7.00092077255249 + ], + "logprob_reward": -0.0009205871028825641, + "yes_no_margin_reward": 7.0, + "token_credit_reward": -0.0009205871028825641, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.01961679570376873, + 0.018540969118475914, + 0.022324519231915474, + 0.027371488511562347, + 0.042974766343832016, + 0.07407853752374649, + 0.054343294352293015, + 0.02948749251663685, + 0.3823726773262024, + 0.3288894295692444 + ], + "credit_map": [ + 0.0026250542141497135, + 0.002311727497726679, + 0.003050908213481307, + 0.001717398758046329, + 0.0024376586079597473, + 0.0019920587074011564, + 0.0018680953653529286, + 0.003613896667957306, + 0.0015860553830862045, + 0.0017692864639684558, + 0.003141610650345683, + 0.0013248459435999393, + 0.0028571137227118015, + 0.004075969103723764, + 0.0018262361409142613, + 0.001959853572770953, + 0.002601433079689741, + 0.002067217603325844, + 0.0036470401100814342, + 0.0021190939005464315, + 0.0024678308982402086, + 0.0035670397337526083, + 0.002031371695920825, + 0.003823493840172887, + 0.0029697371646761894, + 0.0035369659308344126, + 0.006980676203966141, + 0.0033086102921515703, + 0.002971008652821183, + 0.002076685195788741, + 0.0028714302461594343, + 0.0026563769206404686, + 0.004435513634234667, + 0.005892605055123568, + 0.00513160414993763, + 0.005117890890687704, + 0.007220699451863766, + 0.004963588435202837, + 0.006222472991794348, + 0.003990395460277796, + 0.007477137725800276, + 0.01232737023383379, + 0.006192644126713276, + 0.01860198937356472, + 0.008605388924479485, + 0.008217763155698776, + 0.006302081048488617, + 0.00635417178273201, + 0.006081318482756615, + 0.008519900031387806, + 0.004135222639888525, + 0.00918862596154213, + 0.007543221116065979, + 0.005215896759182215, + 0.009219485335052013, + 0.0044396258890628815, + 0.0030693961307406425, + 0.0033818287774920464, + 0.0030003513675183058, + 0.005925888195633888, + 0.004242893308401108, + 0.0030563606414943933, + 0.00443625682964921, + 0.002374519594013691, + 0.028989728540182114, + 0.13806381821632385, + 0.1204860731959343, + 0.02653469890356064, + 0.018759513273835182, + 0.012235005386173725, + 0.013168899342417717, + 0.02413494512438774, + 0.029549580067396164, + 0.012653264217078686, + 0.08586238324642181, + 0.022386537864804268, + 0.01706254482269287, + 0.029893558472394943, + 0.030341217294335365, + 0.10114037245512009 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 5 + ], + "selected_reward": -0.0009205871028825641 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "stunning close-up shot, shot from an extremely low angle, almost touching the ground, focuses on a massive rocket engine nozzle. Its complex metal structure, thermal insulation tiles, and dense network of pipes create an overwhelming industrial aesthetic, with parts of the surface still blackened by soot. In an instant, a blinding white spark erupts from the nozzle's center. Immediately, a dense column of fire, composed of a blazing white core and swirling orange-yellow flames, erupts violently. This unstoppable torrent of energy rapidly expands, crashing down and completely engulfing the concrete launch platform below. The flames violently impact the rough-hewn guide ducts, sending sparks flying through the sky, while billowing white steam and gray smoke billow out from all sides, rapidly filling the frame. The dazzling light of this raw power illuminates every metallic detail of the engine, casting violently dancing shadows. The entire frame vibrates violently with the immense energy release, creating an awe-inspiring atmosphere of raw power and technological wonder.", + "seed": 27433, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000121/group_10/index_6300_seed_27433_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": -0.020421205821912737, + "vlm_logprob": -0.020421205821912737, + "vlm_yes_no_margin": 4.9, + "vlm_token_credit": -0.020421205821912737, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 massive rocket engine nozzle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle extremely low, almost touching the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the nozzle show complex metal structure, thermal insulation tiles, and soot-blackened surfaces?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a blinding white spark erupt from the nozzle center followed by a dense column of fire with a white core and orange-yellow flames?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do billowing white steam and gray smoke rapidly fill the frame while the frame vibrates violently?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.015749940648674965 + ], + "opposite_logprob_tokens": [ + -4.640749931335449 + ], + "logprob_reward": -0.015749940648674965, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.015749940648674965, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18887166678905487, + 0.04766378551721573, + 0.042039498686790466, + 0.04489678144454956, + 0.07291805744171143, + 0.055115364491939545, + 0.11667027324438095, + 0.23590202629566193, + 0.11452735215425491, + 0.08139526844024658 + ], + "credit_map": [ + 0.0237136073410511, + 0.024929547682404518, + 0.06485044956207275, + 0.01673693023622036, + 0.011892536655068398, + 0.014603632502257824, + 0.014911100268363953, + 0.01723385788500309, + 0.0036163674667477608, + 0.008179198950529099, + 0.015583982691168785, + 0.004499519243836403, + 0.0036077958066016436, + 0.0029814518056809902, + 0.0033763120882213116, + 0.005819159559905529, + 0.005142747424542904, + 0.005658166017383337, + 0.006674600299447775, + 0.00503517035394907, + 0.005830526351928711, + 0.00387825770303607, + 0.00553162069991231, + 0.004288407973945141, + 0.009242217056453228, + 0.006500982213765383, + 0.009125783108174801, + 0.006258457433432341, + 0.003878518706187606, + 0.0027153000701218843, + 0.003775278339162469, + 0.0034002435859292746, + 0.012420624494552612, + 0.009566373191773891, + 0.010897465981543064, + 0.008863345719873905, + 0.009309908375144005, + 0.005050978157669306, + 0.010992968454957008, + 0.0058163851499557495, + 0.017395664006471634, + 0.017261581495404243, + 0.004393960814923048, + 0.004307473078370094, + 0.0036755679175257683, + 0.002619189443066716, + 0.002783464267849922, + 0.0026784625370055437, + 0.015708325430750847, + 0.018626535311341286, + 0.013080825097858906, + 0.016842644661664963, + 0.020715387538075447, + 0.008460699580609798, + 0.014462688937783241, + 0.008773165754973888, + 0.05838119611144066, + 0.032399971038103104, + 0.012567932717502117, + 0.0734044760465622, + 0.02122146263718605, + 0.009620481170713902, + 0.017132170498371124, + 0.011174308136105537, + 0.01863822154700756, + 0.014645657502114773, + 0.014132682234048843, + 0.012237867340445518, + 0.015902630984783173, + 0.015281316824257374, + 0.012374688871204853, + 0.01131428312510252, + 0.015497763641178608, + 0.010342951864004135, + 0.01209593378007412, + 0.0105040417984128, + 0.009260783903300762, + 0.006859372369945049, + 0.011107562109827995, + 0.0057268510572612286 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 6 + ], + "selected_reward": -0.015749940648674965 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 massive rocket engine nozzle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle extremely low, almost touching the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the nozzle show complex metal structure, thermal insulation tiles, and soot-blackened surfaces?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a blinding white spark erupt from the nozzle center followed by a dense column of fire with a white core and orange-yellow flames?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do billowing white steam and gray smoke rapidly fill the frame while the frame vibrates violently?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07005105167627335 + ], + "opposite_logprob_tokens": [ + -2.6950509548187256 + ], + "logprob_reward": -0.07005105167627335, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07005105167627335, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1757049411535263, + 0.02493933215737343, + 0.033939678221940994, + 0.02586565539240837, + 0.06012952700257301, + 0.05053359270095825, + 0.05663921684026718, + 0.24075648188591003, + 0.2852380573749542, + 0.04625355079770088 + ], + "credit_map": [ + 0.01875854656100273, + 0.012518404051661491, + 0.07045171409845352, + 0.009030952118337154, + 0.008597742766141891, + 0.011647441424429417, + 0.009560816921293736, + 0.035139329731464386, + 0.002458943286910653, + 0.0035124393180012703, + 0.004724747501313686, + 0.003445717738941312, + 0.0020535290241241455, + 0.0028175467159599066, + 0.0029251547530293465, + 0.0030012535862624645, + 0.0035864608362317085, + 0.00317439716309309, + 0.007805144414305687, + 0.00565737672150135, + 0.0029195041861385107, + 0.0035990597680211067, + 0.003562576835975051, + 0.0036351580638438463, + 0.004292057361453772, + 0.0035708737559616566, + 0.004710538778454065, + 0.004244640469551086, + 0.0020502200350165367, + 0.0015988333616405725, + 0.0027816256042569876, + 0.002616866957396269, + 0.009115466848015785, + 0.00788096059113741, + 0.007938834838569164, + 0.006735065020620823, + 0.0078121572732925415, + 0.006724763195961714, + 0.008191725239157677, + 0.005730559583753347, + 0.014457346871495247, + 0.014874878339469433, + 0.0032372423447668552, + 0.004222171846777201, + 0.0038863103836774826, + 0.0035219385754317045, + 0.002941506914794445, + 0.0033921971917152405, + 0.006399185862392187, + 0.00732463039457798, + 0.007843277417123318, + 0.006890113465487957, + 0.006800183095037937, + 0.0055519333109259605, + 0.009401265531778336, + 0.0064286282286047935, + 0.06613676995038986, + 0.025796595960855484, + 0.014825181104242802, + 0.07810790091753006, + 0.016772039234638214, + 0.008330337703227997, + 0.01903478056192398, + 0.011752871796488762, + 0.03704310208559036, + 0.03966110199689865, + 0.026747198775410652, + 0.04067331925034523, + 0.04424304887652397, + 0.032670002430677414, + 0.03515202924609184, + 0.029048249125480652, + 0.005829682573676109, + 0.005183584988117218, + 0.008129225112497807, + 0.005937718320637941, + 0.006775433197617531, + 0.003654721425846219, + 0.006571044214069843, + 0.004172143060714006 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 0 + ], + "selected_reward": -0.07005105167627335 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 massive rocket engine nozzle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle extremely low, almost touching the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the nozzle show complex metal structure, thermal insulation tiles, and soot-blackened surfaces?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a blinding white spark erupt from the nozzle center followed by a dense column of fire with a white core and orange-yellow flames?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do billowing white steam and gray smoke rapidly fill the frame while the frame vibrates violently?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012559957802295685 + ], + "opposite_logprob_tokens": [ + -4.38755989074707 + ], + "logprob_reward": -0.012559957802295685, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012559957802295685, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2501455247402191, + 0.07865239679813385, + 0.07327239960432053, + 0.058976415544748306, + 0.07856494933366776, + 0.08814967423677444, + 0.06993624567985535, + 0.09794994443655014, + 0.16015364229679108, + 0.04419874772429466 + ], + "credit_map": [ + 0.01586524024605751, + 0.013546750880777836, + 0.09101510792970657, + 0.01356945838779211, + 0.01458562072366476, + 0.016829922795295715, + 0.015432514250278473, + 0.06930093467235565, + 0.008859432302415371, + 0.011957108043134212, + 0.021023599430918694, + 0.008293213322758675, + 0.0061723943799734116, + 0.00618479959666729, + 0.0062920087948441505, + 0.009869851171970367, + 0.008550646714866161, + 0.011263785883784294, + 0.013433817774057388, + 0.007913599722087383, + 0.009162282571196556, + 0.006601281464099884, + 0.009768711403012276, + 0.006578278262168169, + 0.009389630518853664, + 0.0073877968825399876, + 0.006043399218469858, + 0.007747096940875053, + 0.006515112705528736, + 0.006285184063017368, + 0.009634268470108509, + 0.005973926279693842, + 0.012374171987175941, + 0.009958217851817608, + 0.013017084449529648, + 0.00903749093413353, + 0.010371621698141098, + 0.0056480951607227325, + 0.011398199014365673, + 0.006760069169104099, + 0.02904518134891987, + 0.030034305527806282, + 0.005298001691699028, + 0.005544428247958422, + 0.005268985405564308, + 0.004276862367987633, + 0.004023630637675524, + 0.004658292047679424, + 0.009064257144927979, + 0.014024443924427032, + 0.008035135455429554, + 0.007793982047587633, + 0.011463211849331856, + 0.006248641759157181, + 0.008298725821077824, + 0.005007859785109758, + 0.01634906604886055, + 0.01696258783340454, + 0.0077322316356003284, + 0.021135782822966576, + 0.012013721279799938, + 0.005888638086616993, + 0.010652226395905018, + 0.007215697783976793, + 0.025814393535256386, + 0.027673885226249695, + 0.01849580928683281, + 0.021990180015563965, + 0.022565603256225586, + 0.013623594306409359, + 0.017033251002430916, + 0.012956949882209301, + 0.005765043664723635, + 0.004417837131768465, + 0.007412561681121588, + 0.005468329414725304, + 0.006456652656197548, + 0.0041130827739834785, + 0.006308270152658224, + 0.004256972577422857 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 7 + ], + "selected_reward": -0.012559957802295685 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 massive rocket engine nozzle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle extremely low, almost touching the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the nozzle show complex metal structure, thermal insulation tiles, and soot-blackened surfaces?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a blinding white spark erupt from the nozzle center followed by a dense column of fire with a white core and orange-yellow flames?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do billowing white steam and gray smoke rapidly fill the frame while the frame vibrates violently?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0009235645993612707 + ], + "opposite_logprob_tokens": [ + -7.0009236335754395 + ], + "logprob_reward": -0.0009235645993612707, + "yes_no_margin_reward": 7.0, + "token_credit_reward": -0.0009235645993612707, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24473083019256592, + 0.0557747408747673, + 0.06180292367935181, + 0.040239252150058746, + 0.08499757945537567, + 0.047111280262470245, + 0.07416833192110062, + 0.07687666267156601, + 0.2563295364379883, + 0.057968951761722565 + ], + "credit_map": [ + 0.023982426151633263, + 0.02178020216524601, + 0.09995106607675552, + 0.013572916388511658, + 0.009827559813857079, + 0.016051456332206726, + 0.013652374036610126, + 0.045912809669971466, + 0.005463744048029184, + 0.01016913540661335, + 0.008631167002022266, + 0.006014237180352211, + 0.005841303151100874, + 0.006878795567899942, + 0.006766460835933685, + 0.006009899079799652, + 0.012132023461163044, + 0.008196516893804073, + 0.013542799279093742, + 0.008494221605360508, + 0.004768952261656523, + 0.0052546109072864056, + 0.004567071795463562, + 0.0048467242158949375, + 0.006547356955707073, + 0.005090043880045414, + 0.008865292184054852, + 0.005281130317598581, + 0.00391283817589283, + 0.0026377462781965733, + 0.0046391962096095085, + 0.003265649313107133, + 0.012872321531176567, + 0.011366762220859528, + 0.016538549214601517, + 0.009211033582687378, + 0.010570003651082516, + 0.006108883768320084, + 0.01223374716937542, + 0.006096282508224249, + 0.006749541498720646, + 0.005708838813006878, + 0.006485917139798403, + 0.007914009504020214, + 0.006544554140418768, + 0.004698478151112795, + 0.005059777293354273, + 0.003950163722038269, + 0.01003626175224781, + 0.013094210997223854, + 0.010937750339508057, + 0.008819812908768654, + 0.008823602460324764, + 0.007189309224486351, + 0.00935985054820776, + 0.005907529499381781, + 0.00814516469836235, + 0.016461791470646858, + 0.006085934583097696, + 0.010110455565154552, + 0.01576155424118042, + 0.004074095748364925, + 0.010728824883699417, + 0.005508839152753353, + 0.032819461077451706, + 0.03682611137628555, + 0.025519495829939842, + 0.04120798036456108, + 0.04793842136859894, + 0.021874187514185905, + 0.029294107109308243, + 0.020849768072366714, + 0.010240674950182438, + 0.0067607080563902855, + 0.009221664629876614, + 0.0067514763213694096, + 0.006747748702764511, + 0.00517632719129324, + 0.007519159931689501, + 0.005551189184188843 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 4 + ], + "selected_reward": -0.0009235645993612707 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 massive rocket engine nozzle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle extremely low, almost touching the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the nozzle show complex metal structure, thermal insulation tiles, and soot-blackened surfaces?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a blinding white spark erupt from the nozzle center followed by a dense column of fire with a white core and orange-yellow flames?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do billowing white steam and gray smoke rapidly fill the frame while the frame vibrates violently?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"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": [ + 0.24270878732204437, + 0.017716573551297188, + 0.02985197678208351, + 0.03886767476797104, + 0.06563779711723328, + 0.025856688618659973, + 0.196865051984787, + 0.12833458185195923, + 0.11265212297439575, + 0.14150874316692352 + ], + "credit_map": [ + 0.021332435309886932, + 0.018851103261113167, + 0.0980064868927002, + 0.016161371022462845, + 0.007848409004509449, + 0.013758472166955471, + 0.012808728031814098, + 0.053941793739795685, + 0.0019701234996318817, + 0.0025748766493052244, + 0.002812828402966261, + 0.001898651011288166, + 0.0014261285541579127, + 0.0020543858408927917, + 0.0026345672085881233, + 0.0023450120352208614, + 0.005645233206450939, + 0.004097366705536842, + 0.005602761637419462, + 0.0032615785021334887, + 0.0036272669676691294, + 0.0019715342205017805, + 0.002911393763497472, + 0.0027348422445356846, + 0.00663489755243063, + 0.005156444851309061, + 0.007141816895455122, + 0.0052575902082026005, + 0.004207043908536434, + 0.003284567967057228, + 0.004359041806310415, + 0.0028262725099921227, + 0.01004042848944664, + 0.008425899781286716, + 0.011038080789148808, + 0.00755572784692049, + 0.008129049092531204, + 0.004712095018476248, + 0.009712986648082733, + 0.00602352898567915, + 0.005069419741630554, + 0.004351369570940733, + 0.0034129091072827578, + 0.0032474712934345007, + 0.003187262685969472, + 0.0020233604591339827, + 0.0025225470308214426, + 0.0020423501264303923, + 0.03169798105955124, + 0.050409696996212006, + 0.020037706941366196, + 0.017520373687148094, + 0.026321960613131523, + 0.015514271333813667, + 0.019006546586751938, + 0.016356531530618668, + 0.03142699971795082, + 0.016225541010499, + 0.00824308767914772, + 0.03503121808171272, + 0.010132129304111004, + 0.005429966375231743, + 0.012113978154957294, + 0.009731653146445751, + 0.014704965986311436, + 0.013268495909869671, + 0.010351925157010555, + 0.015668466687202454, + 0.018181417137384415, + 0.011772263795137405, + 0.011713635176420212, + 0.016990959644317627, + 0.022099116817116737, + 0.016565466299653053, + 0.02368331328034401, + 0.017048636451363564, + 0.020662039518356323, + 0.012672554701566696, + 0.017559703439474106, + 0.011217908933758736 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 9 + ], + "selected_reward": -0.002821514382958412 + } + ], + "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/0000121/group_11/index_tensor([1215, 1215])_group_11_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000121/group_11/index_tensor([1215, 1215])_group_11_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..5fe3b4469cca6609aa632f1a56424f385a27bca6 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000121/group_11/index_tensor([1215, 1215])_group_11_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "low-angle, parallel tracking shot follows a historic destroyer, a steel behemoth slicing through the raging, almost black, deep blue waters with unstoppable force. Its weathered, dark gray hull is mottled with rust and crisscrossed with scratches. Its towering bridge and silent turrets appear majestic against the gloomy sky, the slowly rotating radar antenna on the bridge the only steady movement in the frame. Mountainous waves crash violently against the ship's side, then burst into a thousand crystalline droplets, each spray carving a brief, powerful arc through the air. Thick, dark clouds threaten to crush the sea, but a solemn, stage-like Jesus light pierces the clouds, casting a precise, dramatic glow on the waters beneath the ship's path, illuminating the swirling white foam and drifting mist with a tense, dramatic sheen, profoundly carving out the ship's hard metal contours and damp, reflective texture. The steady and leisurely camera movement combined with this powerful scene creates an unparalleled sense of epic heaviness in the cold-toned picture, as if telling a grand narrative about steel, ocean and indomitable will.", + "seed": 22724, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000121/group_11/index_1215_seed_22724_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": -0.06373293586075306, + "vlm_logprob": -0.06373293586075306, + "vlm_yes_no_margin": 3.125, + "vlm_token_credit": -0.06373293586075306, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a historic destroyer with a weathered, dark gray hull showing rust and scratches?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the ship moving through deep blue, almost black waters with mountainous waves crashing violently against its side?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene illuminated by a solemn, stage-like Jesus light piercing through thick, dark clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera executing a steady, low-angle parallel tracking shot following the ship?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the radar antenna on the ship's bridge shown as the only steady movement in 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": [ + -0.06347551941871643 + ], + "opposite_logprob_tokens": [ + -2.9384756088256836 + ], + "logprob_reward": -0.06347551941871643, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.06347551941871643, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.35961052775382996, + 0.1273566037416458, + 0.1534857153892517, + 0.07013802975416183, + 0.040322739630937576, + 0.13741301000118256, + 0.02130473591387272, + 0.04771914705634117, + 0.018493488430976868, + 0.02415599301457405 + ], + "credit_map": [ + 0.05597689375281334, + 0.028953436762094498, + 0.045567587018013, + 0.0375669039785862, + 0.04489514231681824, + 0.04567471891641617, + 0.04957778379321098, + 0.051398079842329025, + 0.018178008496761322, + 0.012362929992377758, + 0.017931701615452766, + 0.02167888730764389, + 0.01606190949678421, + 0.013450910337269306, + 0.013795428909361362, + 0.013896814547479153, + 0.0141317592933774, + 0.01829688996076584, + 0.011329143308103085, + 0.01834181323647499, + 0.04197553172707558, + 0.02478804811835289, + 0.01353664230555296, + 0.01108588557690382, + 0.00888022594153881, + 0.01316271536052227, + 0.009336913004517555, + 0.009299620054662228, + 0.008008943870663643, + 0.007672556675970554, + 0.006418123841285706, + 0.007358933798968792, + 0.0065264576114714146, + 0.003663754789158702, + 0.003938221372663975, + 0.0037879906594753265, + 0.004966362379491329, + 0.009603583253920078, + 0.0032129872124642134, + 0.004623384680598974, + 0.01570294424891472, + 0.010917911306023598, + 0.019348740577697754, + 0.01594802737236023, + 0.014418745413422585, + 0.03058173693716526, + 0.014454827643930912, + 0.016040071845054626, + 0.002937283366918564, + 0.0026754371356219053, + 0.002046209294348955, + 0.003201911458745599, + 0.0030145079363137484, + 0.0025176142808049917, + 0.0023878533393144608, + 0.002523917704820633, + 0.008718333207070827, + 0.007500912062823772, + 0.003351607359945774, + 0.003939582034945488, + 0.0069923303090035915, + 0.008828310295939445, + 0.0030307958368211985, + 0.005357278510928154, + 0.002646220149472356, + 0.00284769874997437, + 0.0021847616881132126, + 0.002143959514796734, + 0.0016875275177881122, + 0.002184974728152156, + 0.0028623417019844055, + 0.0019360044971108437, + 0.00304085249081254, + 0.0026888318825513124, + 0.0023615604732185602, + 0.003832621034234762, + 0.0027028268668800592, + 0.003095153719186783, + 0.0034381875302642584, + 0.002995958784595132 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 5 + ], + "selected_reward": -0.06347551941871643 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a historic destroyer with a weathered, dark gray hull showing rust and scratches?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the ship moving through deep blue, almost black waters with mountainous waves crashing violently against its side?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene illuminated by a solemn, stage-like Jesus light piercing through thick, dark clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera executing a steady, low-angle parallel tracking shot following the ship?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the radar antenna on the ship's bridge shown as the only steady movement in 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": [ + -0.011092543601989746 + ], + "opposite_logprob_tokens": [ + -4.511092662811279 + ], + "logprob_reward": -0.011092543601989746, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011092543601989746, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14163349568843842, + 0.11691813915967941, + 0.26481202244758606, + 0.11090116202831268, + 0.030917279422283173, + 0.12112148106098175, + 0.07129258662462234, + 0.08227162063121796, + 0.01619863696396351, + 0.043933555483818054 + ], + "credit_map": [ + 0.016145437955856323, + 0.011197011917829514, + 0.013978200033307076, + 0.01699279434978962, + 0.015504557639360428, + 0.025224540382623672, + 0.0174478217959404, + 0.025143150240182877, + 0.01158641092479229, + 0.012957831844687462, + 0.015967408195137978, + 0.020620064809918404, + 0.013496494852006435, + 0.014698173850774765, + 0.01581990160048008, + 0.011771856807172298, + 0.021398594602942467, + 0.03998110815882683, + 0.0178673192858696, + 0.027045782655477524, + 0.0813676118850708, + 0.03812020644545555, + 0.02247978188097477, + 0.016551632434129715, + 0.013318751938641071, + 0.01276661828160286, + 0.014672095887362957, + 0.011528277769684792, + 0.013835215009748936, + 0.02313738316297531, + 0.011065570637583733, + 0.01057726051658392, + 0.003955469932407141, + 0.0036578665021806955, + 0.003599312389269471, + 0.002833126112818718, + 0.005075715947896242, + 0.0033759253565222025, + 0.0033895676024258137, + 0.005030300002545118, + 0.01292923279106617, + 0.014314710162580013, + 0.02115955948829651, + 0.02454664744436741, + 0.010719460435211658, + 0.016581179574131966, + 0.010231881402432919, + 0.010638820007443428, + 0.011738178320229053, + 0.010048211552202702, + 0.007869122549891472, + 0.005305935628712177, + 0.009238187223672867, + 0.01071647647768259, + 0.007443520240485668, + 0.008932963944971561, + 0.01562754437327385, + 0.01066578458994627, + 0.005846591666340828, + 0.006408786866813898, + 0.011563467793166637, + 0.01618076302111149, + 0.005012457258999348, + 0.010966235771775246, + 0.002084138337522745, + 0.001663357368670404, + 0.002128905849531293, + 0.002068036003038287, + 0.0019298875704407692, + 0.0018857012037187815, + 0.0016026573721319437, + 0.0028359531424939632, + 0.005715684965252876, + 0.005825860425829887, + 0.005177214276045561, + 0.005976071115583181, + 0.005142487585544586, + 0.005375652574002743, + 0.005886751227080822, + 0.004833841230720282 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 5 + ], + "selected_reward": -0.011092543601989746 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a historic destroyer with a weathered, dark gray hull showing rust and scratches?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the ship moving through deep blue, almost black waters with mountainous waves crashing violently against its side?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene illuminated by a solemn, stage-like Jesus light piercing through thick, dark clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera executing a steady, low-angle parallel tracking shot following the ship?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the radar antenna on the ship's bridge shown as the only steady movement in 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": [ + -0.018195616081357002 + ], + "opposite_logprob_tokens": [ + -4.018195629119873 + ], + "logprob_reward": -0.018195616081357002, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018195616081357002, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.078069768846035, + 0.0332353413105011, + 0.26276519894599915, + 0.06498149782419205, + 0.3686668872833252, + 0.05463802069425583, + 0.0193861722946167, + 0.023419609293341637, + 0.04169902205467224, + 0.053138524293899536 + ], + "credit_map": [ + 0.011063673533499241, + 0.0053049540147185326, + 0.009804870933294296, + 0.007541558239609003, + 0.010650020092725754, + 0.011713081039488316, + 0.008993390947580338, + 0.012998221442103386, + 0.00557759590446949, + 0.004374263808131218, + 0.0032654437236487865, + 0.003926202654838562, + 0.005400901194661856, + 0.003730434225872159, + 0.0037896777503192425, + 0.0031708243768662214, + 0.02137007564306259, + 0.04054773598909378, + 0.01905539631843567, + 0.026896517723798752, + 0.08122918009757996, + 0.03762397542595863, + 0.021225756034255028, + 0.01481654867529869, + 0.00858564767986536, + 0.008022534660995007, + 0.007177081890404224, + 0.010324802249670029, + 0.007714591454714537, + 0.009689807891845703, + 0.007220472674816847, + 0.006246557924896479, + 0.07700416445732117, + 0.043970219790935516, + 0.03327906131744385, + 0.016872264444828033, + 0.036199573427438736, + 0.12179995328187943, + 0.015913881361484528, + 0.023627765476703644, + 0.005793902091681957, + 0.006459169089794159, + 0.00897328183054924, + 0.009791607968509197, + 0.005073712673038244, + 0.007680121809244156, + 0.005011065397411585, + 0.005855162162333727, + 0.003944270312786102, + 0.0024292191956192255, + 0.002374792005866766, + 0.0012385507579892874, + 0.00162687455303967, + 0.0027994627598673105, + 0.002277613617479801, + 0.0026953890919685364, + 0.004317550454288721, + 0.004055605735629797, + 0.002389113884419203, + 0.0027135706041008234, + 0.0032498554792255163, + 0.0024094795808196068, + 0.0019411664688959718, + 0.002343268832191825, + 0.007875115610659122, + 0.008721502497792244, + 0.0035451173316687346, + 0.003816759679466486, + 0.0071070436388254166, + 0.005293855909258127, + 0.002343451138585806, + 0.002996176015585661, + 0.010624888353049755, + 0.008556174114346504, + 0.004079817328602076, + 0.0047538382932543755, + 0.005958858877420425, + 0.010956956073641777, + 0.0041053867898881435, + 0.004102607723325491 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 2, + 0 + ], + "selected_reward": -0.018195616081357002 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a historic destroyer with a weathered, dark gray hull showing rust and scratches?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the ship moving through deep blue, almost black waters with mountainous waves crashing violently against its side?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene illuminated by a solemn, stage-like Jesus light piercing through thick, dark clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera executing a steady, low-angle parallel tracking shot following the ship?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the radar antenna on the ship's bridge shown as the only steady movement in 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": [ + -0.11293245106935501 + ], + "opposite_logprob_tokens": [ + -2.2379324436187744 + ], + "logprob_reward": -0.11293245106935501, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11293245106935501, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10002080351114273, + 0.06210293248295784, + 0.49618303775787354, + 0.04579539969563484, + 0.060470134019851685, + 0.08788996189832687, + 0.046317391097545624, + 0.07169166952371597, + 0.011189276352524757, + 0.01833944208920002 + ], + "credit_map": [ + 0.01240441296249628, + 0.006355461664497852, + 0.025419654324650764, + 0.01672697626054287, + 0.008684465661644936, + 0.00964045338332653, + 0.01293228194117546, + 0.00785708986222744, + 0.008992807939648628, + 0.008541908115148544, + 0.008374606259167194, + 0.008799836039543152, + 0.007517102174460888, + 0.006684752646833658, + 0.00641719251871109, + 0.006774732377380133, + 0.03908398374915123, + 0.07628990709781647, + 0.03436199948191643, + 0.049548324197530746, + 0.15407010912895203, + 0.07105588912963867, + 0.04159502312541008, + 0.03017779253423214, + 0.0072212074883282185, + 0.0061094374395906925, + 0.006067298352718353, + 0.0041854954324662685, + 0.008280741050839424, + 0.00519714318215847, + 0.0038891497533768415, + 0.004844929091632366, + 0.010949143208563328, + 0.00743749737739563, + 0.0053031956776976585, + 0.003419078653678298, + 0.0073061417788267136, + 0.01789209060370922, + 0.0033355485647916794, + 0.004827439319342375, + 0.008307025767862797, + 0.009538194164633751, + 0.014984361827373505, + 0.017314311116933823, + 0.010499211959540844, + 0.0120846601203084, + 0.008869530633091927, + 0.00629266444593668, + 0.008112246170639992, + 0.008626646362245083, + 0.005582411307841539, + 0.004019860178232193, + 0.007741614710539579, + 0.004422908183187246, + 0.004499075002968311, + 0.0033126291818916798, + 0.012697093188762665, + 0.010763706639409065, + 0.00540776876732707, + 0.006330171599984169, + 0.009674428962171078, + 0.012652778998017311, + 0.005649823695421219, + 0.008515900000929832, + 0.002069807145744562, + 0.0013109487481415272, + 0.0011814795434474945, + 0.0009037120034918189, + 0.001491329981945455, + 0.0019614878110587597, + 0.0012137498706579208, + 0.0010567604331299663, + 0.0030390634201467037, + 0.003484999993816018, + 0.0014958131359890103, + 0.0023033791221678257, + 0.0021920008584856987, + 0.0021453923545777798, + 0.0016644611023366451, + 0.0020143319852650166 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 5 + ], + "selected_reward": -0.11293245106935501 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a historic destroyer with a weathered, dark gray hull showing rust and scratches?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the ship moving through deep blue, almost black waters with mountainous waves crashing violently against its side?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene illuminated by a solemn, stage-like Jesus light piercing through thick, dark clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera executing a steady, low-angle parallel tracking shot following the ship?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the radar antenna on the ship's bridge shown as the only steady movement in 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": [ + -0.1129685491323471 + ], + "opposite_logprob_tokens": [ + -2.2379684448242188 + ], + "logprob_reward": -0.1129685491323471, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.1129685491323471, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.201214000582695, + 0.131547749042511, + 0.27162277698516846, + 0.06226878985762596, + 0.12967592477798462, + 0.06103935465216637, + 0.03323834016919136, + 0.04981401190161705, + 0.02438356727361679, + 0.03519543632864952 + ], + "credit_map": [ + 0.027956847101449966, + 0.01827836222946644, + 0.04126432165503502, + 0.0323486365377903, + 0.01960344985127449, + 0.020454445853829384, + 0.026487072929739952, + 0.014820878393948078, + 0.01687055639922619, + 0.015110432170331478, + 0.020261896774172783, + 0.022814027965068817, + 0.014687802642583847, + 0.015141073614358902, + 0.015302575193345547, + 0.01135937962681055, + 0.02202192321419716, + 0.04320836439728737, + 0.021224332973361015, + 0.02627517469227314, + 0.08218485116958618, + 0.0376436822116375, + 0.023121435195207596, + 0.015943000093102455, + 0.007210835814476013, + 0.008475813083350658, + 0.007277741562575102, + 0.006552709266543388, + 0.010838273912668228, + 0.008181113749742508, + 0.007848214358091354, + 0.005884089507162571, + 0.02636881172657013, + 0.015232568606734276, + 0.010977812111377716, + 0.006801785435527563, + 0.012141933664679527, + 0.04121549427509308, + 0.006843710783869028, + 0.010093798860907555, + 0.007677864748984575, + 0.0054852343164384365, + 0.011042293161153793, + 0.005507828202098608, + 0.00533259054645896, + 0.014583430252969265, + 0.005545796360820532, + 0.005864315666258335, + 0.004867304116487503, + 0.004558070562779903, + 0.003352588042616844, + 0.0030910326167941093, + 0.005588062107563019, + 0.004205847159028053, + 0.004138875752687454, + 0.0034365623723715544, + 0.009581505320966244, + 0.005948317237198353, + 0.0028456614818423986, + 0.0031425426714122295, + 0.00729437405243516, + 0.010695138014853, + 0.003215125063434243, + 0.007091346196830273, + 0.004045811481773853, + 0.0035504153929650784, + 0.0015458459965884686, + 0.002362769329920411, + 0.0031789804343134165, + 0.004559924826025963, + 0.0029082242399454117, + 0.002231596503406763, + 0.006758227478712797, + 0.006127773784101009, + 0.002317743841558695, + 0.004120729397982359, + 0.003919908776879311, + 0.0051880814135074615, + 0.0028982425574213266, + 0.0038647293113172054 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 1 + ], + "selected_reward": -0.1129685491323471 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle, parallel tracking shot follows a historic destroyer, a steel behemoth slicing through the raging, almost black, deep blue waters with unstoppable force. Its weathered, dark gray hull is mottled with rust and crisscrossed with scratches. Its towering bridge and silent turrets appear majestic against the gloomy sky, the slowly rotating radar antenna on the bridge the only steady movement in the frame. Mountainous waves crash violently against the ship's side, then burst into a thousand crystalline droplets, each spray carving a brief, powerful arc through the air. Thick, dark clouds threaten to crush the sea, but a solemn, stage-like Jesus light pierces the clouds, casting a precise, dramatic glow on the waters beneath the ship's path, illuminating the swirling white foam and drifting mist with a tense, dramatic sheen, profoundly carving out the ship's hard metal contours and damp, reflective texture. The steady and leisurely camera movement combined with this powerful scene creates an unparalleled sense of epic heaviness in the cold-toned picture, as if telling a grand narrative about steel, ocean and indomitable will.", + "seed": 22725, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000121/group_11/index_1215_seed_22725_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": -0.06826418619602918, + "vlm_logprob": -0.06826418619602918, + "vlm_yes_no_margin": 3.025, + "vlm_token_credit": -0.06826418619602918, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a historic destroyer with a weathered, dark gray hull showing rust and scratches?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the ship moving through deep blue, almost black waters with mountainous waves crashing violently against its side?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene illuminated by a solemn, stage-like Jesus light piercing through thick, dark clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera executing a steady, low-angle parallel tracking shot following the ship?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the radar antenna on the ship's bridge shown as the only steady movement in 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": [ + -0.08639468252658844 + ], + "opposite_logprob_tokens": [ + -2.586394786834717 + ], + "logprob_reward": -0.08639468252658844, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.08639468252658844, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21681863069534302, + 0.05190165340900421, + 0.05320585146546364, + 0.03187573701143265, + 0.0842190831899643, + 0.04893922060728073, + 0.07063400745391846, + 0.059405624866485596, + 0.30593496561050415, + 0.07706515491008759 + ], + "credit_map": [ + 0.02051791548728943, + 0.022340044379234314, + 0.01792943850159645, + 0.021675636991858482, + 0.043021414428949356, + 0.02724541909992695, + 0.03500517085194588, + 0.02908359281718731, + 0.008111072704195976, + 0.005402631126344204, + 0.004405233543366194, + 0.005258817691355944, + 0.008167502470314503, + 0.007105017546564341, + 0.007110488601028919, + 0.006340888794511557, + 0.0042646839283406734, + 0.004320655949413776, + 0.003817369230091572, + 0.0054685198701918125, + 0.009636075235903263, + 0.0062510548159480095, + 0.012920227833092213, + 0.00652726786211133, + 0.005310384090989828, + 0.003226369619369507, + 0.0028392011299729347, + 0.0031689598690718412, + 0.0042858789674937725, + 0.002820312976837158, + 0.005664096213877201, + 0.004560535773634911, + 0.011086443439126015, + 0.007325204089283943, + 0.007137436419725418, + 0.0057584308087825775, + 0.01259192731231451, + 0.008109922520816326, + 0.015489968471229076, + 0.016719752922654152, + 0.003922218456864357, + 0.0039140041917562485, + 0.003415756393224001, + 0.004129666835069656, + 0.011556224897503853, + 0.005812832619994879, + 0.007977875880897045, + 0.008210642263293266, + 0.008470545522868633, + 0.008606921881437302, + 0.004196748603135347, + 0.004591255448758602, + 0.014399166218936443, + 0.009012904018163681, + 0.010120182298123837, + 0.01123628206551075, + 0.007564957719296217, + 0.0061499206349253654, + 0.004218382760882378, + 0.005270511377602816, + 0.011134414002299309, + 0.006076571997255087, + 0.010149840265512466, + 0.00884102564305067, + 0.06311293691396713, + 0.01845281571149826, + 0.012715009972453117, + 0.011096588335931301, + 0.023297913372516632, + 0.016513360664248466, + 0.05321464315056801, + 0.10753168910741806, + 0.008212505839765072, + 0.014333289116621017, + 0.004968665540218353, + 0.00815388560295105, + 0.014843710698187351, + 0.00962477084249258, + 0.007277257274836302, + 0.009651066735386848 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 4 + ], + "selected_reward": -0.08639468252658844 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a historic destroyer with a weathered, dark gray hull showing rust and scratches?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the ship moving through deep blue, almost black waters with mountainous waves crashing violently against its side?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene illuminated by a solemn, stage-like Jesus light piercing through thick, dark clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera executing a steady, low-angle parallel tracking shot following the ship?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the radar antenna on the ship's bridge shown as the only steady movement in 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": [ + -0.01255148183554411 + ], + "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": [ + 0.21082362532615662, + 0.10434091091156006, + 0.03442838042974472, + 0.019468098878860474, + 0.03499212861061096, + 0.02941468358039856, + 0.055628541857004166, + 0.09598174691200256, + 0.22893883287906647, + 0.18598297238349915 + ], + "credit_map": [ + 0.02498152107000351, + 0.023281946778297424, + 0.023771245032548904, + 0.019289655610919, + 0.033946726471185684, + 0.022295016795396805, + 0.03849653899669647, + 0.02476099506020546, + 0.009746252559125423, + 0.008296896703541279, + 0.018125608563423157, + 0.013511997647583485, + 0.012798020616173744, + 0.007268163841217756, + 0.021351318806409836, + 0.013242646120488644, + 0.0036314784083515406, + 0.0028185141272842884, + 0.0033533915411680937, + 0.0031661021057516336, + 0.005446206778287888, + 0.003284349339082837, + 0.008082130923867226, + 0.0046462067402899265, + 0.0023176143877208233, + 0.0016081691719591618, + 0.0022732578217983246, + 0.0017005109693855047, + 0.004552152939140797, + 0.0019499334739521146, + 0.002712312387302518, + 0.0023541483096778393, + 0.004486409947276115, + 0.00323202321305871, + 0.0032428742852061987, + 0.0033743069507181644, + 0.005624407436698675, + 0.0027618790045380592, + 0.006975469179451466, + 0.005294760223478079, + 0.0030877580866217613, + 0.0026366207748651505, + 0.0021025585010647774, + 0.002489614300429821, + 0.005357394460588694, + 0.003556035226210952, + 0.004707013256847858, + 0.0054776896722614765, + 0.00641587283462286, + 0.008523431606590748, + 0.004353339318186045, + 0.005522832274436951, + 0.009794259443879128, + 0.005789895076304674, + 0.007281883154064417, + 0.007947031408548355, + 0.006387526169419289, + 0.0077257417142391205, + 0.005809035152196884, + 0.007141838315874338, + 0.016460027545690536, + 0.01863754354417324, + 0.017190825194120407, + 0.016629207879304886, + 0.04058636352419853, + 0.015613740310072899, + 0.010155484080314636, + 0.009304285049438477, + 0.018629152327775955, + 0.013879499398171902, + 0.04248631000518799, + 0.07828401029109955, + 0.01869550533592701, + 0.038292452692985535, + 0.015127778984606266, + 0.017664005979895592, + 0.03981202095746994, + 0.024176178500056267, + 0.01789567992091179, + 0.014319350942969322 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 9 + ], + "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 main subject a historic destroyer with a weathered, dark gray hull showing rust and scratches?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the ship moving through deep blue, almost black waters with mountainous waves crashing violently against its side?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene illuminated by a solemn, stage-like Jesus light piercing through thick, dark clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera executing a steady, low-angle parallel tracking shot following the ship?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the radar antenna on the ship's bridge shown as the only steady movement in 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": [ + -0.020585577934980392 + ], + "opposite_logprob_tokens": [ + -3.895585536956787 + ], + "logprob_reward": -0.020585577934980392, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020585577934980392, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12068787962198257, + 0.04468080773949623, + 0.07663362473249435, + 0.07604461163282394, + 0.42622390389442444, + 0.02856249362230301, + 0.0648854523897171, + 0.08857522159814835, + 0.035989295691251755, + 0.03771670162677765 + ], + "credit_map": [ + 0.011221697553992271, + 0.012071410194039345, + 0.008340870030224323, + 0.013233627192676067, + 0.026001982390880585, + 0.018680954352021217, + 0.01761741191148758, + 0.013519921340048313, + 0.005011612083762884, + 0.003454255871474743, + 0.0031177690252661705, + 0.00421163160353899, + 0.008608085103332996, + 0.0060959262773394585, + 0.009078987874090672, + 0.005102538503706455, + 0.011087965220212936, + 0.008083345368504524, + 0.008076499216258526, + 0.007975206710398197, + 0.007709205616265535, + 0.008097123354673386, + 0.017275717109441757, + 0.008328567259013653, + 0.011160066351294518, + 0.008288831450045109, + 0.006366833578795195, + 0.011915608309209347, + 0.010350456461310387, + 0.005584338214248419, + 0.012109492905437946, + 0.010268987156450748, + 0.06760969012975693, + 0.03438572213053703, + 0.02704566903412342, + 0.026186075061559677, + 0.05799537152051926, + 0.03359941765666008, + 0.07234617322683334, + 0.10705576837062836, + 0.004612536635249853, + 0.0032063026446849108, + 0.0021606653463095427, + 0.002521099988371134, + 0.0026140064001083374, + 0.0024802773259580135, + 0.004634617827832699, + 0.0063329883851110935, + 0.008588946424424648, + 0.007436523679643869, + 0.006745232734829187, + 0.006182814482599497, + 0.007557433098554611, + 0.005993721075356007, + 0.012381906621158123, + 0.009998880326747894, + 0.016862304881215096, + 0.00888631772249937, + 0.008918404579162598, + 0.00638731149956584, + 0.012418401427567005, + 0.010502406395971775, + 0.013842283748090267, + 0.010757798328995705, + 0.0057057482190430164, + 0.0035990646574646235, + 0.005236437078565359, + 0.003741668537259102, + 0.004085223656147718, + 0.003954706713557243, + 0.00471687875688076, + 0.004949567373842001, + 0.006260327994823456, + 0.004146228544414043, + 0.00337027944624424, + 0.0036526506301015615, + 0.007899518124759197, + 0.003942164592444897, + 0.004584076814353466, + 0.003861456410959363 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 7 + ], + "selected_reward": -0.020585577934980392 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a historic destroyer with a weathered, dark gray hull showing rust and scratches?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the ship moving through deep blue, almost black waters with mountainous waves crashing violently against its side?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene illuminated by a solemn, stage-like Jesus light piercing through thick, dark clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera executing a steady, low-angle parallel tracking shot following the ship?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the radar antenna on the ship's bridge shown as the only steady movement in 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": [ + -0.07896807044744492 + ], + "opposite_logprob_tokens": [ + -2.578968048095703 + ], + "logprob_reward": -0.07896807044744492, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07896807044744492, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.050811804831027985, + 0.028591783717274666, + 0.024929946288466454, + 0.027355486527085304, + 0.13171738386154175, + 0.04122886434197426, + 0.0535084530711174, + 0.16939383745193481, + 0.3862297236919403, + 0.08623263984918594 + ], + "credit_map": [ + 0.006971335504204035, + 0.004938628990203142, + 0.006061519030481577, + 0.0036460079718381166, + 0.01075669378042221, + 0.005341117735952139, + 0.007388880010694265, + 0.005707623437047005, + 0.00319220800884068, + 0.0027465636376291513, + 0.0025434454437345266, + 0.0019771100487560034, + 0.006153763737529516, + 0.002944471314549446, + 0.003886015387251973, + 0.005148207303136587, + 0.0023490923922508955, + 0.0025175742339342833, + 0.003587275743484497, + 0.0027635064907372, + 0.004173848778009415, + 0.002823193557560444, + 0.004194189328700304, + 0.0025212676264345646, + 0.0042413147166371346, + 0.002671837108209729, + 0.001990508520975709, + 0.0030720382928848267, + 0.004774652887135744, + 0.002120639430359006, + 0.004750215448439121, + 0.003734281286597252, + 0.01609226129949093, + 0.009729897603392601, + 0.007404254749417305, + 0.008424216881394386, + 0.02255609631538391, + 0.011137530207633972, + 0.02697087824344635, + 0.02940225787460804, + 0.004480323754251003, + 0.002831166610121727, + 0.002551248064264655, + 0.0034940114710479975, + 0.005482537671923637, + 0.005526545457541943, + 0.009874740615487099, + 0.0069882930256426334, + 0.012118535116314888, + 0.009745897725224495, + 0.004565322306007147, + 0.004345843568444252, + 0.006348509807139635, + 0.003352567320689559, + 0.006324905902147293, + 0.006706869229674339, + 0.011194998398423195, + 0.014518018811941147, + 0.01013955194503069, + 0.011286729015409946, + 0.029364662244915962, + 0.03451627492904663, + 0.03135904297232628, + 0.027014564722776413, + 0.07590828835964203, + 0.025049543008208275, + 0.01274027954787016, + 0.008066604845225811, + 0.02427341789007187, + 0.017514538019895554, + 0.07822481542825699, + 0.14445224404335022, + 0.008223984390497208, + 0.014996632002294064, + 0.007648190017789602, + 0.007443600799888372, + 0.018996506929397583, + 0.012625962495803833, + 0.008705459535121918, + 0.007592292036861181 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 4 + ], + "selected_reward": -0.07896807044744492 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a historic destroyer with a weathered, dark gray hull showing rust and scratches?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the ship moving through deep blue, almost black waters with mountainous waves crashing violently against its side?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene illuminated by a solemn, stage-like Jesus light piercing through thick, dark clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera executing a steady, low-angle parallel tracking shot following the ship?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the radar antenna on the ship's bridge shown as the only steady movement in 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": [ + -0.14282111823558807 + ], + "opposite_logprob_tokens": [ + -2.0178210735321045 + ], + "logprob_reward": -0.14282111823558807, + "yes_no_margin_reward": 1.875, + "token_credit_reward": -0.14282111823558807, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17992235720157623, + 0.10404492914676666, + 0.08243511617183685, + 0.05303605645895004, + 0.08367496728897095, + 0.12162034213542938, + 0.05353853851556778, + 0.057741593569517136, + 0.10989250987768173, + 0.15409362316131592 + ], + "credit_map": [ + 0.021801093593239784, + 0.01305344607681036, + 0.018684960901737213, + 0.017163369804620743, + 0.03628808259963989, + 0.01633073017001152, + 0.029634391888976097, + 0.026966281235218048, + 0.01271716970950365, + 0.010091406293213367, + 0.006383365951478481, + 0.010082933120429516, + 0.02878495678305626, + 0.00919960904866457, + 0.011375218629837036, + 0.015410265885293484, + 0.007624672260135412, + 0.007435931824147701, + 0.006515853106975555, + 0.007648508530110121, + 0.01674206182360649, + 0.008057293482124805, + 0.01703951507806778, + 0.011371280066668987, + 0.006901292130351067, + 0.004772753454744816, + 0.004101872444152832, + 0.0057047209702432156, + 0.00930576678365469, + 0.0034156939946115017, + 0.01057918556034565, + 0.008254768326878548, + 0.008189832791686058, + 0.006226481404155493, + 0.004309842362999916, + 0.006349322386085987, + 0.01921955868601799, + 0.009244192391633987, + 0.017608782276511192, + 0.012526962906122208, + 0.009856891818344593, + 0.008401281200349331, + 0.007173492107540369, + 0.01041503343731165, + 0.025085357949137688, + 0.015282535925507545, + 0.0235641747713089, + 0.02184157259762287, + 0.007601572200655937, + 0.006589408498257399, + 0.0038966326974332333, + 0.004433642607182264, + 0.0072817252948880196, + 0.006055754609405994, + 0.009058007970452309, + 0.00862179510295391, + 0.008220135234296322, + 0.006481816992163658, + 0.004121646750718355, + 0.0055249519646167755, + 0.009252876974642277, + 0.0069372523576021194, + 0.008693356066942215, + 0.008509555831551552, + 0.018876591697335243, + 0.008189047686755657, + 0.006460289470851421, + 0.005709508899599314, + 0.01331431046128273, + 0.007451881654560566, + 0.016540836542844772, + 0.033350054174661636, + 0.015409248881042004, + 0.03055417723953724, + 0.011790207587182522, + 0.0134406303986907, + 0.030992398038506508, + 0.02152726612985134, + 0.016354717314243317, + 0.014024985954165459 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 5 + ], + "selected_reward": -0.14282111823558807 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle, parallel tracking shot follows a historic destroyer, a steel behemoth slicing through the raging, almost black, deep blue waters with unstoppable force. Its weathered, dark gray hull is mottled with rust and crisscrossed with scratches. Its towering bridge and silent turrets appear majestic against the gloomy sky, the slowly rotating radar antenna on the bridge the only steady movement in the frame. Mountainous waves crash violently against the ship's side, then burst into a thousand crystalline droplets, each spray carving a brief, powerful arc through the air. Thick, dark clouds threaten to crush the sea, but a solemn, stage-like Jesus light pierces the clouds, casting a precise, dramatic glow on the waters beneath the ship's path, illuminating the swirling white foam and drifting mist with a tense, dramatic sheen, profoundly carving out the ship's hard metal contours and damp, reflective texture. The steady and leisurely camera movement combined with this powerful scene creates an unparalleled sense of epic heaviness in the cold-toned picture, as if telling a grand narrative about steel, ocean and indomitable will.", + "seed": 22726, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000121/group_11/index_1215_seed_22726_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": -0.07381700668483973, + "vlm_logprob": -0.07381700668483973, + "vlm_yes_no_margin": 2.899999976158142, + "vlm_token_credit": -0.07381700668483973, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a historic destroyer with a weathered, dark gray hull showing rust and scratches?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the ship moving through deep blue, almost black waters with mountainous waves crashing violently against its side?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene illuminated by a solemn, stage-like Jesus light piercing through thick, dark clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera executing a steady, low-angle parallel tracking shot following the ship?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the radar antenna on the ship's bridge shown as the only steady movement in 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": [ + -0.06079436093568802 + ], + "opposite_logprob_tokens": [ + -2.9357943534851074 + ], + "logprob_reward": -0.06079436093568802, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.06079436093568802, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.632802426815033, + 0.03884162753820419, + 0.01844714768230915, + 0.021100087091326714, + 0.037004537880420685, + 0.0563749223947525, + 0.05822377651929855, + 0.07812223583459854, + 0.03212324157357216, + 0.02696002647280693 + ], + "credit_map": [ + 0.06127186119556427, + 0.0272725448012352, + 0.029895469546318054, + 0.017273928970098495, + 0.032073210924863815, + 0.2659204602241516, + 0.14182350039482117, + 0.05727145075798035, + 0.003596746362745762, + 0.004306387156248093, + 0.0043038236908614635, + 0.004596286453306675, + 0.006165619008243084, + 0.006080139894038439, + 0.0051941401325166225, + 0.0045984843745827675, + 0.001910868100821972, + 0.0012273145839571953, + 0.0013869862305000424, + 0.0030763400718569756, + 0.0026234870310872793, + 0.0028392209205776453, + 0.0029051934834569693, + 0.002477736910805106, + 0.002275345381349325, + 0.0015094474656507373, + 0.0018549116794019938, + 0.0030192367266863585, + 0.003213710617274046, + 0.0032654842361807823, + 0.003443371271714568, + 0.002518579363822937, + 0.002902624662965536, + 0.0025943752843886614, + 0.003949133213609457, + 0.005003457888960838, + 0.0064057824201881886, + 0.005421205423772335, + 0.006607630755752325, + 0.004120327532291412, + 0.004542543552815914, + 0.003747300012037158, + 0.005936911795288324, + 0.005004604812711477, + 0.009131974540650845, + 0.00916339922696352, + 0.01125091128051281, + 0.0075972722843289375, + 0.004726741462945938, + 0.0037330621853470802, + 0.006046755239367485, + 0.0071456399746239185, + 0.0096517875790596, + 0.00913267396390438, + 0.008380042389035225, + 0.009407068602740765, + 0.012386437505483627, + 0.004998763091862202, + 0.006823924835771322, + 0.005194027442485094, + 0.004672301933169365, + 0.0263233482837677, + 0.007000611629337072, + 0.010722815990447998, + 0.006241340655833483, + 0.002190198050811887, + 0.0015843163710087538, + 0.0027015344239771366, + 0.0019270620541647077, + 0.005472768098115921, + 0.001808720058761537, + 0.010197300463914871, + 0.0025374568067491055, + 0.0020831238944083452, + 0.002782569034025073, + 0.004417054820805788, + 0.004501135088503361, + 0.004159206990152597, + 0.004223859868943691, + 0.0022556185722351074 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 6 + ], + "selected_reward": -0.06079436093568802 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a historic destroyer with a weathered, dark gray hull showing rust and scratches?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the ship moving through deep blue, almost black waters with mountainous waves crashing violently against its side?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene illuminated by a solemn, stage-like Jesus light piercing through thick, dark clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera executing a steady, low-angle parallel tracking shot following the ship?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the radar antenna on the ship's bridge shown as the only steady movement in 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": [ + -0.011093133129179478 + ], + "opposite_logprob_tokens": [ + -4.5110931396484375 + ], + "logprob_reward": -0.011093133129179478, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011093133129179478, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.6648814082145691, + 0.021948345005512238, + 0.01706394925713539, + 0.013345100916922092, + 0.026206398382782936, + 0.048585448414087296, + 0.024955693632364273, + 0.04245639964938164, + 0.09004057943820953, + 0.05051672086119652 + ], + "credit_map": [ + 0.06802082061767578, + 0.029628155753016472, + 0.030282363295555115, + 0.01848836988210678, + 0.03319809213280678, + 0.2778341472148895, + 0.14745678007602692, + 0.05997268855571747, + 0.0019014956196770072, + 0.0022774781100451946, + 0.002562503796070814, + 0.0026113532949239016, + 0.004384475294500589, + 0.0032830513082444668, + 0.002668496686965227, + 0.002259490778669715, + 0.0014371161814779043, + 0.0014409595169126987, + 0.0016937495674937963, + 0.0021592602133750916, + 0.0026640337891876698, + 0.002450175816193223, + 0.0022428419906646013, + 0.002975812880322337, + 0.0009647343540564179, + 0.0009046732448041439, + 0.0015457647386938334, + 0.0020872463937848806, + 0.0018837039824575186, + 0.002324635861441493, + 0.001920424634590745, + 0.001713917008601129, + 0.0018425190355628729, + 0.0015038677956908941, + 0.0035167718306183815, + 0.003973175305873156, + 0.0018541117897257209, + 0.005056141410022974, + 0.004093383904546499, + 0.004366426263004541, + 0.0041846781969070435, + 0.0032531653996556997, + 0.006121782120317221, + 0.00430750846862793, + 0.005675524938851595, + 0.008368189446628094, + 0.010338179767131805, + 0.006336420774459839, + 0.003807089291512966, + 0.0018464135937392712, + 0.0042494190856814384, + 0.003085256088525057, + 0.0020397671032696962, + 0.0033871494233608246, + 0.0031942350324243307, + 0.0033463644795119762, + 0.004380162805318832, + 0.002437070943415165, + 0.0020686215721070766, + 0.0032181802671402693, + 0.004847229458391666, + 0.01393149420619011, + 0.003472426673397422, + 0.008101212792098522, + 0.01990498974919319, + 0.004887512419372797, + 0.003780961036682129, + 0.003322743810713291, + 0.00433533638715744, + 0.015813473612070084, + 0.0037380170542746782, + 0.034257546067237854, + 0.004521791357547045, + 0.003418687265366316, + 0.007680446375161409, + 0.00408139219507575, + 0.0070872618816792965, + 0.006439284887164831, + 0.009797882288694382, + 0.0074899750761687756 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 9 + ], + "selected_reward": -0.011093133129179478 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a historic destroyer with a weathered, dark gray hull showing rust and scratches?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the ship moving through deep blue, almost black waters with mountainous waves crashing violently against its side?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene illuminated by a solemn, stage-like Jesus light piercing through thick, dark clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera executing a steady, low-angle parallel tracking shot following the ship?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the radar antenna on the ship's bridge shown as the only steady movement in 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": [ + -0.043072763830423355 + ], + "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": [ + 0.43801265954971313, + 0.23630906641483307, + 0.018691973760724068, + 0.016915665939450264, + 0.019286181777715683, + 0.03982555866241455, + 0.016744835302233696, + 0.03361726552248001, + 0.10285713523626328, + 0.07773967832326889 + ], + "credit_map": [ + 0.03875938802957535, + 0.019529830664396286, + 0.021389082074165344, + 0.015689535066485405, + 0.022402901202440262, + 0.18139295279979706, + 0.09895524382591248, + 0.039893705397844315, + 0.031506527215242386, + 0.030711328610777855, + 0.031048357486724854, + 0.022737950086593628, + 0.0225986260920763, + 0.04168766364455223, + 0.03242480382323265, + 0.023593764752149582, + 0.0030972177628427744, + 0.0026049180887639523, + 0.0021946774795651436, + 0.0022711327765136957, + 0.0015530280070379376, + 0.002962445840239525, + 0.0019507782999426126, + 0.002057773759588599, + 0.0012604525545611978, + 0.0013304802123457193, + 0.003648810787126422, + 0.002697218209505081, + 0.0012583520729094744, + 0.0024914441164582968, + 0.0017965045990422368, + 0.002432403387501836, + 0.0017214497784152627, + 0.00103241135366261, + 0.0019362909952178597, + 0.0022238476667553186, + 0.0019264743896201253, + 0.0031869562808424234, + 0.004801345057785511, + 0.002457404974848032, + 0.0026586055755615234, + 0.002222174545750022, + 0.003329455852508545, + 0.00451867887750268, + 0.0038833334110677242, + 0.006769281346350908, + 0.0113926250487566, + 0.0050514014437794685, + 0.0016494238516315818, + 0.0010249775368720293, + 0.0020768416579812765, + 0.002124411053955555, + 0.002744942205026746, + 0.0026688033249229193, + 0.0023968431632965803, + 0.002058589830994606, + 0.0041942051611840725, + 0.0020517962984740734, + 0.0021036695688962936, + 0.003388255136087537, + 0.0039191232062876225, + 0.010449821129441261, + 0.002529220888391137, + 0.004981172736734152, + 0.022730065509676933, + 0.0054916562512516975, + 0.00453915586695075, + 0.005042366683483124, + 0.005068568978458643, + 0.018040519207715988, + 0.004263180308043957, + 0.037681613117456436, + 0.009962785057723522, + 0.010371841490268707, + 0.00869077816605568, + 0.009832167997956276, + 0.01216548029333353, + 0.010793542489409447, + 0.00909358449280262, + 0.006829496007412672 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 8 + ], + "selected_reward": -0.043072763830423355 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a historic destroyer with a weathered, dark gray hull showing rust and scratches?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the ship moving through deep blue, almost black waters with mountainous waves crashing violently against its side?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene illuminated by a solemn, stage-like Jesus light piercing through thick, dark clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera executing a steady, low-angle parallel tracking shot following the ship?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the radar antenna on the ship's bridge shown as the only steady movement in 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": [ + -0.12704144418239594 + ], + "opposite_logprob_tokens": [ + -2.1270413398742676 + ], + "logprob_reward": -0.12704144418239594, + "yes_no_margin_reward": 1.9999998807907104, + "token_credit_reward": -0.12704144418239594, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.40163055062294006, + 0.06656742095947266, + 0.035382192581892014, + 0.029432261362671852, + 0.05705228075385094, + 0.10656125098466873, + 0.09086281061172485, + 0.07791714370250702, + 0.055667027831077576, + 0.07892706990242004 + ], + "credit_map": [ + 0.04013538360595703, + 0.017069801688194275, + 0.02101602591574192, + 0.02210000716149807, + 0.01920202746987343, + 0.1610996276140213, + 0.08328496664762497, + 0.03772271052002907, + 0.004471360240131617, + 0.003711903002113104, + 0.005500636994838715, + 0.011905885301530361, + 0.010473302565515041, + 0.010202468372881413, + 0.010920390486717224, + 0.009381472133100033, + 0.0028508310206234455, + 0.0021093974355608225, + 0.004876464605331421, + 0.004154730588197708, + 0.005176473408937454, + 0.006097027100622654, + 0.004419033881276846, + 0.005698235705494881, + 0.001990277785807848, + 0.0016295216046273708, + 0.002487810328602791, + 0.004881309345364571, + 0.004043815657496452, + 0.004603839945048094, + 0.005723124369978905, + 0.004072563722729683, + 0.00516845565289259, + 0.0034050391986966133, + 0.007618780713528395, + 0.007914370857179165, + 0.005143384449183941, + 0.009828799404203892, + 0.010674851015210152, + 0.007298601325601339, + 0.007582611404359341, + 0.0059864153154194355, + 0.008399955928325653, + 0.011797603219747543, + 0.010520268231630325, + 0.018742259591817856, + 0.028137288987636566, + 0.01539484690874815, + 0.009628581814467907, + 0.004604583606123924, + 0.014345146715641022, + 0.010428913868963718, + 0.011892087757587433, + 0.014089949429035187, + 0.011411438696086407, + 0.014462107792496681, + 0.007354400120675564, + 0.005184441804885864, + 0.00545844528824091, + 0.008547693490982056, + 0.007685166317969561, + 0.022049555554986, + 0.007036036346107721, + 0.014601403847336769, + 0.012208851985633373, + 0.0037358365952968597, + 0.0042847865261137486, + 0.004587657283991575, + 0.0036836164072155952, + 0.007756248116493225, + 0.0026729872915893793, + 0.01673704758286476, + 0.007629713509231806, + 0.007490683812648058, + 0.009002736769616604, + 0.007685071788728237, + 0.014004855416715145, + 0.009803342632949352, + 0.013186552561819553, + 0.010124110616743565 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 6 + ], + "selected_reward": -0.12704144418239594 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a historic destroyer with a weathered, dark gray hull showing rust and scratches?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the ship moving through deep blue, almost black waters with mountainous waves crashing violently against its side?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene illuminated by a solemn, stage-like Jesus light piercing through thick, dark clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera executing a steady, low-angle parallel tracking shot following the ship?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the radar antenna on the ship's bridge shown as the only steady movement in 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": [ + -0.12708333134651184 + ], + "opposite_logprob_tokens": [ + -2.1270833015441895 + ], + "logprob_reward": -0.12708333134651184, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.12708333134651184, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.7186735272407532, + 0.04879254475235939, + 0.03804680332541466, + 0.014525827020406723, + 0.013815967366099358, + 0.01682225428521633, + 0.02690378949046135, + 0.03038913570344448, + 0.0711674913764, + 0.020862724632024765 + ], + "credit_map": [ + 0.07322279363870621, + 0.031853605061769485, + 0.031370311975479126, + 0.01835298165678978, + 0.034798141568899155, + 0.30200740694999695, + 0.16140054166316986, + 0.06566764414310455, + 0.0032860138453543186, + 0.0020992346107959747, + 0.003772559342905879, + 0.006586726289242506, + 0.007664830423891544, + 0.007394048850983381, + 0.0102784913033247, + 0.007710631471127272, + 0.003221917664632201, + 0.0027607649099081755, + 0.002910956973209977, + 0.004606543108820915, + 0.006216728128492832, + 0.006080271676182747, + 0.005357652436941862, + 0.006891963072121143, + 0.0009615541202947497, + 0.0008602207526564598, + 0.002507180441170931, + 0.0024162407498806715, + 0.0016873142449185252, + 0.002100185723975301, + 0.0020320876501500607, + 0.0019610414747148752, + 0.0008661099127493799, + 0.0009487839415669441, + 0.0011483108391985297, + 0.0014770828420296311, + 0.002643316052854061, + 0.0023174635134637356, + 0.0020888906437903643, + 0.0023260079324245453, + 0.001076967571862042, + 0.001132630044594407, + 0.0012173275463283062, + 0.002502035116776824, + 0.0027226379606872797, + 0.0027529485523700714, + 0.0035430416464805603, + 0.0018746642163023353, + 0.0037170203868299723, + 0.0017736199079081416, + 0.005797108635306358, + 0.0026346396189182997, + 0.0018992742989212275, + 0.0038296286948025227, + 0.003942982759326696, + 0.003309511346742511, + 0.004614852834492922, + 0.0018557988805696368, + 0.0016277986578643322, + 0.0018466695910319686, + 0.0031573281157761812, + 0.010344055481255054, + 0.002498015994206071, + 0.004444611258804798, + 0.01629818044602871, + 0.0038043262902647257, + 0.0035403312649577856, + 0.0016277343966066837, + 0.0029298667795956135, + 0.012738359160721302, + 0.002770704682916403, + 0.027457980439066887, + 0.0023980741389095783, + 0.001629236969165504, + 0.002413379494100809, + 0.0034571874421089888, + 0.0027425303123891354, + 0.0034425987396389246, + 0.002304786816239357, + 0.002474927343428135 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 1 + ], + "selected_reward": -0.12708333134651184 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle, parallel tracking shot follows a historic destroyer, a steel behemoth slicing through the raging, almost black, deep blue waters with unstoppable force. Its weathered, dark gray hull is mottled with rust and crisscrossed with scratches. Its towering bridge and silent turrets appear majestic against the gloomy sky, the slowly rotating radar antenna on the bridge the only steady movement in the frame. Mountainous waves crash violently against the ship's side, then burst into a thousand crystalline droplets, each spray carving a brief, powerful arc through the air. Thick, dark clouds threaten to crush the sea, but a solemn, stage-like Jesus light pierces the clouds, casting a precise, dramatic glow on the waters beneath the ship's path, illuminating the swirling white foam and drifting mist with a tense, dramatic sheen, profoundly carving out the ship's hard metal contours and damp, reflective texture. The steady and leisurely camera movement combined with this powerful scene creates an unparalleled sense of epic heaviness in the cold-toned picture, as if telling a grand narrative about steel, ocean and indomitable will.", + "seed": 22727, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000121/group_11/index_1215_seed_22727_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": -0.06092195156961679, + "vlm_logprob": -0.06092195156961679, + "vlm_yes_no_margin": 3.125, + "vlm_token_credit": -0.06092195156961679, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a historic destroyer with a weathered, dark gray hull showing rust and scratches?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the ship moving through deep blue, almost black waters with mountainous waves crashing violently against its side?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene illuminated by a solemn, stage-like Jesus light piercing through thick, dark clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera executing a steady, low-angle parallel tracking shot following the ship?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the radar antenna on the ship's bridge shown as the only steady movement in 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": [ + -0.05622130632400513 + ], + "opposite_logprob_tokens": [ + -3.0562212467193604 + ], + "logprob_reward": -0.05622130632400513, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.05622130632400513, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.438720703125, + 0.08565454185009003, + 0.04801154509186745, + 0.03704577684402466, + 0.0239973496645689, + 0.014853078871965408, + 0.07381261885166168, + 0.09885843843221664, + 0.055780086666345596, + 0.12326592206954956 + ], + "credit_map": [ + 0.028238140046596527, + 0.018726229667663574, + 0.012761412188410759, + 0.005132729187607765, + 0.02592981792986393, + 0.19518719613552094, + 0.024461351335048676, + 0.12828382849693298, + 0.009476196952164173, + 0.009294048883020878, + 0.005908820312470198, + 0.007404266390949488, + 0.010943076573312283, + 0.01425887644290924, + 0.01602926105260849, + 0.012339991517364979, + 0.004209778271615505, + 0.006818733643740416, + 0.002895286539569497, + 0.006468838546425104, + 0.006162423174828291, + 0.006000989116728306, + 0.00609825411811471, + 0.009357238188385963, + 0.0052469270303845406, + 0.00483693927526474, + 0.002232816070318222, + 0.0036126438062638044, + 0.005934113170951605, + 0.007285908795893192, + 0.003608141792938113, + 0.004288285970687866, + 0.002148362109437585, + 0.002771672559902072, + 0.0018720175139606, + 0.004391568712890148, + 0.0033327476121485233, + 0.0040028709918260574, + 0.0030298202764242887, + 0.0024482908193022013, + 0.0014030024176463485, + 0.0025179199874401093, + 0.001255957642570138, + 0.0018511225935071707, + 0.002828290220350027, + 0.0021811106707900763, + 0.0014617610722780228, + 0.0013539135688915849, + 0.005234336014837027, + 0.007487923838198185, + 0.0054166680201888084, + 0.008660443127155304, + 0.008634025231003761, + 0.010483425110578537, + 0.01570272631943226, + 0.01219306793063879, + 0.009021761827170849, + 0.009899742901325226, + 0.006713138893246651, + 0.011530548334121704, + 0.014753961004316807, + 0.016239121556282043, + 0.018039528280496597, + 0.012660644017159939, + 0.006747073959559202, + 0.007177275605499744, + 0.004028588533401489, + 0.006056261248886585, + 0.012055731378495693, + 0.00801375787705183, + 0.007357493042945862, + 0.0043439012952148914, + 0.017082437872886658, + 0.0168751273304224, + 0.008777406066656113, + 0.014521912671625614, + 0.023096097633242607, + 0.017357295379042625, + 0.012161667458713055, + 0.013393969275057316 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 7 + ], + "selected_reward": -0.05622130632400513 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a historic destroyer with a weathered, dark gray hull showing rust and scratches?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the ship moving through deep blue, almost black waters with mountainous waves crashing violently against its side?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene illuminated by a solemn, stage-like Jesus light piercing through thick, dark clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera executing a steady, low-angle parallel tracking shot following the ship?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the radar antenna on the ship's bridge shown as the only steady movement in 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": [ + -0.00980123970657587 + ], + "opposite_logprob_tokens": [ + -4.634801387786865 + ], + "logprob_reward": -0.00980123970657587, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.00980123970657587, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.6598717570304871, + 0.03682566434144974, + 0.07454171776771545, + 0.05051314830780029, + 0.0600682869553566, + 0.022889966145157814, + 0.00884973630309105, + 0.02543213777244091, + 0.030701937153935432, + 0.0303056538105011 + ], + "credit_map": [ + 0.043414268642663956, + 0.025608962401747704, + 0.019845064729452133, + 0.009677966125309467, + 0.03918558731675148, + 0.2939118444919586, + 0.036775462329387665, + 0.19145263731479645, + 0.004617007914930582, + 0.0040409802459180355, + 0.0020624594762921333, + 0.0024402295239269733, + 0.0041829608380794525, + 0.00867505557835102, + 0.005470840726047754, + 0.005336132366210222, + 0.006249531637877226, + 0.006973606534302235, + 0.004377459175884724, + 0.004294277634471655, + 0.007464772555977106, + 0.013008092530071735, + 0.012095927260816097, + 0.020078051835298538, + 0.004835998639464378, + 0.005983696319162846, + 0.002961392980068922, + 0.005967407021671534, + 0.007984093390405178, + 0.006790889427065849, + 0.009731840342283249, + 0.006257834378629923, + 0.00515019753947854, + 0.009270938113331795, + 0.005524020176380873, + 0.009901398792862892, + 0.01132846437394619, + 0.006900361273437738, + 0.00756683899089694, + 0.004426073282957077, + 0.0026603590231388807, + 0.0036360134836286306, + 0.0018264084355905652, + 0.0023922519758343697, + 0.005299433600157499, + 0.003397891065105796, + 0.0021800899412482977, + 0.0014975189696997404, + 0.0009383663418702781, + 0.001062240800820291, + 0.0010321458103135228, + 0.0009622785146348178, + 0.0013221529079601169, + 0.0010273244697600603, + 0.0013341954909265041, + 0.001171032665297389, + 0.0032895291224122047, + 0.0023989796172827482, + 0.002287503331899643, + 0.0022443989291787148, + 0.004423787351697683, + 0.005093802697956562, + 0.0026771218981593847, + 0.003017017850652337, + 0.0035009398125112057, + 0.0029018453788012266, + 0.003716046456247568, + 0.003710200311616063, + 0.005938165821135044, + 0.004404508508741856, + 0.003698163665831089, + 0.002832067897543311, + 0.0033948186319321394, + 0.0033484227024018764, + 0.0038138392847031355, + 0.0038901485968381166, + 0.00476375175639987, + 0.004013889469206333, + 0.003388263052329421, + 0.003692520782351494 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 4 + ], + "selected_reward": -0.00980123970657587 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a historic destroyer with a weathered, dark gray hull showing rust and scratches?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the ship moving through deep blue, almost black waters with mountainous waves crashing violently against its side?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene illuminated by a solemn, stage-like Jesus light piercing through thick, dark clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera executing a steady, low-angle parallel tracking shot following the ship?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the radar antenna on the ship's bridge shown as the only steady movement in 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": [ + -0.033709973096847534 + ], + "opposite_logprob_tokens": [ + -3.40871000289917 + ], + "logprob_reward": -0.033709973096847534, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.033709973096847534, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4289231598377228, + 0.02738676220178604, + 0.04068969562649727, + 0.026193559169769287, + 0.19301247596740723, + 0.01811586134135723, + 0.10809259861707687, + 0.08430223166942596, + 0.03334597870707512, + 0.03993760049343109 + ], + "credit_map": [ + 0.030579812824726105, + 0.022080840542912483, + 0.014033661223948002, + 0.012550113722682, + 0.02538178116083145, + 0.1829693764448166, + 0.025730382651090622, + 0.11559724807739258, + 0.0029576129745692015, + 0.005135925021022558, + 0.003599405288696289, + 0.002649616217240691, + 0.003545936429873109, + 0.0030958515126258135, + 0.002789522288367152, + 0.003612895729020238, + 0.002515484346076846, + 0.004969634581357241, + 0.0033806869760155678, + 0.005726983305066824, + 0.004176680464297533, + 0.005394939333200455, + 0.00812293216586113, + 0.006402361206710339, + 0.002304847352206707, + 0.0031395575497299433, + 0.0025291768833994865, + 0.0024938832502812147, + 0.005787975620478392, + 0.003965245094150305, + 0.002994390670210123, + 0.0029784850776195526, + 0.016438158228993416, + 0.032200492918491364, + 0.015764771029353142, + 0.03326236456632614, + 0.029962003231048584, + 0.023460978642106056, + 0.027224555611610413, + 0.014699174091219902, + 0.002328817266970873, + 0.0020213457755744457, + 0.0022267980966717005, + 0.0018375158542767167, + 0.002546690171584487, + 0.002055772813037038, + 0.0027994359843432903, + 0.0022994880564510822, + 0.011853412725031376, + 0.01296114083379507, + 0.010922162793576717, + 0.01099312398582697, + 0.02627277374267578, + 0.01496254000812769, + 0.010221798904240131, + 0.009905651211738586, + 0.011265762150287628, + 0.010101807303726673, + 0.009467869997024536, + 0.008282351307570934, + 0.012968654744327068, + 0.015071279369294643, + 0.0072561828419566154, + 0.009888331405818462, + 0.005206183530390263, + 0.003042376134544611, + 0.0033307578414678574, + 0.003139847656711936, + 0.006579283624887466, + 0.005187089089304209, + 0.003721583401784301, + 0.003138861386105418, + 0.006169280968606472, + 0.005782297346740961, + 0.0033938295673578978, + 0.0037734282668679953, + 0.0076433285139501095, + 0.005668079014867544, + 0.003932471852749586, + 0.0035748914815485477 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 6 + ], + "selected_reward": -0.033709973096847534 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a historic destroyer with a weathered, dark gray hull showing rust and scratches?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the ship moving through deep blue, almost black waters with mountainous waves crashing violently against its side?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene illuminated by a solemn, stage-like Jesus light piercing through thick, dark clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera executing a steady, low-angle parallel tracking shot following the ship?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the radar antenna on the ship's bridge shown as the only steady movement in 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": [ + -0.062038760632276535 + ], + "opposite_logprob_tokens": [ + -2.8120386600494385 + ], + "logprob_reward": -0.062038760632276535, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.062038760632276535, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3252420425415039, + 0.059209153056144714, + 0.12122959643602371, + 0.05079633370041847, + 0.09173141419887543, + 0.0419754832983017, + 0.03265509009361267, + 0.11030670255422592, + 0.0540735088288784, + 0.11278067529201508 + ], + "credit_map": [ + 0.02598566561937332, + 0.01668490655720234, + 0.013714166358113289, + 0.0113139059394598, + 0.021517084911465645, + 0.13227896392345428, + 0.02432049624621868, + 0.07942686229944229, + 0.007247910834848881, + 0.006843902636319399, + 0.005825154483318329, + 0.0065613859333097935, + 0.00693309772759676, + 0.012230119667947292, + 0.007399815134704113, + 0.006167765241116285, + 0.008511469699442387, + 0.013235179707407951, + 0.0077681043185293674, + 0.01475986186414957, + 0.014503690414130688, + 0.016716646030545235, + 0.02850586734712124, + 0.017228776589035988, + 0.007115091197192669, + 0.008123692125082016, + 0.005885380320250988, + 0.005868616513907909, + 0.005444144364446402, + 0.005786154884845018, + 0.007508442737162113, + 0.005064806900918484, + 0.008839148096740246, + 0.016460023820400238, + 0.006573474500328302, + 0.015971427783370018, + 0.009693045169115067, + 0.011999771930277348, + 0.012424220331013203, + 0.009770304895937443, + 0.0028085384983569384, + 0.006797438953071833, + 0.0031716141384094954, + 0.0072112311609089375, + 0.0058900765143334866, + 0.005006765015423298, + 0.006245631258934736, + 0.004844187758862972, + 0.0018583553610369563, + 0.003124604932963848, + 0.001635297667235136, + 0.004543771035969257, + 0.004018695559352636, + 0.004917565267533064, + 0.006205196026712656, + 0.006351600866764784, + 0.01291132066398859, + 0.009770972654223442, + 0.008903696201741695, + 0.010428331792354584, + 0.019311899319291115, + 0.019595656543970108, + 0.016131283715367317, + 0.01325353141874075, + 0.005991427693516016, + 0.008081437088549137, + 0.004576627165079117, + 0.005424798466265202, + 0.011017056182026863, + 0.007415653672069311, + 0.0059618521481752396, + 0.0056046550162136555, + 0.013902585953474045, + 0.011675330810248852, + 0.009050070308148861, + 0.01583881489932537, + 0.018716901540756226, + 0.016540657728910446, + 0.013292244635522366, + 0.01376407127827406 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 9 + ], + "selected_reward": -0.062038760632276535 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a historic destroyer with a weathered, dark gray hull showing rust and scratches?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the ship moving through deep blue, almost black waters with mountainous waves crashing violently against its side?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene illuminated by a solemn, stage-like Jesus light piercing through thick, dark clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera executing a steady, low-angle parallel tracking shot following the ship?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the radar antenna on the ship's bridge shown as the only steady movement in 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": [ + -0.1428384780883789 + ], + "opposite_logprob_tokens": [ + -2.017838478088379 + ], + "logprob_reward": -0.1428384780883789, + "yes_no_margin_reward": 1.875, + "token_credit_reward": -0.1428384780883789, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.6582561135292053, + 0.03558344766497612, + 0.04998147487640381, + 0.034612663090229034, + 0.040266454219818115, + 0.01492620725184679, + 0.030093220993876457, + 0.05790746212005615, + 0.043924398720264435, + 0.03444855287671089 + ], + "credit_map": [ + 0.038440361618995667, + 0.026348499581217766, + 0.018893729895353317, + 0.008177480660378933, + 0.03810720518231392, + 0.2920920252799988, + 0.034863609820604324, + 0.2013331800699234, + 0.004936115816235542, + 0.004293873440474272, + 0.0031767981126904488, + 0.00344682764261961, + 0.003916341811418533, + 0.006370934657752514, + 0.004956445656716824, + 0.0044861119240522385, + 0.0028499895706772804, + 0.00408377218991518, + 0.0025364982429891825, + 0.0037768157199025154, + 0.008189598098397255, + 0.006684842519462109, + 0.011097545735538006, + 0.0107624102383852, + 0.003156092017889023, + 0.0035119473468512297, + 0.0028739955741912127, + 0.0036616483703255653, + 0.005324202124029398, + 0.0047159865498542786, + 0.005760826636105776, + 0.0056079658679664135, + 0.0036165358033031225, + 0.006418445613235235, + 0.003817650955170393, + 0.0074170841835439205, + 0.0040820143185555935, + 0.005172475706785917, + 0.006035932339727879, + 0.003706317860633135, + 0.0013689654879271984, + 0.0024896604008972645, + 0.0007754211546853185, + 0.001738810446113348, + 0.0026261743623763323, + 0.002374788746237755, + 0.0016435710713267326, + 0.0019088154658675194, + 0.002142632845789194, + 0.002782512689009309, + 0.0014832548331469297, + 0.005229761824011803, + 0.005519035737961531, + 0.004018549341708422, + 0.0048446315340697765, + 0.004072841722518206, + 0.008316190913319588, + 0.006568355485796928, + 0.00447336258366704, + 0.00475557753816247, + 0.009657803922891617, + 0.011263031512498856, + 0.005781305022537708, + 0.007091831415891647, + 0.005246711894869804, + 0.005102992989122868, + 0.004095165990293026, + 0.004841837100684643, + 0.008768610656261444, + 0.007008473388850689, + 0.004732792731374502, + 0.004127816762775183, + 0.0042791846208274364, + 0.0044076708145439625, + 0.002906333189457655, + 0.004326439928263426, + 0.006301924120634794, + 0.004489825572818518, + 0.0038178632967174053, + 0.0039193094708025455 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 2 + ], + "selected_reward": -0.1428384780883789 + } + ], + "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/0000121/group_12/index_tensor([3563, 3563])_group_12_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000121/group_12/index_tensor([3563, 3563])_group_12_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..bc64f6f80219e80866c4fb21e168e547a6ea8b63 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000121/group_12/index_tensor([3563, 3563])_group_12_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "slightly high-angle macro lens triggers a very slow, circular tracking motion, immersing the viewer in the delicate courtship of two stunning pink dragon millipedes. Their vibrant, segmented bodies, a stunning hue of purple, are covered in intricate, spiny protrusions that give them an alien texture. They move with hypnotic grace, slowly circling each other over a miniature landscape of lush, emerald-green moss, still shimmering with tiny droplets of morning dew. As they orbit, their delicate, silky antennae delicately tap against each other in an exploratory dance, while the rhythmic, undulating waves of their countless tiny legs propel them at a steady, deliberate pace. Dappled sunlight filters through the unseen canopy, casting a shifting pattern of warm light and cool shadow across the scene and creating bright, mirror-like highlights on their smooth, carapace-like exoskeletons. The intense color contrasts, the extremely shallow depth of field that blurs the background into a creamy bokeh, and the unhurried pace of the camera all combine to create an intimate, focused, and utterly otherworldly atmosphere that emphasizes every minute textural detail and transforms the encounter into a moment of mesmerizing natural beauty.", + "seed": 7369, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000121/group_12/index_3563_seed_7369_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": -0.032323095481842756, + "vlm_logprob": -0.032323095481842756, + "vlm_yes_no_margin": 3.875, + "vlm_token_credit": -0.032323095481842756, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" 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 pink dragon millipedes visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the millipedes have vibrant purple segmented bodies with intricate spiny protrusions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the millipedes moving in a slow circular orbit over a landscape of emerald-green moss with dew droplets?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video feature a high-angle macro shot with extremely shallow depth of field creating a creamy bokeh background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera executing a very slow, circular tracking motion while dappled sunlight creates shifting light and shadow patterns?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04479487985372543 + ], + "opposite_logprob_tokens": [ + -3.169794797897339 + ], + "logprob_reward": -0.04479487985372543, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04479487985372543, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21685175597667694, + 0.09315560758113861, + 0.06334163248538971, + 0.05655974894762039, + 0.08387427031993866, + 0.2009703665971756, + 0.13329561054706573, + 0.06287986785173416, + 0.05048482492566109, + 0.03858628869056702 + ], + "credit_map": [ + 0.030729318037629128, + 0.034473687410354614, + 0.020748944953083992, + 0.018901996314525604, + 0.037227563560009, + 0.02010280452668667, + 0.03269520029425621, + 0.021972233429551125, + 0.009631993249058723, + 0.017932424321770668, + 0.009942601434886456, + 0.012316049076616764, + 0.011637886054813862, + 0.01214705128222704, + 0.008807315491139889, + 0.010740285739302635, + 0.00582081126049161, + 0.0051795560866594315, + 0.007157224230468273, + 0.0064923083409667015, + 0.014754789881408215, + 0.007588325999677181, + 0.008474919013679028, + 0.00787369441241026, + 0.006109891925007105, + 0.008488561026751995, + 0.006688747089356184, + 0.005044666118919849, + 0.012499778531491756, + 0.005617658607661724, + 0.005155659280717373, + 0.0069547840394079685, + 0.023880135267972946, + 0.005524363834410906, + 0.006938987877219915, + 0.0034390867222100496, + 0.004435813054442406, + 0.007946011610329151, + 0.003925912082195282, + 0.027783960103988647, + 0.022640977054834366, + 0.025583025068044662, + 0.04034583643078804, + 0.026919348165392876, + 0.026964087039232254, + 0.020791659131646156, + 0.019980819895863533, + 0.01774461753666401, + 0.016043227165937424, + 0.011125500313937664, + 0.02023891918361187, + 0.027461426332592964, + 0.014510445296764374, + 0.01634971611201763, + 0.013787196017801762, + 0.0137791708111763, + 0.017154481261968613, + 0.00492267357185483, + 0.005735486280173063, + 0.005312765017151833, + 0.004728915635496378, + 0.0052376980893313885, + 0.003648185869678855, + 0.01613965816795826, + 0.005850685760378838, + 0.008279784582555294, + 0.008821147494018078, + 0.006160506047308445, + 0.007475154474377632, + 0.003822864731773734, + 0.0064622219651937485, + 0.0036124589387327433, + 0.004150393418967724, + 0.0028661577962338924, + 0.004544158931821585, + 0.003452551318332553, + 0.004075821489095688, + 0.002953288145363331, + 0.012609175406396389, + 0.003934742882847786 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 6 + ], + "selected_reward": -0.04479487985372543 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" 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 pink dragon millipedes visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the millipedes have vibrant purple segmented bodies with intricate spiny protrusions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the millipedes moving in a slow circular orbit over a landscape of emerald-green moss with dew droplets?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video feature a high-angle macro shot with extremely shallow depth of field creating a creamy bokeh background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera executing a very slow, circular tracking motion while dappled sunlight creates shifting light and shadow patterns?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01607411541044712 + ], + "opposite_logprob_tokens": [ + -4.141074180603027 + ], + "logprob_reward": -0.01607411541044712, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.01607411541044712, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3935421407222748, + 0.1148562878370285, + 0.023001085966825485, + 0.09732259809970856, + 0.026121996343135834, + 0.048816949129104614, + 0.07222927361726761, + 0.037611640989780426, + 0.05596008524298668, + 0.13053786754608154 + ], + "credit_map": [ + 0.033762864768505096, + 0.04772873595356941, + 0.05505627393722534, + 0.044090066105127335, + 0.06574362516403198, + 0.04812900722026825, + 0.061596013605594635, + 0.03743554651737213, + 0.011600143276154995, + 0.012320993468165398, + 0.01377051044255495, + 0.02370329573750496, + 0.014867263846099377, + 0.013822787441313267, + 0.012593282386660576, + 0.01217801682651043, + 0.0021796575747430325, + 0.0020009318832308054, + 0.00259884144179523, + 0.0031947840470820665, + 0.003839306067675352, + 0.0023985656443983316, + 0.002916621509939432, + 0.003872377099469304, + 0.00899691041558981, + 0.007851172238588333, + 0.007843920961022377, + 0.012024455703794956, + 0.030927017331123352, + 0.010540146380662918, + 0.009100311435759068, + 0.010038667358458042, + 0.0036620430182665586, + 0.002737868344411254, + 0.0033684861846268177, + 0.003479590406641364, + 0.0023384487722069025, + 0.003215618897229433, + 0.001979254884645343, + 0.005340686533600092, + 0.008874154649674892, + 0.0067436289973556995, + 0.009549228474497795, + 0.0055138543248176575, + 0.006264300551265478, + 0.004100316669791937, + 0.00367525452747941, + 0.004096208605915308, + 0.00840976182371378, + 0.005903247743844986, + 0.011680263094604015, + 0.014143968932330608, + 0.007544041611254215, + 0.009230797179043293, + 0.0061707994900643826, + 0.009146394208073616, + 0.009655424393713474, + 0.002951190108433366, + 0.0032152452040463686, + 0.0034510621335357428, + 0.003681605914607644, + 0.0030900267884135246, + 0.0026942105032503605, + 0.00887287501245737, + 0.0059868451207876205, + 0.008600397035479546, + 0.006823881529271603, + 0.006781646050512791, + 0.008299642242491245, + 0.005894520319998264, + 0.008951246738433838, + 0.004621908534318209, + 0.018767619505524635, + 0.02121683396399021, + 0.031382013112306595, + 0.013339318335056305, + 0.01252274215221405, + 0.011996951885521412, + 0.011967095546424389, + 0.009345298632979393 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 1 + ], + "selected_reward": -0.01607411541044712 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" 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 pink dragon millipedes visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the millipedes have vibrant purple segmented bodies with intricate spiny protrusions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the millipedes moving in a slow circular orbit over a landscape of emerald-green moss with dew droplets?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video feature a high-angle macro shot with extremely shallow depth of field creating a creamy bokeh background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera executing a very slow, circular tracking motion while dappled sunlight creates shifting light and shadow patterns?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07895860075950623 + ], + "opposite_logprob_tokens": [ + -2.578958511352539 + ], + "logprob_reward": -0.07895860075950623, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07895860075950623, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.45451563596725464, + 0.06230306997895241, + 0.01782500371336937, + 0.04613036289811134, + 0.040050819516181946, + 0.1617603749036789, + 0.06080817058682442, + 0.09773295372724533, + 0.02786909230053425, + 0.03100455179810524 + ], + "credit_map": [ + 0.036167435348033905, + 0.04285270720720291, + 0.06615086644887924, + 0.051174793392419815, + 0.07486601918935776, + 0.05985035002231598, + 0.07853856682777405, + 0.04491492360830307, + 0.005939886439591646, + 0.004913599696010351, + 0.005868331529200077, + 0.007403024937957525, + 0.005440045613795519, + 0.011182422749698162, + 0.013508052565157413, + 0.008047706447541714, + 0.0016297186957672238, + 0.0015959906158968806, + 0.003229977795854211, + 0.0038156076334416866, + 0.001572824432514608, + 0.0016481123166158795, + 0.002350570634007454, + 0.001982202287763357, + 0.005247514229267836, + 0.003987632226198912, + 0.0031628175638616085, + 0.003974274732172489, + 0.01492262352257967, + 0.005396179389208555, + 0.004573190584778786, + 0.004866131581366062, + 0.008909422904253006, + 0.003583563957363367, + 0.004645777866244316, + 0.002606408204883337, + 0.004225854761898518, + 0.0037366372998803854, + 0.0026472103781998158, + 0.009695948101580143, + 0.024158261716365814, + 0.019892454147338867, + 0.032971277832984924, + 0.020020101219415665, + 0.021129343658685684, + 0.014643408358097076, + 0.015227710828185081, + 0.013717811554670334, + 0.00976886972784996, + 0.005772549193352461, + 0.007127109915018082, + 0.010873330757021904, + 0.006085703615099192, + 0.006853792350739241, + 0.007037520408630371, + 0.00728929927572608, + 0.028719065710902214, + 0.008188373409211636, + 0.0064020282588899136, + 0.006631402764469385, + 0.00819079577922821, + 0.0077237519435584545, + 0.0050370036624372005, + 0.026840535923838615, + 0.0039835041388869286, + 0.003998341970145702, + 0.003367012832313776, + 0.00361213949508965, + 0.004657703451812267, + 0.002648408990353346, + 0.002575075952336192, + 0.003026907565072179, + 0.003936988301575184, + 0.0034931940026581287, + 0.004398479126393795, + 0.0029425392858684063, + 0.0028839632868766785, + 0.0025554669555276632, + 0.008214030414819717, + 0.0025798925198614597 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 7 + ], + "selected_reward": -0.07895860075950623 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" 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 pink dragon millipedes visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the millipedes have vibrant purple segmented bodies with intricate spiny protrusions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the millipedes moving in a slow circular orbit over a landscape of emerald-green moss with dew droplets?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video feature a high-angle macro shot with extremely shallow depth of field creating a creamy bokeh background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera executing a very slow, circular tracking motion while dappled sunlight creates shifting light and shadow patterns?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0036108787171542645 + ], + "opposite_logprob_tokens": [ + -5.628611087799072 + ], + "logprob_reward": -0.0036108787171542645, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.0036108787171542645, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1576816439628601, + 0.06139868497848511, + 0.05465685948729515, + 0.06947686523199081, + 0.13918709754943848, + 0.21583257615566254, + 0.16739504039287567, + 0.06948114931583405, + 0.014324887655675411, + 0.05056523159146309 + ], + "credit_map": [ + 0.01219017431139946, + 0.012843352742493153, + 0.023008301854133606, + 0.018535299226641655, + 0.02560393325984478, + 0.019902464002370834, + 0.03085765428841114, + 0.014740458689630032, + 0.004450083244591951, + 0.00528574176132679, + 0.00915492419153452, + 0.01109382975846529, + 0.004640590865164995, + 0.009386762045323849, + 0.00997846107929945, + 0.00740828737616539, + 0.0052497535943984985, + 0.0049962797202169895, + 0.007077727932482958, + 0.006230430211871862, + 0.008157444186508656, + 0.00809802208095789, + 0.008242273703217506, + 0.00660492479801178, + 0.007849969901144505, + 0.004625760950148106, + 0.005550770089030266, + 0.006160207558423281, + 0.023010799661278725, + 0.00780306663364172, + 0.007182521745562553, + 0.007293761242181063, + 0.03461150452494621, + 0.012292715720832348, + 0.01627759449183941, + 0.010412098839879036, + 0.010810931213200092, + 0.011762612499296665, + 0.007950096391141415, + 0.035069528967142105, + 0.038850169628858566, + 0.027236558496952057, + 0.041867904365062714, + 0.024712955579161644, + 0.026608766987919807, + 0.01904396153986454, + 0.018886089324951172, + 0.018626166507601738, + 0.024348970502614975, + 0.017117979004979134, + 0.024022389203310013, + 0.026222025975584984, + 0.021402869373559952, + 0.019320884719491005, + 0.013126636855304241, + 0.021833280101418495, + 0.020865488797426224, + 0.005916486959904432, + 0.0044980738312006, + 0.003403882961720228, + 0.00654944684356451, + 0.006470049731433392, + 0.0030867489986121655, + 0.018690969794988632, + 0.002100791083648801, + 0.0021434309892356396, + 0.0020144707523286343, + 0.0017247663345187902, + 0.001954434672370553, + 0.001404335955157876, + 0.0014036798384040594, + 0.0015789770986884832, + 0.005500522907823324, + 0.0036423264537006617, + 0.005464574322104454, + 0.0041962615214288235, + 0.004145135637372732, + 0.0031523939687758684, + 0.019112814217805862, + 0.005351196974515915 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 6, + 0 + ], + "selected_reward": -0.0036108787171542645 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" 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 pink dragon millipedes visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the millipedes have vibrant purple segmented bodies with intricate spiny protrusions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the millipedes moving in a slow circular orbit over a landscape of emerald-green moss with dew droplets?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video feature a high-angle macro shot with extremely shallow depth of field creating a creamy bokeh background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera executing a very slow, circular tracking motion while dappled sunlight creates shifting light and shadow patterns?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018177002668380737 + ], + "opposite_logprob_tokens": [ + -4.018177032470703 + ], + "logprob_reward": -0.018177002668380737, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018177002668380737, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20639458298683167, + 0.1078319400548935, + 0.03298158571124077, + 0.03490225970745087, + 0.056681305170059204, + 0.3563793897628784, + 0.07910509407520294, + 0.061174120754003525, + 0.02930619940161705, + 0.0352434404194355 + ], + "credit_map": [ + 0.025962913408875465, + 0.040452741086483, + 0.023530276492238045, + 0.019301755353808403, + 0.025673551484942436, + 0.02110978029668331, + 0.029017068445682526, + 0.02134651318192482, + 0.011852818541228771, + 0.009699937887489796, + 0.013977475464344025, + 0.024858757853507996, + 0.012858757749199867, + 0.010928701609373093, + 0.01276642456650734, + 0.01088907290250063, + 0.002826034789904952, + 0.002596792997792363, + 0.003738994477316737, + 0.003772488096728921, + 0.0059043485671281815, + 0.004488125443458557, + 0.005022623110562563, + 0.004632178694009781, + 0.0038235217798501253, + 0.003323374781757593, + 0.00327423051930964, + 0.003727759001776576, + 0.009460918605327606, + 0.004373285453766584, + 0.003284422680735588, + 0.0036347468849271536, + 0.01077196467667818, + 0.004616767633706331, + 0.010090077295899391, + 0.0058356099762022495, + 0.00411115400493145, + 0.006044090259820223, + 0.003569344524294138, + 0.011642304249107838, + 0.038482196629047394, + 0.0451914444565773, + 0.06819440424442291, + 0.04903077706694603, + 0.048860255628824234, + 0.03785385191440582, + 0.03760101646184921, + 0.031165463849902153, + 0.008403225801885128, + 0.007497056387364864, + 0.012282627634704113, + 0.017797907814383507, + 0.008488931693136692, + 0.008780799806118011, + 0.009294520132243633, + 0.006560030858963728, + 0.016194988042116165, + 0.005928382743149996, + 0.00708498852327466, + 0.004142362158745527, + 0.00530906580388546, + 0.004729896783828735, + 0.003217345103621483, + 0.01456709485501051, + 0.0036664726212620735, + 0.0042524696327745914, + 0.003885357640683651, + 0.004174632951617241, + 0.004385555163025856, + 0.002699863398447633, + 0.003081670729443431, + 0.003160178428515792, + 0.0043024905025959015, + 0.0041152918711304665, + 0.006532816216349602, + 0.002697051502764225, + 0.003713792422786355, + 0.0029522785916924477, + 0.008245982229709625, + 0.00268374290317297 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 1 + ], + "selected_reward": -0.018177002668380737 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "slightly high-angle macro lens triggers a very slow, circular tracking motion, immersing the viewer in the delicate courtship of two stunning pink dragon millipedes. Their vibrant, segmented bodies, a stunning hue of purple, are covered in intricate, spiny protrusions that give them an alien texture. They move with hypnotic grace, slowly circling each other over a miniature landscape of lush, emerald-green moss, still shimmering with tiny droplets of morning dew. As they orbit, their delicate, silky antennae delicately tap against each other in an exploratory dance, while the rhythmic, undulating waves of their countless tiny legs propel them at a steady, deliberate pace. Dappled sunlight filters through the unseen canopy, casting a shifting pattern of warm light and cool shadow across the scene and creating bright, mirror-like highlights on their smooth, carapace-like exoskeletons. The intense color contrasts, the extremely shallow depth of field that blurs the background into a creamy bokeh, and the unhurried pace of the camera all combine to create an intimate, focused, and utterly otherworldly atmosphere that emphasizes every minute textural detail and transforms the encounter into a moment of mesmerizing natural beauty.", + "seed": 7370, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000121/group_12/index_3563_seed_7370_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": -0.02766597382724285, + "vlm_logprob": -0.02766597382724285, + "vlm_yes_no_margin": 4.075, + "vlm_token_credit": -0.02766597382724285, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" 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 pink dragon millipedes visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the millipedes have vibrant purple segmented bodies with intricate spiny protrusions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the millipedes moving in a slow circular orbit over a landscape of emerald-green moss with dew droplets?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video feature a high-angle macro shot with extremely shallow depth of field creating a creamy bokeh background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera executing a very slow, circular tracking motion while dappled sunlight creates shifting light and shadow patterns?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08046817779541016 + ], + "opposite_logprob_tokens": [ + -2.58046817779541 + ], + "logprob_reward": -0.08046817779541016, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.08046817779541016, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11581628024578094, + 0.08126715570688248, + 0.15816813707351685, + 0.18140891194343567, + 0.11604399234056473, + 0.08769159764051437, + 0.07073500007390976, + 0.032612916082143784, + 0.026848042383790016, + 0.12940798699855804 + ], + "credit_map": [ + 0.020745716989040375, + 0.01037067361176014, + 0.010289437137544155, + 0.009956995025277138, + 0.009272420778870583, + 0.014562639407813549, + 0.013780388049781322, + 0.02683800272643566, + 0.013735642656683922, + 0.009957049041986465, + 0.017846914008259773, + 0.008315476588904858, + 0.007497177924960852, + 0.006662908475846052, + 0.007603503298014402, + 0.009648478589951992, + 0.021411191672086716, + 0.029333030804991722, + 0.017245745286345482, + 0.016496727243065834, + 0.011161684058606625, + 0.011885147541761398, + 0.020898424088954926, + 0.02973618544638157, + 0.02252940461039543, + 0.012198066338896751, + 0.03906736895442009, + 0.02263358049094677, + 0.012016364373266697, + 0.03610354661941528, + 0.011808635666966438, + 0.025051943957805634, + 0.013379284180700779, + 0.0102908406406641, + 0.009538033045828342, + 0.02871910110116005, + 0.003596558002755046, + 0.006584653165191412, + 0.005525870248675346, + 0.038409650325775146, + 0.03125620260834694, + 0.014270029030740261, + 0.00838000699877739, + 0.007650382816791534, + 0.005615061614662409, + 0.0061675203032791615, + 0.007872728630900383, + 0.006479661446064711, + 0.00867748074233532, + 0.017633143812417984, + 0.007349591236561537, + 0.007015688810497522, + 0.009231179021298885, + 0.008975958451628685, + 0.0059700473211705685, + 0.005881916731595993, + 0.004813368432223797, + 0.003155656158924103, + 0.005499304737895727, + 0.002959625329822302, + 0.004089120775461197, + 0.00488671101629734, + 0.0031871439423412085, + 0.0040219868533313274, + 0.005616046022623777, + 0.0030721602961421013, + 0.00282447412610054, + 0.003114635357633233, + 0.002219795249402523, + 0.004492857959121466, + 0.0020328352693468332, + 0.0034752399660646915, + 0.02019428461790085, + 0.010887187905609608, + 0.011418919079005718, + 0.009139337576925755, + 0.040779829025268555, + 0.012074735015630722, + 0.012952437624335289, + 0.011961258016526699 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 2, + 9 + ], + "selected_reward": -0.08046817779541016 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" 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 pink dragon millipedes visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the millipedes have vibrant purple segmented bodies with intricate spiny protrusions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the millipedes moving in a slow circular orbit over a landscape of emerald-green moss with dew droplets?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video feature a high-angle macro shot with extremely shallow depth of field creating a creamy bokeh background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera executing a very slow, circular tracking motion while dappled sunlight creates shifting light and shadow patterns?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009789552539587021 + ], + "opposite_logprob_tokens": [ + -4.63478946685791 + ], + "logprob_reward": -0.009789552539587021, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009789552539587021, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1254090964794159, + 0.09782623499631882, + 0.02855391427874565, + 0.12161849439144135, + 0.17502008378505707, + 0.052753034979104996, + 0.12378038465976715, + 0.1324334442615509, + 0.05452132225036621, + 0.08808401226997375 + ], + "credit_map": [ + 0.01849370449781418, + 0.009825232438743114, + 0.021265733987092972, + 0.018312759697437286, + 0.011072670109570026, + 0.01458743680268526, + 0.014285819604992867, + 0.01756572723388672, + 0.015146682970225811, + 0.012788765132427216, + 0.014865263365209103, + 0.01229178998619318, + 0.01184086874127388, + 0.008653038181364536, + 0.009011873044073582, + 0.013227947987616062, + 0.004165546968579292, + 0.005679109133780003, + 0.00407611858099699, + 0.003858159063383937, + 0.002557017607614398, + 0.002535994164645672, + 0.0026325879152864218, + 0.003049379913136363, + 0.02497587539255619, + 0.012377860024571419, + 0.02100754901766777, + 0.012467886321246624, + 0.013285739347338676, + 0.014673781581223011, + 0.00958053395152092, + 0.013249261304736137, + 0.02043738029897213, + 0.020015206187963486, + 0.030353017151355743, + 0.047964222729206085, + 0.007348816841840744, + 0.009571307338774204, + 0.018023531883955002, + 0.021306583657860756, + 0.010141003876924515, + 0.01394131314009428, + 0.00485854921862483, + 0.006579256150871515, + 0.003545149462297559, + 0.003174156416207552, + 0.005197471939027309, + 0.005316129885613918, + 0.018795613199472427, + 0.016549279913306236, + 0.018999522551894188, + 0.009545878507196903, + 0.012034508399665356, + 0.026195934042334557, + 0.010300093330442905, + 0.011359550058841705, + 0.020606588572263718, + 0.013222343288362026, + 0.02610497735440731, + 0.011357050389051437, + 0.01387182530015707, + 0.01874574087560177, + 0.0127413934096694, + 0.015783509239554405, + 0.012468271888792515, + 0.0056228432804346085, + 0.005731641780585051, + 0.0055039809085428715, + 0.0033104007598012686, + 0.0075193908996880054, + 0.0036775474436581135, + 0.010687236674129963, + 0.008122279308736324, + 0.008140540681779385, + 0.01844261772930622, + 0.005442723166197538, + 0.027828142046928406, + 0.007130099926143885, + 0.006716851610690355, + 0.006260755471885204 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 7, + 0 + ], + "selected_reward": -0.009789552539587021 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" 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 pink dragon millipedes visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the millipedes have vibrant purple segmented bodies with intricate spiny protrusions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the millipedes moving in a slow circular orbit over a landscape of emerald-green moss with dew droplets?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video feature a high-angle macro shot with extremely shallow depth of field creating a creamy bokeh background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera executing a very slow, circular tracking motion while dappled sunlight creates shifting light and shadow patterns?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02979137934744358 + ], + "opposite_logprob_tokens": [ + -3.5297913551330566 + ], + "logprob_reward": -0.02979137934744358, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.02979137934744358, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12458819895982742, + 0.046382978558540344, + 0.15522518754005432, + 0.09943003207445145, + 0.06998305767774582, + 0.04519305378198624, + 0.09713547676801682, + 0.06746940314769745, + 0.043164126574993134, + 0.251428484916687 + ], + "credit_map": [ + 0.021360065788030624, + 0.010443616658449173, + 0.011401692405343056, + 0.009563653729856014, + 0.008511041291058064, + 0.015477542765438557, + 0.014064256101846695, + 0.033766329288482666, + 0.007321413606405258, + 0.004382921848446131, + 0.007092985790222883, + 0.005738291889429092, + 0.004299772437661886, + 0.004076427314430475, + 0.004361359402537346, + 0.009109809063374996, + 0.035661082714796066, + 0.011740543879568577, + 0.047368425875902176, + 0.01172963622957468, + 0.005359945818781853, + 0.00735695194453001, + 0.01176969800144434, + 0.024238893762230873, + 0.017444642260670662, + 0.00886608101427555, + 0.017627529799938202, + 0.010418297722935677, + 0.01380577776581049, + 0.011645233258605003, + 0.00803662370890379, + 0.01158584002405405, + 0.005633547902107239, + 0.00901237316429615, + 0.010716911405324936, + 0.021869994699954987, + 0.0038468027487397194, + 0.005378175061196089, + 0.00649663433432579, + 0.007028618827462196, + 0.013623197562992573, + 0.004778358153998852, + 0.004671531263738871, + 0.005285919178277254, + 0.002460027113556862, + 0.0026244402397423983, + 0.005988911259919405, + 0.005760669242590666, + 0.013099931180477142, + 0.007036922499537468, + 0.014778180047869682, + 0.006944551598280668, + 0.010059713385999203, + 0.027267251163721085, + 0.008710896596312523, + 0.009238036349415779, + 0.007625499274581671, + 0.01190987043082714, + 0.01126919873058796, + 0.0070112538523972034, + 0.007213448639959097, + 0.007360728457570076, + 0.006270849611610174, + 0.008808552287518978, + 0.004466512240469456, + 0.004574065562337637, + 0.003895236412063241, + 0.010512172244489193, + 0.003911913838237524, + 0.004524350166320801, + 0.006593275815248489, + 0.004686599597334862, + 0.023897744715213776, + 0.020841991528868675, + 0.04233089089393616, + 0.01939135417342186, + 0.08840760588645935, + 0.021406657993793488, + 0.017195990309119225, + 0.017956268042325974 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 2, + 0 + ], + "selected_reward": -0.02979137934744358 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" 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 pink dragon millipedes visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the millipedes have vibrant purple segmented bodies with intricate spiny protrusions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the millipedes moving in a slow circular orbit over a landscape of emerald-green moss with dew droplets?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video feature a high-angle macro shot with extremely shallow depth of field creating a creamy bokeh background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera executing a very slow, circular tracking motion while dappled sunlight creates shifting light and shadow patterns?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0040903957560658455 + ], + "opposite_logprob_tokens": [ + -5.504090309143066 + ], + "logprob_reward": -0.0040903957560658455, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.0040903957560658455, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0712396427989006, + 0.08590985834598541, + 0.053213875740766525, + 0.13925348222255707, + 0.1893346607685089, + 0.15300625562667847, + 0.02156001329421997, + 0.029390405863523483, + 0.03481081500649452, + 0.22228094935417175 + ], + "credit_map": [ + 0.012502606026828289, + 0.006077218800783157, + 0.00798701960593462, + 0.007744441740214825, + 0.0058922539465129375, + 0.00867489818483591, + 0.007958809845149517, + 0.014402397908270359, + 0.014829563908278942, + 0.005146874580532312, + 0.008476287126541138, + 0.009548134170472622, + 0.005732505116611719, + 0.006380685139447451, + 0.010169625282287598, + 0.025626186281442642, + 0.009865985251963139, + 0.005962484050542116, + 0.013115541078150272, + 0.00484295841306448, + 0.002332641975954175, + 0.002465662080794573, + 0.005450638942420483, + 0.009177964180707932, + 0.03298540413379669, + 0.015345186926424503, + 0.013266190886497498, + 0.009232559241354465, + 0.017576782032847404, + 0.03279951214790344, + 0.00606928113847971, + 0.011978560127317905, + 0.025606397539377213, + 0.017882630228996277, + 0.034892890602350235, + 0.05852144584059715, + 0.006022351328283548, + 0.008058363571763039, + 0.01918020285665989, + 0.019170386716723442, + 0.017906568944454193, + 0.027755828574299812, + 0.01967730186879635, + 0.01916741393506527, + 0.012190843001008034, + 0.010442033410072327, + 0.021880073472857475, + 0.023986179381608963, + 0.0029676882550120354, + 0.004188825376331806, + 0.003094213083386421, + 0.002962663536891341, + 0.0020901053212583065, + 0.0022616058122366667, + 0.002088711131364107, + 0.0019062019418925047, + 0.00413513882085681, + 0.002485537901520729, + 0.004892919212579727, + 0.0026427379343658686, + 0.004450374748557806, + 0.003926228731870651, + 0.0024035414680838585, + 0.004453927278518677, + 0.004363040905445814, + 0.003172138473019004, + 0.005033263936638832, + 0.003959606401622295, + 0.0030582512263208628, + 0.005406619515269995, + 0.004133269656449556, + 0.005684626288712025, + 0.017354045063257217, + 0.012093855068087578, + 0.04740368574857712, + 0.015661729499697685, + 0.09093287587165833, + 0.013646779581904411, + 0.011243153363466263, + 0.013944814912974834 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 4, + 5 + ], + "selected_reward": -0.0040903957560658455 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" 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 pink dragon millipedes visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the millipedes have vibrant purple segmented bodies with intricate spiny protrusions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the millipedes moving in a slow circular orbit over a landscape of emerald-green moss with dew droplets?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video feature a high-angle macro shot with extremely shallow depth of field creating a creamy bokeh background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera executing a very slow, circular tracking motion while dappled sunlight creates shifting light and shadow patterns?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014190363697707653 + ], + "opposite_logprob_tokens": [ + -4.264190196990967 + ], + "logprob_reward": -0.014190363697707653, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014190363697707653, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.204899862408638, + 0.04160623997449875, + 0.14465665817260742, + 0.1564236283302307, + 0.049451667815446854, + 0.06526796519756317, + 0.1338578760623932, + 0.08026956021785736, + 0.06595727056264877, + 0.05760933831334114 + ], + "credit_map": [ + 0.04718966782093048, + 0.018234698101878166, + 0.01650836504995823, + 0.016503088176250458, + 0.014222349971532822, + 0.025093184784054756, + 0.019957467913627625, + 0.047191016376018524, + 0.007187190931290388, + 0.003405837109312415, + 0.0061590662226080894, + 0.007189528550952673, + 0.0035040942020714283, + 0.0032430374994874, + 0.004558462183922529, + 0.006359024904668331, + 0.039572350680828094, + 0.009185027331113815, + 0.04691363871097565, + 0.009636717848479748, + 0.004813298583030701, + 0.005294555798172951, + 0.009342866018414497, + 0.019898198544979095, + 0.01649465411901474, + 0.014009024947881699, + 0.02029135823249817, + 0.012851007282733917, + 0.021663213148713112, + 0.027642004191875458, + 0.014140261337161064, + 0.029332106932997704, + 0.005326942075043917, + 0.004437481053173542, + 0.006776740308851004, + 0.013169734738767147, + 0.005123800598084927, + 0.0037470809184014797, + 0.0055462634190917015, + 0.005323624238371849, + 0.011266760528087616, + 0.011932742781937122, + 0.007551310118287802, + 0.00982577446848154, + 0.005518289282917976, + 0.00456745270639658, + 0.007306351326406002, + 0.007299288176000118, + 0.01889030449092388, + 0.018903130665421486, + 0.020853480324149132, + 0.010839109309017658, + 0.012471592053771019, + 0.029626931995153427, + 0.012205656617879868, + 0.01006767526268959, + 0.008346864953637123, + 0.011023195460438728, + 0.01436634175479412, + 0.008549697697162628, + 0.007122521288692951, + 0.009590431116521358, + 0.007864403538405895, + 0.013406103476881981, + 0.011524897068738937, + 0.006144395098090172, + 0.006698195822536945, + 0.008896706625819206, + 0.006255037616938353, + 0.01180504821240902, + 0.007894912734627724, + 0.006738079246133566, + 0.009120479226112366, + 0.006553678307682276, + 0.007237236015498638, + 0.00484005780890584, + 0.013736862689256668, + 0.004863090347498655, + 0.004952870775014162, + 0.006305061746388674 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 2 + ], + "selected_reward": -0.014190363697707653 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "slightly high-angle macro lens triggers a very slow, circular tracking motion, immersing the viewer in the delicate courtship of two stunning pink dragon millipedes. Their vibrant, segmented bodies, a stunning hue of purple, are covered in intricate, spiny protrusions that give them an alien texture. They move with hypnotic grace, slowly circling each other over a miniature landscape of lush, emerald-green moss, still shimmering with tiny droplets of morning dew. As they orbit, their delicate, silky antennae delicately tap against each other in an exploratory dance, while the rhythmic, undulating waves of their countless tiny legs propel them at a steady, deliberate pace. Dappled sunlight filters through the unseen canopy, casting a shifting pattern of warm light and cool shadow across the scene and creating bright, mirror-like highlights on their smooth, carapace-like exoskeletons. The intense color contrasts, the extremely shallow depth of field that blurs the background into a creamy bokeh, and the unhurried pace of the camera all combine to create an intimate, focused, and utterly otherworldly atmosphere that emphasizes every minute textural detail and transforms the encounter into a moment of mesmerizing natural beauty.", + "seed": 7371, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000121/group_12/index_3563_seed_7371_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": -0.02191411042585969, + "vlm_logprob": -0.02191411042585969, + "vlm_yes_no_margin": 4.425, + "vlm_token_credit": -0.02191411042585969, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" 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 pink dragon millipedes visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the millipedes have vibrant purple segmented bodies with intricate spiny protrusions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the millipedes moving in a slow circular orbit over a landscape of emerald-green moss with dew droplets?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video feature a high-angle macro shot with extremely shallow depth of field creating a creamy bokeh background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera executing a very slow, circular tracking motion while dappled sunlight creates shifting light and shadow patterns?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02464374154806137 + ], + "opposite_logprob_tokens": [ + -3.774643659591675 + ], + "logprob_reward": -0.02464374154806137, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.02464374154806137, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16317446529865265, + 0.039703864604234695, + 0.03911159560084343, + 0.033740438520908356, + 0.05613778904080391, + 0.09660861641168594, + 0.30406099557876587, + 0.05505501106381416, + 0.08557826280593872, + 0.12682892382144928 + ], + "credit_map": [ + 0.012573165819048882, + 0.010069656185805798, + 0.03663010895252228, + 0.013539542444050312, + 0.05060277506709099, + 0.019085682928562164, + 0.008892648853361607, + 0.011780885048210621, + 0.006361857056617737, + 0.0035732940305024385, + 0.0029527884908020496, + 0.002894040895625949, + 0.012734167277812958, + 0.004678498022258282, + 0.0032372032292187214, + 0.003272012574598193, + 0.005368851125240326, + 0.005942159797996283, + 0.004166497383266687, + 0.004985400475561619, + 0.0032728929072618484, + 0.003263771766796708, + 0.005916147958487272, + 0.006195876747369766, + 0.004789137747138739, + 0.006015208549797535, + 0.003667782759293914, + 0.00292672379873693, + 0.0026160033885389566, + 0.004098590929061174, + 0.00584948156028986, + 0.003777510952204466, + 0.010105160996317863, + 0.008250494487583637, + 0.00425646873190999, + 0.011405116878449917, + 0.005013644229620695, + 0.006152909714728594, + 0.005949423648416996, + 0.005004572216421366, + 0.010499868541955948, + 0.013447409495711327, + 0.009491361677646637, + 0.008267325349152088, + 0.020573832094669342, + 0.011113524436950684, + 0.012112866155803204, + 0.011102426797151566, + 0.12213742733001709, + 0.019065456464886665, + 0.01864306628704071, + 0.009824616834521294, + 0.0826236680150032, + 0.014305425807833672, + 0.012987912632524967, + 0.024473434314131737, + 0.006844567134976387, + 0.014646760188043118, + 0.004817560315132141, + 0.006551842670887709, + 0.005768482107669115, + 0.0032122652046382427, + 0.008837848901748657, + 0.004375681281089783, + 0.01050562597811222, + 0.013537393882870674, + 0.007468056865036488, + 0.01035978738218546, + 0.009921460412442684, + 0.006670073606073856, + 0.016690444201231003, + 0.010425415821373463, + 0.011937878094613552, + 0.017732059583067894, + 0.017104968428611755, + 0.0217316634953022, + 0.015344355255365372, + 0.011411999352276325, + 0.014400775544345379, + 0.01716523803770542 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 9 + ], + "selected_reward": -0.02464374154806137 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" 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 pink dragon millipedes visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the millipedes have vibrant purple segmented bodies with intricate spiny protrusions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the millipedes moving in a slow circular orbit over a landscape of emerald-green moss with dew droplets?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video feature a high-angle macro shot with extremely shallow depth of field creating a creamy bokeh background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera executing a very slow, circular tracking motion while dappled sunlight creates shifting light and shadow patterns?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004647050518542528 + ], + "opposite_logprob_tokens": [ + -5.379647254943848 + ], + "logprob_reward": -0.004647050518542528, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004647050518542528, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2840242087841034, + 0.0696229413151741, + 0.028193321079015732, + 0.021955199539661407, + 0.03374265134334564, + 0.05888982489705086, + 0.2414398044347763, + 0.020901625975966454, + 0.06116415187716484, + 0.1800662875175476 + ], + "credit_map": [ + 0.018370017409324646, + 0.017202522605657578, + 0.051482200622558594, + 0.01809077151119709, + 0.10783583670854568, + 0.04234323278069496, + 0.013111937791109085, + 0.015587692148983479, + 0.011141514405608177, + 0.007367067504674196, + 0.005647640209645033, + 0.005481957457959652, + 0.015887033194303513, + 0.007211450487375259, + 0.009960009716451168, + 0.006926271133124828, + 0.003458995372056961, + 0.003592453431338072, + 0.0035239930730313063, + 0.0030532663222402334, + 0.0033902875147759914, + 0.00251691578887403, + 0.004934894386678934, + 0.0037225144915282726, + 0.0035723650362342596, + 0.0032097860239446163, + 0.002991898451000452, + 0.003420928493142128, + 0.001436150399968028, + 0.0028331235516816378, + 0.0023747452069073915, + 0.0021162028424441814, + 0.003912545740604401, + 0.0037544986698776484, + 0.004383694846183062, + 0.005062458571046591, + 0.0034264656715095043, + 0.004463711753487587, + 0.004579318221658468, + 0.004159955773502588, + 0.008352608419954777, + 0.00687982514500618, + 0.006519552320241928, + 0.003955827094614506, + 0.013799208216369152, + 0.005443415138870478, + 0.007855549454689026, + 0.006083841435611248, + 0.09651725739240646, + 0.015507029369473457, + 0.015941966325044632, + 0.008047928102314472, + 0.06954734027385712, + 0.010788067243993282, + 0.011058527044951916, + 0.014031690545380116, + 0.0020986769814044237, + 0.003154803765937686, + 0.0022770657669752836, + 0.003963913302868605, + 0.002003974048420787, + 0.002303163520991802, + 0.00326920161023736, + 0.0018308256985619664, + 0.0079457126557827, + 0.012730916030704975, + 0.006145532242953777, + 0.009031160734593868, + 0.005754482466727495, + 0.005780673120170832, + 0.00724346237257123, + 0.006532213184982538, + 0.021054252982139587, + 0.019994841888546944, + 0.01651758700609207, + 0.03291170671582222, + 0.019998090341687202, + 0.022445134818553925, + 0.019253287464380264, + 0.027891414240002632 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 9 + ], + "selected_reward": -0.004647050518542528 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" 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 pink dragon millipedes visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the millipedes have vibrant purple segmented bodies with intricate spiny protrusions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the millipedes moving in a slow circular orbit over a landscape of emerald-green moss with dew droplets?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video feature a high-angle macro shot with extremely shallow depth of field creating a creamy bokeh background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera executing a very slow, circular tracking motion while dappled sunlight creates shifting light and shadow patterns?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06202464550733566 + ], + "opposite_logprob_tokens": [ + -2.8120245933532715 + ], + "logprob_reward": -0.06202464550733566, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06202464550733566, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1309499740600586, + 0.09814317524433136, + 0.013784446753561497, + 0.02601318061351776, + 0.09357395023107529, + 0.09461812674999237, + 0.4611973166465759, + 0.042258258908987045, + 0.01710984855890274, + 0.022351743653416634 + ], + "credit_map": [ + 0.009658101946115494, + 0.011996359564363956, + 0.029090747237205505, + 0.011976772919297218, + 0.01255856640636921, + 0.023459194228053093, + 0.011824460700154305, + 0.02038576453924179, + 0.016774795949459076, + 0.010641506873071194, + 0.007705210242420435, + 0.006136453710496426, + 0.03162987530231476, + 0.009415967389941216, + 0.008564859628677368, + 0.007274506147950888, + 0.0017630690708756447, + 0.0017776412423700094, + 0.001770039671100676, + 0.0016116941114887595, + 0.0019064791267737746, + 0.0017031674506142735, + 0.0017785689560696483, + 0.0014737878227606416, + 0.0035161736886948347, + 0.0040307315066456795, + 0.0035245686303824186, + 0.0037453793920576572, + 0.0015314664924517274, + 0.002938061486929655, + 0.002846562070772052, + 0.0038802355993539095, + 0.010688144713640213, + 0.012260714545845985, + 0.008535155095160007, + 0.023164168000221252, + 0.009814002551138401, + 0.010486982762813568, + 0.010718991048634052, + 0.007905792444944382, + 0.014559048227965832, + 0.010989575646817684, + 0.013259690254926682, + 0.007028758525848389, + 0.019945776090025902, + 0.008775092661380768, + 0.01105028111487627, + 0.009009899571537971, + 0.18452219665050507, + 0.029674770310521126, + 0.029980625957250595, + 0.015963751822710037, + 0.13178062438964844, + 0.020734479650855064, + 0.021213380619883537, + 0.027327507734298706, + 0.005622665397822857, + 0.009741689078509808, + 0.003678594483062625, + 0.005192342679947615, + 0.003955240827053785, + 0.002571594901382923, + 0.007522185333073139, + 0.003973945043981075, + 0.0021277237683534622, + 0.002339589176699519, + 0.001442323555238545, + 0.0016867307713255286, + 0.001579364063218236, + 0.001466030371375382, + 0.004262788686901331, + 0.0022052964195609093, + 0.002648688154295087, + 0.002923489548265934, + 0.002576641971245408, + 0.002648161258548498, + 0.002963811857625842, + 0.0019180609378963709, + 0.004117211792618036, + 0.0025556785985827446 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 1 + ], + "selected_reward": -0.06202464550733566 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" 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 pink dragon millipedes visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the millipedes have vibrant purple segmented bodies with intricate spiny protrusions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the millipedes moving in a slow circular orbit over a landscape of emerald-green moss with dew droplets?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video feature a high-angle macro shot with extremely shallow depth of field creating a creamy bokeh background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera executing a very slow, circular tracking motion while dappled sunlight creates shifting light and shadow patterns?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002193903550505638 + ], + "opposite_logprob_tokens": [ + -6.127193927764893 + ], + "logprob_reward": -0.002193903550505638, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.002193903550505638, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24948926270008087, + 0.2195197492837906, + 0.07271123677492142, + 0.04248678311705589, + 0.15081162750720978, + 0.04703550040721893, + 0.09973226487636566, + 0.019870847463607788, + 0.04592576622962952, + 0.05241701379418373 + ], + "credit_map": [ + 0.014718781225383282, + 0.012358162552118301, + 0.05117638781666756, + 0.018149808049201965, + 0.0784195140004158, + 0.038131218403577805, + 0.01294818613678217, + 0.02358718030154705, + 0.03954171761870384, + 0.024995436891913414, + 0.018296319991350174, + 0.018078841269016266, + 0.05186520144343376, + 0.018030622974038124, + 0.026931922882795334, + 0.021779663860797882, + 0.01448247954249382, + 0.010142672806978226, + 0.005591943860054016, + 0.006467950064688921, + 0.007239482365548611, + 0.013672015629708767, + 0.007849731482565403, + 0.007264959160238504, + 0.006602893117815256, + 0.006181982345879078, + 0.004892199300229549, + 0.005059433169662952, + 0.003559347242116928, + 0.00876509491354227, + 0.004854904022067785, + 0.002570927841588855, + 0.031757865101099014, + 0.015963442623615265, + 0.01234983466565609, + 0.04142485931515694, + 0.008724153973162174, + 0.01423549372702837, + 0.009778164327144623, + 0.01657780446112156, + 0.004817655775696039, + 0.0058329058811068535, + 0.004220986273139715, + 0.0040603699162602425, + 0.010971703566610813, + 0.005719879642128944, + 0.005984889343380928, + 0.005427103955298662, + 0.037522684782743454, + 0.008405719883739948, + 0.006229627877473831, + 0.0040766713209450245, + 0.012925663962960243, + 0.005938631482422352, + 0.00495557114481926, + 0.01967768557369709, + 0.002814208623021841, + 0.003104640170931816, + 0.0019623655825853348, + 0.002307793591171503, + 0.002345507964491844, + 0.0015118537703529, + 0.003859807038679719, + 0.00196466944180429, + 0.005555242300033569, + 0.005834599025547504, + 0.004769482649862766, + 0.004651346709579229, + 0.005229387432336807, + 0.0032563211861997843, + 0.010964889079332352, + 0.005664495751261711, + 0.006213306449353695, + 0.004811881575733423, + 0.005425720475614071, + 0.009960233233869076, + 0.005644857417792082, + 0.006278876215219498, + 0.0052878642454743385, + 0.008794267661869526 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 4 + ], + "selected_reward": -0.002193903550505638 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" 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 pink dragon millipedes visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the millipedes have vibrant purple segmented bodies with intricate spiny protrusions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the millipedes moving in a slow circular orbit over a landscape of emerald-green moss with dew droplets?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video feature a high-angle macro shot with extremely shallow depth of field creating a creamy bokeh background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera executing a very slow, circular tracking motion while dappled sunlight creates shifting light and shadow patterns?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01606121100485325 + ], + "opposite_logprob_tokens": [ + -4.141061305999756 + ], + "logprob_reward": -0.01606121100485325, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.01606121100485325, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14532975852489471, + 0.12476897984743118, + 0.027710868045687675, + 0.009388014674186707, + 0.04207305610179901, + 0.03649788722395897, + 0.5713066458702087, + 0.013783736154437065, + 0.009415704756975174, + 0.019725341349840164 + ], + "credit_map": [ + 0.011711988598108292, + 0.008385999128222466, + 0.027832098305225372, + 0.009865542873740196, + 0.043236859142780304, + 0.021648194640874863, + 0.007484892383217812, + 0.015164203941822052, + 0.021349197253584862, + 0.01330942939966917, + 0.00828577671200037, + 0.008905614726245403, + 0.03793121874332428, + 0.01155154500156641, + 0.01242710929363966, + 0.011009098030626774, + 0.002766731195151806, + 0.00262615573592484, + 0.0023471571039408445, + 0.0032880466897040606, + 0.002877313643693924, + 0.0027090085204690695, + 0.002510322257876396, + 0.008586138486862183, + 0.0012299668742343783, + 0.000951167952734977, + 0.0015610295813530684, + 0.001171125448308885, + 0.0006057514692656696, + 0.001287382561713457, + 0.0015743450494483113, + 0.0010072465520352125, + 0.004347856156527996, + 0.0033896970562636852, + 0.004796574357897043, + 0.011790413409471512, + 0.0036005652509629726, + 0.004998696036636829, + 0.004756794776767492, + 0.004392461385577917, + 0.0062034474685788155, + 0.0052954223938286304, + 0.0034653388429433107, + 0.0025705278385430574, + 0.008383173495531082, + 0.00277370004914701, + 0.004560999106615782, + 0.0032452826853841543, + 0.2272835522890091, + 0.037869445979595184, + 0.03790506720542908, + 0.020294884219765663, + 0.1653263121843338, + 0.025760801509022713, + 0.02720380388200283, + 0.029662800952792168, + 0.0021717343479394913, + 0.002831549383699894, + 0.0013172455364838243, + 0.0013240146217867732, + 0.0013886449160054326, + 0.00088665442308411, + 0.0024174326099455357, + 0.0014464607229456306, + 0.0012775007635354996, + 0.0015736839268356562, + 0.0008000843226909637, + 0.0012267675483599305, + 0.0008612758829258382, + 0.0006172528956085443, + 0.0021250818390399218, + 0.0009340584510937333, + 0.0023471710737794638, + 0.003385868389159441, + 0.002199085196480155, + 0.0017801594221964478, + 0.00217347196303308, + 0.0019908309914171696, + 0.002943608909845352, + 0.0029051462188363075 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 1 + ], + "selected_reward": -0.01606121100485325 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "slightly high-angle macro lens triggers a very slow, circular tracking motion, immersing the viewer in the delicate courtship of two stunning pink dragon millipedes. Their vibrant, segmented bodies, a stunning hue of purple, are covered in intricate, spiny protrusions that give them an alien texture. They move with hypnotic grace, slowly circling each other over a miniature landscape of lush, emerald-green moss, still shimmering with tiny droplets of morning dew. As they orbit, their delicate, silky antennae delicately tap against each other in an exploratory dance, while the rhythmic, undulating waves of their countless tiny legs propel them at a steady, deliberate pace. Dappled sunlight filters through the unseen canopy, casting a shifting pattern of warm light and cool shadow across the scene and creating bright, mirror-like highlights on their smooth, carapace-like exoskeletons. The intense color contrasts, the extremely shallow depth of field that blurs the background into a creamy bokeh, and the unhurried pace of the camera all combine to create an intimate, focused, and utterly otherworldly atmosphere that emphasizes every minute textural detail and transforms the encounter into a moment of mesmerizing natural beauty.", + "seed": 7372, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000121/group_12/index_3563_seed_7372_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": -0.015069936588406562, + "vlm_logprob": -0.015069936588406562, + "vlm_yes_no_margin": 4.4, + "vlm_token_credit": -0.015069936588406562, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" 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 pink dragon millipedes visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the millipedes have vibrant purple segmented bodies with intricate spiny protrusions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the millipedes moving in a slow circular orbit over a landscape of emerald-green moss with dew droplets?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video feature a high-angle macro shot with extremely shallow depth of field creating a creamy bokeh background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera executing a very slow, circular tracking motion while dappled sunlight creates shifting light and shadow patterns?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01938037760555744 + ], + "opposite_logprob_tokens": [ + -4.019380569458008 + ], + "logprob_reward": -0.01938037760555744, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.01938037760555744, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06317650526762009, + 0.06130193918943405, + 0.25815778970718384, + 0.07564712315797806, + 0.026122326031327248, + 0.29717618227005005, + 0.10753165930509567, + 0.03564145416021347, + 0.03357741981744766, + 0.04166765883564949 + ], + "credit_map": [ + 0.008438482880592346, + 0.007767346687614918, + 0.012757550925016403, + 0.008261088281869888, + 0.006050822790712118, + 0.0063063218258321285, + 0.005331309512257576, + 0.008263574913144112, + 0.008565136231482029, + 0.00763776944950223, + 0.009455286897718906, + 0.006049337796866894, + 0.004561767913401127, + 0.007705832365900278, + 0.006167182233184576, + 0.011159623041749, + 0.012273963540792465, + 0.09039228409528732, + 0.015376005321741104, + 0.013609944842755795, + 0.012383963912725449, + 0.07809697091579437, + 0.01607588492333889, + 0.019948752596974373, + 0.008244028314948082, + 0.00677897036075592, + 0.00892562884837389, + 0.008168939501047134, + 0.00817101076245308, + 0.01464484166353941, + 0.011520369909703732, + 0.009193329140543938, + 0.002409852109849453, + 0.0027699158526957035, + 0.002235028427094221, + 0.002922865329310298, + 0.0023064042907208204, + 0.005133800208568573, + 0.005564155988395214, + 0.0027803026605397463, + 0.028180284425616264, + 0.024185430258512497, + 0.02251926064491272, + 0.024322040379047394, + 0.021184133365750313, + 0.10371231287717819, + 0.025989988818764687, + 0.04708272963762283, + 0.030547451227903366, + 0.008271067403256893, + 0.00711579667404294, + 0.008133095689117908, + 0.006382757797837257, + 0.0121343107894063, + 0.008184765465557575, + 0.02676241844892502, + 0.0071665882132947445, + 0.004608000162988901, + 0.00461201136931777, + 0.0032046830747276545, + 0.0026865480467677116, + 0.0037836171686649323, + 0.003648769808933139, + 0.005931233987212181, + 0.00320653454400599, + 0.002825987059623003, + 0.004822553601115942, + 0.005545563995838165, + 0.004826453980058432, + 0.003780423430725932, + 0.0034474136773496866, + 0.005122488830238581, + 0.004869161639362574, + 0.004341376479715109, + 0.0051340158097445965, + 0.0064063225872814655, + 0.004389328882098198, + 0.004383338615298271, + 0.006949865724891424, + 0.005194245371967554 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 2, + 6 + ], + "selected_reward": -0.01938037760555744 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" 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 pink dragon millipedes visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the millipedes have vibrant purple segmented bodies with intricate spiny protrusions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the millipedes moving in a slow circular orbit over a landscape of emerald-green moss with dew droplets?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video feature a high-angle macro shot with extremely shallow depth of field creating a creamy bokeh background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera executing a very slow, circular tracking motion while dappled sunlight creates shifting light and shadow patterns?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"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": [ + 0.07313573360443115, + 0.043665122240781784, + 0.33450600504875183, + 0.04913611710071564, + 0.04007323086261749, + 0.09418592602014542, + 0.10211876034736633, + 0.0638013482093811, + 0.0628095269203186, + 0.13656820356845856 + ], + "credit_map": [ + 0.010449212044477463, + 0.009213272482156754, + 0.011969883926212788, + 0.007016514427959919, + 0.006467803381383419, + 0.0095669599249959, + 0.007039748132228851, + 0.01141233742237091, + 0.004869879223406315, + 0.005254329647868872, + 0.007562830578535795, + 0.005200158804655075, + 0.004067009314894676, + 0.0036518704146146774, + 0.004345523193478584, + 0.00871351920068264, + 0.014190055429935455, + 0.12844136357307434, + 0.01409168727695942, + 0.012934263795614243, + 0.01629967801272869, + 0.11100173741579056, + 0.013231714256107807, + 0.02431550994515419, + 0.005610223859548569, + 0.005227698013186455, + 0.004945813678205013, + 0.0049459063448011875, + 0.003780773840844631, + 0.012683250941336155, + 0.006198025308549404, + 0.00574442557990551, + 0.0044099111109972, + 0.003709198674187064, + 0.003963003866374493, + 0.004149843938648701, + 0.005083160474896431, + 0.007588016334921122, + 0.0058839693665504456, + 0.005286125931888819, + 0.007907786406576633, + 0.0078085800632834435, + 0.006560838781297207, + 0.006184879224747419, + 0.0077405464835464954, + 0.03378857299685478, + 0.0061974977143108845, + 0.017997222021222115, + 0.027906475588679314, + 0.009502867236733437, + 0.0067865303717553616, + 0.007188627030700445, + 0.008980142883956432, + 0.01082634087651968, + 0.005853278562426567, + 0.025074496865272522, + 0.013243828900158405, + 0.009059942327439785, + 0.00783148780465126, + 0.004803377203643322, + 0.004537853877991438, + 0.00644860602915287, + 0.006381690036505461, + 0.011494566686451435, + 0.007681720890104771, + 0.00766771100461483, + 0.0074468255043029785, + 0.009346180595457554, + 0.008398802950978279, + 0.009520270861685276, + 0.005365277174860239, + 0.007382736075669527, + 0.015246832743287086, + 0.01651754043996334, + 0.019863879308104515, + 0.01711849868297577, + 0.015916280448436737, + 0.0173832755535841, + 0.020041488111019135, + 0.01448040921241045 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 9, + 6 + ], + "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\": \"Are two pink dragon millipedes visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the millipedes have vibrant purple segmented bodies with intricate spiny protrusions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the millipedes moving in a slow circular orbit over a landscape of emerald-green moss with dew droplets?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video feature a high-angle macro shot with extremely shallow depth of field creating a creamy bokeh background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera executing a very slow, circular tracking motion while dappled sunlight creates shifting light and shadow patterns?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014186250045895576 + ], + "opposite_logprob_tokens": [ + -4.264186382293701 + ], + "logprob_reward": -0.014186250045895576, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014186250045895576, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07667268067598343, + 0.22549137473106384, + 0.3563907742500305, + 0.02457500249147415, + 0.021938005462288857, + 0.05733948573470116, + 0.07017886638641357, + 0.03388097882270813, + 0.08353020995855331, + 0.050002600997686386 + ], + "credit_map": [ + 0.010971570387482643, + 0.010987450368702412, + 0.01115480437874794, + 0.01070469245314598, + 0.007941659539937973, + 0.006717432290315628, + 0.0076272631995379925, + 0.010567812249064445, + 0.02704235352575779, + 0.02796146087348461, + 0.042225949466228485, + 0.019841283559799194, + 0.03186718747019768, + 0.022212356328964233, + 0.019484201446175575, + 0.03485659882426262, + 0.01773860678076744, + 0.13703355193138123, + 0.011052366346120834, + 0.010708583518862724, + 0.013712344691157341, + 0.1329660415649414, + 0.009069493971765041, + 0.024109797552227974, + 0.003075894434005022, + 0.0026509477756917477, + 0.0037889175582677126, + 0.0024980236776173115, + 0.001873751520179212, + 0.0048902505077421665, + 0.003065696218982339, + 0.0027315218467265368, + 0.0023913385812193155, + 0.003126756753772497, + 0.0022668333258479834, + 0.002999795600771904, + 0.002518551191315055, + 0.0031125610694289207, + 0.003107351018115878, + 0.002414819085970521, + 0.004617896396666765, + 0.004633834585547447, + 0.00453295698389411, + 0.005013544578105211, + 0.0068340664729475975, + 0.014272624626755714, + 0.006406983360648155, + 0.011027581058442593, + 0.018969591706991196, + 0.005685112439095974, + 0.005581904202699661, + 0.004080086015164852, + 0.005435651633888483, + 0.007372209802269936, + 0.004285839386284351, + 0.018768470734357834, + 0.004808072000741959, + 0.005511001218110323, + 0.0036895019002258778, + 0.003954019863158464, + 0.0030406287405639887, + 0.0045846570283174515, + 0.0030384918209165335, + 0.0052546062506735325, + 0.006665227934718132, + 0.007696547545492649, + 0.009703065268695354, + 0.012229511514306068, + 0.021516339853405952, + 0.010413396172225475, + 0.00685400515794754, + 0.008452119305729866, + 0.005677653010934591, + 0.005583977792412043, + 0.0047515034675598145, + 0.006551169324666262, + 0.004809924867004156, + 0.007769711781293154, + 0.007085565011948347, + 0.00777309900149703 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 1, + 8 + ], + "selected_reward": -0.014186250045895576 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" 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 pink dragon millipedes visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the millipedes have vibrant purple segmented bodies with intricate spiny protrusions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the millipedes moving in a slow circular orbit over a landscape of emerald-green moss with dew droplets?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video feature a high-angle macro shot with extremely shallow depth of field creating a creamy bokeh background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera executing a very slow, circular tracking motion while dappled sunlight creates shifting light and shadow patterns?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006730150431394577 + ], + "opposite_logprob_tokens": [ + -5.006730079650879 + ], + "logprob_reward": -0.006730150431394577, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006730150431394577, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07249017059803009, + 0.14250506460666656, + 0.389317125082016, + 0.059364162385463715, + 0.022658828645944595, + 0.11213277280330658, + 0.026583975180983543, + 0.08720312267541885, + 0.04350300505757332, + 0.04424180090427399 + ], + "credit_map": [ + 0.010896041058003902, + 0.007641558535397053, + 0.017035655677318573, + 0.006643894128501415, + 0.0062087541446089745, + 0.007537292316555977, + 0.007455355953425169, + 0.009071619249880314, + 0.02112763561308384, + 0.013069369830191135, + 0.033211320638656616, + 0.01289331167936325, + 0.011086839251220226, + 0.016227776184678078, + 0.012827242724597454, + 0.02206156775355339, + 0.017512766644358635, + 0.1569153517484665, + 0.01116877980530262, + 0.012379494495689869, + 0.016274768859148026, + 0.13683383166790009, + 0.010567985475063324, + 0.02766413986682892, + 0.007450324948877096, + 0.006533750332891941, + 0.007664921227842569, + 0.00576664786785841, + 0.0043967957608401775, + 0.011808990500867367, + 0.008994342759251595, + 0.006748391781002283, + 0.0027237795293331146, + 0.002171118510887027, + 0.0023216649424284697, + 0.0021121688187122345, + 0.0018226910615339875, + 0.003779236227273941, + 0.004206222016364336, + 0.003521946258842945, + 0.012512508779764175, + 0.008063949644565582, + 0.005726293660700321, + 0.002806880511343479, + 0.007670372724533081, + 0.05017811432480812, + 0.004725951235741377, + 0.020448703318834305, + 0.005646448116749525, + 0.002996824448928237, + 0.0025456463918089867, + 0.002057157689705491, + 0.001962878741323948, + 0.002924388274550438, + 0.002220909111201763, + 0.006229724269360304, + 0.02020880952477455, + 0.008114554919302464, + 0.011094079352915287, + 0.007700064685195684, + 0.005751264281570911, + 0.010880796238780022, + 0.006205520126968622, + 0.017248030751943588, + 0.007217474281787872, + 0.003478158265352249, + 0.006581534165889025, + 0.004594113677740097, + 0.004714461043477058, + 0.005867763422429562, + 0.004543981049209833, + 0.006505520548671484, + 0.005930696148425341, + 0.003458146471530199, + 0.008028182201087475, + 0.00728659937158227, + 0.004301873035728931, + 0.00438725808635354, + 0.00548975495621562, + 0.005359292030334473 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 1, + 5 + ], + "selected_reward": -0.006730150431394577 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" 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 pink dragon millipedes visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the millipedes have vibrant purple segmented bodies with intricate spiny protrusions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the millipedes moving in a slow circular orbit over a landscape of emerald-green moss with dew droplets?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video feature a high-angle macro shot with extremely shallow depth of field creating a creamy bokeh background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera executing a very slow, circular tracking motion while dappled sunlight creates shifting light and shadow patterns?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"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": [ + 0.11411688476800919, + 0.09609048068523407, + 0.21438953280448914, + 0.09474116563796997, + 0.04215167462825775, + 0.0279263686388731, + 0.19046449661254883, + 0.0491284616291523, + 0.08838742226362228, + 0.08260359615087509 + ], + "credit_map": [ + 0.01697714626789093, + 0.016369689255952835, + 0.017506450414657593, + 0.013845735229551792, + 0.013329382985830307, + 0.014835813082754612, + 0.009179634042084217, + 0.012073024176061153, + 0.011903071776032448, + 0.013184838928282261, + 0.012961735017597675, + 0.007021300494670868, + 0.011155396699905396, + 0.011373433284461498, + 0.007903749123215675, + 0.020586933940649033, + 0.01527318824082613, + 0.06414033472537994, + 0.01514973770827055, + 0.008496423251926899, + 0.008537216112017632, + 0.07963453978300095, + 0.009741918183863163, + 0.013416148722171783, + 0.011345305480062962, + 0.009532982483506203, + 0.014846495352685452, + 0.009526037611067295, + 0.00809849426150322, + 0.018266847357153893, + 0.012238294817507267, + 0.010886700823903084, + 0.005486773326992989, + 0.003675668267533183, + 0.004355315119028091, + 0.004048608709126711, + 0.004386911168694496, + 0.006843313109129667, + 0.007277090102434158, + 0.006077991332858801, + 0.0031984057277441025, + 0.0030278980266302824, + 0.0036726759281009436, + 0.0021496410481631756, + 0.0026765461079776287, + 0.006659925915300846, + 0.0030724930111318827, + 0.0034687798470258713, + 0.06432126462459564, + 0.01288190484046936, + 0.011112289503216743, + 0.007778107188642025, + 0.00974203459918499, + 0.019936542958021164, + 0.009738307446241379, + 0.05495402216911316, + 0.009202077053487301, + 0.0092674121260643, + 0.005047048442065716, + 0.0036991233937442303, + 0.0042281984351575375, + 0.0067907837219536304, + 0.0036410989705473185, + 0.007252712268382311, + 0.012467036955058575, + 0.00808402057737112, + 0.010350852273404598, + 0.010721092112362385, + 0.01690889708697796, + 0.007623300887644291, + 0.009799307212233543, + 0.012432901188731194, + 0.012367219664156437, + 0.009508587419986725, + 0.01227644830942154, + 0.013744705356657505, + 0.010304495692253113, + 0.006155854091048241, + 0.012175102718174458, + 0.006071168929338455 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 6, + 0 + ], + "selected_reward": -0.029790453612804413 + } + ], + "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/0000121/group_13/index_tensor([9596, 9596])_group_13_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000121/group_13/index_tensor([9596, 9596])_group_13_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..5d06ad3a19f92b458779294df4e7fe1214d630e8 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000121/group_13/index_tensor([9596, 9596])_group_13_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "Shot in an eerie, cartoonish horror style, extreme close-up static footage at a high frame rate focuses on a shimmering, blood-red gelatin dessert. The translucent substance teeters on a grimy white porcelain plate, dotted with protruding black fragments and faint, hairline cracks, when it suddenly begins to twitch violently. Each unnatural, arrhythmic tremor and spasm is rendered with unsettling clarity, the gelatinous form teetering and contracting as if it were a living organ in agony. The entire scene is set against a grimy, stained laminate kitchen counter, a rust-stained fork resting ominously beside the plate, its prong pointed at the trembling dessert. Under the harsh, flickering light of a single overhead bulb, the gelatin's surface gleams with a sickly, damp sheen, and as it vibrates, tiny droplets of the crimson substance spurt outward, captured in sharp detail as they arc through the air. The vibrant yet irritating red of the dessert contrasts starkly with the grimy yellows and browns of the grimy background, creating a deeply unsettling atmosphere of body horror and disgust.", + "seed": 22234, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000121/group_13/index_9596_seed_22234_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": -0.06615202408283949, + "vlm_logprob": -0.06615202408283949, + "vlm_yes_no_margin": 3.224999952316284, + "vlm_token_credit": -0.06615202408283949, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a shimmering, blood-red gelatin dessert?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the dessert placed on a grimy white porcelain plate with black fragments and hairline cracks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the gelatin dessert exhibit violent, unnatural twitching and spasm motions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a rust-stained fork with a pointed prong visible resting beside the plate on a grimy laminate counter?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature tiny droplets of crimson substance spurting outward from the vibrating dessert?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011223279871046543 + ], + "opposite_logprob_tokens": [ + -5.136223316192627 + ], + "logprob_reward": -0.011223279871046543, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.011223279871046543, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14149555563926697, + 0.14810386300086975, + 0.17563693225383759, + 0.09011194854974747, + 0.07030203938484192, + 0.061787959188222885, + 0.05620158091187477, + 0.11458611488342285, + 0.0970853939652443, + 0.04468856751918793 + ], + "credit_map": [ + 0.020381605252623558, + 0.02289767749607563, + 0.020184772089123726, + 0.018598506227135658, + 0.019939303398132324, + 0.014879733324050903, + 0.010736657306551933, + 0.013877302408218384, + 0.020278098061680794, + 0.03444254398345947, + 0.014511721208691597, + 0.01602950692176819, + 0.027150554582476616, + 0.014090059325098991, + 0.011508525349199772, + 0.010092854499816895, + 0.022005559876561165, + 0.021528443321585655, + 0.04171016812324524, + 0.017666498199105263, + 0.016708794981241226, + 0.0165698379278183, + 0.011978943832218647, + 0.02746870368719101, + 0.016056543216109276, + 0.011725072748959064, + 0.007399106398224831, + 0.013277477584779263, + 0.01874036155641079, + 0.010607544332742691, + 0.007329312153160572, + 0.0049765282310545444, + 0.011986997909843922, + 0.007183213718235493, + 0.005529669113457203, + 0.008802623488008976, + 0.01884625479578972, + 0.00821718294173479, + 0.005608043167740107, + 0.004128049593418837, + 0.0076417503878474236, + 0.00908239372074604, + 0.009031638503074646, + 0.007084437180310488, + 0.009021388366818428, + 0.007470165379345417, + 0.0069944956339895725, + 0.005461693741381168, + 0.007606259547173977, + 0.004893017932772636, + 0.011179707013070583, + 0.0064823217689991, + 0.004586800001561642, + 0.007768742740154266, + 0.004360605496913195, + 0.009324124082922935, + 0.023885641247034073, + 0.009638250805437565, + 0.012290850281715393, + 0.007790416479110718, + 0.031117668375372887, + 0.011257447302341461, + 0.01105419173836708, + 0.007551657967269421, + 0.020852312445640564, + 0.005817597731947899, + 0.006526846904307604, + 0.011919302865862846, + 0.02943740040063858, + 0.00860522035509348, + 0.009410553611814976, + 0.00451615871861577, + 0.006945246364921331, + 0.0050704991444945335, + 0.003032340668141842, + 0.006408158224076033, + 0.009093765169382095, + 0.005736013874411583, + 0.004490844905376434, + 0.0039116996340453625 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 1, + 0 + ], + "selected_reward": -0.011223279871046543 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a shimmering, blood-red gelatin dessert?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the dessert placed on a grimy white porcelain plate with black fragments and hairline cracks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the gelatin dessert exhibit violent, unnatural twitching and spasm motions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a rust-stained fork with a pointed prong visible resting beside the plate on a grimy laminate counter?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature tiny droplets of crimson substance spurting outward from the vibrating dessert?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018203457817435265 + ], + "opposite_logprob_tokens": [ + -4.018203258514404 + ], + "logprob_reward": -0.018203457817435265, + "yes_no_margin_reward": 3.999999761581421, + "token_credit_reward": -0.018203457817435265, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11400093883275986, + 0.0920611023902893, + 0.3148403465747833, + 0.02485533058643341, + 0.0761951208114624, + 0.04139842465519905, + 0.027313409373164177, + 0.13067160546779633, + 0.15858159959316254, + 0.0200821366161108 + ], + "credit_map": [ + 0.014751724898815155, + 0.014420543797314167, + 0.019918812438845634, + 0.01699673756957054, + 0.01494806818664074, + 0.0129014877602458, + 0.011476658284664154, + 0.008586892858147621, + 0.015033776871860027, + 0.010193942114710808, + 0.015175444073975086, + 0.01280607283115387, + 0.008131200447678566, + 0.008699430152773857, + 0.00782153569161892, + 0.014199689961969852, + 0.040584906935691833, + 0.027123285457491875, + 0.08777285367250443, + 0.031712841242551804, + 0.026076478883624077, + 0.029311522841453552, + 0.02055337280035019, + 0.05170506611466408, + 0.0031621805392205715, + 0.0031174933537840843, + 0.003307378152385354, + 0.0035648965276777744, + 0.0027794046327471733, + 0.004297618754208088, + 0.001988845644518733, + 0.0026375120505690575, + 0.011153820902109146, + 0.011803447268903255, + 0.008018325082957745, + 0.008015213534235954, + 0.017591649666428566, + 0.008927088230848312, + 0.006565706804394722, + 0.004119862802326679, + 0.004934929311275482, + 0.005362316034734249, + 0.004801670555025339, + 0.007847386412322521, + 0.004563875496387482, + 0.005450823809951544, + 0.004640008322894573, + 0.0037974142469465733, + 0.003969515673816204, + 0.003374916035681963, + 0.004479711409658194, + 0.0037662028335034847, + 0.0036404719576239586, + 0.0022270777262747288, + 0.0029555412475019693, + 0.002899970393627882, + 0.03234988451004028, + 0.009632882662117481, + 0.011159494519233704, + 0.007251795846968889, + 0.037529148161411285, + 0.012464163824915886, + 0.013377094641327858, + 0.0069071208126842976, + 0.04015757143497467, + 0.009974606335163116, + 0.008234156295657158, + 0.011884042993187904, + 0.05685439333319664, + 0.011424148455262184, + 0.013797353021800518, + 0.0062553356401622295, + 0.002265038201585412, + 0.003122634021565318, + 0.002372132847085595, + 0.003897698363289237, + 0.002187272533774376, + 0.0026688487268984318, + 0.0020725615322589874, + 0.0014959491090849042 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 8, + 7 + ], + "selected_reward": -0.018203457817435265 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a shimmering, blood-red gelatin dessert?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the dessert placed on a grimy white porcelain plate with black fragments and hairline cracks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the gelatin dessert exhibit violent, unnatural twitching and spasm motions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a rust-stained fork with a pointed prong visible resting beside the plate on a grimy laminate counter?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature tiny droplets of crimson substance spurting outward from the vibrating dessert?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.16032707691192627 + ], + "opposite_logprob_tokens": [ + -1.9103270769119263 + ], + "logprob_reward": -0.16032707691192627, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.16032707691192627, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11750490218400955, + 0.09163279086351395, + 0.10978465527296066, + 0.17452143132686615, + 0.23786772787570953, + 0.061732854694128036, + 0.03488263487815857, + 0.04213397204875946, + 0.10772687196731567, + 0.022212129086256027 + ], + "credit_map": [ + 0.013093813322484493, + 0.01999688148498535, + 0.013738621026277542, + 0.011145883239805698, + 0.021975189447402954, + 0.009156499058008194, + 0.014973915182054043, + 0.013424096629023552, + 0.011691080406308174, + 0.017270756885409355, + 0.00914074294269085, + 0.011546737514436245, + 0.013550515286624432, + 0.008867213502526283, + 0.007367902435362339, + 0.012197831645607948, + 0.015692176297307014, + 0.008427347987890244, + 0.026857635006308556, + 0.011530776508152485, + 0.009957385249435902, + 0.013897127471864223, + 0.007616935297846794, + 0.015805264934897423, + 0.04134593531489372, + 0.019055962562561035, + 0.013400780968368053, + 0.014653563499450684, + 0.04892570897936821, + 0.01708786003291607, + 0.00968716200441122, + 0.010364465415477753, + 0.03647369146347046, + 0.03033440001308918, + 0.02574385702610016, + 0.02633000537753105, + 0.05649051070213318, + 0.03071373514831066, + 0.017793824896216393, + 0.013987712562084198, + 0.008733942173421383, + 0.010608906857669353, + 0.008080966770648956, + 0.0067680212669074535, + 0.009151709266006947, + 0.008089094422757626, + 0.005865422077476978, + 0.004434795584529638, + 0.00952734425663948, + 0.0022797316778451204, + 0.0023322822526097298, + 0.001829864806495607, + 0.010323927737772465, + 0.0037941010668873787, + 0.002383972518146038, + 0.0024114118423312902, + 0.006436975672841072, + 0.005396511871367693, + 0.0036940358113497496, + 0.006024726666510105, + 0.009201460517942905, + 0.004458826035261154, + 0.004353977274149656, + 0.002567456802353263, + 0.025437429547309875, + 0.0074167437851428986, + 0.005558311939239502, + 0.008191924542188644, + 0.03949495404958725, + 0.008321897126734257, + 0.009313325397670269, + 0.003992286045104265, + 0.002595519181340933, + 0.0033180727623403072, + 0.0025558776687830687, + 0.0035873407032340765, + 0.0023756055161356926, + 0.002401895122602582, + 0.003644472686573863, + 0.0017333460273221135 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 3, + 0 + ], + "selected_reward": -0.16032707691192627 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a shimmering, blood-red gelatin dessert?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the dessert placed on a grimy white porcelain plate with black fragments and hairline cracks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the gelatin dessert exhibit violent, unnatural twitching and spasm motions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a rust-stained fork with a pointed prong visible resting beside the plate on a grimy laminate counter?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature tiny droplets of crimson substance spurting outward from the vibrating dessert?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0789726972579956 + ], + "opposite_logprob_tokens": [ + -2.578972816467285 + ], + "logprob_reward": -0.0789726972579956, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.0789726972579956, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07065761089324951, + 0.05675999075174332, + 0.4817589223384857, + 0.07575379312038422, + 0.03778534382581711, + 0.017975911498069763, + 0.0726160928606987, + 0.11356744170188904, + 0.0488424152135849, + 0.024282479658722878 + ], + "credit_map": [ + 0.010530882515013218, + 0.011113103479146957, + 0.010946680791676044, + 0.008927498012781143, + 0.007939032278954983, + 0.008459947071969509, + 0.006852257531136274, + 0.005888218060135841, + 0.009625812061131, + 0.008064660243690014, + 0.009440572001039982, + 0.008154702372848988, + 0.007128899917006493, + 0.005415625870227814, + 0.003911493346095085, + 0.005018227733671665, + 0.06031225249171257, + 0.04254073649644852, + 0.12995733320713043, + 0.05079684033989906, + 0.04053609073162079, + 0.04698152095079422, + 0.03254568576812744, + 0.07808849215507507, + 0.022016525268554688, + 0.004971743561327457, + 0.00418996624648571, + 0.0034646010026335716, + 0.028484083712100983, + 0.005845892708748579, + 0.004089581314474344, + 0.0026914109475910664, + 0.005440180189907551, + 0.006689615547657013, + 0.003949949983507395, + 0.0037141835782676935, + 0.007747999392449856, + 0.0047193109057843685, + 0.0032803730573505163, + 0.002243735594674945, + 0.0026842160150408745, + 0.0023937576916068792, + 0.001659109373576939, + 0.0021588397212326527, + 0.004058728460222483, + 0.0020233357790857553, + 0.0016551883891224861, + 0.0013427369995042682, + 0.01899380423128605, + 0.005354413762688637, + 0.004468670580536127, + 0.004415811039507389, + 0.024750428274273872, + 0.005990887060761452, + 0.005562628153711557, + 0.0030794525519013405, + 0.021121220663189888, + 0.008764885365962982, + 0.0216970257461071, + 0.012275236658751965, + 0.01621481217443943, + 0.011358343064785004, + 0.011973366141319275, + 0.010162567719817162, + 0.011864670552313328, + 0.0034513066057115793, + 0.0032100260723382235, + 0.005519209895282984, + 0.013106374070048332, + 0.005259980447590351, + 0.003964817151427269, + 0.0024660383351147175, + 0.002687998814508319, + 0.0032536038197577, + 0.0028548745904117823, + 0.004426436498761177, + 0.002815409330651164, + 0.0034190050791949034, + 0.0024581565521657467, + 0.0023669993970543146 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 7, + 3 + ], + "selected_reward": -0.0789726972579956 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a shimmering, blood-red gelatin dessert?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the dessert placed on a grimy white porcelain plate with black fragments and hairline cracks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the gelatin dessert exhibit violent, unnatural twitching and spasm motions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a rust-stained fork with a pointed prong visible resting beside the plate on a grimy laminate counter?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature tiny droplets of crimson substance spurting outward from the vibrating dessert?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06203360855579376 + ], + "opposite_logprob_tokens": [ + -2.8120336532592773 + ], + "logprob_reward": -0.06203360855579376, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06203360855579376, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08191483467817307, + 0.051666781306266785, + 0.054112620651721954, + 0.23609700798988342, + 0.14896853268146515, + 0.09895758330821991, + 0.09992492198944092, + 0.020647739991545677, + 0.1927974373102188, + 0.01491247396916151 + ], + "credit_map": [ + 0.010404089465737343, + 0.012931901030242443, + 0.011841768398880959, + 0.009196111932396889, + 0.013966374099254608, + 0.007579762488603592, + 0.009543327614665031, + 0.00645150663331151, + 0.006829720456153154, + 0.008710765279829502, + 0.006655060686171055, + 0.0067013949155807495, + 0.006910183001309633, + 0.005000048782676458, + 0.004210198298096657, + 0.006649416405707598, + 0.00691479817032814, + 0.00708769029006362, + 0.009081372059881687, + 0.007311347406357527, + 0.0066065541468560696, + 0.005614669062197208, + 0.004288061521947384, + 0.0072081382386386395, + 0.0681610256433487, + 0.01433519646525383, + 0.014142190106213093, + 0.011632471345365047, + 0.08846614509820938, + 0.016797129064798355, + 0.014264972880482674, + 0.00829790998250246, + 0.0238204225897789, + 0.019843928515911102, + 0.015594694763422012, + 0.01587682031095028, + 0.03642398118972778, + 0.018695460632443428, + 0.011087179183959961, + 0.007626064587384462, + 0.01953905262053013, + 0.010862432420253754, + 0.01078557688742876, + 0.009232566691935062, + 0.019495636224746704, + 0.010506289079785347, + 0.009312055073678493, + 0.009223978966474533, + 0.02566101774573326, + 0.007706219330430031, + 0.005724010989069939, + 0.007072011474519968, + 0.035257913172245026, + 0.007346498314291239, + 0.007531663402915001, + 0.0036255924496799707, + 0.003427365329116583, + 0.0018003233708441257, + 0.0029934043996036053, + 0.002770356833934784, + 0.003582337638363242, + 0.0028751767240464687, + 0.001558696385473013, + 0.0016400811728090048, + 0.050868671387434006, + 0.01212253700941801, + 0.010166521184146404, + 0.0073805940337479115, + 0.07510195672512054, + 0.013673064298927784, + 0.016493938863277435, + 0.006990174762904644, + 0.001812310772947967, + 0.0017812543082982302, + 0.002198431408032775, + 0.002317230449989438, + 0.0016481890343129635, + 0.0021145823411643505, + 0.0018087896751239896, + 0.0012316873762756586 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 8, + 4 + ], + "selected_reward": -0.06203360855579376 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "Shot in an eerie, cartoonish horror style, extreme close-up static footage at a high frame rate focuses on a shimmering, blood-red gelatin dessert. The translucent substance teeters on a grimy white porcelain plate, dotted with protruding black fragments and faint, hairline cracks, when it suddenly begins to twitch violently. Each unnatural, arrhythmic tremor and spasm is rendered with unsettling clarity, the gelatinous form teetering and contracting as if it were a living organ in agony. The entire scene is set against a grimy, stained laminate kitchen counter, a rust-stained fork resting ominously beside the plate, its prong pointed at the trembling dessert. Under the harsh, flickering light of a single overhead bulb, the gelatin's surface gleams with a sickly, damp sheen, and as it vibrates, tiny droplets of the crimson substance spurt outward, captured in sharp detail as they arc through the air. The vibrant yet irritating red of the dessert contrasts starkly with the grimy yellows and browns of the grimy background, creating a deeply unsettling atmosphere of body horror and disgust.", + "seed": 22235, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000121/group_13/index_9596_seed_22235_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": -0.08865856602787972, + "vlm_logprob": -0.08865856602787972, + "vlm_yes_no_margin": 3.075, + "vlm_token_credit": -0.08865856602787972, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a shimmering, blood-red gelatin dessert?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the dessert placed on a grimy white porcelain plate with black fragments and hairline cracks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the gelatin dessert exhibit violent, unnatural twitching and spasm motions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a rust-stained fork with a pointed prong visible resting beside the plate on a grimy laminate counter?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature tiny droplets of crimson substance spurting outward from the vibrating dessert?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011920761317014694 + ], + "opposite_logprob_tokens": [ + -5.136920928955078 + ], + "logprob_reward": -0.011920761317014694, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.011920761317014694, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10257136076688766, + 0.05237932875752449, + 0.059401072561740875, + 0.10691490769386292, + 0.18341337144374847, + 0.25387686491012573, + 0.06034541502594948, + 0.0501176081597805, + 0.038750793784856796, + 0.09222923219203949 + ], + "credit_map": [ + 0.009360197931528091, + 0.018343200907111168, + 0.01195740420371294, + 0.02583497390151024, + 0.01071147620677948, + 0.007679260801523924, + 0.011326158419251442, + 0.007358689326792955, + 0.0046431440860033035, + 0.009706583805382252, + 0.0074525149539113045, + 0.006432516500353813, + 0.005936817731708288, + 0.00549360690638423, + 0.0062157753854990005, + 0.006498368456959724, + 0.005221553146839142, + 0.010421608574688435, + 0.007677617482841015, + 0.006355689838528633, + 0.006772689055651426, + 0.005664947908371687, + 0.006895852275192738, + 0.010391118004918098, + 0.011671244166791439, + 0.01775852032005787, + 0.009910671040415764, + 0.026020243763923645, + 0.01004953682422638, + 0.01006514672189951, + 0.01304076798260212, + 0.008398775942623615, + 0.019971538335084915, + 0.02188175357878208, + 0.02771923504769802, + 0.020953398197889328, + 0.022868918254971504, + 0.034425072371959686, + 0.014668496325612068, + 0.020924976095557213, + 0.013798513449728489, + 0.10796712338924408, + 0.06298407167196274, + 0.015079562552273273, + 0.017680881544947624, + 0.014142059721052647, + 0.011440904811024666, + 0.010783781297504902, + 0.004823749884963036, + 0.008563259616494179, + 0.008307338692247868, + 0.012361996807157993, + 0.009756884537637234, + 0.005116094835102558, + 0.0056056310422718525, + 0.005810456816107035, + 0.005102168302983046, + 0.007885359227657318, + 0.007477556820958853, + 0.012941701337695122, + 0.005657704547047615, + 0.003183254972100258, + 0.0036403576377779245, + 0.0042295074090361595, + 0.0038744553457945585, + 0.006263568066060543, + 0.003781817387789488, + 0.006121267098933458, + 0.00714414706453681, + 0.003570694476366043, + 0.004112067632377148, + 0.003882778575643897, + 0.008355258964002132, + 0.015312407165765762, + 0.013772523030638695, + 0.01271500438451767, + 0.01925664022564888, + 0.00768969627097249, + 0.00836980901658535, + 0.00675789313390851 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 4, + 3 + ], + "selected_reward": -0.011920761317014694 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a shimmering, blood-red gelatin dessert?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the dessert placed on a grimy white porcelain plate with black fragments and hairline cracks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the gelatin dessert exhibit violent, unnatural twitching and spasm motions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a rust-stained fork with a pointed prong visible resting beside the plate on a grimy laminate counter?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature tiny droplets of crimson substance spurting outward from the vibrating dessert?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04307139292359352 + ], + "opposite_logprob_tokens": [ + -3.1680715084075928 + ], + "logprob_reward": -0.04307139292359352, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04307139292359352, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1596374362707138, + 0.10171708464622498, + 0.13071472942829132, + 0.16024178266525269, + 0.09108626842498779, + 0.06613415479660034, + 0.07439353317022324, + 0.10325799137353897, + 0.06932500749826431, + 0.04349207878112793 + ], + "credit_map": [ + 0.014598323963582516, + 0.022414378821849823, + 0.018031058833003044, + 0.033826135098934174, + 0.016161572188138962, + 0.014222273603081703, + 0.0201219841837883, + 0.02026170678436756, + 0.011117292568087578, + 0.013109628111124039, + 0.010506202466785908, + 0.010023516602814198, + 0.01817844808101654, + 0.017432747408747673, + 0.011494727805256844, + 0.009854521602392197, + 0.012667682953178883, + 0.020215125754475594, + 0.017038438469171524, + 0.01452036015689373, + 0.025548256933689117, + 0.019883863627910614, + 0.010340578854084015, + 0.010500418022274971, + 0.014929433353245258, + 0.026626404374837875, + 0.013986336067318916, + 0.03691674396395683, + 0.02531057596206665, + 0.014243099838495255, + 0.01879124529659748, + 0.009437935426831245, + 0.008541766554117203, + 0.011920738033950329, + 0.011198323220014572, + 0.01804598979651928, + 0.011129955761134624, + 0.011785490438342094, + 0.008866088464856148, + 0.009597914293408394, + 0.004588178358972073, + 0.014503628946840763, + 0.012875651940703392, + 0.009865514002740383, + 0.005543386563658714, + 0.007804010063409805, + 0.005372489802539349, + 0.005581298377364874, + 0.0071081919595599174, + 0.010613585822284222, + 0.011602945625782013, + 0.01708519645035267, + 0.009822147898375988, + 0.006532222963869572, + 0.00693219481036067, + 0.004697042517364025, + 0.009066158905625343, + 0.01355519238859415, + 0.015406211838126183, + 0.014766762033104897, + 0.018562979996204376, + 0.015779832378029823, + 0.007406735327094793, + 0.008714117109775543, + 0.006166710052639246, + 0.007544567808508873, + 0.006650034803897142, + 0.015464224852621555, + 0.011189230717718601, + 0.007309818174690008, + 0.008513535372912884, + 0.006486888043582439, + 0.002932487055659294, + 0.006588413845747709, + 0.010565283708274364, + 0.005903932731598616, + 0.008902822621166706, + 0.0026317364536225796, + 0.003169009229168296, + 0.0027983924373984337 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 2 + ], + "selected_reward": -0.04307139292359352 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a shimmering, blood-red gelatin dessert?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the dessert placed on a grimy white porcelain plate with black fragments and hairline cracks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the gelatin dessert exhibit violent, unnatural twitching and spasm motions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a rust-stained fork with a pointed prong visible resting beside the plate on a grimy laminate counter?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature tiny droplets of crimson substance spurting outward from the vibrating dessert?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2520829439163208 + ], + "opposite_logprob_tokens": [ + -1.5020829439163208 + ], + "logprob_reward": -0.2520829439163208, + "yes_no_margin_reward": 1.25, + "token_credit_reward": -0.2520829439163208, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1643792688846588, + 0.06867829710245132, + 0.029227782040834427, + 0.11835455149412155, + 0.22999659180641174, + 0.23739489912986755, + 0.03599850460886955, + 0.024051936343312263, + 0.04384559765458107, + 0.04807252436876297 + ], + "credit_map": [ + 0.016538646072149277, + 0.02911483310163021, + 0.020146936178207397, + 0.03485407307744026, + 0.019188040867447853, + 0.01406747568398714, + 0.018282020464539528, + 0.012187257409095764, + 0.006427417974919081, + 0.0118556534871459, + 0.007797902449965477, + 0.011808594688773155, + 0.011831168085336685, + 0.007006902247667313, + 0.007214684970676899, + 0.004735974129289389, + 0.0030286216642707586, + 0.005531672853976488, + 0.004837559070438147, + 0.004457942675799131, + 0.0033576395362615585, + 0.0024197231978178024, + 0.0033517766278237104, + 0.002242845483124256, + 0.013324841856956482, + 0.015943404287099838, + 0.01207805797457695, + 0.022219160571694374, + 0.011908356100320816, + 0.0132077531889081, + 0.016475817188620567, + 0.013197170570492744, + 0.023139825090765953, + 0.05153941735625267, + 0.03226426616311073, + 0.020164353772997856, + 0.02472241036593914, + 0.03538263216614723, + 0.019616395235061646, + 0.023167280480265617, + 0.01329486258327961, + 0.10208432376384735, + 0.060822054743766785, + 0.016683993861079216, + 0.016721582040190697, + 0.011507328599691391, + 0.010006856173276901, + 0.00627389270812273, + 0.004035836085677147, + 0.003142438130453229, + 0.004123789258301258, + 0.003392278216779232, + 0.00761144096031785, + 0.008040168322622776, + 0.0023267557844519615, + 0.0033257966861128807, + 0.002319077728316188, + 0.002652737544849515, + 0.0022117746993899345, + 0.004035502206534147, + 0.0033394605852663517, + 0.004785343073308468, + 0.001973671605810523, + 0.002734370296820998, + 0.003185782115906477, + 0.0052542658522725105, + 0.006015143357217312, + 0.009555387310683727, + 0.008832794614136219, + 0.004056351725012064, + 0.004376270342618227, + 0.0025696002412587404, + 0.00459734071046114, + 0.005016101989895105, + 0.008780356496572495, + 0.0049216775223612785, + 0.0137672433629632, + 0.0045201750472188, + 0.003019484458491206, + 0.003450141055509448 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 4, + 0 + ], + "selected_reward": -0.2520829439163208 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a shimmering, blood-red gelatin dessert?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the dessert placed on a grimy white porcelain plate with black fragments and hairline cracks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the gelatin dessert exhibit violent, unnatural twitching and spasm motions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a rust-stained fork with a pointed prong visible resting beside the plate on a grimy laminate counter?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature tiny droplets of crimson substance spurting outward from the vibrating dessert?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02328772470355034 + ], + "opposite_logprob_tokens": [ + -3.773287773132324 + ], + "logprob_reward": -0.02328772470355034, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.02328772470355034, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08743247389793396, + 0.05550843104720116, + 0.08523435145616531, + 0.24423520267009735, + 0.06687746196985245, + 0.031557273119688034, + 0.20361150801181793, + 0.12463679164648056, + 0.029564563184976578, + 0.07134192436933517 + ], + "credit_map": [ + 0.0075360932387411594, + 0.016182366758584976, + 0.00904133915901184, + 0.012176324613392353, + 0.014028317295014858, + 0.009845835156738758, + 0.009717771783471107, + 0.008904431946575642, + 0.006092585623264313, + 0.00834461860358715, + 0.0044367965310812, + 0.0066357203759253025, + 0.008922509849071503, + 0.010060237720608711, + 0.005522291176021099, + 0.005493673030287027, + 0.009362660348415375, + 0.011628118343651295, + 0.01305755041539669, + 0.009059404022991657, + 0.009987243451178074, + 0.011099154129624367, + 0.007607918698340654, + 0.013432304374873638, + 0.025408992543816566, + 0.04213123396039009, + 0.024684296920895576, + 0.06204672530293465, + 0.02330041490495205, + 0.024452243000268936, + 0.027206867933273315, + 0.015004429034888744, + 0.0053574396297335625, + 0.014456769451498985, + 0.008686776272952557, + 0.011938938871026039, + 0.006214807741343975, + 0.006322531960904598, + 0.005343361757695675, + 0.008556841872632504, + 0.002125048078596592, + 0.009805608540773392, + 0.006414738483726978, + 0.0037409940268844366, + 0.0031456430442631245, + 0.00254233181476593, + 0.0021043047308921814, + 0.0016786045162007213, + 0.014882463030517101, + 0.026943104341626167, + 0.023793265223503113, + 0.039817970246076584, + 0.046061139553785324, + 0.016499554738402367, + 0.019189080223441124, + 0.016424939036369324, + 0.016403600573539734, + 0.01359157357364893, + 0.012742344290018082, + 0.010217608883976936, + 0.0322764590382576, + 0.021276813000440598, + 0.008257314562797546, + 0.00987108051776886, + 0.0035380860790610313, + 0.004802567884325981, + 0.003301203018054366, + 0.005015656817704439, + 0.0046029649674892426, + 0.002875021193176508, + 0.0029388347174972296, + 0.002490229671820998, + 0.0058265188708901405, + 0.015018315985798836, + 0.006457305047661066, + 0.011042950674891472, + 0.015326429158449173, + 0.005764756817370653, + 0.008053303696215153, + 0.0038523427210748196 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 6, + 7 + ], + "selected_reward": -0.02328772470355034 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a shimmering, blood-red gelatin dessert?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the dessert placed on a grimy white porcelain plate with black fragments and hairline cracks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the gelatin dessert exhibit violent, unnatural twitching and spasm motions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a rust-stained fork with a pointed prong visible resting beside the plate on a grimy laminate counter?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature tiny droplets of crimson substance spurting outward from the vibrating dessert?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11293000727891922 + ], + "opposite_logprob_tokens": [ + -2.2379300594329834 + ], + "logprob_reward": -0.11293000727891922, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11293000727891922, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.117343969643116, + 0.028712604194879532, + 0.027845174074172974, + 0.07368133217096329, + 0.11058509349822998, + 0.21861813962459564, + 0.06967280805110931, + 0.1542678326368332, + 0.07446504384279251, + 0.12480799108743668 + ], + "credit_map": [ + 0.009426127187907696, + 0.024036014452576637, + 0.011388867162168026, + 0.03030703403055668, + 0.015794092789292336, + 0.00781589187681675, + 0.010055892169475555, + 0.008520044386386871, + 0.0029697692953050137, + 0.005086512304842472, + 0.0028415615670382977, + 0.003568517044186592, + 0.005901715252548456, + 0.0030676014721393585, + 0.002853331621736288, + 0.0024235937744379044, + 0.0022814928088337183, + 0.004975161515176296, + 0.0029991911724209785, + 0.0035878221970051527, + 0.00390652334317565, + 0.0030249948613345623, + 0.0029660752043128014, + 0.00410391204059124, + 0.009137623012065887, + 0.01081237941980362, + 0.007357093971222639, + 0.013654991053044796, + 0.007129006087779999, + 0.007504976354539394, + 0.009340066462755203, + 0.00874520093202591, + 0.01221098005771637, + 0.01817590370774269, + 0.014114218764007092, + 0.010516502894461155, + 0.01455132756382227, + 0.02016863040626049, + 0.008607711642980576, + 0.012239813804626465, + 0.012892022728919983, + 0.09487921744585037, + 0.05822019651532173, + 0.012270179577171803, + 0.01600951887667179, + 0.009306322783231735, + 0.008372609503567219, + 0.006668075453490019, + 0.008426024578511715, + 0.00835480634123087, + 0.0068646264262497425, + 0.011905056424438953, + 0.008255716413259506, + 0.012222078628838062, + 0.005465060938149691, + 0.00817943550646305, + 0.019081363454461098, + 0.020118072628974915, + 0.012779341079294682, + 0.027867503464221954, + 0.02336684986948967, + 0.026382319629192352, + 0.009884392842650414, + 0.014787986874580383, + 0.006549579091370106, + 0.010215765796601772, + 0.006589546799659729, + 0.01392513606697321, + 0.014187710359692574, + 0.00814833678305149, + 0.008951284922659397, + 0.005897682625800371, + 0.010319942608475685, + 0.02613077685236931, + 0.03314529359340668, + 0.009787340648472309, + 0.01909804716706276, + 0.010734844021499157, + 0.008598245680332184, + 0.006993493996560574 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 7, + 9 + ], + "selected_reward": -0.11293000727891922 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "Shot in an eerie, cartoonish horror style, extreme close-up static footage at a high frame rate focuses on a shimmering, blood-red gelatin dessert. The translucent substance teeters on a grimy white porcelain plate, dotted with protruding black fragments and faint, hairline cracks, when it suddenly begins to twitch violently. Each unnatural, arrhythmic tremor and spasm is rendered with unsettling clarity, the gelatinous form teetering and contracting as if it were a living organ in agony. The entire scene is set against a grimy, stained laminate kitchen counter, a rust-stained fork resting ominously beside the plate, its prong pointed at the trembling dessert. Under the harsh, flickering light of a single overhead bulb, the gelatin's surface gleams with a sickly, damp sheen, and as it vibrates, tiny droplets of the crimson substance spurt outward, captured in sharp detail as they arc through the air. The vibrant yet irritating red of the dessert contrasts starkly with the grimy yellows and browns of the grimy background, creating a deeply unsettling atmosphere of body horror and disgust.", + "seed": 22236, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000121/group_13/index_9596_seed_22236_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": -0.05564295239746571, + "vlm_logprob": -0.05564295239746571, + "vlm_yes_no_margin": 3.349999952316284, + "vlm_token_credit": -0.05564295239746571, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a shimmering, blood-red gelatin dessert?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the dessert placed on a grimy white porcelain plate with black fragments and hairline cracks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the gelatin dessert exhibit violent, unnatural twitching and spasm motions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a rust-stained fork with a pointed prong visible resting beside the plate on a grimy laminate counter?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature tiny droplets of crimson substance spurting outward from the vibrating dessert?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.015036454424262047 + ], + "opposite_logprob_tokens": [ + -5.265036582946777 + ], + "logprob_reward": -0.015036454424262047, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.015036454424262047, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2142418473958969, + 0.23534128069877625, + 0.08166718482971191, + 0.021073253825306892, + 0.029188983142375946, + 0.06701882183551788, + 0.17171794176101685, + 0.05338607355952263, + 0.05580559000372887, + 0.0705590769648552 + ], + "credit_map": [ + 0.04121715947985649, + 0.01435741689056158, + 0.012024580501019955, + 0.022379372268915176, + 0.027170009911060333, + 0.012151592411100864, + 0.06587909162044525, + 0.01906261220574379, + 0.0679556280374527, + 0.011806161142885685, + 0.013962291181087494, + 0.01782248355448246, + 0.016538241878151894, + 0.010787519626319408, + 0.07963643223047256, + 0.016832489520311356, + 0.017697995528578758, + 0.006570412311702967, + 0.005501758307218552, + 0.006012542638927698, + 0.010427681729197502, + 0.004480633419007063, + 0.02568664401769638, + 0.0052895075641572475, + 0.0031389519572257996, + 0.0015676285838708282, + 0.0033073839731514454, + 0.001849862397648394, + 0.003696356900036335, + 0.0018818979151546955, + 0.003353473264724016, + 0.0022776960395276546, + 0.003909239079803228, + 0.002129031578078866, + 0.005851199850440025, + 0.002426732564345002, + 0.00471620773896575, + 0.002471442800015211, + 0.0055352444760501385, + 0.0021498845890164375, + 0.009047970175743103, + 0.0047372193075716496, + 0.015409428626298904, + 0.0053174737840890884, + 0.009817391633987427, + 0.007341628894209862, + 0.00894453190267086, + 0.006403171457350254, + 0.031045356765389442, + 0.010405020788311958, + 0.015629209578037262, + 0.008603636175394058, + 0.018336158245801926, + 0.020138902589678764, + 0.05385631322860718, + 0.013703327625989914, + 0.012250115163624287, + 0.003712845267727971, + 0.00360425328835845, + 0.0035791872069239616, + 0.007753077428787947, + 0.005035687237977982, + 0.012979643419384956, + 0.004471256397664547, + 0.00965805258601904, + 0.009271930903196335, + 0.007075934205204248, + 0.004532734863460064, + 0.007454037666320801, + 0.00474986806511879, + 0.00859790202230215, + 0.004465125966817141, + 0.009791900403797626, + 0.009390858933329582, + 0.013678533025085926, + 0.007495534140616655, + 0.007525231689214706, + 0.008004955016076565, + 0.008754348382353783, + 0.005917709320783615 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 6 + ], + "selected_reward": -0.015036454424262047 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a shimmering, blood-red gelatin dessert?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the dessert placed on a grimy white porcelain plate with black fragments and hairline cracks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the gelatin dessert exhibit violent, unnatural twitching and spasm motions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a rust-stained fork with a pointed prong visible resting beside the plate on a grimy laminate counter?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature tiny droplets of crimson substance spurting outward from the vibrating dessert?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018204864114522934 + ], + "opposite_logprob_tokens": [ + -4.018204689025879 + ], + "logprob_reward": -0.018204864114522934, + "yes_no_margin_reward": 3.999999761581421, + "token_credit_reward": -0.018204864114522934, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1191607117652893, + 0.17806683480739594, + 0.02416614070534706, + 0.05195446312427521, + 0.06752242892980576, + 0.05423973128199577, + 0.050615403801202774, + 0.016713252291083336, + 0.38872668147087097, + 0.04883432760834694 + ], + "credit_map": [ + 0.022801056504249573, + 0.007867972366511822, + 0.008603795431554317, + 0.012488308362662792, + 0.010418894700706005, + 0.013162679970264435, + 0.03342144936323166, + 0.010396545752882957, + 0.04738351330161095, + 0.00797255989164114, + 0.010563859716057777, + 0.013285022228956223, + 0.014462760649621487, + 0.008703436702489853, + 0.06101837754249573, + 0.01467729639261961, + 0.00396683206781745, + 0.0027982082683593035, + 0.003051561303436756, + 0.002151012886315584, + 0.004519190173596144, + 0.002583016874268651, + 0.003323014359921217, + 0.0017733037238940597, + 0.006278280168771744, + 0.004006695002317429, + 0.005772237200289965, + 0.0045285881496965885, + 0.014539853669703007, + 0.0044267065823078156, + 0.008070283569395542, + 0.004331819713115692, + 0.0070912777446210384, + 0.005637321155518293, + 0.00929327867925167, + 0.006244112737476826, + 0.01130157895386219, + 0.0105275334790349, + 0.010914054699242115, + 0.006513272877782583, + 0.0076872860081493855, + 0.0038271716330200434, + 0.014092570170760155, + 0.004032924771308899, + 0.005727304145693779, + 0.005907682701945305, + 0.008010448887944221, + 0.004954339936375618, + 0.007119869813323021, + 0.00325832050293684, + 0.008275415748357773, + 0.0033139041624963284, + 0.005117119289934635, + 0.006807259749621153, + 0.012509064748883247, + 0.0042144497856497765, + 0.002504238858819008, + 0.0013551641022786498, + 0.0033292826265096664, + 0.0014890398597344756, + 0.002161131240427494, + 0.0017244457267224789, + 0.0027569252997636795, + 0.0013930252753198147, + 0.09670994430780411, + 0.14025598764419556, + 0.01669039949774742, + 0.02158372476696968, + 0.04248619079589844, + 0.01218762993812561, + 0.04146235063672066, + 0.017350422218441963, + 0.005156646016985178, + 0.006279942113906145, + 0.00868596788495779, + 0.0062560648657381535, + 0.004909045062959194, + 0.0071012359112501144, + 0.005624992307275534, + 0.0048204343765974045 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 1, + 0 + ], + "selected_reward": -0.018204864114522934 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a shimmering, blood-red gelatin dessert?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the dessert placed on a grimy white porcelain plate with black fragments and hairline cracks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the gelatin dessert exhibit violent, unnatural twitching and spasm motions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a rust-stained fork with a pointed prong visible resting beside the plate on a grimy laminate counter?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature tiny droplets of crimson substance spurting outward from the vibrating dessert?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11291296780109406 + ], + "opposite_logprob_tokens": [ + -2.237912893295288 + ], + "logprob_reward": -0.11291296780109406, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11291296780109406, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18008500337600708, + 0.21668384969234467, + 0.05228984355926514, + 0.05980970710515976, + 0.06392689049243927, + 0.0670582577586174, + 0.09422432631254196, + 0.09236491471529007, + 0.08399586379528046, + 0.08956136554479599 + ], + "credit_map": [ + 0.024915486574172974, + 0.0127129927277565, + 0.015025908127427101, + 0.015587531961500645, + 0.031152620911598206, + 0.031099636107683182, + 0.033827297389507294, + 0.01576351560652256, + 0.06047589331865311, + 0.011020983569324017, + 0.012611454352736473, + 0.016892990097403526, + 0.015822863206267357, + 0.011674143373966217, + 0.07211218029260635, + 0.016073329374194145, + 0.00606051180511713, + 0.005064807366579771, + 0.0047329505905508995, + 0.005545307882130146, + 0.00925474613904953, + 0.00605497183278203, + 0.010382439941167831, + 0.005194106139242649, + 0.00719904201105237, + 0.00462453393265605, + 0.004266227129846811, + 0.007243246305733919, + 0.016280004754662514, + 0.00460328022018075, + 0.010236685164272785, + 0.005356692709028721, + 0.009000726975500584, + 0.004033183678984642, + 0.012172233313322067, + 0.005344111006706953, + 0.007442072499543428, + 0.005975807085633278, + 0.013568886555731297, + 0.006389870774000883, + 0.009049045853316784, + 0.004488894250243902, + 0.010798217728734016, + 0.004991899710148573, + 0.009484286420047283, + 0.006057255435734987, + 0.015140931122004986, + 0.007047720719128847, + 0.013150387443602085, + 0.006043590139597654, + 0.010076901875436306, + 0.005053891334682703, + 0.01286560483276844, + 0.011610274203121662, + 0.02617555484175682, + 0.009248130954802036, + 0.01245497539639473, + 0.006194845773279667, + 0.006748623214662075, + 0.005890140309929848, + 0.017907503992319107, + 0.00931820273399353, + 0.023300012573599815, + 0.010550613515079021, + 0.014343587681651115, + 0.025226043537259102, + 0.006043565459549427, + 0.004576527047902346, + 0.010318591259419918, + 0.007247866131365299, + 0.01210835948586464, + 0.004131324589252472, + 0.014426622539758682, + 0.02366238459944725, + 0.008007382974028587, + 0.006234966218471527, + 0.010960997082293034, + 0.008624999783933163, + 0.010177316144108772, + 0.007466698996722698 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 6 + ], + "selected_reward": -0.11291296780109406 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a shimmering, blood-red gelatin dessert?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the dessert placed on a grimy white porcelain plate with black fragments and hairline cracks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the gelatin dessert exhibit violent, unnatural twitching and spasm motions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a rust-stained fork with a pointed prong visible resting beside the plate on a grimy laminate counter?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature tiny droplets of crimson substance spurting outward from the vibrating dessert?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"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": [ + 0.07469339668750763, + 0.09518513083457947, + 0.06967541575431824, + 0.06532144546508789, + 0.23851525783538818, + 0.04570820555090904, + 0.059937626123428345, + 0.03490934893488884, + 0.1929367631673813, + 0.12311751395463943 + ], + "credit_map": [ + 0.00971731636673212, + 0.005757317412644625, + 0.009152805432677269, + 0.008040739223361015, + 0.00893395859748125, + 0.011661709286272526, + 0.014916008338332176, + 0.006513537839055061, + 0.013410634361207485, + 0.011320389807224274, + 0.010311468504369259, + 0.007725297007709742, + 0.01571323163807392, + 0.013207849115133286, + 0.012356147170066833, + 0.011140103451907635, + 0.017809467390179634, + 0.004096621181815863, + 0.004915835335850716, + 0.004488730803132057, + 0.007837595418095589, + 0.003842054633423686, + 0.02232305519282818, + 0.004362049046903849, + 0.008020845241844654, + 0.004201211966574192, + 0.013700268231332302, + 0.006532875820994377, + 0.010352515615522861, + 0.005454854108393192, + 0.01117573119699955, + 0.005883142352104187, + 0.03148682788014412, + 0.021255576983094215, + 0.05572562664747238, + 0.0160536989569664, + 0.024260958656668663, + 0.022304557263851166, + 0.04490644857287407, + 0.022521529346704483, + 0.005625746678560972, + 0.003197717247530818, + 0.01164768636226654, + 0.003849337575957179, + 0.004970742389559746, + 0.005418387707322836, + 0.006873003207147121, + 0.004125584382563829, + 0.00974436104297638, + 0.003886300139129162, + 0.007942278869450092, + 0.003767724148929119, + 0.005431788973510265, + 0.006971518509089947, + 0.017031505703926086, + 0.005162141285836697, + 0.006105125416070223, + 0.0019709363114088774, + 0.006719321943819523, + 0.00272695766761899, + 0.003657872788608074, + 0.0039476025849580765, + 0.007011341862380505, + 0.002770187333226204, + 0.04822118207812309, + 0.06600602716207504, + 0.009637857787311077, + 0.010630766861140728, + 0.021412627771496773, + 0.006867905147373676, + 0.021035294979810715, + 0.009125086478888988, + 0.018986299633979797, + 0.035660162568092346, + 0.014057482592761517, + 0.010828730650246143, + 0.012650604359805584, + 0.010249068029224873, + 0.011298482306301594, + 0.009386679157614708 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 8, + 9 + ], + "selected_reward": -0.0700213760137558 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a shimmering, blood-red gelatin dessert?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the dessert placed on a grimy white porcelain plate with black fragments and hairline cracks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the gelatin dessert exhibit violent, unnatural twitching and spasm motions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a rust-stained fork with a pointed prong visible resting beside the plate on a grimy laminate counter?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature tiny droplets of crimson substance spurting outward from the vibrating dessert?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.062039099633693695 + ], + "opposite_logprob_tokens": [ + -2.8120391368865967 + ], + "logprob_reward": -0.062039099633693695, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.062039099633693695, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.049189578741788864, + 0.5032230019569397, + 0.0432305634021759, + 0.03624073415994644, + 0.02741185389459133, + 0.04848885536193848, + 0.03461058810353279, + 0.03685549274086952, + 0.15657342970371246, + 0.06417597085237503 + ], + "credit_map": [ + 0.007575392257422209, + 0.004502600524574518, + 0.00494014797732234, + 0.00562400883063674, + 0.006889170501381159, + 0.009622634388506413, + 0.0060883634723722935, + 0.003947258926928043, + 0.14836153388023376, + 0.021297330036759377, + 0.028883256018161774, + 0.036021191626787186, + 0.03017326444387436, + 0.020761635154485703, + 0.1780463606119156, + 0.0396784208714962, + 0.008815662004053593, + 0.0028848969377577305, + 0.0030828164890408516, + 0.003771387040615082, + 0.007833834737539291, + 0.002101307036355138, + 0.011949636973440647, + 0.002791026374325156, + 0.004318345803767443, + 0.003094006795436144, + 0.0030687639955431223, + 0.003873667446896434, + 0.009139382280409336, + 0.003376506734639406, + 0.006285085808485746, + 0.003084974130615592, + 0.0023970049805939198, + 0.0023527860175818205, + 0.004202489275485277, + 0.0037790201604366302, + 0.004349150229245424, + 0.003347990568727255, + 0.00358755886554718, + 0.003395853331312537, + 0.00599465100094676, + 0.003906627651304007, + 0.01112421415746212, + 0.0032679049763828516, + 0.005496854428201914, + 0.004629123955965042, + 0.00923170056194067, + 0.004837776999920607, + 0.0055487803183496, + 0.0024105478078126907, + 0.005948427598923445, + 0.0023927774745970964, + 0.0032636423129588366, + 0.004609983414411545, + 0.007401361130177975, + 0.0030350664164870977, + 0.0052339802496135235, + 0.0031687724404037, + 0.004306293558329344, + 0.003052199026569724, + 0.006832861341536045, + 0.003474223194643855, + 0.006391983479261398, + 0.004395177587866783, + 0.03949647396802902, + 0.04271911829710007, + 0.006242104806005955, + 0.009683281183242798, + 0.020437534898519516, + 0.006991437636315823, + 0.021668849512934685, + 0.009334624744951725, + 0.013911614194512367, + 0.018515588715672493, + 0.004864963702857494, + 0.004506226163357496, + 0.0064474837854504585, + 0.00515542970970273, + 0.006300153210759163, + 0.004474511370062828 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 8, + 9 + ], + "selected_reward": -0.062039099633693695 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "Shot in an eerie, cartoonish horror style, extreme close-up static footage at a high frame rate focuses on a shimmering, blood-red gelatin dessert. The translucent substance teeters on a grimy white porcelain plate, dotted with protruding black fragments and faint, hairline cracks, when it suddenly begins to twitch violently. Each unnatural, arrhythmic tremor and spasm is rendered with unsettling clarity, the gelatinous form teetering and contracting as if it were a living organ in agony. The entire scene is set against a grimy, stained laminate kitchen counter, a rust-stained fork resting ominously beside the plate, its prong pointed at the trembling dessert. Under the harsh, flickering light of a single overhead bulb, the gelatin's surface gleams with a sickly, damp sheen, and as it vibrates, tiny droplets of the crimson substance spurt outward, captured in sharp detail as they arc through the air. The vibrant yet irritating red of the dessert contrasts starkly with the grimy yellows and browns of the grimy background, creating a deeply unsettling atmosphere of body horror and disgust.", + "seed": 22237, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000121/group_13/index_9596_seed_22237_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": -0.03125288113951683, + "vlm_logprob": -0.03125288113951683, + "vlm_yes_no_margin": 3.775, + "vlm_token_credit": -0.03125288113951683, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a shimmering, blood-red gelatin dessert?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the dessert placed on a grimy white porcelain plate with black fragments and hairline cracks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the gelatin dessert exhibit violent, unnatural twitching and spasm motions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a rust-stained fork with a pointed prong visible resting beside the plate on a grimy laminate counter?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature tiny droplets of crimson substance spurting outward from the vibrating dessert?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012002626433968544 + ], + "opposite_logprob_tokens": [ + -5.012002468109131 + ], + "logprob_reward": -0.012002626433968544, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.012002626433968544, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15345856547355652, + 0.1689756214618683, + 0.26114192605018616, + 0.060855671763420105, + 0.07898980379104614, + 0.04479639604687691, + 0.04011962190270424, + 0.049281030893325806, + 0.03157011419534683, + 0.11081115901470184 + ], + "credit_map": [ + 0.014289389364421368, + 0.010523825883865356, + 0.017946120351552963, + 0.015345191583037376, + 0.029828693717718124, + 0.018009722232818604, + 0.026682069525122643, + 0.020833559334278107, + 0.010740025900304317, + 0.007522881031036377, + 0.009849965572357178, + 0.008801509626209736, + 0.009949949570000172, + 0.06690943241119385, + 0.04477901756763458, + 0.010422861203551292, + 0.015886779874563217, + 0.009318403899669647, + 0.007737613748759031, + 0.012193865142762661, + 0.016711879521608353, + 0.10843263566493988, + 0.07599668949842453, + 0.014864102937281132, + 0.007425864227116108, + 0.005688767414540052, + 0.007769864052534103, + 0.00643055560067296, + 0.00757558224722743, + 0.006553360726684332, + 0.010021882131695747, + 0.009389800950884819, + 0.004276461899280548, + 0.00886264257133007, + 0.004907120019197464, + 0.00920795276761055, + 0.021445443853735924, + 0.006730782333761454, + 0.012464352883398533, + 0.01109505258500576, + 0.0040563251823186874, + 0.004020141903311014, + 0.004764281213283539, + 0.007209621835500002, + 0.006185863167047501, + 0.004772381391376257, + 0.006741164717823267, + 0.0070466226898133755, + 0.004274630919098854, + 0.0025768636260181665, + 0.003040939336642623, + 0.005053006578236818, + 0.00484366063028574, + 0.008657531812787056, + 0.007391731254756451, + 0.004281261470168829, + 0.005862078629434109, + 0.0031433668918907642, + 0.0065009454265236855, + 0.006326308473944664, + 0.0041810572147369385, + 0.006718598306179047, + 0.008740536868572235, + 0.007808144204318523, + 0.004893387667834759, + 0.003266169922426343, + 0.004033337347209454, + 0.004584703128784895, + 0.0032661757431924343, + 0.0035744113847613335, + 0.0038991994224488735, + 0.004052732139825821, + 0.014963598921895027, + 0.010936371050775051, + 0.013637973926961422, + 0.018620314076542854, + 0.011545532383024693, + 0.014458105899393559, + 0.017069226130843163, + 0.009580044075846672 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 1, + 0 + ], + "selected_reward": -0.012002626433968544 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a shimmering, blood-red gelatin dessert?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the dessert placed on a grimy white porcelain plate with black fragments and hairline cracks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the gelatin dessert exhibit violent, unnatural twitching and spasm motions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a rust-stained fork with a pointed prong visible resting beside the plate on a grimy laminate counter?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature tiny droplets of crimson substance spurting outward from the vibrating dessert?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014218922704458237 + ], + "opposite_logprob_tokens": [ + -4.264218807220459 + ], + "logprob_reward": -0.014218922704458237, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014218922704458237, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.27516406774520874, + 0.17728938162326813, + 0.08096790313720703, + 0.09093822538852692, + 0.11037567257881165, + 0.0817008763551712, + 0.03838317096233368, + 0.03908227011561394, + 0.05663188919425011, + 0.04946647211909294 + ], + "credit_map": [ + 0.01723572425544262, + 0.015830982476472855, + 0.020476043224334717, + 0.01764177344739437, + 0.04165001958608627, + 0.07285384833812714, + 0.07050757110118866, + 0.01896812580525875, + 0.012078206986188889, + 0.009368343278765678, + 0.012730048969388008, + 0.013043424114584923, + 0.011027460917830467, + 0.06119542196393013, + 0.04321425408124924, + 0.014632229693233967, + 0.007646077778190374, + 0.00450694328173995, + 0.01100149191915989, + 0.006018179003149271, + 0.007516957353800535, + 0.023056291043758392, + 0.015066275373101234, + 0.00615569157525897, + 0.009840880520641804, + 0.009013257920742035, + 0.010108347050845623, + 0.009664739482104778, + 0.012851540930569172, + 0.009877282194793224, + 0.014531035907566547, + 0.015051154419779778, + 0.0080454396083951, + 0.010027662850916386, + 0.011294661089777946, + 0.011052262037992477, + 0.023719118908047676, + 0.014181356877088547, + 0.018236421048641205, + 0.013818761333823204, + 0.007139294408261776, + 0.008391746319830418, + 0.00719325290992856, + 0.012918302789330482, + 0.009878364391624928, + 0.00888758897781372, + 0.015699177980422974, + 0.011593155562877655, + 0.0025549493730068207, + 0.0035743536427617073, + 0.004115783143788576, + 0.006136762909591198, + 0.00576422456651926, + 0.0052494509145617485, + 0.006213475484400988, + 0.004774173256009817, + 0.005710366647690535, + 0.0034235124476253986, + 0.007050410844385624, + 0.004165430553257465, + 0.0033126729540526867, + 0.006135998759418726, + 0.005546738859266043, + 0.0037371402140706778, + 0.009415396489202976, + 0.005842993501573801, + 0.0070493207313120365, + 0.007029568310827017, + 0.00533851096406579, + 0.008355588652193546, + 0.006938988342881203, + 0.006661526393145323, + 0.007686239201575518, + 0.0039055864326655865, + 0.007405194453895092, + 0.006398008204996586, + 0.0046877129934728146, + 0.00838608480989933, + 0.00687207980081439, + 0.00412557041272521 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 4 + ], + "selected_reward": -0.014218922704458237 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a shimmering, blood-red gelatin dessert?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the dessert placed on a grimy white porcelain plate with black fragments and hairline cracks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the gelatin dessert exhibit violent, unnatural twitching and spasm motions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a rust-stained fork with a pointed prong visible resting beside the plate on a grimy laminate counter?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature tiny droplets of crimson substance spurting outward from the vibrating dessert?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07000703364610672 + ], + "opposite_logprob_tokens": [ + -2.695007085800171 + ], + "logprob_reward": -0.07000703364610672, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07000703364610672, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.355086088180542, + 0.13253489136695862, + 0.06836364418268204, + 0.04533229395747185, + 0.12309727072715759, + 0.06531219929456711, + 0.04662343114614487, + 0.03092109225690365, + 0.04018349200487137, + 0.09254556894302368 + ], + "credit_map": [ + 0.0604044646024704, + 0.022012174129486084, + 0.03802306205034256, + 0.027184341102838516, + 0.04703766107559204, + 0.0664689689874649, + 0.0539948008954525, + 0.039960648864507675, + 0.009477801620960236, + 0.0067785680294036865, + 0.0076794931665062904, + 0.008314843289554119, + 0.011139143258333206, + 0.044764988124370575, + 0.03345901146531105, + 0.010921061038970947, + 0.005642366595566273, + 0.004496637266129255, + 0.0054572452791035175, + 0.004150881431996822, + 0.009257441386580467, + 0.020294198766350746, + 0.012508640065789223, + 0.006556239444762468, + 0.003663369920104742, + 0.004393485374748707, + 0.0047012981958687305, + 0.00535209383815527, + 0.007980231195688248, + 0.004750717431306839, + 0.007518047001212835, + 0.006973055191338062, + 0.006862827576696873, + 0.01564975641667843, + 0.009522783569991589, + 0.020269956439733505, + 0.027958426624536514, + 0.011025937274098396, + 0.01937689445912838, + 0.012430693954229355, + 0.004845275077968836, + 0.007047584746032953, + 0.005926314741373062, + 0.011500634253025055, + 0.009272761642932892, + 0.007290632463991642, + 0.010561834089457989, + 0.008867170661687851, + 0.0031271432526409626, + 0.0047577000223100185, + 0.004555969964712858, + 0.008230787701904774, + 0.006281422451138496, + 0.006038432940840721, + 0.007819426245987415, + 0.005812556017190218, + 0.0031872112303972244, + 0.0020871167071163654, + 0.003713991492986679, + 0.004173734225332737, + 0.003241196507588029, + 0.004684600979089737, + 0.006027617491781712, + 0.0038056245539337397, + 0.005257614888250828, + 0.0033406198490411043, + 0.004923759028315544, + 0.005656799301505089, + 0.0036996190901845694, + 0.00394354248419404, + 0.005502038169652224, + 0.007859504781663418, + 0.010752205736935139, + 0.008265508338809013, + 0.010083905421197414, + 0.018104316666722298, + 0.008131178095936775, + 0.012851507402956486, + 0.014257320202887058, + 0.010099628008902073 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 4 + ], + "selected_reward": -0.07000703364610672 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a shimmering, blood-red gelatin dessert?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the dessert placed on a grimy white porcelain plate with black fragments and hairline cracks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the gelatin dessert exhibit violent, unnatural twitching and spasm motions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a rust-stained fork with a pointed prong visible resting beside the plate on a grimy laminate counter?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature tiny droplets of crimson substance spurting outward from the vibrating dessert?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026338299736380577 + ], + "opposite_logprob_tokens": [ + -3.6513383388519287 + ], + "logprob_reward": -0.026338299736380577, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026338299736380577, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.25279077887535095, + 0.13654986023902893, + 0.10974204540252686, + 0.07046770304441452, + 0.15502537786960602, + 0.09562347829341888, + 0.027490921318531036, + 0.04902079701423645, + 0.042704027146101, + 0.060584940016269684 + ], + "credit_map": [ + 0.02364308573305607, + 0.015933463349938393, + 0.01834246888756752, + 0.013280980288982391, + 0.03622850030660629, + 0.06029367074370384, + 0.05484243109822273, + 0.030226213857531548, + 0.016704972833395004, + 0.007739367429167032, + 0.01396922767162323, + 0.008329622447490692, + 0.015118144452571869, + 0.03709283098578453, + 0.021757328882813454, + 0.015838366001844406, + 0.015387462452054024, + 0.0065679969266057014, + 0.011881598271429539, + 0.006133404094725847, + 0.009689661674201488, + 0.02934849262237549, + 0.016900697723031044, + 0.013832734897732735, + 0.008648529648780823, + 0.00791851431131363, + 0.011316517367959023, + 0.0075989291071891785, + 0.008342036046087742, + 0.007943554781377316, + 0.009384987875819206, + 0.009314636699855328, + 0.016173765063285828, + 0.011325621977448463, + 0.024811770766973495, + 0.019070668146014214, + 0.019927384331822395, + 0.015952985733747482, + 0.025311678647994995, + 0.022451521828770638, + 0.012153605930507183, + 0.009837640449404716, + 0.015364116989076138, + 0.012930761091411114, + 0.007486975751817226, + 0.011027294211089611, + 0.013293217867612839, + 0.013529877178370953, + 0.002673619193956256, + 0.0016253822250291705, + 0.003036021487787366, + 0.0029925385024398565, + 0.002653588308021426, + 0.0066619738936424255, + 0.005310476291924715, + 0.0025373243261128664, + 0.007631692104041576, + 0.003407440846785903, + 0.006152761168777943, + 0.006750340107828379, + 0.004652146715670824, + 0.006894866935908794, + 0.008174092508852482, + 0.005357458256185055, + 0.00580016290768981, + 0.0038594764191657305, + 0.0040984018705785275, + 0.006599221378564835, + 0.004437122959643602, + 0.005326594691723585, + 0.006038566119968891, + 0.006544485688209534, + 0.00786602683365345, + 0.005117494147270918, + 0.0073804366402328014, + 0.011122078634798527, + 0.00674542598426342, + 0.007431577891111374, + 0.00928658340126276, + 0.005635318346321583 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 1 + ], + "selected_reward": -0.026338299736380577 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a shimmering, blood-red gelatin dessert?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the dessert placed on a grimy white porcelain plate with black fragments and hairline cracks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the gelatin dessert exhibit violent, unnatural twitching and spasm motions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a rust-stained fork with a pointed prong visible resting beside the plate on a grimy laminate counter?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature tiny droplets of crimson substance spurting outward from the vibrating dessert?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"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": [ + 0.2124844342470169, + 0.24985404312610626, + 0.0693417564034462, + 0.0385095588862896, + 0.15930111706256866, + 0.08349420130252838, + 0.04048888757824898, + 0.028610432520508766, + 0.03411538526415825, + 0.08380023390054703 + ], + "credit_map": [ + 0.014131366275250912, + 0.02083575166761875, + 0.017439499497413635, + 0.021143244579434395, + 0.058287639170885086, + 0.021058447659015656, + 0.03636498376727104, + 0.02322346903383732, + 0.015608973801136017, + 0.010935208760201931, + 0.009557338431477547, + 0.014873315580189228, + 0.015157355926930904, + 0.09954626858234406, + 0.0686613991856575, + 0.015514177270233631, + 0.007255027070641518, + 0.0061625223606824875, + 0.008175033144652843, + 0.0076184384524822235, + 0.015632662922143936, + 0.007086421363055706, + 0.010238103568553925, + 0.007173541467636824, + 0.0034538700710982084, + 0.004098931793123484, + 0.0035928781144320965, + 0.004626262933015823, + 0.008426857180893421, + 0.0038746835198253393, + 0.00559638487175107, + 0.004839684348553419, + 0.015523293986916542, + 0.016386697068810463, + 0.013532391749322414, + 0.01870064064860344, + 0.034298233687877655, + 0.018018705770373344, + 0.025870174169540405, + 0.016970958560705185, + 0.006238963920623064, + 0.008112375624477863, + 0.007641893345862627, + 0.01309529785066843, + 0.011600235477089882, + 0.009102198295295238, + 0.016444476321339607, + 0.011258753947913647, + 0.0033076838590204716, + 0.004203520715236664, + 0.0038607255555689335, + 0.007284457795321941, + 0.006163101643323898, + 0.004892162047326565, + 0.006860286928713322, + 0.003916943911463022, + 0.002992927562445402, + 0.0027679100167006254, + 0.002806674689054489, + 0.003956256899982691, + 0.003372574457898736, + 0.003436438739299774, + 0.0046274238266050816, + 0.004650221671909094, + 0.005212641786783934, + 0.0030792593024671078, + 0.003648290643468499, + 0.004837410058826208, + 0.003832110669463873, + 0.003923683427274227, + 0.0047769625671207905, + 0.004805021919310093, + 0.008947867900133133, + 0.00570724718272686, + 0.010002545081079006, + 0.018492890521883965, + 0.009449784643948078, + 0.00942439865320921, + 0.012418609112501144, + 0.00935688428580761 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 4 + ], + "selected_reward": -0.033697523176670074 + } + ], + "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/0000121/group_14/index_tensor([5107, 5107])_group_14_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000121/group_14/index_tensor([5107, 5107])_group_14_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..4da0cb823d9165cb7d0011bbe31eb7a48b0ef1ce --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000121/group_14/index_tensor([5107, 5107])_group_14_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "wide shot, shot from behind and slightly upward, slowly pans forward, gradually moving over their shoulders, revealing a breathtaking scene. Atop a skyscraper, a metal platform crisscrossed with pipes and ventilation systems, a slender, sharply defined figure stands quietly at the edge. They wear a high-collared, waterproof black trench coat, its hem fluttering gently in the wind and rain. Below, a boundless cyberpunk metropolis unfolds. Dense rain, illuminated by countless holographic billboards and giant neon signs, transforms into flickering curtains of light. Deep blue, voluptuous purple, and piercing magenta cast upward, outlining the figure's resolute profile and dampened shoulders. They remain motionless, leaning slightly forward, gazing intently at the steel jungle below, a woven fabric of light, shadow, and rain. Between the towering buildings, a constant stream of cars crisscrosses, leaving long trails of light. The wet metal platform and the character's waterproof windbreaker reflect the ever-changing neon halo, forming an extremely realistic and high-contrast light and shadow effect. The whole picture is full of futuristic coldness, loneliness and magnificent cinematic atmosphere.", + "seed": 5227, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000121/group_14/index_5107_seed_5227_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": -0.09322640188038349, + "vlm_logprob": -0.09322640188038349, + "vlm_yes_no_margin": 3.0, + "vlm_token_credit": -0.09322640188038349, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a wide shot taken from behind and slightly upward of the character?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character wearing a high-collared, waterproof black trench coat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a metal platform atop a skyscraper with pipes and ventilation systems?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a cyberpunk metropolis with dense rain, holographic billboards, and neon signs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly pan forward, gradually moving over the character's shoulders?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04389100894331932 + ], + "opposite_logprob_tokens": [ + -3.918890953063965 + ], + "logprob_reward": -0.04389100894331932, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.04389100894331932, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1313248723745346, + 0.09110582619905472, + 0.12416508048772812, + 0.035959724336862564, + 0.06630763411521912, + 0.05334898829460144, + 0.051845941692590714, + 0.1783151924610138, + 0.22481295466423035, + 0.04281378909945488 + ], + "credit_map": [ + 0.013372096233069897, + 0.009443665854632854, + 0.01298204530030489, + 0.027049217373132706, + 0.011656411923468113, + 0.011270644143223763, + 0.03128571808338165, + 0.014265072531998158, + 0.010121718049049377, + 0.007194338832050562, + 0.012087821029126644, + 0.014111336320638657, + 0.009063183329999447, + 0.011496764607727528, + 0.01697598211467266, + 0.010054678656160831, + 0.009056768380105495, + 0.014211405999958515, + 0.010469562374055386, + 0.018370484933257103, + 0.030799785628914833, + 0.008921868167817593, + 0.023277893662452698, + 0.009057301096618176, + 0.00271215639077127, + 0.002436653710901737, + 0.003062148578464985, + 0.005693772342056036, + 0.00793096050620079, + 0.0030469310004264116, + 0.007758027408272028, + 0.00331907719373703, + 0.0054396684281528, + 0.008236975409090519, + 0.008875813335180283, + 0.01146281324326992, + 0.0072672078385949135, + 0.0053215837106108665, + 0.01045782770961523, + 0.009245743043720722, + 0.004441204015165567, + 0.005684554111212492, + 0.006631386931985617, + 0.00827461015433073, + 0.008917591534554958, + 0.0037935194559395313, + 0.010969433933496475, + 0.004636684432625771, + 0.0025687541346997023, + 0.00344552518799901, + 0.00944559182971716, + 0.007408656179904938, + 0.007983562536537647, + 0.003214538097381592, + 0.012478003278374672, + 0.005301311146467924, + 0.016727382317185402, + 0.017686188220977783, + 0.019534049555659294, + 0.02731332555413246, + 0.021079987287521362, + 0.018671564757823944, + 0.04171938821673393, + 0.01558330561965704, + 0.0216493122279644, + 0.023623788729310036, + 0.028597578406333923, + 0.023628635331988335, + 0.0336412750184536, + 0.03904986381530762, + 0.03674009069800377, + 0.01788240298628807, + 0.0034074201248586178, + 0.004191450774669647, + 0.0062959082424640656, + 0.006452394649386406, + 0.009078413248062134, + 0.0029904600232839584, + 0.0065357438288629055, + 0.003861995879560709 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 0 + ], + "selected_reward": -0.04389100894331932 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a wide shot taken from behind and slightly upward of the character?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character wearing a high-collared, waterproof black trench coat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a metal platform atop a skyscraper with pipes and ventilation systems?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a cyberpunk metropolis with dense rain, holographic billboards, and neon signs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly pan forward, gradually moving over the character's shoulders?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10041717439889908 + ], + "opposite_logprob_tokens": [ + -2.350417137145996 + ], + "logprob_reward": -0.10041717439889908, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10041717439889908, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1387653797864914, + 0.08516200631856918, + 0.16577966511249542, + 0.02374502830207348, + 0.11219912767410278, + 0.05333379656076431, + 0.0884542390704155, + 0.1450471132993698, + 0.09720882773399353, + 0.09030484408140182 + ], + "credit_map": [ + 0.008922277018427849, + 0.010585849173367023, + 0.014787430875003338, + 0.03658765181899071, + 0.012275267392396927, + 0.009282235987484455, + 0.028167523443698883, + 0.018157146871089935, + 0.00866310391575098, + 0.0065127150155603886, + 0.013526008464396, + 0.014988405629992485, + 0.007382992655038834, + 0.006134936586022377, + 0.013610884547233582, + 0.01434295903891325, + 0.012555000372231007, + 0.01717628352344036, + 0.014237670227885246, + 0.026065168902277946, + 0.03775821998715401, + 0.0134954284876585, + 0.034177184104919434, + 0.010314695537090302, + 0.0019167353166267276, + 0.001634469022974372, + 0.002271686214953661, + 0.0046785627491772175, + 0.002690717577934265, + 0.0016061156056821346, + 0.00578860379755497, + 0.0031581379007548094, + 0.009038389660418034, + 0.013432284817099571, + 0.017170002683997154, + 0.020810959860682487, + 0.01318101491779089, + 0.00808674842119217, + 0.019473183900117874, + 0.011006548069417477, + 0.0032467462588101625, + 0.004879494663327932, + 0.004053018055856228, + 0.010672244243323803, + 0.010549970902502537, + 0.0025950430426746607, + 0.010478897020220757, + 0.006858387030661106, + 0.002989209722727537, + 0.004757193382829428, + 0.015752259641885757, + 0.015229005366563797, + 0.007649310864508152, + 0.003963822964578867, + 0.02981480397284031, + 0.008298623375594616, + 0.014602716080844402, + 0.013518496416509151, + 0.01806287094950676, + 0.02768273837864399, + 0.017523208633065224, + 0.014623126946389675, + 0.029046835377812386, + 0.009987123310565948, + 0.009555322118103504, + 0.008648308925330639, + 0.015693577006459236, + 0.013287014327943325, + 0.014318753965198994, + 0.015478073619306087, + 0.011337182484567165, + 0.008890588767826557, + 0.007576954085379839, + 0.009485989809036255, + 0.013127942569553852, + 0.01161739882081747, + 0.022028716281056404, + 0.005098906811326742, + 0.013872028328478336, + 0.007496915757656097 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 7, + 0 + ], + "selected_reward": -0.10041717439889908 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a wide shot taken from behind and slightly upward of the character?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character wearing a high-collared, waterproof black trench coat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a metal platform atop a skyscraper with pipes and ventilation systems?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a cyberpunk metropolis with dense rain, holographic billboards, and neon signs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly pan forward, gradually moving over the character's shoulders?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2521490752696991 + ], + "opposite_logprob_tokens": [ + -1.5021491050720215 + ], + "logprob_reward": -0.2521490752696991, + "yes_no_margin_reward": 1.25, + "token_credit_reward": -0.2521490752696991, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19632619619369507, + 0.1285104751586914, + 0.2081085443496704, + 0.045588113367557526, + 0.1103658378124237, + 0.03887864947319031, + 0.045723140239715576, + 0.06376238912343979, + 0.10923327505588531, + 0.05350339412689209 + ], + "credit_map": [ + 0.017956925556063652, + 0.0160499457269907, + 0.02092377468943596, + 0.044465210288763046, + 0.016497323289513588, + 0.019750278443098068, + 0.039393339306116104, + 0.021289397031068802, + 0.013744652271270752, + 0.008405715227127075, + 0.014687160030007362, + 0.021330446004867554, + 0.01618429273366928, + 0.015175383538007736, + 0.023048026487231255, + 0.01593480072915554, + 0.015386736951768398, + 0.02273349091410637, + 0.020608294755220413, + 0.02746736817061901, + 0.053153619170188904, + 0.017764423042535782, + 0.03812875226140022, + 0.012865857221186161, + 0.00612694351002574, + 0.003946666605770588, + 0.004027253948152065, + 0.005719088017940521, + 0.007637038826942444, + 0.008346169255673885, + 0.005383285228163004, + 0.0044016726315021515, + 0.009171160869300365, + 0.014675348065793514, + 0.017536938190460205, + 0.020728880539536476, + 0.01199403963983059, + 0.008180806413292885, + 0.018599404022097588, + 0.00947925541549921, + 0.0032268355134874582, + 0.0040162294171750546, + 0.0039020925760269165, + 0.00892659742385149, + 0.004761966411024332, + 0.0031824777834117413, + 0.0071976156905293465, + 0.003664835589006543, + 0.003933699801564217, + 0.005018905736505985, + 0.00905595626682043, + 0.005498221609741449, + 0.007883957587182522, + 0.0031464749481528997, + 0.007623415905982256, + 0.0035625058226287365, + 0.008543426170945168, + 0.0072122071869671345, + 0.007838306948542595, + 0.01023796945810318, + 0.005970829166471958, + 0.0076668113470077515, + 0.011076772585511208, + 0.005216063465923071, + 0.018269145861268044, + 0.013870533555746078, + 0.0114287743344903, + 0.01081970613449812, + 0.012080112472176552, + 0.021794430911540985, + 0.012423070147633553, + 0.008547494187951088, + 0.005285369697958231, + 0.00505840452387929, + 0.00741244712844491, + 0.007393063977360725, + 0.009224587120115757, + 0.0046503981575369835, + 0.0086312685161829, + 0.0058478545397520065 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 1 + ], + "selected_reward": -0.2521490752696991 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a wide shot taken from behind and slightly upward of the character?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character wearing a high-collared, waterproof black trench coat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a metal platform atop a skyscraper with pipes and ventilation systems?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a cyberpunk metropolis with dense rain, holographic billboards, and neon signs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly pan forward, gradually moving over the character's shoulders?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007632853463292122 + ], + "opposite_logprob_tokens": [ + -4.882632732391357 + ], + "logprob_reward": -0.007632853463292122, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007632853463292122, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1284128874540329, + 0.07222151756286621, + 0.1704447716474533, + 0.028269317001104355, + 0.07674861699342728, + 0.024650197476148605, + 0.04993387311697006, + 0.13358192145824432, + 0.2593821585178375, + 0.05635472759604454 + ], + "credit_map": [ + 0.011810199357569218, + 0.010082889348268509, + 0.012725715525448322, + 0.02256520465016365, + 0.01340113952755928, + 0.015402245335280895, + 0.026581158861517906, + 0.015844333916902542, + 0.007454821839928627, + 0.004922882188111544, + 0.01005628239363432, + 0.011476624757051468, + 0.005481850355863571, + 0.008750151842832565, + 0.01198893878608942, + 0.012089959345757961, + 0.01053507998585701, + 0.018637191504240036, + 0.01822543703019619, + 0.019611936062574387, + 0.0493525005877018, + 0.013716539368033409, + 0.028800765052437782, + 0.011565318331122398, + 0.0028930711559951305, + 0.0024636010639369488, + 0.0038270107470452785, + 0.0041610789485275745, + 0.004732523579150438, + 0.0025648996233940125, + 0.004669253248721361, + 0.0029578793328255415, + 0.006560529116541147, + 0.010008417069911957, + 0.011330400593578815, + 0.014547936618328094, + 0.009252774529159069, + 0.006009016651660204, + 0.012639055959880352, + 0.0064004831947386265, + 0.0016540288925170898, + 0.002662315499037504, + 0.002805093303322792, + 0.004472218919545412, + 0.003826986998319626, + 0.0012927926145493984, + 0.004632406402379274, + 0.003304354613646865, + 0.002128301654011011, + 0.00356499501504004, + 0.009486137889325619, + 0.008019394241273403, + 0.0036102794110774994, + 0.002807401353493333, + 0.015791529789566994, + 0.004525833297520876, + 0.014399182051420212, + 0.010488481260836124, + 0.015920160338282585, + 0.023306725546717644, + 0.019341900944709778, + 0.010213266126811504, + 0.029360057786107063, + 0.010552153922617435, + 0.04308347776532173, + 0.03282255306839943, + 0.02688990719616413, + 0.024394521489739418, + 0.03125539422035217, + 0.04830942675471306, + 0.033716049045324326, + 0.018910840153694153, + 0.00554501311853528, + 0.006033985409885645, + 0.006743111647665501, + 0.009347779676318169, + 0.00842138472944498, + 0.005478712730109692, + 0.008697877638041973, + 0.006086861714720726 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 2, + 7 + ], + "selected_reward": -0.007632853463292122 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a wide shot taken from behind and slightly upward of the character?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character wearing a high-collared, waterproof black trench coat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a metal platform atop a skyscraper with pipes and ventilation systems?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a cyberpunk metropolis with dense rain, holographic billboards, and neon signs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly pan forward, gradually moving over the character's shoulders?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06204189732670784 + ], + "opposite_logprob_tokens": [ + -2.812041997909546 + ], + "logprob_reward": -0.06204189732670784, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06204189732670784, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0386759378015995, + 0.03871776908636093, + 0.2625456154346466, + 0.02826259844005108, + 0.04343109205365181, + 0.06509052217006683, + 0.07481837272644043, + 0.17233498394489288, + 0.2606857419013977, + 0.015437301248311996 + ], + "credit_map": [ + 0.0029646167531609535, + 0.004189790226519108, + 0.0038658245466649532, + 0.00928652286529541, + 0.005099417641758919, + 0.0029179940465837717, + 0.006327718961983919, + 0.004024054389446974, + 0.002861666027456522, + 0.003132839687168598, + 0.007062891963869333, + 0.00762313324958086, + 0.0037251722533255816, + 0.0029667008202522993, + 0.006532185710966587, + 0.00481317937374115, + 0.01711598038673401, + 0.029394254088401794, + 0.027056217193603516, + 0.031846653670072556, + 0.07665291428565979, + 0.0201185941696167, + 0.04346912354230881, + 0.01689191348850727, + 0.0021552727557718754, + 0.002146465005353093, + 0.005852143280208111, + 0.004709326662123203, + 0.002634707372635603, + 0.0021172957494854927, + 0.0064210970886051655, + 0.0022262930870056152, + 0.002889445284381509, + 0.002954225754365325, + 0.00769874406978488, + 0.007698907516896725, + 0.005287659354507923, + 0.0021380085963755846, + 0.00753434281796217, + 0.007229764945805073, + 0.005205575376749039, + 0.00810148473829031, + 0.007500490173697472, + 0.009588751941919327, + 0.011212829500436783, + 0.005727511364966631, + 0.011596424505114555, + 0.006157458294183016, + 0.0031416029669344425, + 0.005530375521630049, + 0.011988624930381775, + 0.008451633155345917, + 0.013448314741253853, + 0.003473194083198905, + 0.019700635224580765, + 0.009084001183509827, + 0.011896900832653046, + 0.014600654132664204, + 0.01993228681385517, + 0.03195676580071449, + 0.022635692730545998, + 0.012774351052939892, + 0.04405831918120384, + 0.01448002178221941, + 0.03065752238035202, + 0.02720414660871029, + 0.03274811804294586, + 0.02463439106941223, + 0.035937245935201645, + 0.047245025634765625, + 0.04314607009291649, + 0.01911322772502899, + 0.0011535787489265203, + 0.00149241101462394, + 0.0017840617801994085, + 0.0020028247963637114, + 0.004367631394416094, + 0.0009786303853616118, + 0.002138282172381878, + 0.0015198815381154418 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 8, + 7 + ], + "selected_reward": -0.06204189732670784 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "wide shot, shot from behind and slightly upward, slowly pans forward, gradually moving over their shoulders, revealing a breathtaking scene. Atop a skyscraper, a metal platform crisscrossed with pipes and ventilation systems, a slender, sharply defined figure stands quietly at the edge. They wear a high-collared, waterproof black trench coat, its hem fluttering gently in the wind and rain. Below, a boundless cyberpunk metropolis unfolds. Dense rain, illuminated by countless holographic billboards and giant neon signs, transforms into flickering curtains of light. Deep blue, voluptuous purple, and piercing magenta cast upward, outlining the figure's resolute profile and dampened shoulders. They remain motionless, leaning slightly forward, gazing intently at the steel jungle below, a woven fabric of light, shadow, and rain. Between the towering buildings, a constant stream of cars crisscrosses, leaving long trails of light. The wet metal platform and the character's waterproof windbreaker reflect the ever-changing neon halo, forming an extremely realistic and high-contrast light and shadow effect. The whole picture is full of futuristic coldness, loneliness and magnificent cinematic atmosphere.", + "seed": 5228, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000121/group_14/index_5107_seed_5228_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": -0.06622334644198417, + "vlm_logprob": -0.06622334644198417, + "vlm_yes_no_margin": 3.1, + "vlm_token_credit": -0.06622334644198417, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a wide shot taken from behind and slightly upward of the character?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character wearing a high-collared, waterproof black trench coat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a metal platform atop a skyscraper with pipes and ventilation systems?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a cyberpunk metropolis with dense rain, holographic billboards, and neon signs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly pan forward, gradually moving over the character's shoulders?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04389888420701027 + ], + "opposite_logprob_tokens": [ + -3.793898820877075 + ], + "logprob_reward": -0.04389888420701027, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.04389888420701027, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3243449330329895, + 0.056545425206422806, + 0.02724241465330124, + 0.1314443051815033, + 0.07763843983411789, + 0.04306327551603317, + 0.023517979308962822, + 0.27744540572166443, + 0.02178376354277134, + 0.016974087804555893 + ], + "credit_map": [ + 0.06482244282960892, + 0.019946768879890442, + 0.023002352565526962, + 0.047643158584833145, + 0.023322194814682007, + 0.015382979065179825, + 0.09616594761610031, + 0.0340590737760067, + 0.009745740331709385, + 0.004077727906405926, + 0.004916179459542036, + 0.008415643125772476, + 0.007714652456343174, + 0.0034898915328085423, + 0.012155831791460514, + 0.006029762327671051, + 0.004322477150708437, + 0.0021818068344146013, + 0.003632243722677231, + 0.0036357506178319454, + 0.004201117902994156, + 0.0015756512293592095, + 0.004591708071529865, + 0.003101660870015621, + 0.0235275998711586, + 0.006552502978593111, + 0.007454054895788431, + 0.018778376281261444, + 0.011818742379546165, + 0.018441442400217056, + 0.0326501727104187, + 0.01222142018377781, + 0.014067557640373707, + 0.0038488213904201984, + 0.006913683842867613, + 0.005470339208841324, + 0.0042076329700648785, + 0.005285503808408976, + 0.01399809680879116, + 0.023846805095672607, + 0.003983973991125822, + 0.01030769944190979, + 0.006353213917464018, + 0.004914441611617804, + 0.0028786365874111652, + 0.003960598260164261, + 0.006600557826459408, + 0.004064153414219618, + 0.002878333441913128, + 0.002677622716873884, + 0.002974549075588584, + 0.003497994039207697, + 0.002897468861192465, + 0.003062581177800894, + 0.0034425293561071157, + 0.0020868999417871237, + 0.018456704914569855, + 0.015325398184359074, + 0.009044063277542591, + 0.0904659554362297, + 0.010998939163982868, + 0.05391755327582359, + 0.05295014753937721, + 0.0262866523116827, + 0.0020473264157772064, + 0.005650461185723543, + 0.002451602602377534, + 0.0031063160859048367, + 0.0019219688838347793, + 0.0019892838317900896, + 0.0031215616036206484, + 0.0014952437486499548, + 0.0016206137370318174, + 0.002937331795692444, + 0.002422144403681159, + 0.002857874147593975, + 0.001365970354527235, + 0.0015039563877508044, + 0.0028282329440116882, + 0.0014379644999280572 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 3 + ], + "selected_reward": -0.04389888420701027 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a wide shot taken from behind and slightly upward of the character?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character wearing a high-collared, waterproof black trench coat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a metal platform atop a skyscraper with pipes and ventilation systems?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a cyberpunk metropolis with dense rain, holographic billboards, and neon signs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly pan forward, gradually moving over the character's shoulders?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04871983826160431 + ], + "opposite_logprob_tokens": [ + -3.048719882965088 + ], + "logprob_reward": -0.04871983826160431, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04871983826160431, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.137220561504364, + 0.02450183406472206, + 0.019115721806883812, + 0.01991172693669796, + 0.03186732158064842, + 0.24547909200191498, + 0.007988380268216133, + 0.4792594015598297, + 0.02086828462779522, + 0.013787698931992054 + ], + "credit_map": [ + 0.029226958751678467, + 0.010792000219225883, + 0.01050624344497919, + 0.018262311816215515, + 0.01002980675548315, + 0.006220897659659386, + 0.038531169295310974, + 0.013651165179908276, + 0.003948010038584471, + 0.0016585427802056074, + 0.00228486699052155, + 0.0030255012679845095, + 0.0038298931904137135, + 0.0025649224407970905, + 0.004044129978865385, + 0.0031459678430110216, + 0.0032568383030593395, + 0.0016312535153701901, + 0.001972553785890341, + 0.002732099499553442, + 0.0022714093793183565, + 0.001374684157781303, + 0.003192730015143752, + 0.0026841505896300077, + 0.0031393400859087706, + 0.0013284757733345032, + 0.001597113092429936, + 0.0022671371698379517, + 0.002383436542004347, + 0.003258582204580307, + 0.004008378833532333, + 0.001929262070916593, + 0.005018874071538448, + 0.0018154897261410952, + 0.0032888681162148714, + 0.0029351082630455494, + 0.0027905732858926058, + 0.0029451868031173944, + 0.005071844439953566, + 0.008001375012099743, + 0.018368607386946678, + 0.0944526270031929, + 0.009214683435857296, + 0.019495338201522827, + 0.012515503913164139, + 0.01750098541378975, + 0.027826163917779922, + 0.046105172485113144, + 0.0009281969396397471, + 0.000929331174120307, + 0.0010172434849664569, + 0.0010885456576943398, + 0.0011461841640993953, + 0.0010747237829491496, + 0.0010242587886750698, + 0.0007798962178640068, + 0.03618345409631729, + 0.023970579728484154, + 0.013599910773336887, + 0.17361332476139069, + 0.016548456624150276, + 0.08419118821620941, + 0.1012052372097969, + 0.029947221279144287, + 0.001797020435333252, + 0.0031586019322276115, + 0.0024732621386647224, + 0.0037499910686165094, + 0.0015819669933989644, + 0.003306978615000844, + 0.003173055825755, + 0.0016274049412459135, + 0.0012090892996639013, + 0.0020049288868904114, + 0.0023437209893018007, + 0.002233130857348442, + 0.0010117592755705118, + 0.0012911439407616854, + 0.0027213068678975105, + 0.0009726177086122334 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 5, + 0 + ], + "selected_reward": -0.04871983826160431 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a wide shot taken from behind and slightly upward of the character?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character wearing a high-collared, waterproof black trench coat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a metal platform atop a skyscraper with pipes and ventilation systems?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a cyberpunk metropolis with dense rain, holographic billboards, and neon signs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly pan forward, gradually moving over the character's shoulders?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10037297010421753 + ], + "opposite_logprob_tokens": [ + -2.3503730297088623 + ], + "logprob_reward": -0.10037297010421753, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10037297010421753, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08562019467353821, + 0.031580645591020584, + 0.03917428106069565, + 0.016153428703546524, + 0.06604285538196564, + 0.35895049571990967, + 0.008419902995228767, + 0.3780147433280945, + 0.004534771665930748, + 0.011508714407682419 + ], + "credit_map": [ + 0.01912771537899971, + 0.008023632690310478, + 0.00805412046611309, + 0.01064808014780283, + 0.006013205274939537, + 0.004099080804735422, + 0.02179344743490219, + 0.007860912941396236, + 0.005190132185816765, + 0.002248137490823865, + 0.0030101125594228506, + 0.004374551121145487, + 0.004252644255757332, + 0.002414148300886154, + 0.00600332859903574, + 0.004087592475116253, + 0.005517472978681326, + 0.003208741545677185, + 0.0030160322785377502, + 0.008189043030142784, + 0.004911977332085371, + 0.0029850404243916273, + 0.0069090151228010654, + 0.00443695904687047, + 0.002655711257830262, + 0.0008652880787849426, + 0.0014015637570992112, + 0.0017971884226426482, + 0.0013153734616935253, + 0.0030869143083691597, + 0.0034266721922904253, + 0.0016047174576669931, + 0.012275857850909233, + 0.003963634371757507, + 0.006674964912235737, + 0.0037009113002568483, + 0.0034614740870893, + 0.005306864157319069, + 0.010476111434400082, + 0.020183037966489792, + 0.02716556005179882, + 0.13956902921199799, + 0.014665222726762295, + 0.029021013528108597, + 0.018617630004882812, + 0.025308290496468544, + 0.04343263432383537, + 0.06117112189531326, + 0.0012456076219677925, + 0.001541760633699596, + 0.0010021980851888657, + 0.0011813656892627478, + 0.0007458829204551876, + 0.0011105965822935104, + 0.0008849156438373029, + 0.0007075759931467474, + 0.028745006769895554, + 0.01892932876944542, + 0.010720350779592991, + 0.13692979514598846, + 0.013180045410990715, + 0.06639372557401657, + 0.07979893684387207, + 0.023317527025938034, + 0.00043105974327772856, + 0.0006977675948292017, + 0.0006651735166087747, + 0.0006912696408107877, + 0.00039022424607537687, + 0.0005649427766911685, + 0.0007949565770104527, + 0.0002993775124195963, + 0.0011301751947030425, + 0.0015046607004478574, + 0.0017086281441152096, + 0.002203752286732197, + 0.0008259124006144702, + 0.0010366732021793723, + 0.0021417748648673296, + 0.0009571376140229404 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 5, + 0 + ], + "selected_reward": -0.10037297010421753 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a wide shot taken from behind and slightly upward of the character?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character wearing a high-collared, waterproof black trench coat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a metal platform atop a skyscraper with pipes and ventilation systems?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a cyberpunk metropolis with dense rain, holographic billboards, and neon signs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly pan forward, gradually moving over the character's shoulders?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011086530983448029 + ], + "opposite_logprob_tokens": [ + -4.511086463928223 + ], + "logprob_reward": -0.011086530983448029, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011086530983448029, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09506542980670929, + 0.033838775008916855, + 0.024113891646265984, + 0.01893886923789978, + 0.045322105288505554, + 0.39833277463912964, + 0.005190471187233925, + 0.3522244095802307, + 0.014587664976716042, + 0.012385539710521698 + ], + "credit_map": [ + 0.024679940193891525, + 0.013522705994546413, + 0.010442777536809444, + 0.00904519110918045, + 0.007779279723763466, + 0.005145345348864794, + 0.0157333817332983, + 0.008716823533177376, + 0.006026996299624443, + 0.0029849475249648094, + 0.0032990670297294855, + 0.005871068220585585, + 0.004520224407315254, + 0.0027998003643006086, + 0.0042648110538721085, + 0.004071865230798721, + 0.0044859787449240685, + 0.00201915530487895, + 0.002834980608895421, + 0.0032091259490698576, + 0.0025692221242934465, + 0.0017961428966373205, + 0.003439301857724786, + 0.003759985789656639, + 0.002931719645857811, + 0.0009025165345519781, + 0.0016488193068653345, + 0.0030618368182331324, + 0.0025565274991095066, + 0.0021772817708551884, + 0.00404399586841464, + 0.001616174355149269, + 0.009574253112077713, + 0.002532260725274682, + 0.004066500812768936, + 0.003443160094320774, + 0.002392763737589121, + 0.0035536966752260923, + 0.0064615365117788315, + 0.013297936879098415, + 0.030715562403202057, + 0.15331880748271942, + 0.012598180212080479, + 0.03087223507463932, + 0.020188236609101295, + 0.029462847858667374, + 0.043634261935949326, + 0.07754269242286682, + 0.0006004328606650233, + 0.0005928071332164109, + 0.0007590129971504211, + 0.0007474500453099608, + 0.00039795442717149854, + 0.000993420835584402, + 0.0005876207142136991, + 0.0005117725813761353, + 0.027042968198657036, + 0.01784050464630127, + 0.010013998486101627, + 0.12971735000610352, + 0.011977863498032093, + 0.060202956199645996, + 0.07558432966470718, + 0.019844483584165573, + 0.001457777339965105, + 0.002107899636030197, + 0.0012578392634168267, + 0.0020079948008060455, + 0.001089267316274345, + 0.0034130809362977743, + 0.0018636207096278667, + 0.0013901867205277085, + 0.0011374612804502249, + 0.0014498865930363536, + 0.0021773246116936207, + 0.002118073869496584, + 0.0007839103345759213, + 0.001809063833206892, + 0.0019136910559609532, + 0.0009961285395547748 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 7, + 0 + ], + "selected_reward": -0.011086530983448029 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a wide shot taken from behind and slightly upward of the character?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character wearing a high-collared, waterproof black trench coat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a metal platform atop a skyscraper with pipes and ventilation systems?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a cyberpunk metropolis with dense rain, holographic billboards, and neon signs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly pan forward, gradually moving over the character's shoulders?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.12703850865364075 + ], + "opposite_logprob_tokens": [ + -2.1270384788513184 + ], + "logprob_reward": -0.12703850865364075, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.12703850865364075, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.036591894924640656, + 0.02045491151511669, + 0.03457826375961304, + 0.04456577077507973, + 0.07881154119968414, + 0.5213829278945923, + 0.021518468856811523, + 0.18940205872058868, + 0.03991072624921799, + 0.012783491052687168 + ], + "credit_map": [ + 0.003600961994379759, + 0.004880284424871206, + 0.004239019937813282, + 0.006980271078646183, + 0.004820866510272026, + 0.0031545429956167936, + 0.005084967706352472, + 0.0038309793453663588, + 0.002157128183171153, + 0.0015838148538023233, + 0.0016555807087570429, + 0.004184189718216658, + 0.003132975660264492, + 0.001562182093039155, + 0.0025204461999237537, + 0.003658593399450183, + 0.0035015856847167015, + 0.003503284649923444, + 0.004003255162388086, + 0.00696307048201561, + 0.004961535334587097, + 0.0025798967108130455, + 0.004953653551638126, + 0.004111981950700283, + 0.010279687121510506, + 0.0027150644455105066, + 0.0034514344297349453, + 0.006162792444229126, + 0.004667043220251799, + 0.002633528783917427, + 0.010754511691629887, + 0.0039017065428197384, + 0.012736178934574127, + 0.006105832755565643, + 0.007203694898635149, + 0.006667216308414936, + 0.009829598478972912, + 0.0037600311916321516, + 0.013818430714309216, + 0.018690554425120354, + 0.039138469845056534, + 0.19878825545310974, + 0.020929114893078804, + 0.04131181165575981, + 0.026480508968234062, + 0.037316203117370605, + 0.057527147233486176, + 0.09989137947559357, + 0.0018348388839513063, + 0.002356181852519512, + 0.0018460664432495832, + 0.0027785610873252153, + 0.0023691928945481777, + 0.004711303859949112, + 0.003477666061371565, + 0.0021446566097438335, + 0.012425452470779419, + 0.00980992428958416, + 0.006194765213876963, + 0.05951964482665062, + 0.007628695573657751, + 0.03820132836699486, + 0.03494446352124214, + 0.020677782595157623, + 0.0033903825096786022, + 0.006859866436570883, + 0.005058376118540764, + 0.006402166560292244, + 0.00322475703433156, + 0.00618695467710495, + 0.006238214205950499, + 0.00255000707693398, + 0.0014445612905547023, + 0.002171306638047099, + 0.002066256944090128, + 0.0024316213093698025, + 0.0008730570552870631, + 0.0009939523879438639, + 0.0019682953134179115, + 0.0008344397647306323 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 7, + 4 + ], + "selected_reward": -0.12703850865364075 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "wide shot, shot from behind and slightly upward, slowly pans forward, gradually moving over their shoulders, revealing a breathtaking scene. Atop a skyscraper, a metal platform crisscrossed with pipes and ventilation systems, a slender, sharply defined figure stands quietly at the edge. They wear a high-collared, waterproof black trench coat, its hem fluttering gently in the wind and rain. Below, a boundless cyberpunk metropolis unfolds. Dense rain, illuminated by countless holographic billboards and giant neon signs, transforms into flickering curtains of light. Deep blue, voluptuous purple, and piercing magenta cast upward, outlining the figure's resolute profile and dampened shoulders. They remain motionless, leaning slightly forward, gazing intently at the steel jungle below, a woven fabric of light, shadow, and rain. Between the towering buildings, a constant stream of cars crisscrosses, leaving long trails of light. The wet metal platform and the character's waterproof windbreaker reflect the ever-changing neon halo, forming an extremely realistic and high-contrast light and shadow effect. The whole picture is full of futuristic coldness, loneliness and magnificent cinematic atmosphere.", + "seed": 5229, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000121/group_14/index_5107_seed_5229_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": -0.06011172123253346, + "vlm_logprob": -0.06011172123253346, + "vlm_yes_no_margin": 3.25, + "vlm_token_credit": -0.06011172123253346, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a wide shot taken from behind and slightly upward of the character?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character wearing a high-collared, waterproof black trench coat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a metal platform atop a skyscraper with pipes and ventilation systems?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a cyberpunk metropolis with dense rain, holographic billboards, and neon signs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly pan forward, gradually moving over the character's shoulders?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06449700146913528 + ], + "opposite_logprob_tokens": [ + -3.3144969940185547 + ], + "logprob_reward": -0.06449700146913528, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.06449700146913528, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5047427415847778, + 0.02631985954940319, + 0.01894594542682171, + 0.023991702124476433, + 0.04525889828801155, + 0.06712943315505981, + 0.052804119884967804, + 0.09718070924282074, + 0.033450718969106674, + 0.13017594814300537 + ], + "credit_map": [ + 0.14851102232933044, + 0.01935044303536415, + 0.02112504653632641, + 0.03963446244597435, + 0.038547445088624954, + 0.02201220579445362, + 0.1895974725484848, + 0.025964612141251564, + 0.0030755766201764345, + 0.0016487255925312638, + 0.002497304929420352, + 0.004351694602519274, + 0.004199322313070297, + 0.0030591425020247698, + 0.004549211356788874, + 0.002938880119472742, + 0.0015430374769493937, + 0.001454931334592402, + 0.0021780922543257475, + 0.003497489495202899, + 0.0030483058653771877, + 0.0017775435699149966, + 0.003600328229367733, + 0.0018462154548615217, + 0.0014903402188792825, + 0.0022046782542020082, + 0.0020281921606510878, + 0.002773848595097661, + 0.006108845118433237, + 0.002827421296387911, + 0.0036071124486625195, + 0.0029512622859328985, + 0.004762038588523865, + 0.0054017324000597, + 0.0038819541223347187, + 0.004530341364443302, + 0.006462433375418186, + 0.004804934840649366, + 0.006657135207206011, + 0.008758322335779667, + 0.019733691588044167, + 0.003158221021294594, + 0.007182975299656391, + 0.0031319214031100273, + 0.007120511494576931, + 0.016719093546271324, + 0.005533750168979168, + 0.004549264907836914, + 0.0037276691291481256, + 0.005584322847425938, + 0.0025739013217389584, + 0.005802369676530361, + 0.01662043295800686, + 0.005887807812541723, + 0.008115963079035282, + 0.004491650033742189, + 0.020622769370675087, + 0.009457286447286606, + 0.010034143924713135, + 0.007719438523054123, + 0.013973050750792027, + 0.014662042260169983, + 0.009915792383253574, + 0.01079617440700531, + 0.002757545094937086, + 0.0055958605371415615, + 0.0018581869080662727, + 0.005813080817461014, + 0.004956832155585289, + 0.0037725139409303665, + 0.006651232019066811, + 0.0020454656332731247, + 0.040916286408901215, + 0.017743784934282303, + 0.012063005939126015, + 0.013363407924771309, + 0.008451947011053562, + 0.018008606508374214, + 0.010455516166985035, + 0.009173388592898846 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 7 + ], + "selected_reward": -0.06449700146913528 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a wide shot taken from behind and slightly upward of the character?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character wearing a high-collared, waterproof black trench coat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a metal platform atop a skyscraper with pipes and ventilation systems?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a cyberpunk metropolis with dense rain, holographic billboards, and neon signs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly pan forward, gradually moving over the character's shoulders?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"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": [ + 0.37750786542892456, + 0.038596268743276596, + 0.022915396839380264, + 0.018513059243559837, + 0.0704193264245987, + 0.09631867706775665, + 0.0436055026948452, + 0.18781214952468872, + 0.0698975920677185, + 0.07441408932209015 + ], + "credit_map": [ + 0.10589290410280228, + 0.016441363841295242, + 0.01812276616692543, + 0.03413934260606766, + 0.02912943996489048, + 0.0175388902425766, + 0.13574178516864777, + 0.020501388236880302, + 0.003523350926116109, + 0.0026865238323807716, + 0.0036068714689463377, + 0.005772472359240055, + 0.006682473234832287, + 0.005033985711634159, + 0.008223900571465492, + 0.003066691569983959, + 0.0021923298481851816, + 0.00431999983265996, + 0.0020943491254001856, + 0.0025514126755297184, + 0.0035524673294276, + 0.002499818569049239, + 0.0029284677002578974, + 0.0027765529230237007, + 0.0013895798474550247, + 0.0017969135660678148, + 0.0014224555343389511, + 0.002139283576980233, + 0.00432334840297699, + 0.0025784054305404425, + 0.0027218153700232506, + 0.0021412570495158434, + 0.005978633649647236, + 0.004357853904366493, + 0.00509123969823122, + 0.009305708110332489, + 0.015114092268049717, + 0.007753840181976557, + 0.014970146119594574, + 0.007847817614674568, + 0.02384886145591736, + 0.004906289745122194, + 0.004957455210387707, + 0.005347115453332663, + 0.014620820060372353, + 0.026100637391209602, + 0.010176531970500946, + 0.006360969040542841, + 0.005319775547832251, + 0.0039826976135373116, + 0.0020224961917847395, + 0.0031740055419504642, + 0.01194018591195345, + 0.008114958181977272, + 0.0057911379262804985, + 0.0032602455466985703, + 0.0411430187523365, + 0.01975370943546295, + 0.019183846190571785, + 0.015045041218400002, + 0.02306680753827095, + 0.028064223006367683, + 0.018249185755848885, + 0.023306315764784813, + 0.010540509596467018, + 0.01203411165624857, + 0.005545240826904774, + 0.007213805336505175, + 0.010343801230192184, + 0.00887686014175415, + 0.00944268424063921, + 0.0059005748480558395, + 0.024259107187390327, + 0.011742786504328251, + 0.004004734102636576, + 0.006700442172586918, + 0.006994958035647869, + 0.011631608009338379, + 0.004580262582749128, + 0.004500194918364286 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 5 + ], + "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 shot a wide shot taken from behind and slightly upward of the character?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character wearing a high-collared, waterproof black trench coat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a metal platform atop a skyscraper with pipes and ventilation systems?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a cyberpunk metropolis with dense rain, holographic billboards, and neon signs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly pan forward, gradually moving over the character's shoulders?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11300485581159592 + ], + "opposite_logprob_tokens": [ + -2.2380049228668213 + ], + "logprob_reward": -0.11300485581159592, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11300485581159592, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.49872350692749023, + 0.04979795590043068, + 0.03908532112836838, + 0.025582870468497276, + 0.02830611728131771, + 0.1741591840982437, + 0.030400006100535393, + 0.05713166296482086, + 0.02421940676867962, + 0.07259398698806763 + ], + "credit_map": [ + 0.14606446027755737, + 0.02004983276128769, + 0.020840642973780632, + 0.03760765492916107, + 0.0374768003821373, + 0.025377634912729263, + 0.1843332201242447, + 0.026973238214850426, + 0.008130960166454315, + 0.005025816150009632, + 0.004467832390218973, + 0.004840389359742403, + 0.006248298101127148, + 0.004896717611700296, + 0.010252469219267368, + 0.0059354668483138084, + 0.004894569516181946, + 0.010737652890384197, + 0.0026802332140505314, + 0.00302347750402987, + 0.005139793269336224, + 0.004624138120561838, + 0.00417938781902194, + 0.003806063672527671, + 0.0026907827705144882, + 0.004920432809740305, + 0.0024718239437788725, + 0.0027371933683753014, + 0.0033521007280796766, + 0.0033916737884283066, + 0.0029581990092992783, + 0.003060661954805255, + 0.0025385774206370115, + 0.0022163172252476215, + 0.002943475963547826, + 0.003164714900776744, + 0.0043682218529284, + 0.004860392305999994, + 0.004973881412297487, + 0.0032405355013906956, + 0.05998189002275467, + 0.007944058626890182, + 0.009667479433119297, + 0.004764400888234377, + 0.011102618649601936, + 0.058839038014411926, + 0.01136381272226572, + 0.010495873168110847, + 0.006384715437889099, + 0.001991183264181018, + 0.0028675587382167578, + 0.002090946538373828, + 0.0049851080402731895, + 0.007060523610562086, + 0.0024884557351469994, + 0.002531510079279542, + 0.010823328979313374, + 0.00527181476354599, + 0.005335414316505194, + 0.004702336620539427, + 0.009507142938673496, + 0.008179687894880772, + 0.005469457246363163, + 0.007842475548386574, + 0.003238270292058587, + 0.003306793048977852, + 0.002213001251220703, + 0.0028385124169290066, + 0.0023846947588026524, + 0.002546017523854971, + 0.0037427316419780254, + 0.003949383273720741, + 0.018321679905056953, + 0.011342865414917469, + 0.011159262619912624, + 0.006819755304604769, + 0.007860234007239342, + 0.008241642266511917, + 0.0046532656997442245, + 0.0041952794417738914 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 9 + ], + "selected_reward": -0.11300485581159592 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a wide shot taken from behind and slightly upward of the character?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character wearing a high-collared, waterproof black trench coat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a metal platform atop a skyscraper with pipes and ventilation systems?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a cyberpunk metropolis with dense rain, holographic billboards, and neon signs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly pan forward, gradually moving over the character's shoulders?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007637467235326767 + ], + "opposite_logprob_tokens": [ + -4.8826375007629395 + ], + "logprob_reward": -0.007637467235326767, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007637467235326767, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.6027950644493103, + 0.04672546312212944, + 0.021203283220529556, + 0.011868800967931747, + 0.033394865691661835, + 0.08192386478185654, + 0.03978218883275986, + 0.03518686443567276, + 0.01449607778340578, + 0.11262350529432297 + ], + "credit_map": [ + 0.17827284336090088, + 0.025467079132795334, + 0.02406439185142517, + 0.04863516986370087, + 0.04379653185606003, + 0.029422594234347343, + 0.22213949263095856, + 0.030997006222605705, + 0.007548341993242502, + 0.004550397861748934, + 0.005639552604407072, + 0.0047567556612193584, + 0.004981295671314001, + 0.006780547089874744, + 0.007603139616549015, + 0.004865438677370548, + 0.0023163296282291412, + 0.0044939336366951466, + 0.0018958512227982283, + 0.0020657023414969444, + 0.003155739977955818, + 0.002495719585567713, + 0.002667549531906843, + 0.0021124607883393764, + 0.0008057548548094928, + 0.0015514206606894732, + 0.001305695390328765, + 0.0012875328538939357, + 0.0022452285047620535, + 0.0014352956786751747, + 0.00171732809394598, + 0.0015205470845103264, + 0.003009376348927617, + 0.002999769989401102, + 0.002698762807995081, + 0.0038299455773085356, + 0.005217467434704304, + 0.0037873920518904924, + 0.005793346092104912, + 0.006058809347450733, + 0.02964596264064312, + 0.003311817767098546, + 0.004879985004663467, + 0.0022691369522362947, + 0.004447857849299908, + 0.027376649901270866, + 0.00508157117292285, + 0.004910893738269806, + 0.010675113648176193, + 0.002466977573931217, + 0.004327267874032259, + 0.0018935168627649546, + 0.003900402458384633, + 0.010817809030413628, + 0.0028312387876212597, + 0.002869868651032448, + 0.005377001129090786, + 0.0035516172647476196, + 0.0024746302515268326, + 0.002788857789710164, + 0.007606033701449633, + 0.0038987533189356327, + 0.004656848963350058, + 0.004833125974982977, + 0.00198910990729928, + 0.0017365994863212109, + 0.0018183090724050999, + 0.0016395654529333115, + 0.001770894741639495, + 0.0016384754562750459, + 0.0020253777038306, + 0.0018777477089315653, + 0.03858394920825958, + 0.018211815506219864, + 0.005237875506281853, + 0.011388289742171764, + 0.009590040892362595, + 0.016930900514125824, + 0.006386803463101387, + 0.006293835584074259 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 5 + ], + "selected_reward": -0.007637467235326767 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a wide shot taken from behind and slightly upward of the character?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character wearing a high-collared, waterproof black trench coat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a metal platform atop a skyscraper with pipes and ventilation systems?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a cyberpunk metropolis with dense rain, holographic billboards, and neon signs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly pan forward, gradually moving over the character's shoulders?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08903302997350693 + ], + "opposite_logprob_tokens": [ + -2.4640331268310547 + ], + "logprob_reward": -0.08903302997350693, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08903302997350693, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06378661841154099, + 0.024219399318099022, + 0.0187855064868927, + 0.029903756454586983, + 0.028895586729049683, + 0.1510564684867859, + 0.03599492087960243, + 0.11422690749168396, + 0.11032601445913315, + 0.4228048324584961 + ], + "credit_map": [ + 0.006334242876619101, + 0.0073158955201506615, + 0.006752494256943464, + 0.010671699419617653, + 0.009208183735609055, + 0.006185450591146946, + 0.011828172020614147, + 0.005490479990839958, + 0.001909438637085259, + 0.0022391926031559706, + 0.002022381639108062, + 0.002985941478982568, + 0.005417544394731522, + 0.0034621059894561768, + 0.004499917384237051, + 0.0016828732332214713, + 0.0017042800318449736, + 0.0014893495244905353, + 0.0021820261608809233, + 0.00234589702449739, + 0.003360037226229906, + 0.0022504362277686596, + 0.0037392384838312864, + 0.0017142397118732333, + 0.0020462442189455032, + 0.0029973229393363, + 0.0027762497775256634, + 0.0031087982933968306, + 0.007085589691996574, + 0.003661682829260826, + 0.004933760967105627, + 0.003294103778898716, + 0.0023666350170969963, + 0.0032374418806284666, + 0.0026388466358184814, + 0.003206458408385515, + 0.005174084100872278, + 0.002587464638054371, + 0.005843692924827337, + 0.0038409617263823748, + 0.050061456859111786, + 0.005830972455441952, + 0.013414507731795311, + 0.005024772137403488, + 0.012046690098941326, + 0.04657623544335365, + 0.009981859475374222, + 0.008119955658912659, + 0.005380160640925169, + 0.004078546538949013, + 0.0037941981572657824, + 0.002681449754163623, + 0.0077522206120193005, + 0.0053869285620749, + 0.003791470779106021, + 0.0031299430411309004, + 0.023260299116373062, + 0.013328669592738152, + 0.011553035117685795, + 0.008410300128161907, + 0.013111019507050514, + 0.017134366557002068, + 0.013034597039222717, + 0.01439460925757885, + 0.009279855526983738, + 0.02021843194961548, + 0.00763547420501709, + 0.013084054924547672, + 0.017584390938282013, + 0.01461105328053236, + 0.019462810829281807, + 0.008449927903711796, + 0.13359582424163818, + 0.046649616211652756, + 0.04444922134280205, + 0.04782353341579437, + 0.03707576543092728, + 0.061647072434425354, + 0.027706244960427284, + 0.023857541382312775 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 5, + 7 + ], + "selected_reward": -0.08903302997350693 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "wide shot, shot from behind and slightly upward, slowly pans forward, gradually moving over their shoulders, revealing a breathtaking scene. Atop a skyscraper, a metal platform crisscrossed with pipes and ventilation systems, a slender, sharply defined figure stands quietly at the edge. They wear a high-collared, waterproof black trench coat, its hem fluttering gently in the wind and rain. Below, a boundless cyberpunk metropolis unfolds. Dense rain, illuminated by countless holographic billboards and giant neon signs, transforms into flickering curtains of light. Deep blue, voluptuous purple, and piercing magenta cast upward, outlining the figure's resolute profile and dampened shoulders. They remain motionless, leaning slightly forward, gazing intently at the steel jungle below, a woven fabric of light, shadow, and rain. Between the towering buildings, a constant stream of cars crisscrosses, leaving long trails of light. The wet metal platform and the character's waterproof windbreaker reflect the ever-changing neon halo, forming an extremely realistic and high-contrast light and shadow effect. The whole picture is full of futuristic coldness, loneliness and magnificent cinematic atmosphere.", + "seed": 5230, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000121/group_14/index_5107_seed_5230_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": -0.061001983284950254, + "vlm_logprob": -0.061001983284950254, + "vlm_yes_no_margin": 3.35, + "vlm_token_credit": -0.061001983284950254, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a wide shot taken from behind and slightly upward of the character?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character wearing a high-collared, waterproof black trench coat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a metal platform atop a skyscraper with pipes and ventilation systems?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a cyberpunk metropolis with dense rain, holographic billboards, and neon signs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly pan forward, gradually moving over the character's shoulders?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04696591570973396 + ], + "opposite_logprob_tokens": [ + -3.9219658374786377 + ], + "logprob_reward": -0.04696591570973396, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.04696591570973396, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.161565363407135, + 0.07356011122465134, + 0.10416725277900696, + 0.13499601185321808, + 0.10518106073141098, + 0.10344331711530685, + 0.10639922320842743, + 0.04808041825890541, + 0.08715220540761948, + 0.07545506209135056 + ], + "credit_map": [ + 0.01475547719746828, + 0.01649527996778488, + 0.011602578684687614, + 0.024131527170538902, + 0.026032354682683945, + 0.01969684287905693, + 0.025705663487315178, + 0.0231456458568573, + 0.006691006477922201, + 0.01025998592376709, + 0.004382527433335781, + 0.006920859217643738, + 0.013603278435766697, + 0.010670822113752365, + 0.01030492503196001, + 0.010726706124842167, + 0.01883927918970585, + 0.011159291490912437, + 0.007654809858649969, + 0.010052379220724106, + 0.011144605465233326, + 0.014431496150791645, + 0.014739769510924816, + 0.016145624220371246, + 0.007234813179820776, + 0.006145319435745478, + 0.006968270055949688, + 0.007651586085557938, + 0.01316535659134388, + 0.007860034704208374, + 0.03984099254012108, + 0.046129632741212845, + 0.01197096984833479, + 0.009657522663474083, + 0.009313651360571384, + 0.007457366678863764, + 0.011955618858337402, + 0.0261030662804842, + 0.01611272618174553, + 0.012610135599970818, + 0.00888573843985796, + 0.008976614102721214, + 0.012315531261265278, + 0.010616251267492771, + 0.01740298978984356, + 0.018487343564629555, + 0.015582476742565632, + 0.011176375672221184, + 0.019825704395771027, + 0.02218802645802498, + 0.010373751632869244, + 0.009609725326299667, + 0.010700837709009647, + 0.011239472776651382, + 0.011641017161309719, + 0.010820687748491764, + 0.0030756790656596422, + 0.006113736890256405, + 0.00272552203387022, + 0.004781287629157305, + 0.009255927056074142, + 0.008990200236439705, + 0.0072487336583435535, + 0.0058893305249512196, + 0.0069557419046759605, + 0.007631769869476557, + 0.008053293451666832, + 0.013679172843694687, + 0.008688081987202168, + 0.02195795066654682, + 0.009943794459104538, + 0.010242396965622902, + 0.006298907101154327, + 0.005402885377407074, + 0.01280371192842722, + 0.01717466674745083, + 0.005372999235987663, + 0.0053707207553088665, + 0.006849801633507013, + 0.016181370243430138 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 6 + ], + "selected_reward": -0.04696591570973396 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a wide shot taken from behind and slightly upward of the character?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character wearing a high-collared, waterproof black trench coat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a metal platform atop a skyscraper with pipes and ventilation systems?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a cyberpunk metropolis with dense rain, holographic billboards, and neon signs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly pan forward, gradually moving over the character's shoulders?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05502605810761452 + ], + "opposite_logprob_tokens": [ + -2.930026054382324 + ], + "logprob_reward": -0.05502605810761452, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05502605810761452, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1919650137424469, + 0.06188590079545975, + 0.09666598588228226, + 0.037386681884527206, + 0.03550608456134796, + 0.09051652997732162, + 0.05361255258321762, + 0.2088155299425125, + 0.12737907469272614, + 0.09626670181751251 + ], + "credit_map": [ + 0.017047209665179253, + 0.014034625142812729, + 0.013747408986091614, + 0.022023359313607216, + 0.03409401327371597, + 0.021592920646071434, + 0.037739742547273636, + 0.0316857285797596, + 0.004242924973368645, + 0.006033406592905521, + 0.004026301670819521, + 0.008568010292947292, + 0.008565148338675499, + 0.01095538679510355, + 0.007858439348638058, + 0.011636275798082352, + 0.008643523789942265, + 0.007497723214328289, + 0.014493812806904316, + 0.0169157013297081, + 0.009847144596278667, + 0.014043458737432957, + 0.013239171355962753, + 0.011985443532466888, + 0.002839496824890375, + 0.0026008225977420807, + 0.0020779818296432495, + 0.0030250560957938433, + 0.007651599124073982, + 0.0030730345752090216, + 0.008008740842342377, + 0.008109945803880692, + 0.004327073693275452, + 0.0040453458204865456, + 0.003732638666406274, + 0.0023147487081587315, + 0.003991920035332441, + 0.007140956353396177, + 0.004998428747057915, + 0.004954969044774771, + 0.010155600495636463, + 0.008884605020284653, + 0.007385861128568649, + 0.007463900838047266, + 0.01699214056134224, + 0.008996190503239632, + 0.0125136012211442, + 0.018124619498848915, + 0.004531925544142723, + 0.007151241414248943, + 0.0056266156025230885, + 0.007300149649381638, + 0.0070328605361282825, + 0.006230337545275688, + 0.006607872899621725, + 0.009131546132266521, + 0.012788379564881325, + 0.02910887636244297, + 0.012759004719555378, + 0.018528740853071213, + 0.04815632104873657, + 0.04165628179907799, + 0.023632679134607315, + 0.022185219451785088, + 0.008927233517169952, + 0.013984255492687225, + 0.008614130318164825, + 0.02332662232220173, + 0.01003817655146122, + 0.0317767858505249, + 0.018034981563687325, + 0.012676887214183807, + 0.006800438743084669, + 0.009812754578888416, + 0.012754484079778194, + 0.02063089795410633, + 0.00795335229486227, + 0.00807339046150446, + 0.008287389762699604, + 0.02195398136973381 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 8 + ], + "selected_reward": -0.05502605810761452 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a wide shot taken from behind and slightly upward of the character?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character wearing a high-collared, waterproof black trench coat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a metal platform atop a skyscraper with pipes and ventilation systems?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a cyberpunk metropolis with dense rain, holographic billboards, and neon signs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly pan forward, gradually moving over the character's shoulders?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05496693402528763 + ], + "opposite_logprob_tokens": [ + -2.929966926574707 + ], + "logprob_reward": -0.05496693402528763, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05496693402528763, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.28997716307640076, + 0.0544147714972496, + 0.12165544182062149, + 0.07703221589326859, + 0.08646060526371002, + 0.07016342878341675, + 0.19636450707912445, + 0.03506797179579735, + 0.03248276188969612, + 0.03638111427426338 + ], + "credit_map": [ + 0.029631689190864563, + 0.03507061302661896, + 0.017447454854846, + 0.03212691470980644, + 0.05735793337225914, + 0.03354863449931145, + 0.04114944487810135, + 0.04364448040723801, + 0.004390465561300516, + 0.005962301976978779, + 0.003179048653692007, + 0.0063640400767326355, + 0.006071728188544512, + 0.008904955349862576, + 0.007539068348705769, + 0.012003164738416672, + 0.011720177717506886, + 0.00912853516638279, + 0.024562444537878036, + 0.017147643491625786, + 0.014730614610016346, + 0.012687403708696365, + 0.01709470897912979, + 0.014583917334675789, + 0.007674094755202532, + 0.0068519082851707935, + 0.00649351766332984, + 0.0065962583757936954, + 0.011600080877542496, + 0.005914241075515747, + 0.014346090145409107, + 0.017556024715304375, + 0.008018133230507374, + 0.011643897742033005, + 0.006731055211275816, + 0.007832418195903301, + 0.008603726513683796, + 0.014172147028148174, + 0.011850051581859589, + 0.017609182745218277, + 0.00754247372969985, + 0.01114670466631651, + 0.007648614700883627, + 0.007662383373826742, + 0.010579809546470642, + 0.009696788154542446, + 0.010000736452639103, + 0.0058859181590378284, + 0.028130266815423965, + 0.038045987486839294, + 0.020925180986523628, + 0.018900888040661812, + 0.026179617270827293, + 0.022888945415616035, + 0.019299747422337532, + 0.021993890404701233, + 0.002456553280353546, + 0.004336195066571236, + 0.002725140657275915, + 0.003338585142046213, + 0.005489119328558445, + 0.005342832300812006, + 0.004338906612247229, + 0.007040643598884344, + 0.003209516406059265, + 0.002976084593683481, + 0.003480469109490514, + 0.004754623398184776, + 0.002960840007290244, + 0.007902475073933601, + 0.0032444510143250227, + 0.003954301588237286, + 0.002724324818700552, + 0.002537085907533765, + 0.008234511129558086, + 0.0060430788435041904, + 0.006352324970066547, + 0.0032218496780842543, + 0.0022575906477868557, + 0.005010346882045269 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 2 + ], + "selected_reward": -0.05496693402528763 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a wide shot taken from behind and slightly upward of the character?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character wearing a high-collared, waterproof black trench coat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a metal platform atop a skyscraper with pipes and ventilation systems?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a cyberpunk metropolis with dense rain, holographic billboards, and neon signs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly pan forward, gradually moving over the character's shoulders?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005259960889816284 + ], + "opposite_logprob_tokens": [ + -5.255259990692139 + ], + "logprob_reward": -0.005259960889816284, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005259960889816284, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1437663584947586, + 0.05365591496229172, + 0.2727576792240143, + 0.17420227825641632, + 0.041505053639411926, + 0.03548264130949974, + 0.03850656747817993, + 0.1569247543811798, + 0.032903388142585754, + 0.05029541254043579 + ], + "credit_map": [ + 0.013189254328608513, + 0.016091229394078255, + 0.008438574150204659, + 0.011915692128241062, + 0.03457840532064438, + 0.01608004979789257, + 0.022451531141996384, + 0.021021602675318718, + 0.009895769879221916, + 0.0070844595320522785, + 0.0032919382210820913, + 0.004650938790291548, + 0.004844785667955875, + 0.007591540459543467, + 0.00649002380669117, + 0.009806453250348568, + 0.023146813735365868, + 0.03332049399614334, + 0.01467788964509964, + 0.03136800229549408, + 0.039014432579278946, + 0.04079108685255051, + 0.047411151230335236, + 0.043027788400650024, + 0.015899449586868286, + 0.013642805628478527, + 0.009203042834997177, + 0.012771260924637318, + 0.027373166754841805, + 0.01816105842590332, + 0.039227697998285294, + 0.037923794239759445, + 0.004274285864084959, + 0.003152105025947094, + 0.004300429951399565, + 0.0028665834106504917, + 0.0042804451659321785, + 0.009749562479555607, + 0.0060252961702644825, + 0.0068563357926905155, + 0.0036366537678986788, + 0.007007780019193888, + 0.004580273758620024, + 0.004751360509544611, + 0.0035929277073591948, + 0.003605398815125227, + 0.004662978462874889, + 0.003645261051133275, + 0.003505575004965067, + 0.004662840161472559, + 0.003584959078580141, + 0.004463128745555878, + 0.005243209190666676, + 0.005171570461243391, + 0.004459688905626535, + 0.0074155936017632484, + 0.011708788573741913, + 0.017700478434562683, + 0.01042075827717781, + 0.015960989519953728, + 0.028399907052516937, + 0.023426886647939682, + 0.020665666088461876, + 0.02864125743508339, + 0.0032347608357667923, + 0.003264190163463354, + 0.00355106801725924, + 0.00441299844533205, + 0.006182873155921698, + 0.004465301055461168, + 0.0028363827150315046, + 0.004955808166414499, + 0.004689156077802181, + 0.0034836542326956987, + 0.007887938059866428, + 0.007352916058152914, + 0.011422946117818356, + 0.00423725601285696, + 0.0029286730568856, + 0.008292865939438343 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 3, + 7 + ], + "selected_reward": -0.005259960889816284 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a wide shot taken from behind and slightly upward of the character?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character wearing a high-collared, waterproof black trench coat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a metal platform atop a skyscraper with pipes and ventilation systems?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a cyberpunk metropolis with dense rain, holographic billboards, and neon signs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly pan forward, gradually moving over the character's shoulders?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1427910476922989 + ], + "opposite_logprob_tokens": [ + -2.0177910327911377 + ], + "logprob_reward": -0.1427910476922989, + "yes_no_margin_reward": 1.875, + "token_credit_reward": -0.1427910476922989, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05298200622200966, + 0.031337350606918335, + 0.06979884952306747, + 0.048192113637924194, + 0.025626154616475105, + 0.037988584488630295, + 0.024264391511678696, + 0.621534526348114, + 0.050243936479091644, + 0.038032080978155136 + ], + "credit_map": [ + 0.007179516367614269, + 0.0056271860376000404, + 0.003622861811891198, + 0.0036412046756595373, + 0.010851124301552773, + 0.006101458333432674, + 0.008890089578926563, + 0.0070685651153326035, + 0.002038131467998028, + 0.002906860550865531, + 0.0011506432201713324, + 0.002002461114898324, + 0.004845883697271347, + 0.004833599086850882, + 0.0062272995710372925, + 0.007332470268011093, + 0.004665822256356478, + 0.007970588281750679, + 0.004269315395504236, + 0.007948795333504677, + 0.012217888608574867, + 0.008161230944097042, + 0.013907982036471367, + 0.010657228529453278, + 0.002517649671062827, + 0.0019945960957556963, + 0.001835172763094306, + 0.0020714164711534977, + 0.0038807515520602465, + 0.002669266425073147, + 0.015262345783412457, + 0.017960909754037857, + 0.002724338322877884, + 0.002224559197202325, + 0.00213766866363585, + 0.0019095753086730838, + 0.003223166335374117, + 0.006149544846266508, + 0.003696341998875141, + 0.003560959827154875, + 0.0030768250580877066, + 0.0031564938835799694, + 0.0030505224131047726, + 0.003998666536062956, + 0.004886275622993708, + 0.006987735163420439, + 0.004977343138307333, + 0.007854719646275043, + 0.0038680643774569035, + 0.0038585481233894825, + 0.0027053512167185545, + 0.0027238239999860525, + 0.0025756412651389837, + 0.002702521625906229, + 0.0025567763950675726, + 0.0032736638095229864, + 0.04861524701118469, + 0.06200191751122475, + 0.043566711246967316, + 0.06636415421962738, + 0.09659573435783386, + 0.08658921718597412, + 0.09399528801441193, + 0.12380624562501907, + 0.00577098922803998, + 0.008227149024605751, + 0.005162030458450317, + 0.007583480328321457, + 0.005190091673284769, + 0.005736683029681444, + 0.007005905732512474, + 0.005567607004195452, + 0.0029563256539404392, + 0.003401047782972455, + 0.0051386975683271885, + 0.010398058220744133, + 0.002291061682626605, + 0.0027694886084645987, + 0.0032331650145351887, + 0.007844233885407448 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 2, + 0 + ], + "selected_reward": -0.1427910476922989 + } + ], + "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/0000121/group_15/index_tensor([2158, 2158])_group_15_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000121/group_15/index_tensor([2158, 2158])_group_15_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..eb83c2cff1f874672c484970e920e663b852bb93 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000121/group_15/index_tensor([2158, 2158])_group_15_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "In cinematic slow motion, a smooth, low-angle tracking shot perfectly synchronizes with a weary marathon runner just meters from the finish line. It's a woman in her late twenties, her face a mask of extreme effort and pain. Her dark brown hair, once in a ponytail, is now a wet, tangled mess plastered to her flushed, crimson skin. She wears a light yellow, sleeveless technical running shirt that clings to her lean torso, darkened and heavy with sweat, and black running shorts. As she summons her last reserves of energy and takes one final stride, she exhales sharply, desperate breaths; her cheeks puff, her lips part, and visible beads of sweat spray across her face and hair. As each droplet arcs through the air, the camera captures the harsh golden backlight of the late afternoon sun in exquisite detail, glittering like tiny, scattered diamonds. The background is a beautiful, soft-focus bokeh of the vibrant colors of the cheering crowd and the blurred structure of the finish line arch ahead. The entire scene is steeped in an atmosphere of intense, visceral, and ultimately triumphant intensity, focusing on the raw, powerful moments that push the human body to its absolute limit.", + "seed": 11194, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000121/group_15/index_2158_seed_11194_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": -0.11480081751942635, + "vlm_logprob": -0.11480081751942635, + "vlm_yes_no_margin": 2.7, + "vlm_token_credit": -0.11480081751942635, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 in cinematic slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject a woman in her late twenties wearing a light yellow sleeveless shirt and black shorts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman's hair wet, tangled, and plastered to her flushed crimson skin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a smooth low-angle tracking shot synchronized with the runner's final stride?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are visible beads of sweat shown spraying across the face and hair with droplets arcing 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": [ + -0.27409204840660095 + ], + "opposite_logprob_tokens": [ + -1.8990920782089233 + ], + "logprob_reward": -0.27409204840660095, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.27409204840660095, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.28723371028900146, + 0.047843143343925476, + 0.025388536974787712, + 0.02594323270022869, + 0.04036246985197067, + 0.0946393832564354, + 0.1352965533733368, + 0.05322765186429024, + 0.2635436952114105, + 0.026521584019064903 + ], + "credit_map": [ + 0.015215406194329262, + 0.08989208191633224, + 0.014896226115524769, + 0.017634643241763115, + 0.019435161724686623, + 0.022887319326400757, + 0.013801970519125462, + 0.09347092360258102, + 0.0049073658883571625, + 0.006111543159931898, + 0.0071711549535393715, + 0.006206111051142216, + 0.0048207007348537445, + 0.006247980520129204, + 0.005214904900640249, + 0.007163385860621929, + 0.0038075365591794252, + 0.0021046821493655443, + 0.0034088403917849064, + 0.0038376636803150177, + 0.0032859956845641136, + 0.003155333921313286, + 0.003713109064847231, + 0.0020753764547407627, + 0.001965028466656804, + 0.0032708272337913513, + 0.004887253977358341, + 0.0025150789879262447, + 0.0028996397741138935, + 0.002400313038378954, + 0.0050784628838300705, + 0.002926630899310112, + 0.0046243444085121155, + 0.006947176996618509, + 0.004321693908423185, + 0.004580727778375149, + 0.005624106153845787, + 0.005171266384422779, + 0.0042038047686219215, + 0.004889354109764099, + 0.009697431698441505, + 0.009168243035674095, + 0.009056361392140388, + 0.015466790646314621, + 0.01382411364465952, + 0.008356224745512009, + 0.012315499596297741, + 0.016754724085330963, + 0.025350170210003853, + 0.012257548049092293, + 0.01359407976269722, + 0.011412185616791248, + 0.025934020057320595, + 0.015298440121114254, + 0.013280690647661686, + 0.018169425427913666, + 0.002630917588248849, + 0.001924642245285213, + 0.006825473625212908, + 0.004711124114692211, + 0.0038158739916980267, + 0.01769489236176014, + 0.012682431377470493, + 0.002942304592579603, + 0.013293766416609287, + 0.02533252164721489, + 0.051982201635837555, + 0.020043576136231422, + 0.04686456918716431, + 0.04662732779979706, + 0.024806518107652664, + 0.034593261778354645, + 0.0026178243570029736, + 0.004318903665989637, + 0.002594658173620701, + 0.004490190651267767, + 0.0033545666374266148, + 0.0030071239452809095, + 0.003754805773496628, + 0.0023835147731006145 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 6 + ], + "selected_reward": -0.27409204840660095 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 in cinematic slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject a woman in her late twenties wearing a light yellow sleeveless shirt and black shorts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman's hair wet, tangled, and plastered to her flushed crimson skin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a smooth low-angle tracking shot synchronized with the runner's final stride?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are visible beads of sweat shown spraying across the face and hair with droplets arcing 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": [ + -0.04869576916098595 + ], + "opposite_logprob_tokens": [ + -3.0486958026885986 + ], + "logprob_reward": -0.04869576916098595, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04869576916098595, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12796080112457275, + 0.07056180387735367, + 0.06034894660115242, + 0.17188258469104767, + 0.0995948538184166, + 0.05954481288790703, + 0.04365919157862663, + 0.1615343987941742, + 0.16982901096343994, + 0.0350835956633091 + ], + "credit_map": [ + 0.008653135970234871, + 0.033668290823698044, + 0.012502485886216164, + 0.013244119472801685, + 0.009721939451992512, + 0.011007451452314854, + 0.009843802079558372, + 0.02931957133114338, + 0.006508132442831993, + 0.0066986652091145515, + 0.007360330782830715, + 0.006082998588681221, + 0.010129540227353573, + 0.0067086489871144295, + 0.018427135422825813, + 0.008646353147923946, + 0.007236091420054436, + 0.007078851573169231, + 0.006144245155155659, + 0.004795202519744635, + 0.012302534654736519, + 0.008207731880247593, + 0.008212104439735413, + 0.006372186355292797, + 0.013725809752941132, + 0.023719746619462967, + 0.022922713309526443, + 0.010291899554431438, + 0.014802820980548859, + 0.041335579007864, + 0.03053584136068821, + 0.0145481638610363, + 0.012213259004056454, + 0.015358391217887402, + 0.006430187728255987, + 0.012267276644706726, + 0.015299375168979168, + 0.0167385283857584, + 0.009071326814591885, + 0.012216504663228989, + 0.007034802809357643, + 0.005062244366854429, + 0.004134586546570063, + 0.010315610095858574, + 0.007399931084364653, + 0.006044322624802589, + 0.009866001084446907, + 0.009687310084700584, + 0.005463704466819763, + 0.0034824940375983715, + 0.005964561831206083, + 0.003819257253780961, + 0.008211852982640266, + 0.00369256385602057, + 0.005918023642152548, + 0.0071067349053919315, + 0.009124254807829857, + 0.006874421611428261, + 0.010891348123550415, + 0.045150794088840485, + 0.00855568703263998, + 0.014897177927196026, + 0.054651375859975815, + 0.011389331892132759, + 0.0116187809035182, + 0.01578165590763092, + 0.03963727504014969, + 0.013840471394360065, + 0.021715199574828148, + 0.024344949051737785, + 0.016124935820698738, + 0.02676575258374214, + 0.0023714250419288874, + 0.00625351769849658, + 0.0028746689204126596, + 0.006223404314368963, + 0.003956652712076902, + 0.00359558523632586, + 0.006211278028786182, + 0.003597061848267913 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 8, + 7 + ], + "selected_reward": -0.04869576916098595 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 in cinematic slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject a woman in her late twenties wearing a light yellow sleeveless shirt and black shorts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman's hair wet, tangled, and plastered to her flushed crimson skin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a smooth low-angle tracking shot synchronized with the runner's final stride?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are visible beads of sweat shown spraying across the face and hair with droplets arcing 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": [ + -0.023297857493162155 + ], + "opposite_logprob_tokens": [ + -3.7732977867126465 + ], + "logprob_reward": -0.023297857493162155, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023297857493162155, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16095352172851562, + 0.05869551748037338, + 0.0493917241692543, + 0.05699383094906807, + 0.10388030111789703, + 0.108421191573143, + 0.05260716378688812, + 0.07752469927072525, + 0.2990897595882416, + 0.03244224563241005 + ], + "credit_map": [ + 0.020999258384108543, + 0.016644956544041634, + 0.01495938841253519, + 0.020834418013691902, + 0.021298551931977272, + 0.015452845022082329, + 0.02640233188867569, + 0.024361778050661087, + 0.0065788631327450275, + 0.004749584943056107, + 0.009588786400854588, + 0.005925207864493132, + 0.00986687745898962, + 0.005329776555299759, + 0.010473730973899364, + 0.006182690616697073, + 0.006416196934878826, + 0.00542138097807765, + 0.004700277000665665, + 0.004528683610260487, + 0.0065272049978375435, + 0.008400317281484604, + 0.006836425978690386, + 0.006561243440955877, + 0.004812036175280809, + 0.008955583907663822, + 0.010572087951004505, + 0.006927025970071554, + 0.007141624577343464, + 0.006250615697354078, + 0.006660404149442911, + 0.005674455780535936, + 0.014697750099003315, + 0.012021581642329693, + 0.007939664646983147, + 0.007612460758537054, + 0.02086157724261284, + 0.01276245154440403, + 0.013146651908755302, + 0.01483816560357809, + 0.01000245101749897, + 0.007917211391031742, + 0.010621625930070877, + 0.010708007030189037, + 0.015334062278270721, + 0.00953583512455225, + 0.02209836058318615, + 0.02220364473760128, + 0.011572131887078285, + 0.00657676812261343, + 0.006576408166438341, + 0.006255541928112507, + 0.005312306806445122, + 0.006762133911252022, + 0.004176772199571133, + 0.0053751058876514435, + 0.00434125354513526, + 0.0030378049705177546, + 0.005962529685348272, + 0.015242855995893478, + 0.005249380599707365, + 0.012517011724412441, + 0.025434620678424835, + 0.005739252083003521, + 0.015670355409383774, + 0.02043299563229084, + 0.04044325277209282, + 0.018047530204057693, + 0.06766777485609055, + 0.0776175931096077, + 0.019906336441636086, + 0.03930395096540451, + 0.002421532291918993, + 0.004240460693836212, + 0.0026120846159756184, + 0.003370855236425996, + 0.004793691914528608, + 0.0038831180427223444, + 0.005107904318720102, + 0.006012598518282175 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 5 + ], + "selected_reward": -0.023297857493162155 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 in cinematic slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject a woman in her late twenties wearing a light yellow sleeveless shirt and black shorts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman's hair wet, tangled, and plastered to her flushed crimson skin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a smooth low-angle tracking shot synchronized with the runner's final stride?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are visible beads of sweat shown spraying across the face and hair with droplets arcing 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": [ + -0.20157140493392944 + ], + "opposite_logprob_tokens": [ + -1.7015714645385742 + ], + "logprob_reward": -0.20157140493392944, + "yes_no_margin_reward": 1.5, + "token_credit_reward": -0.20157140493392944, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19514551758766174, + 0.04827836900949478, + 0.03148507699370384, + 0.08806287497282028, + 0.11161430180072784, + 0.11584099382162094, + 0.18697099387645721, + 0.06425205618143082, + 0.137947216629982, + 0.02040261961519718 + ], + "credit_map": [ + 0.014393739402294159, + 0.04722154513001442, + 0.014379887841641903, + 0.015185526572167873, + 0.018688317388296127, + 0.025405166670680046, + 0.014965921640396118, + 0.0449054092168808, + 0.005027522798627615, + 0.005927592515945435, + 0.004350658971816301, + 0.006016516126692295, + 0.006929673254489899, + 0.00644190888851881, + 0.0061327735893428326, + 0.007451722398400307, + 0.004246646538376808, + 0.0031245939899235964, + 0.0038302959874272346, + 0.002799771726131439, + 0.004261757712811232, + 0.0043661040253937244, + 0.004818097222596407, + 0.004037808161228895, + 0.008306014351546764, + 0.01788613200187683, + 0.010292638093233109, + 0.008107488043606281, + 0.009658670984208584, + 0.01661885343492031, + 0.009888180531561375, + 0.007304899860173464, + 0.012995901517570019, + 0.01576823554933071, + 0.009130449965596199, + 0.011001547798514366, + 0.01689283922314644, + 0.020366376265883446, + 0.012487061321735382, + 0.012971892952919006, + 0.012253218330442905, + 0.010009106248617172, + 0.0106532396748662, + 0.016771115362644196, + 0.016282036900520325, + 0.011447515338659286, + 0.016099954023957253, + 0.0223248153924942, + 0.030506743118166924, + 0.017208248376846313, + 0.01429538894444704, + 0.013361507095396519, + 0.04312272369861603, + 0.01990446075797081, + 0.022758591920137405, + 0.02581334114074707, + 0.002709693741053343, + 0.002430783351883292, + 0.005969006568193436, + 0.0060326228849589825, + 0.00416357908397913, + 0.023820528760552406, + 0.015202618204057217, + 0.003923226613551378, + 0.011526858434081078, + 0.009908515959978104, + 0.03170977532863617, + 0.011778944171965122, + 0.019164785742759705, + 0.017591293901205063, + 0.012545263394713402, + 0.023721784353256226, + 0.0014999228296801448, + 0.0025627403520047665, + 0.002757997252047062, + 0.002922095125541091, + 0.0027255944442003965, + 0.0023541615810245275, + 0.0022038021124899387, + 0.003376305103302002 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 8 + ], + "selected_reward": -0.20157140493392944 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 in cinematic slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject a woman in her late twenties wearing a light yellow sleeveless shirt and black shorts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman's hair wet, tangled, and plastered to her flushed crimson skin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a smooth low-angle tracking shot synchronized with the runner's final stride?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are visible beads of sweat shown spraying across the face and hair with droplets arcing 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": [ + -0.026347007602453232 + ], + "opposite_logprob_tokens": [ + -3.6513469219207764 + ], + "logprob_reward": -0.026347007602453232, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026347007602453232, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23829613626003265, + 0.04658203944563866, + 0.02793835662305355, + 0.05248827114701271, + 0.05444185063242912, + 0.2116950899362564, + 0.03948398679494858, + 0.009742706082761288, + 0.29840025305747986, + 0.020931346341967583 + ], + "credit_map": [ + 0.025482231751084328, + 0.053363215178251266, + 0.017739344388246536, + 0.024121945723891258, + 0.028849132359027863, + 0.0281567070633173, + 0.02247626706957817, + 0.03810726851224899, + 0.004903789609670639, + 0.00595858134329319, + 0.005610807333141565, + 0.003838959150016308, + 0.007501460146158934, + 0.005413183011114597, + 0.008036435581743717, + 0.005318817682564259, + 0.003418821608647704, + 0.0031479820609092712, + 0.0028234715573489666, + 0.002021278254687786, + 0.0032874324824661016, + 0.005911934655159712, + 0.003131011500954628, + 0.0041964249685406685, + 0.0047405315563082695, + 0.00799352303147316, + 0.009446064941585064, + 0.006061989348381758, + 0.00687135336920619, + 0.006065967492759228, + 0.006349715404212475, + 0.00495911855250597, + 0.006671488750725985, + 0.007637795526534319, + 0.0032690796069800854, + 0.00479623069986701, + 0.010829895734786987, + 0.00890256930142641, + 0.005557893309742212, + 0.006776892114430666, + 0.019161347299814224, + 0.016853149980306625, + 0.018379731103777885, + 0.019808024168014526, + 0.031024355441331863, + 0.019068051129579544, + 0.04029354080557823, + 0.04710688814520836, + 0.006154086906462908, + 0.004648301284760237, + 0.0042640590108931065, + 0.005345969460904598, + 0.004841213580220938, + 0.004073115531355143, + 0.005000038538128138, + 0.005157200153917074, + 0.0007965576369315386, + 0.0010675498051568866, + 0.0010315411491319537, + 0.001673766877502203, + 0.0016909625846892595, + 0.0007139393710531294, + 0.0017940487014129758, + 0.0009743394912220538, + 0.02293153665959835, + 0.026642488315701485, + 0.03474874049425125, + 0.023375481367111206, + 0.05937182158231735, + 0.07376205176115036, + 0.022442730143666267, + 0.03512538596987724, + 0.001971256220713258, + 0.0028807264752686024, + 0.0016980291111394763, + 0.0029873859602957964, + 0.003162211272865534, + 0.002467599231749773, + 0.0032639678101986647, + 0.002500167116522789 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 5 + ], + "selected_reward": -0.026347007602453232 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In cinematic slow motion, a smooth, low-angle tracking shot perfectly synchronizes with a weary marathon runner just meters from the finish line. It's a woman in her late twenties, her face a mask of extreme effort and pain. Her dark brown hair, once in a ponytail, is now a wet, tangled mess plastered to her flushed, crimson skin. She wears a light yellow, sleeveless technical running shirt that clings to her lean torso, darkened and heavy with sweat, and black running shorts. As she summons her last reserves of energy and takes one final stride, she exhales sharply, desperate breaths; her cheeks puff, her lips part, and visible beads of sweat spray across her face and hair. As each droplet arcs through the air, the camera captures the harsh golden backlight of the late afternoon sun in exquisite detail, glittering like tiny, scattered diamonds. The background is a beautiful, soft-focus bokeh of the vibrant colors of the cheering crowd and the blurred structure of the finish line arch ahead. The entire scene is steeped in an atmosphere of intense, visceral, and ultimately triumphant intensity, focusing on the raw, powerful moments that push the human body to its absolute limit.", + "seed": 11195, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000121/group_15/index_2158_seed_11195_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": -0.11745481006801128, + "vlm_logprob": -0.11745481006801128, + "vlm_yes_no_margin": 2.475, + "vlm_token_credit": -0.11745481006801128, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 in cinematic slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject a woman in her late twenties wearing a light yellow sleeveless shirt and black shorts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman's hair wet, tangled, and plastered to her flushed crimson skin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a smooth low-angle tracking shot synchronized with the runner's final stride?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are visible beads of sweat shown spraying across the face and hair with droplets arcing 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": [ + -0.26679959893226624 + ], + "opposite_logprob_tokens": [ + -1.7667995691299438 + ], + "logprob_reward": -0.26679959893226624, + "yes_no_margin_reward": 1.5, + "token_credit_reward": -0.26679959893226624, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04278044402599335, + 0.02168983779847622, + 0.020908230915665627, + 0.04909465089440346, + 0.05321324244141579, + 0.6735994219779968, + 0.02367476373910904, + 0.06418360024690628, + 0.025808971375226974, + 0.025046827271580696 + ], + "credit_map": [ + 0.0053594377823174, + 0.006407016422599554, + 0.005138175096362829, + 0.00755782425403595, + 0.00490212207660079, + 0.003521943697705865, + 0.005534300580620766, + 0.004359622485935688, + 0.0022469295654445887, + 0.003862631507217884, + 0.0024515215773135424, + 0.0021592527627944946, + 0.0027523133903741837, + 0.0024817059747874737, + 0.003292400622740388, + 0.002443082397803664, + 0.0011289912508800626, + 0.0016556441551074386, + 0.0026683269534260035, + 0.0022516590543091297, + 0.003063945798203349, + 0.0017852516612038016, + 0.0032194936648011208, + 0.005134918261319399, + 0.0036938556004315615, + 0.0037210239097476006, + 0.011765392497181892, + 0.005401580594480038, + 0.005697748623788357, + 0.008872410282492638, + 0.005459641572088003, + 0.004482998978346586, + 0.004243185743689537, + 0.019000161439180374, + 0.003604473080486059, + 0.002765743527561426, + 0.0027906091418117285, + 0.01503014750778675, + 0.0035152819473296404, + 0.0022636419162154198, + 0.060218214988708496, + 0.08824475109577179, + 0.05770226940512657, + 0.040242090821266174, + 0.04559522122144699, + 0.14189408719539642, + 0.193242147564888, + 0.04646066948771477, + 0.0034228351432830095, + 0.0018440884305164218, + 0.004136503674089909, + 0.0018736084457486868, + 0.0014701022300869226, + 0.004583539441227913, + 0.003387149889022112, + 0.002956935903057456, + 0.010161512531340122, + 0.007189153227955103, + 0.009279458783566952, + 0.00632427167147398, + 0.009221266023814678, + 0.009634114801883698, + 0.006314110476523638, + 0.006059713661670685, + 0.0025043589994311333, + 0.0029224776662886143, + 0.002530104713514447, + 0.003122909227386117, + 0.00468165148049593, + 0.002958769677206874, + 0.004094657953828573, + 0.0029940425883978605, + 0.0020012464374303818, + 0.007021279074251652, + 0.0032347533851861954, + 0.0022058156318962574, + 0.0025426424108445644, + 0.0035342720802873373, + 0.0020468165166676044, + 0.0024600017350167036 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 7, + 4 + ], + "selected_reward": -0.26679959893226624 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 in cinematic slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject a woman in her late twenties wearing a light yellow sleeveless shirt and black shorts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman's hair wet, tangled, and plastered to her flushed crimson skin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a smooth low-angle tracking shot synchronized with the runner's final stride?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are visible beads of sweat shown spraying across the face and hair with droplets arcing 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": [ + -0.06209332123398781 + ], + "opposite_logprob_tokens": [ + -2.8120932579040527 + ], + "logprob_reward": -0.06209332123398781, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06209332123398781, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10752861946821213, + 0.04675355553627014, + 0.3683435916900635, + 0.10002756863832474, + 0.08813674002885818, + 0.07946443557739258, + 0.05258332937955856, + 0.06386473774909973, + 0.0532648079097271, + 0.04003264382481575 + ], + "credit_map": [ + 0.013212919235229492, + 0.010129409842193127, + 0.01921016536653042, + 0.01571246050298214, + 0.010634790174663067, + 0.006554191932082176, + 0.016853569075465202, + 0.015221109613776207, + 0.005488554481416941, + 0.004550826735794544, + 0.00470832409337163, + 0.004664754960685968, + 0.00618364242836833, + 0.004629215225577354, + 0.011578766629099846, + 0.004949470981955528, + 0.012336917221546173, + 0.01423828024417162, + 0.11687338352203369, + 0.01318446733057499, + 0.011833319440484047, + 0.014567442238330841, + 0.03421735018491745, + 0.15109242498874664, + 0.006859177257865667, + 0.008031013421714306, + 0.01662193238735199, + 0.02040715701878071, + 0.011839504353702068, + 0.011400432325899601, + 0.01391887478530407, + 0.010949477553367615, + 0.008023166097700596, + 0.030103951692581177, + 0.005353391170501709, + 0.004866170696914196, + 0.005623076111078262, + 0.02148640714585781, + 0.0074395593255758286, + 0.005241020116955042, + 0.008898532949388027, + 0.011055666953325272, + 0.009476780891418457, + 0.008602771908044815, + 0.01371840015053749, + 0.007734532933682203, + 0.011098561808466911, + 0.008879193104803562, + 0.007663090247660875, + 0.006745166145265102, + 0.009788837283849716, + 0.003965415526181459, + 0.0038638219702988863, + 0.008176122792065144, + 0.005489523522555828, + 0.006891353987157345, + 0.005826071836054325, + 0.005400265101343393, + 0.00619107112288475, + 0.006166562903672457, + 0.012552501633763313, + 0.008570299483835697, + 0.009852689690887928, + 0.009305271320044994, + 0.005660668946802616, + 0.006109945010393858, + 0.005353926215320826, + 0.007242737803608179, + 0.00674122991040349, + 0.00615914911031723, + 0.00830084178596735, + 0.007696307729929686, + 0.003888176055625081, + 0.009294455870985985, + 0.005239017307758331, + 0.003214902011677623, + 0.004975670017302036, + 0.005785382352769375, + 0.003999398555606604, + 0.0036356376949697733 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 3 + ], + "selected_reward": -0.06209332123398781 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 in cinematic slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject a woman in her late twenties wearing a light yellow sleeveless shirt and black shorts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman's hair wet, tangled, and plastered to her flushed crimson skin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a smooth low-angle tracking shot synchronized with the runner's final stride?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are visible beads of sweat shown spraying across the face and hair with droplets arcing 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": [ + -0.02981637232005596 + ], + "opposite_logprob_tokens": [ + -3.5298163890838623 + ], + "logprob_reward": -0.02981637232005596, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.02981637232005596, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16054439544677734, + 0.08710113912820816, + 0.12154800444841385, + 0.043686393648386, + 0.16888491809368134, + 0.27876585721969604, + 0.019941801205277443, + 0.03468254208564758, + 0.039433401077985764, + 0.04541158303618431 + ], + "credit_map": [ + 0.02395562268793583, + 0.013526265509426594, + 0.028565365821123123, + 0.017981139943003654, + 0.028909627348184586, + 0.010226254351437092, + 0.0202547125518322, + 0.01712542027235031, + 0.014469237066805363, + 0.00807739607989788, + 0.00942971184849739, + 0.006862926762551069, + 0.00810218881815672, + 0.005439744796603918, + 0.024206368252635002, + 0.010513564571738243, + 0.006997519172728062, + 0.005672823637723923, + 0.04045330733060837, + 0.005260920617729425, + 0.004063791129738092, + 0.004776005167514086, + 0.01041562668979168, + 0.043908003717660904, + 0.005772153846919537, + 0.00463579036295414, + 0.008026722818613052, + 0.005027275066822767, + 0.005289627239108086, + 0.0044054193422198296, + 0.006094081327319145, + 0.004435322247445583, + 0.010480043478310108, + 0.06396391242742538, + 0.012036321684718132, + 0.008032586425542831, + 0.011850138194859028, + 0.042803868651390076, + 0.011864855885505676, + 0.00785318948328495, + 0.023835455998778343, + 0.03439181298017502, + 0.024054646492004395, + 0.014479792676866055, + 0.01808549463748932, + 0.05990634858608246, + 0.08559593558311462, + 0.018416346982121468, + 0.0018763401312753558, + 0.0017718710005283356, + 0.002612865762785077, + 0.0018496185075491667, + 0.001706318580545485, + 0.0026865822728723288, + 0.0035136223305016756, + 0.003924583550542593, + 0.004894646815955639, + 0.0023539895191788673, + 0.0030701535288244486, + 0.003716406412422657, + 0.006205072160810232, + 0.0052247196435928345, + 0.004114923998713493, + 0.005102630704641342, + 0.0029383187647908926, + 0.004020317457616329, + 0.0035980558022856712, + 0.004438106436282396, + 0.008547267876565456, + 0.004085874184966087, + 0.00612833583727479, + 0.005677125882357359, + 0.0038303169421851635, + 0.008226709440350533, + 0.004890636075288057, + 0.004457037430256605, + 0.007042602635920048, + 0.0058245412074029446, + 0.004782145377248526, + 0.006357594858855009 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 4, + 0 + ], + "selected_reward": -0.02981637232005596 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 in cinematic slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject a woman in her late twenties wearing a light yellow sleeveless shirt and black shorts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman's hair wet, tangled, and plastered to her flushed crimson skin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a smooth low-angle tracking shot synchronized with the runner's final stride?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are visible beads of sweat shown spraying across the face and hair with droplets arcing 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": [ + -0.17990589141845703 + ], + "opposite_logprob_tokens": [ + -1.804905891418457 + ], + "logprob_reward": -0.17990589141845703, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.17990589141845703, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.049435876309871674, + 0.03851116821169853, + 0.11482284218072891, + 0.05850986763834953, + 0.15861061215400696, + 0.42263346910476685, + 0.03953445702791214, + 0.02084965445101261, + 0.05432160943746567, + 0.042770471423864365 + ], + "credit_map": [ + 0.0045350296422839165, + 0.006394637282937765, + 0.005291810259222984, + 0.008993118070065975, + 0.006743493489921093, + 0.003878662595525384, + 0.00692303292453289, + 0.006676084361970425, + 0.004586818162351847, + 0.004566625691950321, + 0.004658923484385014, + 0.004253961145877838, + 0.0061424970626831055, + 0.0028196806088089943, + 0.00609405804425478, + 0.005388605408370495, + 0.0041892919689416885, + 0.004366936627775431, + 0.03680776059627533, + 0.004421654157340527, + 0.0036265740636736155, + 0.004463437478989363, + 0.010669074952602386, + 0.04627811163663864, + 0.005099277012050152, + 0.003290402702987194, + 0.01785975694656372, + 0.005686555989086628, + 0.0053848037496209145, + 0.008762944489717484, + 0.008036725223064423, + 0.004389398731291294, + 0.011352787725627422, + 0.06138768792152405, + 0.01771855540573597, + 0.007243403233587742, + 0.009987990371882915, + 0.03400173410773277, + 0.01120449136942625, + 0.005713951773941517, + 0.036689385771751404, + 0.0526963472366333, + 0.03713100403547287, + 0.023449741303920746, + 0.027259837836027145, + 0.09014075249433517, + 0.1262555718421936, + 0.029010822996497154, + 0.004294583573937416, + 0.002982241800054908, + 0.004150296561419964, + 0.003935583867132664, + 0.004096146672964096, + 0.0074437083676457405, + 0.0054981461726129055, + 0.0071337479166686535, + 0.0032158850226551294, + 0.0025721851270645857, + 0.0031605265103280544, + 0.0018946470227092505, + 0.0022186413407325745, + 0.002747781341895461, + 0.0027686054818332195, + 0.0022713823709636927, + 0.004602187778800726, + 0.005254683084785938, + 0.004598828498274088, + 0.007591548841446638, + 0.011258714832365513, + 0.005753993522375822, + 0.007995222695171833, + 0.007266428787261248, + 0.0033240218181163073, + 0.013867919333279133, + 0.004956929013133049, + 0.002733078319579363, + 0.0037802555598318577, + 0.007446032017469406, + 0.003389748279005289, + 0.003272483590990305 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 4, + 2 + ], + "selected_reward": -0.17990589141845703 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 in cinematic slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject a woman in her late twenties wearing a light yellow sleeveless shirt and black shorts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman's hair wet, tangled, and plastered to her flushed crimson skin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a smooth low-angle tracking shot synchronized with the runner's final stride?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are visible beads of sweat shown spraying across the face and hair with droplets arcing 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": [ + -0.04865886643528938 + ], + "opposite_logprob_tokens": [ + -3.048658847808838 + ], + "logprob_reward": -0.04865886643528938, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04865886643528938, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17110231518745422, + 0.03622913733124733, + 0.10951440036296844, + 0.08236794173717499, + 0.06214320287108421, + 0.2855990529060364, + 0.04157280549407005, + 0.066832534968853, + 0.03838132694363594, + 0.10625725984573364 + ], + "credit_map": [ + 0.027299849316477776, + 0.01350540854036808, + 0.05071452260017395, + 0.014046324416995049, + 0.01306961476802826, + 0.011551144532859325, + 0.024110423400998116, + 0.016805032268166542, + 0.006356445606797934, + 0.0034651656169444323, + 0.005056772381067276, + 0.0036419671960175037, + 0.0031713659409433603, + 0.0032628204207867384, + 0.0060644373297691345, + 0.0052101644687354565, + 0.005226355046033859, + 0.005534278694540262, + 0.03311945125460625, + 0.006314495112746954, + 0.004569862503558397, + 0.0046731471084058285, + 0.010369914583861828, + 0.03970690071582794, + 0.007731730584055185, + 0.005479305516928434, + 0.019623897969722748, + 0.010952363722026348, + 0.008165597915649414, + 0.010262696072459221, + 0.012792747467756271, + 0.00735960528254509, + 0.0044543808326125145, + 0.019880780950188637, + 0.008934601210057735, + 0.003027865197509527, + 0.004918798804283142, + 0.011983600445091724, + 0.005677544977515936, + 0.0032656334806233644, + 0.023783627897500992, + 0.03100409172475338, + 0.023139646276831627, + 0.016465192660689354, + 0.01947207935154438, + 0.06374874711036682, + 0.0889158770442009, + 0.01906980387866497, + 0.003596962196752429, + 0.00399447837844491, + 0.003875591093674302, + 0.0038902140222489834, + 0.003285575658082962, + 0.007856135256588459, + 0.007586908061057329, + 0.007486947346478701, + 0.01087521854788065, + 0.00615311274304986, + 0.007626006379723549, + 0.007862363941967487, + 0.007632714230567217, + 0.012027845717966557, + 0.006273135542869568, + 0.008382136933505535, + 0.003305213525891304, + 0.004313075449317694, + 0.004400322679430246, + 0.004444705788046122, + 0.006977536249905825, + 0.0037102210335433483, + 0.005363553762435913, + 0.005866698455065489, + 0.008720749989151955, + 0.0317724272608757, + 0.011394566856324673, + 0.006328857969492674, + 0.010822645388543606, + 0.018788596615195274, + 0.008156532421708107, + 0.010272891260683537 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 2 + ], + "selected_reward": -0.04865886643528938 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In cinematic slow motion, a smooth, low-angle tracking shot perfectly synchronizes with a weary marathon runner just meters from the finish line. It's a woman in her late twenties, her face a mask of extreme effort and pain. Her dark brown hair, once in a ponytail, is now a wet, tangled mess plastered to her flushed, crimson skin. She wears a light yellow, sleeveless technical running shirt that clings to her lean torso, darkened and heavy with sweat, and black running shorts. As she summons her last reserves of energy and takes one final stride, she exhales sharply, desperate breaths; her cheeks puff, her lips part, and visible beads of sweat spray across her face and hair. As each droplet arcs through the air, the camera captures the harsh golden backlight of the late afternoon sun in exquisite detail, glittering like tiny, scattered diamonds. The background is a beautiful, soft-focus bokeh of the vibrant colors of the cheering crowd and the blurred structure of the finish line arch ahead. The entire scene is steeped in an atmosphere of intense, visceral, and ultimately triumphant intensity, focusing on the raw, powerful moments that push the human body to its absolute limit.", + "seed": 11196, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000121/group_15/index_2158_seed_11196_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": -0.091785703971982, + "vlm_logprob": -0.091785703971982, + "vlm_yes_no_margin": 2.825, + "vlm_token_credit": -0.091785703971982, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 in cinematic slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject a woman in her late twenties wearing a light yellow sleeveless shirt and black shorts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman's hair wet, tangled, and plastered to her flushed crimson skin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a smooth low-angle tracking shot synchronized with the runner's final stride?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are visible beads of sweat shown spraying across the face and hair with droplets arcing 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": [ + -0.19846269488334656 + ], + "opposite_logprob_tokens": [ + -2.323462724685669 + ], + "logprob_reward": -0.19846269488334656, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.19846269488334656, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08596444129943848, + 0.43176355957984924, + 0.047225065529346466, + 0.054850827902555466, + 0.07223537564277649, + 0.021486205980181694, + 0.13025470077991486, + 0.04832564666867256, + 0.06143118068575859, + 0.04646297171711922 + ], + "credit_map": [ + 0.012231851927936077, + 0.010093576274812222, + 0.009973430074751377, + 0.013735031709074974, + 0.015552262775599957, + 0.007278142496943474, + 0.010887142270803452, + 0.0062129986472427845, + 0.05089649185538292, + 0.13402417302131653, + 0.16383685171604156, + 0.01985207386314869, + 0.01751272939145565, + 0.012746574357151985, + 0.01923358254134655, + 0.01366108376532793, + 0.007332395296543837, + 0.010374057106673717, + 0.005773274693638086, + 0.0036281971260905266, + 0.004620587453246117, + 0.007694322150200605, + 0.004694860894232988, + 0.0031073635909706354, + 0.005944417789578438, + 0.004465128760784864, + 0.006428082007914782, + 0.007496858015656471, + 0.00776710594072938, + 0.007146218325942755, + 0.008699587546288967, + 0.006903428118675947, + 0.008336080238223076, + 0.008838973939418793, + 0.015575365163385868, + 0.00873955711722374, + 0.008162036538124084, + 0.008880279026925564, + 0.007640464697033167, + 0.006062611006200314, + 0.0022308428306132555, + 0.0031665784772485495, + 0.0024007572792470455, + 0.002933710580691695, + 0.0033047173637896776, + 0.002184830838814378, + 0.0031175781041383743, + 0.0021471893414855003, + 0.011412899009883404, + 0.019230032339692116, + 0.014530918560922146, + 0.012943764217197895, + 0.020740702748298645, + 0.01742960885167122, + 0.017746586352586746, + 0.016220174729824066, + 0.005174012389034033, + 0.006056820973753929, + 0.0073655457235872746, + 0.004309840500354767, + 0.00590013712644577, + 0.007277016062289476, + 0.0057401349768042564, + 0.006502131000161171, + 0.011035683564841747, + 0.005476508289575577, + 0.00993928499519825, + 0.006772868800908327, + 0.006656398065388203, + 0.0072884950786828995, + 0.008828243240714073, + 0.0054336898028850555, + 0.003979694563895464, + 0.00700161000713706, + 0.015448948368430138, + 0.00356365111656487, + 0.0038185645826160908, + 0.002940376987680793, + 0.007131639402359724, + 0.00257848366163671 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 6, + 0 + ], + "selected_reward": -0.19846269488334656 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 in cinematic slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject a woman in her late twenties wearing a light yellow sleeveless shirt and black shorts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman's hair wet, tangled, and plastered to her flushed crimson skin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a smooth low-angle tracking shot synchronized with the runner's final stride?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are visible beads of sweat shown spraying across the face and hair with droplets arcing 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": [ + -0.043095942586660385 + ], + "opposite_logprob_tokens": [ + -3.168095827102661 + ], + "logprob_reward": -0.043095942586660385, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.043095942586660385, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17818598449230194, + 0.23625540733337402, + 0.04918345808982849, + 0.09316537529230118, + 0.03290494158864021, + 0.03322974592447281, + 0.10334055870771408, + 0.08474128693342209, + 0.08363473415374756, + 0.10535851120948792 + ], + "credit_map": [ + 0.027753401547670364, + 0.0237421914935112, + 0.015165064483880997, + 0.020316535606980324, + 0.02971798926591873, + 0.014902357943356037, + 0.030484819784760475, + 0.016103610396385193, + 0.027563422918319702, + 0.053132496774196625, + 0.09990344941616058, + 0.012811185792088509, + 0.010202021338045597, + 0.008952165953814983, + 0.013517167419195175, + 0.01017350610345602, + 0.005690484307706356, + 0.009532894939184189, + 0.007050590123981237, + 0.0038481028750538826, + 0.006304845679551363, + 0.007003304082900286, + 0.007078760769218206, + 0.0026744785718619823, + 0.007477738428860903, + 0.006851173471659422, + 0.006930968724191189, + 0.01361471600830555, + 0.019935693591833115, + 0.01034714374691248, + 0.01750088483095169, + 0.010507061146199703, + 0.003300965065136552, + 0.003493059426546097, + 0.0038250593934208155, + 0.004018288105726242, + 0.004253740888088942, + 0.003325188532471657, + 0.006779592018574476, + 0.003909049089998007, + 0.0032497199717909098, + 0.005396885797381401, + 0.004285522736608982, + 0.00432316679507494, + 0.003719570115208626, + 0.004214457236230373, + 0.0043671573512256145, + 0.0036732680164277554, + 0.00786564126610756, + 0.012613809667527676, + 0.007651845458894968, + 0.009727944619953632, + 0.017725851386785507, + 0.01303914375603199, + 0.02206665277481079, + 0.012649674899876118, + 0.007346672005951405, + 0.009919200092554092, + 0.008506499230861664, + 0.007171029690653086, + 0.013810996897518635, + 0.007410303223878145, + 0.017326336354017258, + 0.013250247575342655, + 0.013568581081926823, + 0.006691265851259232, + 0.010085001587867737, + 0.009114439599215984, + 0.015349105931818485, + 0.009115953929722309, + 0.01230044849216938, + 0.007409938611090183, + 0.008589924313127995, + 0.02130228839814663, + 0.030847804620862007, + 0.009055273607373238, + 0.007763829082250595, + 0.008556265383958817, + 0.010172087699174881, + 0.009071038104593754 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 9 + ], + "selected_reward": -0.043095942586660385 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 in cinematic slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject a woman in her late twenties wearing a light yellow sleeveless shirt and black shorts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman's hair wet, tangled, and plastered to her flushed crimson skin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a smooth low-angle tracking shot synchronized with the runner's final stride?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are visible beads of sweat shown spraying across the face and hair with droplets arcing 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": [ + -0.033707089722156525 + ], + "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": [ + 0.08581466972827911, + 0.3425949215888977, + 0.03297649696469307, + 0.08653785288333893, + 0.035846490412950516, + 0.03328268975019455, + 0.1772894263267517, + 0.047233857214450836, + 0.12740130722522736, + 0.031022317707538605 + ], + "credit_map": [ + 0.012780501507222652, + 0.010499700903892517, + 0.011085200123488903, + 0.00810378696769476, + 0.011096629314124584, + 0.00897054374217987, + 0.01233439240604639, + 0.010943915694952011, + 0.06191176548600197, + 0.11937864124774933, + 0.09724146872758865, + 0.015337269753217697, + 0.01444798894226551, + 0.012016471475362778, + 0.012768197804689407, + 0.00949309766292572, + 0.0042336019687354565, + 0.005960486363619566, + 0.004126001615077257, + 0.003074438776820898, + 0.003684720490127802, + 0.005824380088597536, + 0.0034125091042369604, + 0.002660357626155019, + 0.009458646178245544, + 0.00898226909339428, + 0.008658424019813538, + 0.010945017449557781, + 0.013133336789906025, + 0.00940429326146841, + 0.012463473714888096, + 0.013492392376065254, + 0.0031707293819636106, + 0.007501892279833555, + 0.004250814206898212, + 0.005349376704543829, + 0.004155194852501154, + 0.003338029608130455, + 0.004458935931324959, + 0.003621516516432166, + 0.004016989376395941, + 0.006636371370404959, + 0.004033858422189951, + 0.0036877787206321955, + 0.003817118238657713, + 0.003969600424170494, + 0.004540806170552969, + 0.002580168889835477, + 0.016693949699401855, + 0.032306473702192307, + 0.018113013356924057, + 0.016163697466254234, + 0.01752445101737976, + 0.02288796938955784, + 0.02113226242363453, + 0.032467618584632874, + 0.005121524445712566, + 0.005482083186507225, + 0.0076697878539562225, + 0.004263231065124273, + 0.006337013561278582, + 0.006333450321108103, + 0.005418803542852402, + 0.0066079627722501755, + 0.016543397679924965, + 0.0097366813570261, + 0.018314559012651443, + 0.01540422160178423, + 0.021191049367189407, + 0.014877342619001865, + 0.019353188574314117, + 0.011980860494077206, + 0.002594126621261239, + 0.00284207914955914, + 0.006366095505654812, + 0.0030981688760221004, + 0.004428618121892214, + 0.002378637669607997, + 0.006713639013469219, + 0.002600952982902527 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 6, + 8 + ], + "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 scene shot in cinematic slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject a woman in her late twenties wearing a light yellow sleeveless shirt and black shorts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman's hair wet, tangled, and plastered to her flushed crimson skin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a smooth low-angle tracking shot synchronized with the runner's final stride?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are visible beads of sweat shown spraying across the face and hair with droplets arcing 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": [ + -0.1603722721338272 + ], + "opposite_logprob_tokens": [ + -1.910372257232666 + ], + "logprob_reward": -0.1603722721338272, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.1603722721338272, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1140073910355568, + 0.34731972217559814, + 0.061072781682014465, + 0.06774570047855377, + 0.06363839656114578, + 0.043991945683956146, + 0.09599024057388306, + 0.07253634184598923, + 0.04729427024722099, + 0.08640323579311371 + ], + "credit_map": [ + 0.01517790649086237, + 0.012124672532081604, + 0.01850246824324131, + 0.013623547740280628, + 0.018673285841941833, + 0.010509364306926727, + 0.01328691840171814, + 0.012109226547181606, + 0.045090898871421814, + 0.10974466800689697, + 0.12515756487846375, + 0.015568004921078682, + 0.015133030712604523, + 0.010603461414575577, + 0.014123689383268356, + 0.011898409575223923, + 0.007909436710178852, + 0.009719929657876492, + 0.008599214255809784, + 0.005576586350798607, + 0.008189692161977291, + 0.010223970748484135, + 0.006844249088317156, + 0.004009699448943138, + 0.007031815592199564, + 0.005828752648085356, + 0.007189465221017599, + 0.010222950018942356, + 0.010234855115413666, + 0.00674190791323781, + 0.012423780746757984, + 0.008072171360254288, + 0.0060902549885213375, + 0.00927312858402729, + 0.008522946387529373, + 0.009240463376045227, + 0.007986376993358135, + 0.005640449933707714, + 0.008472648449242115, + 0.008412126451730728, + 0.004929339978843927, + 0.00969777163118124, + 0.006734284572303295, + 0.00580758647993207, + 0.004904974717646837, + 0.003865630365908146, + 0.003586237784475088, + 0.004466119688004255, + 0.008871546946465969, + 0.012598532252013683, + 0.008891643024981022, + 0.013303874991834164, + 0.011390757746994495, + 0.011015205644071102, + 0.0165301151573658, + 0.013388557359576225, + 0.00713952025398612, + 0.00843318272382021, + 0.010113446973264217, + 0.0067858207039535046, + 0.01189530361443758, + 0.00869910977780819, + 0.009629710577428341, + 0.00984024628996849, + 0.00764054711908102, + 0.0047379666939377785, + 0.004990145098417997, + 0.003860803786665201, + 0.007677984423935413, + 0.00807732529938221, + 0.0060122208669781685, + 0.004297276027500629, + 0.008008268661797047, + 0.016392722725868225, + 0.024266479536890984, + 0.009294619783759117, + 0.006163382902741432, + 0.006916872225701809, + 0.007821629755198956, + 0.007539261132478714 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 6 + ], + "selected_reward": -0.1603722721338272 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 in cinematic slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject a woman in her late twenties wearing a light yellow sleeveless shirt and black shorts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman's hair wet, tangled, and plastered to her flushed crimson skin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a smooth low-angle tracking shot synchronized with the runner's final stride?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are visible beads of sweat shown spraying across the face and hair with droplets arcing 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": [ + -0.023290520533919334 + ], + "opposite_logprob_tokens": [ + -3.7732906341552734 + ], + "logprob_reward": -0.023290520533919334, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023290520533919334, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06311056762933731, + 0.22250795364379883, + 0.03237983211874962, + 0.034610968083143234, + 0.07922574877738953, + 0.05559162423014641, + 0.12111818045377731, + 0.033688921481370926, + 0.3003704249858856, + 0.05739579349756241 + ], + "credit_map": [ + 0.008855520747601986, + 0.006570976227521896, + 0.008121101185679436, + 0.010671949945390224, + 0.009687632322311401, + 0.00590789457783103, + 0.007140119094401598, + 0.006155370734632015, + 0.024135004729032516, + 0.06478941440582275, + 0.09032638370990753, + 0.010227374732494354, + 0.008438161574304104, + 0.006378869991749525, + 0.009970675222575665, + 0.008242078125476837, + 0.00392301008105278, + 0.004945800639688969, + 0.003951725084334612, + 0.002320570405572653, + 0.00486315181478858, + 0.005869864020496607, + 0.004174874629825354, + 0.0023308354429900646, + 0.004444381687790155, + 0.0038716483395546675, + 0.0038159294053912163, + 0.0035791455302387476, + 0.00487699918448925, + 0.0043274941854178905, + 0.005346011370420456, + 0.004349358845502138, + 0.009564434178173542, + 0.011072046123445034, + 0.017206747084856033, + 0.009965253062546253, + 0.00607402715831995, + 0.011224497109651566, + 0.008651128970086575, + 0.00546761741861701, + 0.006405162625014782, + 0.011495892889797688, + 0.007137783337384462, + 0.005559380166232586, + 0.006169616244733334, + 0.006247925106436014, + 0.008354631252586842, + 0.004221231676638126, + 0.009580647572875023, + 0.016146888956427574, + 0.00944435503333807, + 0.013997041620314121, + 0.020351232960820198, + 0.012373754754662514, + 0.021458996459841728, + 0.017765259370207787, + 0.0030999104492366314, + 0.0037412953097373247, + 0.003517549717798829, + 0.004173908848315477, + 0.0057718283496797085, + 0.0038346643559634686, + 0.004238922614604235, + 0.00531084043905139, + 0.036936383694410324, + 0.02360139973461628, + 0.04907897859811783, + 0.037056565284729004, + 0.04188796505331993, + 0.03897909075021744, + 0.04486823081970215, + 0.027961814776062965, + 0.005352439358830452, + 0.013240973465144634, + 0.01144978404045105, + 0.004894555546343327, + 0.005185346584767103, + 0.003958108834922314, + 0.009057773277163506, + 0.004256812855601311 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 1, + 6 + ], + "selected_reward": -0.023290520533919334 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In cinematic slow motion, a smooth, low-angle tracking shot perfectly synchronizes with a weary marathon runner just meters from the finish line. It's a woman in her late twenties, her face a mask of extreme effort and pain. Her dark brown hair, once in a ponytail, is now a wet, tangled mess plastered to her flushed, crimson skin. She wears a light yellow, sleeveless technical running shirt that clings to her lean torso, darkened and heavy with sweat, and black running shorts. As she summons her last reserves of energy and takes one final stride, she exhales sharply, desperate breaths; her cheeks puff, her lips part, and visible beads of sweat spray across her face and hair. As each droplet arcs through the air, the camera captures the harsh golden backlight of the late afternoon sun in exquisite detail, glittering like tiny, scattered diamonds. The background is a beautiful, soft-focus bokeh of the vibrant colors of the cheering crowd and the blurred structure of the finish line arch ahead. The entire scene is steeped in an atmosphere of intense, visceral, and ultimately triumphant intensity, focusing on the raw, powerful moments that push the human body to its absolute limit.", + "seed": 11197, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000121/group_15/index_2158_seed_11197_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": -0.14221062660217285, + "vlm_logprob": -0.14221062660217285, + "vlm_yes_no_margin": 2.375, + "vlm_token_credit": -0.14221062660217285, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 in cinematic slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject a woman in her late twenties wearing a light yellow sleeveless shirt and black shorts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman's hair wet, tangled, and plastered to her flushed crimson skin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a smooth low-angle tracking shot synchronized with the runner's final stride?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are visible beads of sweat shown spraying across the face and hair with droplets arcing 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": [ + -0.3171466886997223 + ], + "opposite_logprob_tokens": [ + -1.8171466588974 + ], + "logprob_reward": -0.3171466886997223, + "yes_no_margin_reward": 1.5, + "token_credit_reward": -0.3171466886997223, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14227251708507538, + 0.07478802651166916, + 0.17966648936271667, + 0.2031843513250351, + 0.08625217527151108, + 0.04471947252750397, + 0.05556289479136467, + 0.09385517239570618, + 0.06445968896150589, + 0.055239200592041016 + ], + "credit_map": [ + 0.018570346757769585, + 0.012616508640348911, + 0.022904250770807266, + 0.01790766604244709, + 0.016518929973244667, + 0.017930081114172935, + 0.013128128834068775, + 0.02269659750163555, + 0.0069839730858802795, + 0.018693089485168457, + 0.006597410421818495, + 0.005331886932253838, + 0.00760654965415597, + 0.005264463368803263, + 0.010717595927417278, + 0.013593052513897419, + 0.01381008606404066, + 0.04241267219185829, + 0.02949882298707962, + 0.02701341174542904, + 0.020237954333424568, + 0.017748989164829254, + 0.010640949942171574, + 0.018303601071238518, + 0.020891448482871056, + 0.019768185913562775, + 0.05898773670196533, + 0.012809019535779953, + 0.010281210765242577, + 0.010887697339057922, + 0.008850024081766605, + 0.06070903316140175, + 0.013813083991408348, + 0.01248917169868946, + 0.010294690728187561, + 0.010219288058578968, + 0.010376322083175182, + 0.010953208431601524, + 0.007638469338417053, + 0.01046793907880783, + 0.006027129013091326, + 0.006313614547252655, + 0.005648712627589703, + 0.0034598095808178186, + 0.0052479892037808895, + 0.00675676204264164, + 0.004084418062120676, + 0.0071810404770076275, + 0.008605031296610832, + 0.005868085660040379, + 0.009898342192173004, + 0.005265907384455204, + 0.007058872375637293, + 0.008462594822049141, + 0.0041062734089791775, + 0.006297788117080927, + 0.006458815187215805, + 0.013370096683502197, + 0.020228756591677666, + 0.019912730902433395, + 0.007134357001632452, + 0.00812886469066143, + 0.00849859043955803, + 0.010122966021299362, + 0.00455061998218298, + 0.005490107927471399, + 0.016819031909108162, + 0.007448654621839523, + 0.009047227911651134, + 0.008481964468955994, + 0.005548318848013878, + 0.007073765154927969, + 0.0060210986994206905, + 0.004747631959617138, + 0.007296760566532612, + 0.004890180192887783, + 0.008990867994725704, + 0.00738332187756896, + 0.008338246494531631, + 0.007571091875433922 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 2, + 0 + ], + "selected_reward": -0.3171466886997223 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 in cinematic slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject a woman in her late twenties wearing a light yellow sleeveless shirt and black shorts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman's hair wet, tangled, and plastered to her flushed crimson skin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a smooth low-angle tracking shot synchronized with the runner's final stride?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are visible beads of sweat shown spraying across the face and hair with droplets arcing 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": [ + -0.03814997151494026 + ], + "opposite_logprob_tokens": [ + -3.2881500720977783 + ], + "logprob_reward": -0.03814997151494026, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03814997151494026, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21247994899749756, + 0.0278344564139843, + 0.09718754142522812, + 0.37663498520851135, + 0.032065629959106445, + 0.026632463559508324, + 0.023078447207808495, + 0.0864805281162262, + 0.0816822499036789, + 0.03592376783490181 + ], + "credit_map": [ + 0.022185945883393288, + 0.015700621530413628, + 0.040025580674409866, + 0.0497417151927948, + 0.01589629240334034, + 0.017097139731049538, + 0.02283376082777977, + 0.028998885303735733, + 0.0029212445951998234, + 0.003939731512218714, + 0.004080332815647125, + 0.0027861243579536676, + 0.0034976464230567217, + 0.0028708060272037983, + 0.004053039476275444, + 0.003685530973598361, + 0.008679285645484924, + 0.02346993237733841, + 0.011861843056976795, + 0.014374794438481331, + 0.00977251771837473, + 0.014872751198709011, + 0.00501074455678463, + 0.009145675227046013, + 0.02344692312180996, + 0.019786996766924858, + 0.13003848493099213, + 0.023277077823877335, + 0.01820879429578781, + 0.019253354519605637, + 0.018230829387903214, + 0.12439253181219101, + 0.0032363852951675653, + 0.00331692467443645, + 0.003330792998895049, + 0.0033033506479114294, + 0.004858838394284248, + 0.005077994894236326, + 0.004648199770599604, + 0.0042931451462209225, + 0.0033297722693532705, + 0.002919568680226803, + 0.005398462060838938, + 0.0027359388768672943, + 0.0026354670990258455, + 0.00357558554969728, + 0.002060824306681752, + 0.003976844716817141, + 0.0029064274858683348, + 0.002727779094129801, + 0.004325853660702705, + 0.002435769885778427, + 0.0021375117357820272, + 0.0040082610212266445, + 0.0019675553776323795, + 0.002569289179518819, + 0.009377497248351574, + 0.011095584370195866, + 0.024488642811775208, + 0.014048062264919281, + 0.006482573226094246, + 0.0068937684409320354, + 0.006103276275098324, + 0.007991122081875801, + 0.007448755670338869, + 0.005843584891408682, + 0.010911275632679462, + 0.013064740225672722, + 0.011321738362312317, + 0.008739880286157131, + 0.009743097238242626, + 0.014609167352318764, + 0.004413314163684845, + 0.002455017762258649, + 0.005343636032193899, + 0.003769450355321169, + 0.005152489524334669, + 0.005776476580649614, + 0.004268269520252943, + 0.004745112266391516 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 2 + ], + "selected_reward": -0.03814997151494026 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 in cinematic slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject a woman in her late twenties wearing a light yellow sleeveless shirt and black shorts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman's hair wet, tangled, and plastered to her flushed crimson skin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a smooth low-angle tracking shot synchronized with the runner's final stride?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are visible beads of sweat shown spraying across the face and hair with droplets arcing 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": [ + -0.048670560121536255 + ], + "opposite_logprob_tokens": [ + -3.048670530319214 + ], + "logprob_reward": -0.048670560121536255, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.048670560121536255, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20348697900772095, + 0.040554966777563095, + 0.07050035893917084, + 0.4016996920108795, + 0.050323236733675, + 0.03016355261206627, + 0.02181324176490307, + 0.0220594834536314, + 0.0851297602057457, + 0.07426871359348297 + ], + "credit_map": [ + 0.02335568331182003, + 0.014713572338223457, + 0.04012671485543251, + 0.03191220387816429, + 0.02012186124920845, + 0.018016083166003227, + 0.01896658167243004, + 0.03627428784966469, + 0.0047317007556557655, + 0.006254622712731361, + 0.0062279305420815945, + 0.0036781388334929943, + 0.004343649372458458, + 0.005073742009699345, + 0.004590953700244427, + 0.005654228385537863, + 0.006329638883471489, + 0.016339363530278206, + 0.008864718489348888, + 0.009061777032911777, + 0.0077570704743266106, + 0.011175369843840599, + 0.0040862588211894035, + 0.006886166520416737, + 0.022410353645682335, + 0.015339967794716358, + 0.15075145661830902, + 0.025356831029057503, + 0.017706619575619698, + 0.019230064004659653, + 0.014650190249085426, + 0.1362542062997818, + 0.006547612603753805, + 0.006849979516118765, + 0.005914030596613884, + 0.005780549719929695, + 0.006041325628757477, + 0.006871369201689959, + 0.00505148246884346, + 0.00726688839495182, + 0.003781068604439497, + 0.002989400178194046, + 0.005599259398877621, + 0.003394842380657792, + 0.0035348169039934874, + 0.004226004704833031, + 0.0025993159506469965, + 0.004038845654577017, + 0.003522232873365283, + 0.002309722127392888, + 0.00391069520264864, + 0.00263199838809669, + 0.0026075306814163923, + 0.002473133848980069, + 0.0021458356641232967, + 0.002212093211710453, + 0.0014800684293732047, + 0.0027051642537117004, + 0.004690438508987427, + 0.004562693182379007, + 0.0018884370801970363, + 0.001624589553102851, + 0.0025846175849437714, + 0.002523475792258978, + 0.007215008605271578, + 0.006851429119706154, + 0.013447985984385014, + 0.011887227185070515, + 0.012365876697003841, + 0.006816819775849581, + 0.013625272549688816, + 0.012920141220092773, + 0.008213118650019169, + 0.008227656595408916, + 0.007050223182886839, + 0.005731950514018536, + 0.01628461293876171, + 0.00818430632352829, + 0.012806611135601997, + 0.007770234253257513 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 8 + ], + "selected_reward": -0.048670560121536255 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 in cinematic slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject a woman in her late twenties wearing a light yellow sleeveless shirt and black shorts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman's hair wet, tangled, and plastered to her flushed crimson skin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a smooth low-angle tracking shot synchronized with the runner's final stride?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are visible beads of sweat shown spraying across the face and hair with droplets arcing 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": [ + -0.2521195411682129 + ], + "opposite_logprob_tokens": [ + -1.502119541168213 + ], + "logprob_reward": -0.2521195411682129, + "yes_no_margin_reward": 1.25, + "token_credit_reward": -0.2521195411682129, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06816386431455612, + 0.03949475288391113, + 0.09502900391817093, + 0.4931384325027466, + 0.033214811235666275, + 0.05122942104935646, + 0.030134299769997597, + 0.0835321694612503, + 0.04518432915210724, + 0.06087891384959221 + ], + "credit_map": [ + 0.008250884711742401, + 0.005225490313023329, + 0.013313320465385914, + 0.005531372502446175, + 0.010196774266660213, + 0.006921287160366774, + 0.006618894170969725, + 0.012105845846235752, + 0.0033067145850509405, + 0.0091233616694808, + 0.005191992502659559, + 0.002765919081866741, + 0.0038526097778230906, + 0.00290864915587008, + 0.0058344206772744656, + 0.006511087063699961, + 0.009573802351951599, + 0.02319897897541523, + 0.014520211145281792, + 0.013448142446577549, + 0.007446160074323416, + 0.012205836363136768, + 0.005230164621025324, + 0.009405704215168953, + 0.032088927924633026, + 0.023668577894568443, + 0.1864243447780609, + 0.03410422429442406, + 0.02153543010354042, + 0.019840795546770096, + 0.01737069897353649, + 0.15810541808605194, + 0.004783218260854483, + 0.0036341703962534666, + 0.002860786160454154, + 0.005649893544614315, + 0.004128263797610998, + 0.0032459089998155832, + 0.003999363165348768, + 0.004913207609206438, + 0.004488277714699507, + 0.009642777033150196, + 0.008403787389397621, + 0.004053833894431591, + 0.004385744221508503, + 0.00955270417034626, + 0.0038322722539305687, + 0.006870023440569639, + 0.0036995771806687117, + 0.0025915689766407013, + 0.0058808885514736176, + 0.004692274611443281, + 0.0035504610277712345, + 0.0034547985997051, + 0.0030225215014070272, + 0.0032422090880572796, + 0.008606791496276855, + 0.010744960978627205, + 0.0217899139970541, + 0.015693265944719315, + 0.0064840735867619514, + 0.00615649251267314, + 0.006219109520316124, + 0.00783755537122488, + 0.0027984350454062223, + 0.00377495214343071, + 0.010273762978613377, + 0.006859241519123316, + 0.005961275659501553, + 0.005148126743733883, + 0.0040320721454918385, + 0.006336461286991835, + 0.006148802116513252, + 0.006278632208704948, + 0.007162931375205517, + 0.004394039046019316, + 0.014256652444601059, + 0.005981950554996729, + 0.009135724045336246, + 0.007520181592553854 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 2, + 7 + ], + "selected_reward": -0.2521195411682129 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 in cinematic slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject a woman in her late twenties wearing a light yellow sleeveless shirt and black shorts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman's hair wet, tangled, and plastered to her flushed crimson skin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a smooth low-angle tracking shot synchronized with the runner's final stride?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are visible beads of sweat shown spraying across the face and hair with droplets arcing 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": [ + -0.05496637150645256 + ], + "opposite_logprob_tokens": [ + -2.929966449737549 + ], + "logprob_reward": -0.05496637150645256, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05496637150645256, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23510678112506866, + 0.07745400816202164, + 0.12170910090208054, + 0.059374116361141205, + 0.052901916205883026, + 0.0643724799156189, + 0.03381548449397087, + 0.14712557196617126, + 0.14776860177516937, + 0.06037190183997154 + ], + "credit_map": [ + 0.02716643176972866, + 0.025299228727817535, + 0.04218923673033714, + 0.03971856087446213, + 0.021470192819833755, + 0.020116768777370453, + 0.02058405615389347, + 0.03856232017278671, + 0.011355050839483738, + 0.007412679493427277, + 0.006009294651448727, + 0.0055336919613182545, + 0.005649402737617493, + 0.007126896642148495, + 0.011913412250578403, + 0.02245357632637024, + 0.011771157383918762, + 0.027942920103669167, + 0.019341081380844116, + 0.01804751344025135, + 0.009199066087603569, + 0.013771772384643555, + 0.0074998196214437485, + 0.014135776087641716, + 0.010088542476296425, + 0.013136584311723709, + 0.009644175879657269, + 0.0046626124531030655, + 0.006693203467875719, + 0.00404611648991704, + 0.004041178617626429, + 0.007061705458909273, + 0.007668641861528158, + 0.006224407814443111, + 0.006812634412199259, + 0.006764858961105347, + 0.006004457361996174, + 0.0083570322021842, + 0.005011179484426975, + 0.006058701314032078, + 0.007133680861443281, + 0.005183299537748098, + 0.004917519632726908, + 0.008466950617730618, + 0.01044753473252058, + 0.011345766484737396, + 0.008796912617981434, + 0.008080817759037018, + 0.0036757502239197493, + 0.003364223288372159, + 0.004836945794522762, + 0.005164136178791523, + 0.004069556016474962, + 0.00402491819113493, + 0.004712441936135292, + 0.003967513795942068, + 0.010011691600084305, + 0.01836794801056385, + 0.03664884716272354, + 0.03049539588391781, + 0.01054195687174797, + 0.01245550811290741, + 0.012383845634758472, + 0.016220377758145332, + 0.012933486141264439, + 0.014440339989960194, + 0.022004015743732452, + 0.018004441633820534, + 0.017375536262989044, + 0.025230994448065758, + 0.01939563825726509, + 0.018384166061878204, + 0.0064630755223333836, + 0.007613363675773144, + 0.005235017277300358, + 0.004983352497220039, + 0.008423207327723503, + 0.010786991566419601, + 0.008406119421124458, + 0.008460775949060917 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 7 + ], + "selected_reward": -0.05496637150645256 + } + ], + "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/0000121/group_2/index_tensor([2891, 2891])_group_2_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000121/group_2/index_tensor([2891, 2891])_group_2_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..568464cec057fda0706cf0e2d25c07e0803b308a --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000121/group_2/index_tensor([2891, 2891])_group_2_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "static, extreme close-up from above captures a polished chrome disc magnet, resting motionless on a perfectly smooth, dark, non-reflective surface. From the left edge of the frame, a second polished chrome magnet (a rectangular rod) glides smoothly and deliberately into view. Before any physical contact is made, its invisible magnetic force acts on the disc, causing it to tremble for a microsecond before gliding swiftly and frictionlessly across the vast expanse of darkness. The rod magnet continues its steady advance, maintaining a precise, constant distance from the fleeing disc, its movements coordinating with the other's swift escape. Bright, diffuse studio lighting from above casts clean, elongated highlights across the mirror-like surfaces of both chrome objects. Their perfect forms contrast sharply against the matte black background, creating a minimalist and hypnotic atmosphere of quiet, powerful physical forces at work.", + "seed": 27869, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000121/group_2/index_2891_seed_27869_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": -0.10397846885025501, + "vlm_logprob": -0.10397846885025501, + "vlm_yes_no_margin": 3.2749999880790712, + "vlm_token_credit": -0.10397846885025501, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 an extreme close-up from above?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a polished chrome disc magnet visible resting on a smooth, dark, non-reflective surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a second polished chrome rectangular rod magnet glide into the frame from the left edge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the disc magnet tremble briefly before gliding swiftly away from the approaching rod magnet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do bright, diffuse studio lights from above cast clean, elongated highlights on the chrome objects?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.4165518879890442 + ], + "opposite_logprob_tokens": [ + -1.1665518283843994 + ], + "logprob_reward": -0.4165518879890442, + "yes_no_margin_reward": 0.7499999403953552, + "token_credit_reward": -0.4165518879890442, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3115168809890747, + 0.10127022117376328, + 0.10720488429069519, + 0.025566043332219124, + 0.04377444460988045, + 0.09608194977045059, + 0.163619264960289, + 0.07249141484498978, + 0.02267180196940899, + 0.05580311268568039 + ], + "credit_map": [ + 0.02690713293850422, + 0.031364552676677704, + 0.020645303651690483, + 0.0450003556907177, + 0.03914681449532509, + 0.07570003718137741, + 0.03509389981627464, + 0.03765879198908806, + 0.008502655662596226, + 0.013730809092521667, + 0.006727845408022404, + 0.011644961312413216, + 0.01402344461530447, + 0.018448173999786377, + 0.013615427538752556, + 0.01457690354436636, + 0.006987808272242546, + 0.01132430974394083, + 0.010829403065145016, + 0.015113367699086666, + 0.008748785592615604, + 0.02645786851644516, + 0.013531937263906002, + 0.014211408793926239, + 0.001983128022402525, + 0.0026426080148667097, + 0.0038942985702306032, + 0.0038558165542781353, + 0.0027500346768647432, + 0.003267597872763872, + 0.002682534046471119, + 0.004490027204155922, + 0.00546302879229188, + 0.0022468778770416975, + 0.009389215148985386, + 0.006521217525005341, + 0.004183073062449694, + 0.004368032328784466, + 0.00588740361854434, + 0.00571559788659215, + 0.011458673514425755, + 0.008298095315694809, + 0.015965647995471954, + 0.013376038521528244, + 0.014972114004194736, + 0.009692282415926456, + 0.010880605317652225, + 0.011438501998782158, + 0.02169920690357685, + 0.023175960406661034, + 0.02198052778840065, + 0.024377672001719475, + 0.015030446462333202, + 0.019475596025586128, + 0.020375259220600128, + 0.017504600808024406, + 0.006830040365457535, + 0.007856207899749279, + 0.009893105365335941, + 0.011374888010323048, + 0.009035511873662472, + 0.006117453332990408, + 0.013989027589559555, + 0.007395182736217976, + 0.002990370150655508, + 0.0022756895050406456, + 0.003212806535884738, + 0.0023407761473208666, + 0.002721396740525961, + 0.004717130213975906, + 0.0019597464706748724, + 0.0024538871366530657, + 0.007949604652822018, + 0.00420002918690443, + 0.007029023952782154, + 0.007340693846344948, + 0.004122217185795307, + 0.015958333387970924, + 0.004383280873298645, + 0.004819932859390974 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 2 + ], + "selected_reward": -0.4165518879890442 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 an extreme close-up from above?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a polished chrome disc magnet visible resting on a smooth, dark, non-reflective surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a second polished chrome rectangular rod magnet glide into the frame from the left edge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the disc magnet tremble briefly before gliding swiftly away from the approaching rod magnet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do bright, diffuse studio lights from above cast clean, elongated highlights on the chrome objects?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01424113567918539 + ], + "opposite_logprob_tokens": [ + -4.2642412185668945 + ], + "logprob_reward": -0.01424113567918539, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.01424113567918539, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20675991475582123, + 0.12706084549427032, + 0.11673140525817871, + 0.18995723128318787, + 0.08086271584033966, + 0.07653944194316864, + 0.0992853119969368, + 0.03843275457620621, + 0.017494600266218185, + 0.046875759959220886 + ], + "credit_map": [ + 0.01569456234574318, + 0.021354224532842636, + 0.012692050077021122, + 0.025708982720971107, + 0.022829486057162285, + 0.05468107387423515, + 0.023793688043951988, + 0.030005857348442078, + 0.008638576604425907, + 0.02024861052632332, + 0.009395561181008816, + 0.017207562923431396, + 0.017725450918078423, + 0.02330152690410614, + 0.013807013630867004, + 0.016736555844545364, + 0.0046682581305503845, + 0.007410128600895405, + 0.012406298890709877, + 0.018088292330503464, + 0.006142237689346075, + 0.03375791758298874, + 0.01820450648665428, + 0.01605377532541752, + 0.014023836702108383, + 0.016250409185886383, + 0.02623736672103405, + 0.025615857914090157, + 0.010544161312282085, + 0.02067015878856182, + 0.0226668119430542, + 0.05394866317510605, + 0.003848924534395337, + 0.0019080752972513437, + 0.030148979276418686, + 0.009186393581330776, + 0.003560912562534213, + 0.0032499595545232296, + 0.005616354756057262, + 0.023343121632933617, + 0.007849152199923992, + 0.005053109023720026, + 0.015425738878548145, + 0.01226221676915884, + 0.007305650040507317, + 0.007876792922616005, + 0.01019301638007164, + 0.010573775507509708, + 0.015432745218276978, + 0.011775444261729717, + 0.012362450361251831, + 0.018625533208251, + 0.0069562471471726894, + 0.010822393000125885, + 0.014109397307038307, + 0.009201109409332275, + 0.005251438822597265, + 0.005868986714631319, + 0.004882515873759985, + 0.006241713184863329, + 0.00474532600492239, + 0.0038324692286551, + 0.0034327164757996798, + 0.004177593160420656, + 0.0020109612960368395, + 0.001908581587485969, + 0.002453539287671447, + 0.001919221947900951, + 0.0023348203394562006, + 0.003122233785688877, + 0.0018972910474985838, + 0.0018479526042938232, + 0.004347712267190218, + 0.00402579503133893, + 0.005676012486219406, + 0.007489646784961224, + 0.0033145116176456213, + 0.012644847854971886, + 0.004344257526099682, + 0.005032982211560011 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 1 + ], + "selected_reward": -0.01424113567918539 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle an extreme close-up from above?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a polished chrome disc magnet visible resting on a smooth, dark, non-reflective surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a second polished chrome rectangular rod magnet glide into the frame from the left edge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the disc magnet tremble briefly before gliding swiftly away from the approaching rod magnet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do bright, diffuse studio lights from above cast clean, elongated highlights on the chrome objects?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03370605409145355 + ], + "opposite_logprob_tokens": [ + -3.408705949783325 + ], + "logprob_reward": -0.03370605409145355, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03370605409145355, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0771103948354721, + 0.07255345582962036, + 0.04813910648226738, + 0.060795631259679794, + 0.20246468484401703, + 0.05867254361510277, + 0.06094891205430031, + 0.21682940423488617, + 0.10718098282814026, + 0.09530491381883621 + ], + "credit_map": [ + 0.009452708065509796, + 0.011061725206673145, + 0.005869921296834946, + 0.009230628609657288, + 0.009724650532007217, + 0.013157196342945099, + 0.00871646124869585, + 0.009897089563310146, + 0.004478259477764368, + 0.010050787590444088, + 0.004878491163253784, + 0.007689258083701134, + 0.011469131335616112, + 0.01572006568312645, + 0.007857110351324081, + 0.010410342365503311, + 0.003343252232298255, + 0.005417913664132357, + 0.005694894585758448, + 0.0071610030718147755, + 0.005026220344007015, + 0.006754742935299873, + 0.007955573499202728, + 0.0067855012603104115, + 0.005316163878887892, + 0.006051416974514723, + 0.008263461291790009, + 0.008075673133134842, + 0.0060864523984491825, + 0.006827759090811014, + 0.006440683733671904, + 0.013734017498791218, + 0.009665871039032936, + 0.005990662612020969, + 0.07637539505958557, + 0.017263298854231834, + 0.007845008745789528, + 0.009419210255146027, + 0.008623622357845306, + 0.06728161126375198, + 0.013509788550436497, + 0.004381346050649881, + 0.00887544360011816, + 0.009196393191814423, + 0.006319960113614798, + 0.0049535371363162994, + 0.005606350023299456, + 0.005829721223562956, + 0.00901615060865879, + 0.011103230528533459, + 0.008057170547544956, + 0.00666961632668972, + 0.005451707169413567, + 0.006112988572567701, + 0.005109748803079128, + 0.009428296238183975, + 0.019081348553299904, + 0.019871210679411888, + 0.029153143987059593, + 0.0371817983686924, + 0.022976526990532875, + 0.023602459579706192, + 0.04488090053200722, + 0.02008199691772461, + 0.011396773159503937, + 0.011240560561418533, + 0.0164962038397789, + 0.01815042272210121, + 0.017213275656104088, + 0.013578047975897789, + 0.009809465147554874, + 0.009296225383877754, + 0.014092272147536278, + 0.006835957523435354, + 0.016451919451355934, + 0.013258984312415123, + 0.004819775931537151, + 0.02120071090757847, + 0.007620377000421286, + 0.01102491095662117 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 4, + 8 + ], + "selected_reward": -0.03370605409145355 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 an extreme close-up from above?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a polished chrome disc magnet visible resting on a smooth, dark, non-reflective surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a second polished chrome rectangular rod magnet glide into the frame from the left edge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the disc magnet tremble briefly before gliding swiftly away from the approaching rod magnet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do bright, diffuse studio lights from above cast clean, elongated highlights on the chrome objects?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04865364357829094 + ], + "opposite_logprob_tokens": [ + -3.0486536026000977 + ], + "logprob_reward": -0.04865364357829094, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04865364357829094, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10441894829273224, + 0.06243094056844711, + 0.06052876636385918, + 0.09338383376598358, + 0.20325317978858948, + 0.022237064316868782, + 0.06821295619010925, + 0.1403626799583435, + 0.04111934080719948, + 0.20405228435993195 + ], + "credit_map": [ + 0.015264835208654404, + 0.00824733916670084, + 0.009869391098618507, + 0.01169188879430294, + 0.01110099907964468, + 0.02091110870242119, + 0.013768301345407963, + 0.013565074652433395, + 0.0041468036361038685, + 0.009016172029078007, + 0.006246981211006641, + 0.007925724610686302, + 0.006817439571022987, + 0.010087351314723492, + 0.009149862453341484, + 0.009040603414177895, + 0.004634563345462084, + 0.006533996667712927, + 0.0074757118709385395, + 0.006630681920796633, + 0.006495445501059294, + 0.009242023341357708, + 0.008172009140253067, + 0.011344333179295063, + 0.00656015332788229, + 0.010387216694653034, + 0.013721367344260216, + 0.014052928425371647, + 0.008528020232915878, + 0.011997855268418789, + 0.011112645268440247, + 0.017023639753460884, + 0.011325865983963013, + 0.006805796641856432, + 0.06684666872024536, + 0.023460887372493744, + 0.010267885401844978, + 0.010026616975665092, + 0.015930943191051483, + 0.058588508516550064, + 0.003199442755430937, + 0.0024546952918171883, + 0.0031858079601079226, + 0.003118787659332156, + 0.0023358094040304422, + 0.0027267609257251024, + 0.002403423422947526, + 0.0028123341035097837, + 0.010371128097176552, + 0.011352078057825565, + 0.007810419425368309, + 0.007983111776411533, + 0.006207502447068691, + 0.01041057426482439, + 0.006787667516618967, + 0.007290470879524946, + 0.013598939403891563, + 0.0158066563308239, + 0.0179985873401165, + 0.020874997600913048, + 0.014356247149407864, + 0.014413481578230858, + 0.028485925868153572, + 0.014827825129032135, + 0.00478043407201767, + 0.004291757009923458, + 0.005414047744125128, + 0.005808140616863966, + 0.005084359087049961, + 0.00782355573028326, + 0.004408659879118204, + 0.0035083831753581762, + 0.02489423006772995, + 0.014543081633746624, + 0.026316115632653236, + 0.029646052047610283, + 0.013904722407460213, + 0.058091070502996445, + 0.01733349822461605, + 0.019323496147990227 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 4, + 7 + ], + "selected_reward": -0.04865364357829094 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 an extreme close-up from above?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a polished chrome disc magnet visible resting on a smooth, dark, non-reflective surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a second polished chrome rectangular rod magnet glide into the frame from the left edge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the disc magnet tremble briefly before gliding swiftly away from the approaching rod magnet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do bright, diffuse studio lights from above cast clean, elongated highlights on the chrome objects?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006739622913300991 + ], + "opposite_logprob_tokens": [ + -5.006739616394043 + ], + "logprob_reward": -0.006739622913300991, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006739622913300991, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3288010060787201, + 0.07680308073759079, + 0.18978174030780792, + 0.07722001522779465, + 0.11588642746210098, + 0.032973941415548325, + 0.0724458247423172, + 0.04487486928701401, + 0.024921633303165436, + 0.036291465163230896 + ], + "credit_map": [ + 0.01776967942714691, + 0.0225943922996521, + 0.01925649121403694, + 0.03388358652591705, + 0.02509976364672184, + 0.10790096968412399, + 0.03976764157414436, + 0.06252846121788025, + 0.00467650406062603, + 0.009839506819844246, + 0.0061383298598229885, + 0.014674692414700985, + 0.009711957536637783, + 0.010633409023284912, + 0.012197260744869709, + 0.008931412361562252, + 0.009414417669177055, + 0.017522651702165604, + 0.015424815006554127, + 0.02548176608979702, + 0.024783341214060783, + 0.03914608061313629, + 0.028875945135951042, + 0.029132720082998276, + 0.004866424947977066, + 0.009203179739415646, + 0.00997474230825901, + 0.009635667316615582, + 0.006438967771828175, + 0.013717297464609146, + 0.008773384615778923, + 0.014610356651246548, + 0.006605830974876881, + 0.003198173362761736, + 0.03710563853383064, + 0.01418381929397583, + 0.006170323118567467, + 0.005919288843870163, + 0.011764850467443466, + 0.03093850053846836, + 0.00529597420245409, + 0.0034133458975702524, + 0.0063795107416808605, + 0.00401998171582818, + 0.0030441144481301308, + 0.0037515084259212017, + 0.0032549069728702307, + 0.0038145992439240217, + 0.01372931245714426, + 0.00974066462367773, + 0.009557576850056648, + 0.008164310827851295, + 0.005422740243375301, + 0.009969402104616165, + 0.008009253069758415, + 0.007852574810385704, + 0.005188603885471821, + 0.005182444117963314, + 0.006494601722806692, + 0.007321120239794254, + 0.004484185483306646, + 0.006912724580615759, + 0.004488035570830107, + 0.00480315089225769, + 0.002709574531763792, + 0.0026552064809948206, + 0.0037627718411386013, + 0.00504843657836318, + 0.0026471305172890425, + 0.003281937912106514, + 0.0026193372905254364, + 0.002197239315137267, + 0.002869610209017992, + 0.002866979455575347, + 0.00559812318533659, + 0.007193435914814472, + 0.002257531275972724, + 0.006076416932046413, + 0.005133282393217087, + 0.004296083003282547 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 4 + ], + "selected_reward": -0.006739622913300991 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, extreme close-up from above captures a polished chrome disc magnet, resting motionless on a perfectly smooth, dark, non-reflective surface. From the left edge of the frame, a second polished chrome magnet (a rectangular rod) glides smoothly and deliberately into view. Before any physical contact is made, its invisible magnetic force acts on the disc, causing it to tremble for a microsecond before gliding swiftly and frictionlessly across the vast expanse of darkness. The rod magnet continues its steady advance, maintaining a precise, constant distance from the fleeing disc, its movements coordinating with the other's swift escape. Bright, diffuse studio lighting from above casts clean, elongated highlights across the mirror-like surfaces of both chrome objects. Their perfect forms contrast sharply against the matte black background, creating a minimalist and hypnotic atmosphere of quiet, powerful physical forces at work.", + "seed": 27870, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000121/group_2/index_2891_seed_27870_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": -0.0750770028680563, + "vlm_logprob": -0.0750770028680563, + "vlm_yes_no_margin": 3.3, + "vlm_token_credit": -0.0750770028680563, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 an extreme close-up from above?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a polished chrome disc magnet visible resting on a smooth, dark, non-reflective surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a second polished chrome rectangular rod magnet glide into the frame from the left edge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the disc magnet tremble briefly before gliding swiftly away from the approaching rod magnet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do bright, diffuse studio lights from above cast clean, elongated highlights on the chrome objects?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.27664798498153687 + ], + "opposite_logprob_tokens": [ + -1.5266480445861816 + ], + "logprob_reward": -0.27664798498153687, + "yes_no_margin_reward": 1.25, + "token_credit_reward": -0.27664798498153687, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04767995700240135, + 0.02925325185060501, + 0.01616484485566616, + 0.02694101259112358, + 0.025627847760915756, + 0.13016179203987122, + 0.1348678320646286, + 0.20309707522392273, + 0.31935280561447144, + 0.06685363501310349 + ], + "credit_map": [ + 0.00469618383795023, + 0.004085930995643139, + 0.006272703874856234, + 0.0024262580554932356, + 0.0036509826313704252, + 0.011670813895761967, + 0.0037653690669685602, + 0.01111171767115593, + 0.0023958939127624035, + 0.003419750602915883, + 0.002969892229884863, + 0.002523879986256361, + 0.004223030060529709, + 0.005433776881545782, + 0.003634001826867461, + 0.004653025418519974, + 0.0013640616089105606, + 0.0013605039566755295, + 0.001532857771962881, + 0.0014447268331423402, + 0.0023633483797311783, + 0.003593801986426115, + 0.001962788635864854, + 0.002542756497859955, + 0.004549662582576275, + 0.0017973078647628427, + 0.004852762445807457, + 0.002255410887300968, + 0.0014048932353034616, + 0.004047827795147896, + 0.00269304309040308, + 0.005340105388313532, + 0.0040338789112865925, + 0.004012465011328459, + 0.0037405870389193296, + 0.00522638438269496, + 0.0015298156067728996, + 0.0009687754209153354, + 0.0027127880603075027, + 0.0034031549002975225, + 0.006630694959312677, + 0.03566952794790268, + 0.046476222574710846, + 0.011287105269730091, + 0.008419998921453953, + 0.005139178596436977, + 0.010882691480219364, + 0.005656364373862743, + 0.0056510851718485355, + 0.03056809864938259, + 0.01552003063261509, + 0.01122997235506773, + 0.04195179417729378, + 0.01220184750854969, + 0.012314563617110252, + 0.005430441815406084, + 0.012561571784317493, + 0.025571176782250404, + 0.07468739151954651, + 0.022836335003376007, + 0.0272834450006485, + 0.01567656360566616, + 0.014921819791197777, + 0.009558754973113537, + 0.028742937371134758, + 0.03833220526576042, + 0.07553018629550934, + 0.02751549333333969, + 0.028632350265979767, + 0.03954169899225235, + 0.01894260197877884, + 0.06211533769965172, + 0.004535803571343422, + 0.011460182256996632, + 0.02493337169289589, + 0.007856392301619053, + 0.006103400141000748, + 0.005944832693785429, + 0.004126017913222313, + 0.0018936393316835165 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 6 + ], + "selected_reward": -0.27664798498153687 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 an extreme close-up from above?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a polished chrome disc magnet visible resting on a smooth, dark, non-reflective surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a second polished chrome rectangular rod magnet glide into the frame from the left edge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the disc magnet tremble briefly before gliding swiftly away from the approaching rod magnet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do bright, diffuse studio lights from above cast clean, elongated highlights on the chrome objects?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01823740452528 + ], + "opposite_logprob_tokens": [ + -4.018237590789795 + ], + "logprob_reward": -0.01823740452528, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.01823740452528, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03706032782793045, + 0.08590546995401382, + 0.010416748933494091, + 0.01883259229362011, + 0.009056036360561848, + 0.026114093139767647, + 0.538536787033081, + 0.16322818398475647, + 0.05037274211645126, + 0.06047705560922623 + ], + "credit_map": [ + 0.004097406752407551, + 0.003525596810504794, + 0.005609978921711445, + 0.002396572381258011, + 0.004129039589315653, + 0.007624657824635506, + 0.00237556011416018, + 0.007301516365259886, + 0.007964911870658398, + 0.01132884994149208, + 0.007036002352833748, + 0.0067540002055466175, + 0.011897833086550236, + 0.015016097575426102, + 0.011282791383564472, + 0.014624984934926033, + 0.0008506140438839793, + 0.0008915209909901023, + 0.0009628938278183341, + 0.0008430678281001747, + 0.0016382851172238588, + 0.0022197552025318146, + 0.0016557942144572735, + 0.0013548174174502492, + 0.0025490904226899147, + 0.0012164643267169595, + 0.004158674273639917, + 0.001545637845993042, + 0.0011241990141570568, + 0.0026470189914107323, + 0.002377874683588743, + 0.0032136335503309965, + 0.0014781045028939843, + 0.0013408174272626638, + 0.0012987805530428886, + 0.0013232567580416799, + 0.000677004805766046, + 0.0003848305786959827, + 0.001022761920467019, + 0.0015304795233532786, + 0.0018834962975233793, + 0.003871450899168849, + 0.007251530420035124, + 0.0029493821784853935, + 0.003095332533121109, + 0.002234575105831027, + 0.002398757031187415, + 0.0024295691400766373, + 0.04199903830885887, + 0.12252743542194366, + 0.041268616914749146, + 0.04669239744544029, + 0.10155802220106125, + 0.06251704692840576, + 0.03833348676562309, + 0.08364073932170868, + 0.010941063985228539, + 0.030310269445180893, + 0.060526955872774124, + 0.019501378759741783, + 0.013760560192167759, + 0.01352645829319954, + 0.010016001760959625, + 0.004645482636988163, + 0.005543283652514219, + 0.006680525839328766, + 0.012075891718268394, + 0.00441078282892704, + 0.0032346646767109632, + 0.006006245035678148, + 0.003260793862864375, + 0.00916055403649807, + 0.004182333592325449, + 0.010535565204918385, + 0.023196080699563026, + 0.006927208509296179, + 0.005516136530786753, + 0.004729811102151871, + 0.003667452372610569, + 0.0017224670154973865 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 7, + 1 + ], + "selected_reward": -0.01823740452528 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 an extreme close-up from above?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a polished chrome disc magnet visible resting on a smooth, dark, non-reflective surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a second polished chrome rectangular rod magnet glide into the frame from the left edge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the disc magnet tremble briefly before gliding swiftly away from the approaching rod magnet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do bright, diffuse studio lights from above cast clean, elongated highlights on the chrome objects?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02635815367102623 + ], + "opposite_logprob_tokens": [ + -3.651358127593994 + ], + "logprob_reward": -0.02635815367102623, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.02635815367102623, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03142547607421875, + 0.031990233808755875, + 0.028877196833491325, + 0.0183356162160635, + 0.008256451226770878, + 0.05268135294318199, + 0.5296160578727722, + 0.11678408086299896, + 0.13751068711280823, + 0.04452290013432503 + ], + "credit_map": [ + 0.002391889225691557, + 0.003452450269833207, + 0.003352898173034191, + 0.0024491078220307827, + 0.004346072673797607, + 0.006238043308258057, + 0.0026724233757704496, + 0.006522585172206163, + 0.0032230266369879246, + 0.005209166090935469, + 0.0034761035349220037, + 0.0027883024886250496, + 0.004127016291022301, + 0.005286503117531538, + 0.002727728569880128, + 0.005152382887899876, + 0.002624945482239127, + 0.002072250237688422, + 0.003442036220803857, + 0.002441209740936756, + 0.0033920006826519966, + 0.00561206229031086, + 0.004181860014796257, + 0.005110828671604395, + 0.0025325617752969265, + 0.001130849588662386, + 0.003469520015642047, + 0.0011400956427678466, + 0.0010009543038904667, + 0.002210464561358094, + 0.0027940701693296432, + 0.004057097714394331, + 0.0011869630543515086, + 0.0011529922485351562, + 0.001021496020257473, + 0.001402336172759533, + 0.0006153323920443654, + 0.000502153707202524, + 0.0013299394631758332, + 0.0010452374117448926, + 0.0037370705977082253, + 0.01213433500379324, + 0.018720058724284172, + 0.004711318761110306, + 0.003265807870775461, + 0.0028934103902429342, + 0.003490698290988803, + 0.00372864818200469, + 0.04715406149625778, + 0.11048562824726105, + 0.044613514095544815, + 0.04677548632025719, + 0.07844226807355881, + 0.06643540412187576, + 0.03695144131779671, + 0.09875822067260742, + 0.007733807433396578, + 0.022856852039694786, + 0.041528116911649704, + 0.01379870530217886, + 0.011582877486944199, + 0.009733928367495537, + 0.0063807787373661995, + 0.0031690082978457212, + 0.01346854493021965, + 0.020291350781917572, + 0.02310189977288246, + 0.011671029962599277, + 0.015206865034997463, + 0.01723330281674862, + 0.008755470626056194, + 0.027782201766967773, + 0.0029915382619947195, + 0.008318443782627583, + 0.015407795086503029, + 0.004983821418136358, + 0.003964739851653576, + 0.0035772097762674093, + 0.0030960252042859793, + 0.002183321863412857 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 8, + 7 + ], + "selected_reward": -0.02635815367102623 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 an extreme close-up from above?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a polished chrome disc magnet visible resting on a smooth, dark, non-reflective surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a second polished chrome rectangular rod magnet glide into the frame from the left edge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the disc magnet tremble briefly before gliding swiftly away from the approaching rod magnet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do bright, diffuse studio lights from above cast clean, elongated highlights on the chrome objects?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0430624857544899 + ], + "opposite_logprob_tokens": [ + -3.168062448501587 + ], + "logprob_reward": -0.0430624857544899, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.0430624857544899, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.020896565169095993, + 0.05027736350893974, + 0.02056012488901615, + 0.015076212584972382, + 0.01602444052696228, + 0.06115412339568138, + 0.5476500391960144, + 0.0904880091547966, + 0.15435102581977844, + 0.02352212741971016 + ], + "credit_map": [ + 0.0020505378488451242, + 0.001670211087912321, + 0.003666339674964547, + 0.0010207061422988772, + 0.0021668446715921164, + 0.004473304376006126, + 0.001394324586726725, + 0.0044542960822582245, + 0.005078193731606007, + 0.006388034205883741, + 0.0036525323521345854, + 0.003953791689127684, + 0.0061815353110432625, + 0.01009130384773016, + 0.00519119668751955, + 0.00974077545106411, + 0.0020245409104973078, + 0.0015467358753085136, + 0.003110131248831749, + 0.001923865987919271, + 0.0029612130019813776, + 0.004102897364646196, + 0.002595964353531599, + 0.002294776728376746, + 0.002699750941246748, + 0.0011147804325446486, + 0.002542366972193122, + 0.0014083198038861156, + 0.00106441683601588, + 0.0017327693058177829, + 0.00205364846624434, + 0.002460160292685032, + 0.0021939133293926716, + 0.0035384378861635923, + 0.0022712035570293665, + 0.0028085659723728895, + 0.001042209449224174, + 0.0004710980283562094, + 0.001458483631722629, + 0.0022405285853892565, + 0.003657967783510685, + 0.014736061915755272, + 0.017984021455049515, + 0.007358392234891653, + 0.004880464170128107, + 0.0036071266513317823, + 0.004623536951839924, + 0.004306550603359938, + 0.047381266951560974, + 0.11604342609643936, + 0.04465506970882416, + 0.048330407589673996, + 0.08804614096879959, + 0.06769048422574997, + 0.03796481713652611, + 0.09753844887018204, + 0.006966791581362486, + 0.01683880388736725, + 0.02682272158563137, + 0.009181838482618332, + 0.010592409409582615, + 0.008936489932239056, + 0.0053743342868983746, + 0.005774626042693853, + 0.014909558929502964, + 0.015500016510486603, + 0.03956979513168335, + 0.013299976475536823, + 0.013232025317847729, + 0.01877046748995781, + 0.009843937121331692, + 0.029225245118141174, + 0.0022773814853280783, + 0.004116023890674114, + 0.003970733378082514, + 0.0027037456165999174, + 0.0022432547993957996, + 0.002496804343536496, + 0.0017914875643327832, + 0.0039226943626999855 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 8, + 7 + ], + "selected_reward": -0.0430624857544899 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 an extreme close-up from above?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a polished chrome disc magnet visible resting on a smooth, dark, non-reflective surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a second polished chrome rectangular rod magnet glide into the frame from the left edge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the disc magnet tremble briefly before gliding swiftly away from the approaching rod magnet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do bright, diffuse studio lights from above cast clean, elongated highlights on the chrome objects?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"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": [ + 0.05861159414052963, + 0.05141178146004677, + 0.014955275692045689, + 0.018887514248490334, + 0.018130799755454063, + 0.08335091918706894, + 0.5153003334999084, + 0.07681650668382645, + 0.11299841105937958, + 0.049536846578121185 + ], + "credit_map": [ + 0.0047662630677223206, + 0.008520808070898056, + 0.007528640329837799, + 0.003469348419457674, + 0.006199268624186516, + 0.012107394635677338, + 0.00531908730044961, + 0.010700789280235767, + 0.005122185219079256, + 0.006966353859752417, + 0.0044139232486486435, + 0.004347096662968397, + 0.007348332554101944, + 0.0076251705177128315, + 0.008594281040132046, + 0.006994443945586681, + 0.001483920612372458, + 0.0011192664969712496, + 0.0013398334849625826, + 0.0018860538257285953, + 0.0022888770326972008, + 0.002754292683675885, + 0.0020797690376639366, + 0.0020032627508044243, + 0.0027190931141376495, + 0.0012256345944479108, + 0.0030349132139235735, + 0.002160854171961546, + 0.001294243847951293, + 0.0028068299870938063, + 0.002960716839879751, + 0.0026852295268326998, + 0.0025846741627901793, + 0.0028248007874935865, + 0.00286485836841166, + 0.003672829596325755, + 0.0012329372111707926, + 0.0007349825464189053, + 0.0017672537360340357, + 0.0024484656751155853, + 0.004441483877599239, + 0.016063546761870384, + 0.018505927175283432, + 0.01771850511431694, + 0.008155299350619316, + 0.004975644405931234, + 0.008988690562546253, + 0.004501826129853725, + 0.04263468459248543, + 0.1122155413031578, + 0.041255492717027664, + 0.0451614074409008, + 0.08917517960071564, + 0.061979785561561584, + 0.03616227209568024, + 0.08671600371599197, + 0.006102409213781357, + 0.01386946253478527, + 0.019351212307810783, + 0.00840920489281416, + 0.010560168884694576, + 0.00822377298027277, + 0.004499462898820639, + 0.005800819024443626, + 0.01014260109513998, + 0.019932067021727562, + 0.01664520986378193, + 0.009413676336407661, + 0.015874821692705154, + 0.01345742680132389, + 0.007138767745345831, + 0.020393850281834602, + 0.0015764891868457198, + 0.012276330962777138, + 0.0037339089903980494, + 0.004000099375844002, + 0.016875596717000008, + 0.0057163420133292675, + 0.004007509909570217, + 0.0013505751267075539 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 8, + 5 + ], + "selected_reward": -0.011078985407948494 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, extreme close-up from above captures a polished chrome disc magnet, resting motionless on a perfectly smooth, dark, non-reflective surface. From the left edge of the frame, a second polished chrome magnet (a rectangular rod) glides smoothly and deliberately into view. Before any physical contact is made, its invisible magnetic force acts on the disc, causing it to tremble for a microsecond before gliding swiftly and frictionlessly across the vast expanse of darkness. The rod magnet continues its steady advance, maintaining a precise, constant distance from the fleeing disc, its movements coordinating with the other's swift escape. Bright, diffuse studio lighting from above casts clean, elongated highlights across the mirror-like surfaces of both chrome objects. Their perfect forms contrast sharply against the matte black background, creating a minimalist and hypnotic atmosphere of quiet, powerful physical forces at work.", + "seed": 27871, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000121/group_2/index_2891_seed_27871_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": -0.0533874299377203, + "vlm_logprob": -0.0533874299377203, + "vlm_yes_no_margin": 3.55, + "vlm_token_credit": -0.0533874299377203, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 an extreme close-up from above?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a polished chrome disc magnet visible resting on a smooth, dark, non-reflective surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a second polished chrome rectangular rod magnet glide into the frame from the left edge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the disc magnet tremble briefly before gliding swiftly away from the approaching rod magnet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do bright, diffuse studio lights from above cast clean, elongated highlights on the chrome objects?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.17141106724739075 + ], + "opposite_logprob_tokens": [ + -2.0464110374450684 + ], + "logprob_reward": -0.17141106724739075, + "yes_no_margin_reward": 1.875, + "token_credit_reward": -0.17141106724739075, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14733995497226715, + 0.04560830071568489, + 0.053827572613954544, + 0.14593864977359772, + 0.08473584055900574, + 0.11864359676837921, + 0.11233358085155487, + 0.12114378809928894, + 0.1089990884065628, + 0.06142963841557503 + ], + "credit_map": [ + 0.02001381106674671, + 0.01182581391185522, + 0.012738906778395176, + 0.012956609018146992, + 0.0332721471786499, + 0.024565253406763077, + 0.0201642494648695, + 0.011803166940808296, + 0.002942211227491498, + 0.0027832165360450745, + 0.007537280675023794, + 0.0049254409968853, + 0.006642884574830532, + 0.006668725050985813, + 0.007893403060734272, + 0.006215137429535389, + 0.00513945147395134, + 0.0044656917452812195, + 0.008015792816877365, + 0.005538256373256445, + 0.005947737954556942, + 0.007929910905659199, + 0.00859821867197752, + 0.008192514069378376, + 0.013199965469539165, + 0.008748751133680344, + 0.022777561098337173, + 0.019956154748797417, + 0.00986107625067234, + 0.027277275919914246, + 0.017768675461411476, + 0.026349185034632683, + 0.008566625416278839, + 0.0075562093406915665, + 0.012128851376473904, + 0.02153853513300419, + 0.007447473704814911, + 0.007361569907516241, + 0.010912071913480759, + 0.009224509820342064, + 0.008789605461061, + 0.008228600025177002, + 0.022056668996810913, + 0.025347156450152397, + 0.009734348393976688, + 0.011088569648563862, + 0.017347315326333046, + 0.01605132967233658, + 0.006877121981233358, + 0.006842616479843855, + 0.018413810059428215, + 0.0309499129652977, + 0.007475488353520632, + 0.010882843285799026, + 0.01916116662323475, + 0.011730626225471497, + 0.006966313812881708, + 0.00905123446136713, + 0.02333161234855652, + 0.014725935645401478, + 0.007372166495770216, + 0.011656693182885647, + 0.032514121383428574, + 0.015525715425610542, + 0.006291274446994066, + 0.006149886641651392, + 0.017516160383820534, + 0.020615622401237488, + 0.008246343582868576, + 0.011079076677560806, + 0.021125879138708115, + 0.017974840477108955, + 0.0036094672977924347, + 0.0028798524290323257, + 0.009974672459065914, + 0.0068210125900805, + 0.005686786491423845, + 0.0045595490373671055, + 0.009440954774618149, + 0.01845734566450119 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 7 + ], + "selected_reward": -0.17141106724739075 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 an extreme close-up from above?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a polished chrome disc magnet visible resting on a smooth, dark, non-reflective surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a second polished chrome rectangular rod magnet glide into the frame from the left edge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the disc magnet tremble briefly before gliding swiftly away from the approaching rod magnet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do bright, diffuse studio lights from above cast clean, elongated highlights on the chrome objects?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011128854006528854 + ], + "opposite_logprob_tokens": [ + -4.511128902435303 + ], + "logprob_reward": -0.011128854006528854, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011128854006528854, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10977064073085785, + 0.054659679532051086, + 0.11867032200098038, + 0.35670483112335205, + 0.03382418677210808, + 0.07020276039838791, + 0.02542857825756073, + 0.07534662634134293, + 0.058163292706012726, + 0.09722903370857239 + ], + "credit_map": [ + 0.007953318767249584, + 0.009034417569637299, + 0.009428571909666061, + 0.011256711557507515, + 0.030445624142885208, + 0.01652718149125576, + 0.015968717634677887, + 0.009156105108559132, + 0.002434029011055827, + 0.0026961923576891422, + 0.005651054438203573, + 0.003437172621488571, + 0.011523478664457798, + 0.013908635824918747, + 0.007841652259230614, + 0.007167467847466469, + 0.011984182521700859, + 0.004655900876969099, + 0.024881279096007347, + 0.008700546808540821, + 0.00702765304595232, + 0.025607924908399582, + 0.01298875454813242, + 0.022824101150035858, + 0.0243793036788702, + 0.015722785145044327, + 0.035772550851106644, + 0.04476672038435936, + 0.013069914653897285, + 0.09528402239084244, + 0.02720772475004196, + 0.10050183534622192, + 0.002235815394669771, + 0.0030953206587582827, + 0.004710501525551081, + 0.0065727378241717815, + 0.003685454837977886, + 0.0033200744073837996, + 0.004757146816700697, + 0.00544714042916894, + 0.005553117487579584, + 0.005081633571535349, + 0.010279690846800804, + 0.011160184629261494, + 0.0076934522949159145, + 0.007042448036372662, + 0.012736245058476925, + 0.010655995458364487, + 0.001956401625648141, + 0.001923555275425315, + 0.004176175221800804, + 0.004821467213332653, + 0.0019175768829882145, + 0.0027157540898770094, + 0.004037476144731045, + 0.003880174597725272, + 0.0024474915117025375, + 0.0016825776547193527, + 0.02630852535367012, + 0.006551376078277826, + 0.002641653409227729, + 0.002605121349915862, + 0.02841634303331375, + 0.0046935430727899075, + 0.003327436512336135, + 0.0025142431259155273, + 0.007424553390592337, + 0.00994736049324274, + 0.0042572165839374065, + 0.005247736815363169, + 0.014688655734062195, + 0.010756091214716434, + 0.006955793127417564, + 0.007609743159264326, + 0.012674952857196331, + 0.014314311556518078, + 0.00755385123193264, + 0.011222956702113152, + 0.014826236292719841, + 0.022071203216910362 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 2, + 0 + ], + "selected_reward": -0.011128854006528854 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 an extreme close-up from above?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a polished chrome disc magnet visible resting on a smooth, dark, non-reflective surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a second polished chrome rectangular rod magnet glide into the frame from the left edge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the disc magnet tremble briefly before gliding swiftly away from the approaching rod magnet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do bright, diffuse studio lights from above cast clean, elongated highlights on the chrome objects?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04306294396519661 + ], + "opposite_logprob_tokens": [ + -3.168062925338745 + ], + "logprob_reward": -0.04306294396519661, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04306294396519661, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.028407752513885498, + 0.043781086802482605, + 0.01744130253791809, + 0.05129726976156235, + 0.02685091830790043, + 0.09937653690576553, + 0.1066066101193428, + 0.5937148928642273, + 0.015289864502847195, + 0.017233816906809807 + ], + "credit_map": [ + 0.00207012752071023, + 0.0030726708937436342, + 0.002943272003903985, + 0.003460417501628399, + 0.006065371446311474, + 0.003687411779537797, + 0.004947093315422535, + 0.0021613845601677895, + 0.0018296322086825967, + 0.0020558175165206194, + 0.003214664524421096, + 0.0027366483118385077, + 0.015192748978734016, + 0.01072564348578453, + 0.005333106033504009, + 0.002692823763936758, + 0.0017815008759498596, + 0.0009125586948357522, + 0.003247746266424656, + 0.0015214992454275489, + 0.0016589728184044361, + 0.002853770274668932, + 0.0026851219590753317, + 0.002780131995677948, + 0.0044938623905181885, + 0.0019862521439790726, + 0.008690654300153255, + 0.004222666844725609, + 0.001947938115336001, + 0.013867305591702461, + 0.0032996421214193106, + 0.012788943946361542, + 0.002338303253054619, + 0.002695018658414483, + 0.004796978551894426, + 0.006789828650653362, + 0.002015723381191492, + 0.0026146620512008667, + 0.002704694401472807, + 0.002895706333220005, + 0.008818506263196468, + 0.00633264658972621, + 0.01659131608903408, + 0.016506927087903023, + 0.008274814113974571, + 0.010511208325624466, + 0.014248287305235863, + 0.018092826008796692, + 0.006791879888623953, + 0.006295124068856239, + 0.018452638760209084, + 0.028285598382353783, + 0.007581866811960936, + 0.007968642748892307, + 0.021448036655783653, + 0.009782807901501656, + 0.01979753002524376, + 0.0128862876445055, + 0.19897465407848358, + 0.0456596314907074, + 0.019551334902644157, + 0.01834206096827984, + 0.23819689452648163, + 0.040306467562913895, + 0.0010082682128995657, + 0.0010328780626878142, + 0.0020021279342472553, + 0.001962503185495734, + 0.0020229220390319824, + 0.001830537454225123, + 0.0027740446384996176, + 0.0026565834414213896, + 0.0015401154523715377, + 0.00180173316039145, + 0.0020265665370970964, + 0.003135774051770568, + 0.0018055614782497287, + 0.002042816486209631, + 0.0019123879028484225, + 0.002968860324472189 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 6, + 5 + ], + "selected_reward": -0.04306294396519661 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 an extreme close-up from above?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a polished chrome disc magnet visible resting on a smooth, dark, non-reflective surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a second polished chrome rectangular rod magnet glide into the frame from the left edge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the disc magnet tremble briefly before gliding swiftly away from the approaching rod magnet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do bright, diffuse studio lights from above cast clean, elongated highlights on the chrome objects?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.033701904118061066 + ], + "opposite_logprob_tokens": [ + -3.4087018966674805 + ], + "logprob_reward": -0.033701904118061066, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.033701904118061066, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06829852610826492, + 0.07472162693738937, + 0.03295585885643959, + 0.11681615561246872, + 0.12668149173259735, + 0.03574744239449501, + 0.07095694541931152, + 0.39253395795822144, + 0.03218508139252663, + 0.04910290986299515 + ], + "credit_map": [ + 0.006791026797145605, + 0.0075727785006165504, + 0.006068026646971703, + 0.007562657352536917, + 0.010982700623571873, + 0.013785425573587418, + 0.009339104406535625, + 0.006196809932589531, + 0.002635651733726263, + 0.0023799792397767305, + 0.003664313117042184, + 0.0035210622008889914, + 0.028920244425535202, + 0.019896769896149635, + 0.008833381347358227, + 0.004870239179581404, + 0.003428627736866474, + 0.0030942221637815237, + 0.006621882319450378, + 0.0041644154116511345, + 0.002666245913133025, + 0.0044891019351780415, + 0.0031567560508847237, + 0.00533461244776845, + 0.009353283792734146, + 0.006508463062345982, + 0.015127000398933887, + 0.014188071712851524, + 0.0071635437197983265, + 0.027414556592702866, + 0.012401552870869637, + 0.024659698829054832, + 0.010283255949616432, + 0.009428794495761395, + 0.016919465735554695, + 0.023601621389389038, + 0.010923963040113449, + 0.012540717609226704, + 0.01935141533613205, + 0.02363225631415844, + 0.003215471515432, + 0.0034926841035485268, + 0.005261477082967758, + 0.006936744786798954, + 0.003729310119524598, + 0.002834521932527423, + 0.006464800331741571, + 0.0038124374113976955, + 0.004108203109353781, + 0.004257877357304096, + 0.01221177726984024, + 0.01845356449484825, + 0.004496533889323473, + 0.006033368408679962, + 0.014128012582659721, + 0.0072676148265600204, + 0.013112970627844334, + 0.006786038633435965, + 0.13537022471427917, + 0.030168989673256874, + 0.0124513553455472, + 0.012354020960628986, + 0.15647529065608978, + 0.02581508457660675, + 0.002202157862484455, + 0.001865818747319281, + 0.004416623152792454, + 0.004757474176585674, + 0.002564076567068696, + 0.005475618410855532, + 0.005553801078349352, + 0.005349516402930021, + 0.003042846452444792, + 0.0026432834565639496, + 0.009219975210726261, + 0.007351666688919067, + 0.004647483117878437, + 0.0037886917125433683, + 0.006895433645695448, + 0.01151353307068348 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 4, + 3 + ], + "selected_reward": -0.033701904118061066 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 an extreme close-up from above?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a polished chrome disc magnet visible resting on a smooth, dark, non-reflective surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a second polished chrome rectangular rod magnet glide into the frame from the left edge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the disc magnet tremble briefly before gliding swiftly away from the approaching rod magnet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do bright, diffuse studio lights from above cast clean, elongated highlights on the chrome objects?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007632380351424217 + ], + "opposite_logprob_tokens": [ + -4.882632255554199 + ], + "logprob_reward": -0.007632380351424217, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007632380351424217, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1502983272075653, + 0.0655977800488472, + 0.1002759113907814, + 0.20871685445308685, + 0.07310233265161514, + 0.08543303608894348, + 0.0698937401175499, + 0.12617774307727814, + 0.04413887858390808, + 0.07636537402868271 + ], + "credit_map": [ + 0.011137058027088642, + 0.009973577223718166, + 0.011438912712037563, + 0.01762557215988636, + 0.03787892684340477, + 0.01765672117471695, + 0.02855384536087513, + 0.016033723950386047, + 0.0035174984950572252, + 0.0034374333918094635, + 0.008907184936106205, + 0.00394444540143013, + 0.01933576725423336, + 0.012561499141156673, + 0.007981662638485432, + 0.005912285763770342, + 0.0065828184597194195, + 0.005308845546096563, + 0.018107585608959198, + 0.011550103314220905, + 0.007857877761125565, + 0.014835203997790813, + 0.01967652514576912, + 0.016356946900486946, + 0.01238009613007307, + 0.00677878875285387, + 0.020589804276823997, + 0.022045379504561424, + 0.006370988674461842, + 0.062053002417087555, + 0.015079389326274395, + 0.063419409096241, + 0.005310829728841782, + 0.006668924354016781, + 0.008655653335154057, + 0.01340776588767767, + 0.005285592284053564, + 0.00665220757946372, + 0.01177189126610756, + 0.015349467284977436, + 0.005567722022533417, + 0.004909535404294729, + 0.01562371663749218, + 0.012694615870714188, + 0.009311147034168243, + 0.009045304730534554, + 0.01643289253115654, + 0.011848102323710918, + 0.004021831788122654, + 0.0042803348042070866, + 0.011431102640926838, + 0.019985588267445564, + 0.004187632817775011, + 0.007491967175155878, + 0.01118347980082035, + 0.007311802823096514, + 0.006021552719175816, + 0.010591641999781132, + 0.028784887865185738, + 0.01473221369087696, + 0.007168645039200783, + 0.008507730439305305, + 0.03391161188483238, + 0.016459457576274872, + 0.0024604182690382004, + 0.001967189135029912, + 0.0054524182341992855, + 0.00975643377751112, + 0.0035366059746593237, + 0.0042329453863203526, + 0.00946328416466713, + 0.007269582711160183, + 0.004181572701781988, + 0.0029238248243927956, + 0.019908996298909187, + 0.009461906738579273, + 0.00404026173055172, + 0.006346046458929777, + 0.008085842244327068, + 0.021416915580630302 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 7 + ], + "selected_reward": -0.007632380351424217 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, extreme close-up from above captures a polished chrome disc magnet, resting motionless on a perfectly smooth, dark, non-reflective surface. From the left edge of the frame, a second polished chrome magnet (a rectangular rod) glides smoothly and deliberately into view. Before any physical contact is made, its invisible magnetic force acts on the disc, causing it to tremble for a microsecond before gliding swiftly and frictionlessly across the vast expanse of darkness. The rod magnet continues its steady advance, maintaining a precise, constant distance from the fleeing disc, its movements coordinating with the other's swift escape. Bright, diffuse studio lighting from above casts clean, elongated highlights across the mirror-like surfaces of both chrome objects. Their perfect forms contrast sharply against the matte black background, creating a minimalist and hypnotic atmosphere of quiet, powerful physical forces at work.", + "seed": 27872, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000121/group_2/index_2891_seed_27872_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": -0.10408730935305358, + "vlm_logprob": -0.10408730935305358, + "vlm_yes_no_margin": 2.875, + "vlm_token_credit": -0.10408730935305358, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 an extreme close-up from above?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a polished chrome disc magnet visible resting on a smooth, dark, non-reflective surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a second polished chrome rectangular rod magnet glide into the frame from the left edge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the disc magnet tremble briefly before gliding swiftly away from the approaching rod magnet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do bright, diffuse studio lights from above cast clean, elongated highlights on the chrome objects?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.21145401895046234 + ], + "opposite_logprob_tokens": [ + -1.8364540338516235 + ], + "logprob_reward": -0.21145401895046234, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.21145401895046234, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.7568381428718567, + 0.03715246915817261, + 0.017578115686774254, + 0.06120136380195618, + 0.012752912938594818, + 0.010573848150670528, + 0.029359064996242523, + 0.012249665334820747, + 0.043943312019109726, + 0.018351102247834206 + ], + "credit_map": [ + 0.008270662277936935, + 0.009700865484774113, + 0.004870458971709013, + 0.005488828290253878, + 0.35757648944854736, + 0.27692657709121704, + 0.05905807763338089, + 0.03494618460536003, + 0.0055800434201955795, + 0.0038595576770603657, + 0.003739322302863002, + 0.002405720530077815, + 0.005885471124202013, + 0.0056716687977313995, + 0.0034244093112647533, + 0.0065862745977938175, + 0.001432341756299138, + 0.001984967850148678, + 0.0038585327565670013, + 0.0022485614754259586, + 0.0019036958692595363, + 0.0017095057992264628, + 0.0022558195050805807, + 0.0021846911404281855, + 0.008188815787434578, + 0.004979232791811228, + 0.013597116805613041, + 0.0047534857876598835, + 0.006165377330034971, + 0.005429121665656567, + 0.004877469502389431, + 0.013210736215114594, + 0.002231426304206252, + 0.0015753430780023336, + 0.0016064136289060116, + 0.001103939488530159, + 0.0012332082260400057, + 0.0019417294533923268, + 0.0014076789375394583, + 0.0016531728906556964, + 0.0012138283345848322, + 0.0015021392609924078, + 0.0014281448675319552, + 0.0011384873650968075, + 0.0010754410177469254, + 0.00174245098605752, + 0.001023662742227316, + 0.0014496936928480864, + 0.002661981852725148, + 0.004102280363440514, + 0.003734466852620244, + 0.006247598212212324, + 0.0036905810702592134, + 0.0028767825569957495, + 0.0025537542533129454, + 0.0034916193690150976, + 0.0013120099902153015, + 0.0015986417420208454, + 0.001626278506591916, + 0.0030086361803114414, + 0.0013280637795105577, + 0.0007761492161080241, + 0.0012600061018019915, + 0.001339879585430026, + 0.0035558941308408976, + 0.006067142356187105, + 0.009136868640780449, + 0.007841034792363644, + 0.0059573231264948845, + 0.003294395050033927, + 0.004657275043427944, + 0.003433380974456668, + 0.0011508888565003872, + 0.0028614592738449574, + 0.003118343185633421, + 0.003398793051019311, + 0.002048419089987874, + 0.0014820265350863338, + 0.0024167674127966166, + 0.0018744051922112703 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 8 + ], + "selected_reward": -0.21145401895046234 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 an extreme close-up from above?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a polished chrome disc magnet visible resting on a smooth, dark, non-reflective surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a second polished chrome rectangular rod magnet glide into the frame from the left edge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the disc magnet tremble briefly before gliding swiftly away from the approaching rod magnet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do bright, diffuse studio lights from above cast clean, elongated highlights on the chrome objects?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2256343960762024 + ], + "opposite_logprob_tokens": [ + -1.6006343364715576 + ], + "logprob_reward": -0.2256343960762024, + "yes_no_margin_reward": 1.375, + "token_credit_reward": -0.2256343960762024, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.8567180037498474, + 0.021738925948739052, + 0.008939624764025211, + 0.013810313306748867, + 0.008179862052202225, + 0.012018593959510326, + 0.026068998500704765, + 0.026640716940164566, + 0.014656631276011467, + 0.011228303425014019 + ], + "credit_map": [ + 0.008918723091483116, + 0.009369672276079655, + 0.004377816803753376, + 0.0047623878344893456, + 0.40752148628234863, + 0.3251052498817444, + 0.06406037509441376, + 0.03260236978530884, + 0.0019119856879115105, + 0.0020602892618626356, + 0.0024743240792304277, + 0.001946267089806497, + 0.002339264377951622, + 0.003724538953974843, + 0.0027105221524834633, + 0.004571733996272087, + 0.0005496182129718363, + 0.000863226072397083, + 0.0016694233054295182, + 0.0009671611478552222, + 0.0011524239089339972, + 0.0015773922204971313, + 0.0008722293423488736, + 0.0012881504371762276, + 0.0019376387353986502, + 0.0010236062807962298, + 0.002313367323949933, + 0.001397533225826919, + 0.0015531876124441624, + 0.00147811823990196, + 0.0015724642435088754, + 0.002534399041905999, + 0.0009705831180326641, + 0.0011814548633992672, + 0.0009567678207531571, + 0.0009760584216564894, + 0.0007660940173082054, + 0.0013832688564434648, + 0.0007838725578039885, + 0.001161763328127563, + 0.0013887210516259074, + 0.0017425762489438057, + 0.0015031947987154126, + 0.0019758655689656734, + 0.0008847670396789908, + 0.0016778507269918919, + 0.001432860386557877, + 0.0014127600006759167, + 0.002536188578233123, + 0.0031731370836496353, + 0.0034325735177844763, + 0.004717948380857706, + 0.0030660894699394703, + 0.002692313864827156, + 0.002739660907536745, + 0.003711090888828039, + 0.0020699938759207726, + 0.0028744766023010015, + 0.005051278043538332, + 0.0057357801124453545, + 0.004182093311101198, + 0.0016048866091296077, + 0.0023728322703391314, + 0.00274937623180449, + 0.001052316278219223, + 0.002371148904785514, + 0.002993189264088869, + 0.0026154196821153164, + 0.0015937857097014785, + 0.001324776210822165, + 0.0015103794867172837, + 0.0011956172529608011, + 0.000686071696691215, + 0.0023558144457638264, + 0.001746050431393087, + 0.002263766946271062, + 0.0012903675669804215, + 0.0010343060130253434, + 0.0010680754203349352, + 0.0007838507299311459 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 6 + ], + "selected_reward": -0.2256343960762024 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 an extreme close-up from above?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a polished chrome disc magnet visible resting on a smooth, dark, non-reflective surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a second polished chrome rectangular rod magnet glide into the frame from the left edge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the disc magnet tremble briefly before gliding swiftly away from the approaching rod magnet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do bright, diffuse studio lights from above cast clean, elongated highlights on the chrome objects?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014201059006154537 + ], + "opposite_logprob_tokens": [ + -4.2642011642456055 + ], + "logprob_reward": -0.014201059006154537, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014201059006154537, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.7122121453285217, + 0.03428972139954567, + 0.018520357087254524, + 0.08581014722585678, + 0.053269144147634506, + 0.014594818465411663, + 0.03261875361204147, + 0.016845250502228737, + 0.01435178890824318, + 0.017487861216068268 + ], + "credit_map": [ + 0.012889130041003227, + 0.010583977214992046, + 0.006950635928660631, + 0.007217421196401119, + 0.3309985399246216, + 0.21212831139564514, + 0.07771709561347961, + 0.05372707545757294, + 0.004614981822669506, + 0.002935907104983926, + 0.003718740539625287, + 0.003045082790777087, + 0.004835756029933691, + 0.006119856145232916, + 0.002944961888715625, + 0.006074437405914068, + 0.0013888103421777487, + 0.002285236492753029, + 0.0020226051565259695, + 0.0015884547028690577, + 0.0024987012147903442, + 0.004071815870702267, + 0.0017708452651277184, + 0.002893888158723712, + 0.011016135104000568, + 0.008066191338002682, + 0.020475653931498528, + 0.0061351098120212555, + 0.007473140489310026, + 0.008074624463915825, + 0.006522824987769127, + 0.018046464771032333, + 0.006093749310821295, + 0.009171618148684502, + 0.007528448011726141, + 0.006328699644654989, + 0.005893222987651825, + 0.006576808635145426, + 0.005016257055103779, + 0.006660339888185263, + 0.001776991761289537, + 0.0015342500992119312, + 0.0020032848697155714, + 0.0026584379374980927, + 0.0013913268921896815, + 0.0021173320710659027, + 0.0013423004420474172, + 0.0017708935774862766, + 0.003060277085751295, + 0.004798594396561384, + 0.0042214300483465195, + 0.006170584354549646, + 0.003918774425983429, + 0.0029174110386520624, + 0.003403280396014452, + 0.004128402564674616, + 0.0013847992522642016, + 0.0020881977397948503, + 0.00309711042791605, + 0.00351332756690681, + 0.0019137745257467031, + 0.0012129016686230898, + 0.0012432129587978125, + 0.0023919257801026106, + 0.0009601188357919455, + 0.001696414197795093, + 0.0018923040479421616, + 0.0021283698733896017, + 0.002001118380576372, + 0.0012669325806200504, + 0.0027754565235227346, + 0.0016310750506818295, + 0.0012745041167363524, + 0.002400287427008152, + 0.0032373135909438133, + 0.0036290755961090326, + 0.001922520576044917, + 0.001628840691410005, + 0.0020027433056384325, + 0.00139257637783885 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 4 + ], + "selected_reward": -0.014201059006154537 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle an extreme close-up from above?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a polished chrome disc magnet visible resting on a smooth, dark, non-reflective surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a second polished chrome rectangular rod magnet glide into the frame from the left edge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the disc magnet tremble briefly before gliding swiftly away from the approaching rod magnet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do bright, diffuse studio lights from above cast clean, elongated highlights on the chrome objects?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.054950010031461716 + ], + "opposite_logprob_tokens": [ + -2.929949998855591 + ], + "logprob_reward": -0.054950010031461716, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.054950010031461716, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.9126230478286743, + 0.014821427874267101, + 0.005091774743050337, + 0.011743915267288685, + 0.008061764761805534, + 0.005675772670656443, + 0.015292970463633537, + 0.0061661601066589355, + 0.010617700405418873, + 0.009905523620545864 + ], + "credit_map": [ + 0.0066512213088572025, + 0.009188374504446983, + 0.003832531627267599, + 0.004693349823355675, + 0.43403756618499756, + 0.37799859046936035, + 0.05157895386219025, + 0.024642350152134895, + 0.0017583663575351238, + 0.0021009566262364388, + 0.001441899687051773, + 0.001234856084920466, + 0.0018628021935001016, + 0.0021369096357375383, + 0.0021068225614726543, + 0.002178812865167856, + 0.0003208553825970739, + 0.0005289161927066743, + 0.0005085236043669283, + 0.0005237655714154243, + 0.0009056199342012405, + 0.0011383253149688244, + 0.0004602647677529603, + 0.0007055036257952452, + 0.002179825911298394, + 0.0016584198456257582, + 0.001652320148423314, + 0.0010337483836337924, + 0.0015236320905387402, + 0.0010517536429688334, + 0.0013547916896641254, + 0.0012894223909825087, + 0.0011808241251856089, + 0.00151514308527112, + 0.0010065581882372499, + 0.0008467028965242207, + 0.0007116193301044405, + 0.0010850680992007256, + 0.0007024153601378202, + 0.001013432745821774, + 0.0006437700358219445, + 0.0008225695928558707, + 0.0006860987050458789, + 0.0008107988396659493, + 0.000560406013391912, + 0.0008783231023699045, + 0.000471373088657856, + 0.0008024327107705176, + 0.001395474886521697, + 0.0021019289270043373, + 0.0017775705782696605, + 0.0031622929964214563, + 0.0014624054310843349, + 0.0011582515435293317, + 0.0022177724167704582, + 0.0020172700751572847, + 0.000526684510987252, + 0.0007482383516617119, + 0.0010419227182865143, + 0.0013525196118280292, + 0.000988220563158393, + 0.00045801198575645685, + 0.00040575064485892653, + 0.0006448110798373818, + 0.0009099355665966868, + 0.0014082846464589238, + 0.002146009122952819, + 0.002279468812048435, + 0.001193908043205738, + 0.0008363490924239159, + 0.0011708765523508191, + 0.0006728668813593686, + 0.0008054279023781419, + 0.0016539867501705885, + 0.0014727459056302905, + 0.0018760163802653551, + 0.0010458898032084107, + 0.0008511219639331102, + 0.001480989740230143, + 0.0007193447090685368 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 1 + ], + "selected_reward": -0.054950010031461716 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 an extreme close-up from above?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a polished chrome disc magnet visible resting on a smooth, dark, non-reflective surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a second polished chrome rectangular rod magnet glide into the frame from the left edge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the disc magnet tremble briefly before gliding swiftly away from the approaching rod magnet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do bright, diffuse studio lights from above cast clean, elongated highlights on the chrome objects?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014197062700986862 + ], + "opposite_logprob_tokens": [ + -4.264196872711182 + ], + "logprob_reward": -0.014197062700986862, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014197062700986862, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.6326302289962769, + 0.03317055106163025, + 0.037333257496356964, + 0.040032438933849335, + 0.012972640804946423, + 0.02043275721371174, + 0.08752470463514328, + 0.034364860504865646, + 0.05664166435599327, + 0.04489697143435478 + ], + "credit_map": [ + 0.009879590943455696, + 0.009475658647716045, + 0.005933482199907303, + 0.005980263464152813, + 0.29532966017723083, + 0.20186206698417664, + 0.06448496133089066, + 0.039684515446424484, + 0.0024723780807107687, + 0.003621901385486126, + 0.003910385072231293, + 0.002322685904800892, + 0.003694424405694008, + 0.006132699549198151, + 0.004124270286411047, + 0.006891807075589895, + 0.0024269602727144957, + 0.005499502178281546, + 0.004804940894246101, + 0.00339006120339036, + 0.005022421479225159, + 0.008614339865744114, + 0.002568915719166398, + 0.005006115883588791, + 0.004417475312948227, + 0.0030211620032787323, + 0.010550194419920444, + 0.0031071940902620554, + 0.003698078915476799, + 0.003262398298829794, + 0.0032004977110773325, + 0.008775441907346249, + 0.0018964140908792615, + 0.0012296822387725115, + 0.0017762337811291218, + 0.001787399873137474, + 0.0010948242852464318, + 0.0019480596529319882, + 0.0014423527754843235, + 0.0017976738745346665, + 0.00215703877620399, + 0.003380252979695797, + 0.002816607244312763, + 0.0032368709798902273, + 0.0012232912704348564, + 0.002519705332815647, + 0.0025549973361194134, + 0.002543994225561619, + 0.007277797441929579, + 0.011450760066509247, + 0.010714317671954632, + 0.016600975766777992, + 0.010988399386405945, + 0.00943276658654213, + 0.010312517173588276, + 0.010747173801064491, + 0.0024622136261314154, + 0.004841050133109093, + 0.004610568750649691, + 0.005602575838565826, + 0.005751051474362612, + 0.002590976655483246, + 0.003505687229335308, + 0.00500073516741395, + 0.0034145605750381947, + 0.00821076799184084, + 0.011395367793738842, + 0.01064726896584034, + 0.006765638943761587, + 0.005404728930443525, + 0.006176634691655636, + 0.004626696929335594, + 0.002979711862280965, + 0.005547933746129274, + 0.008042646571993828, + 0.010076572187244892, + 0.004464952275156975, + 0.004721416626125574, + 0.004802853334695101, + 0.004260881803929806 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 8 + ], + "selected_reward": -0.014197062700986862 + } + ], + "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/0000121/group_3/index_tensor([2407, 2407])_group_3_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000121/group_3/index_tensor([2407, 2407])_group_3_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..f976e6c5b4a2e82cd3ba413664e80e22a2447864 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000121/group_3/index_tensor([2407, 2407])_group_3_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "An epic sweeping drone shot, slowly but steadily, descends toward the heart of Niagara Falls, capturing its immense scale against a bright, clear blue sky. The deep, turquoise water accelerates, its surface churning into a white rapid as it inexorably approaches the cliffs. Then, an unimaginably vast torrent of water crashes unstoppably over the jagged, dark, and damp rock shelves in a continuous, thunderous curtain, dropping with immense force into the chaotic cauldron below. The violent impact at the base sends a vast, billowing cloud of dense white mist into the air, drifting and swirling in the wind. Bright sunlight glints off the water's surface and refracts through the atomized spray, creating a vivid, shimmering rainbow that arcs gracefully against the turbulent backdrop, composing an awe-inspiring, primal, majestic atmosphere of natural power.", + "seed": 1103, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000121/group_3/index_2407_seed_1103_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": -0.020112244714982807, + "vlm_logprob": -0.020112244714982807, + "vlm_yes_no_margin": 4.7, + "vlm_token_credit": -0.020112244714982807, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" 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 shot descend slowly and steadily toward the center of Niagara Falls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the sky visible as bright and clear blue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the water depicted as deep turquoise that accelerates and churns into white rapids?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a vast torrent of water shown crashing over jagged, dark, and damp rock shelves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a vivid, shimmering rainbow visible arcing against the turbulent backdrop?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03955841809511185 + ], + "opposite_logprob_tokens": [ + -3.5395584106445312 + ], + "logprob_reward": -0.03955841809511185, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.03955841809511185, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10166364908218384, + 0.0652860477566719, + 0.027328364551067352, + 0.07757629454135895, + 0.08062410354614258, + 0.272116482257843, + 0.09349926561117172, + 0.1269158571958542, + 0.08303385227918625, + 0.07195613533258438 + ], + "credit_map": [ + 0.007529966067522764, + 0.030831562355160713, + 0.006266022566705942, + 0.007345052435994148, + 0.015710655599832535, + 0.008548211306333542, + 0.00765381520614028, + 0.01777835376560688, + 0.003849231405183673, + 0.007254749536514282, + 0.006122464779764414, + 0.008255191147327423, + 0.008300269022583961, + 0.01730225794017315, + 0.006006477866321802, + 0.008195403963327408, + 0.0027746004052460194, + 0.0026507519651204348, + 0.002481280593201518, + 0.00281734811142087, + 0.0035579714458435774, + 0.006522287614643574, + 0.0028176328632980585, + 0.003706489223986864, + 0.005685343407094479, + 0.004831419326364994, + 0.012617912143468857, + 0.00999109260737896, + 0.007456887047737837, + 0.016035497188568115, + 0.009850938804447651, + 0.011107197031378746, + 0.005556208081543446, + 0.0060057565569877625, + 0.009281852282583714, + 0.01083577424287796, + 0.007332970853894949, + 0.011019740253686905, + 0.019652489572763443, + 0.010939306579530239, + 0.018617654219269753, + 0.01847076416015625, + 0.020198609679937363, + 0.028552532196044922, + 0.03968282416462898, + 0.0728539377450943, + 0.03418392688035965, + 0.03955620527267456, + 0.006749470718204975, + 0.009064503014087677, + 0.007652711123228073, + 0.014613351784646511, + 0.009719666093587875, + 0.012707134708762169, + 0.017630673944950104, + 0.01536175049841404, + 0.01472205389291048, + 0.02061469666659832, + 0.01860009878873825, + 0.016907179728150368, + 0.013078917749226093, + 0.01448399294167757, + 0.01469690166413784, + 0.013812015764415264, + 0.0071936557069420815, + 0.00856008194386959, + 0.00664182286709547, + 0.009374083951115608, + 0.008344591595232487, + 0.008374856784939766, + 0.022312398999929428, + 0.012232356704771519, + 0.010397677309811115, + 0.018978547304868698, + 0.004786183126270771, + 0.00642392598092556, + 0.009416376240551472, + 0.007936143316328526, + 0.008411851711571217, + 0.005605424288660288 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 7, + 0 + ], + "selected_reward": -0.03955841809511185 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" 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 shot descend slowly and steadily toward the center of Niagara Falls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the sky visible as bright and clear blue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the water depicted as deep turquoise that accelerates and churns into white rapids?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a vast torrent of water shown crashing over jagged, dark, and damp rock shelves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a vivid, shimmering rainbow visible arcing against the turbulent backdrop?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04309182986617088 + ], + "opposite_logprob_tokens": [ + -3.1680917739868164 + ], + "logprob_reward": -0.04309182986617088, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04309182986617088, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20626698434352875, + 0.02790621481835842, + 0.04913903400301933, + 0.15040138363838196, + 0.06848321110010147, + 0.24309949576854706, + 0.13513913750648499, + 0.050287600606679916, + 0.043662212789058685, + 0.02561473660171032 + ], + "credit_map": [ + 0.014594889245927334, + 0.06267451494932175, + 0.011943813413381577, + 0.012874986976385117, + 0.030737698078155518, + 0.02461717650294304, + 0.012316937558352947, + 0.03650694712996483, + 0.00353102944791317, + 0.0021530406083911657, + 0.0030514427926391363, + 0.004169415682554245, + 0.003222538623958826, + 0.004857703577727079, + 0.0026926076970994473, + 0.0042284345254302025, + 0.00435660732910037, + 0.00438837381079793, + 0.004489846061915159, + 0.005395323038101196, + 0.006224621087312698, + 0.01278003491461277, + 0.004616103135049343, + 0.0068881213665008545, + 0.011310914531350136, + 0.010032186284661293, + 0.021069247275590897, + 0.01933692954480648, + 0.020314110442996025, + 0.02492518536746502, + 0.017963619902729988, + 0.025449169799685478, + 0.005599810741841793, + 0.006044922396540642, + 0.010298837907612324, + 0.009091205894947052, + 0.007230356801301241, + 0.008477550931274891, + 0.010658636689186096, + 0.011081892065703869, + 0.01513557881116867, + 0.016128627583384514, + 0.019947122782468796, + 0.02729876898229122, + 0.03527203947305679, + 0.06476576626300812, + 0.028438491746783257, + 0.03611310198903084, + 0.013024314306676388, + 0.015257089398801327, + 0.010812056250870228, + 0.015594587661325932, + 0.01542514655739069, + 0.019054872915148735, + 0.024683022871613503, + 0.021288052201271057, + 0.005269044544547796, + 0.007718435488641262, + 0.006602732464671135, + 0.006282801274210215, + 0.0056960564106702805, + 0.007120657246559858, + 0.005148380529135466, + 0.006449490785598755, + 0.0053671542555093765, + 0.007114879321306944, + 0.004036678466945887, + 0.005929510574787855, + 0.005441124085336924, + 0.006234636064618826, + 0.004598856437951326, + 0.004939370322972536, + 0.002956363605335355, + 0.004506317432969809, + 0.00281258556060493, + 0.0023281509056687355, + 0.002049171132966876, + 0.0020409487187862396, + 0.0038249001372605562, + 0.0050962986424565315 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 3 + ], + "selected_reward": -0.04309182986617088 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" 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 shot descend slowly and steadily toward the center of Niagara Falls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the sky visible as bright and clear blue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the water depicted as deep turquoise that accelerates and churns into white rapids?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a vast torrent of water shown crashing over jagged, dark, and damp rock shelves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a vivid, shimmering rainbow visible arcing against the turbulent backdrop?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0015200738562271 + ], + "opposite_logprob_tokens": [ + -6.501520156860352 + ], + "logprob_reward": -0.0015200738562271, + "yes_no_margin_reward": 6.5, + "token_credit_reward": -0.0015200738562271, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09326326847076416, + 0.03205376863479614, + 0.01669183187186718, + 0.04584956914186478, + 0.033230025321245193, + 0.5309860110282898, + 0.04749222472310066, + 0.05459064617753029, + 0.042955294251441956, + 0.10288736969232559 + ], + "credit_map": [ + 0.005645702127367258, + 0.02976200357079506, + 0.004814062267541885, + 0.006556776352226734, + 0.01593053713440895, + 0.008789735846221447, + 0.007302113808691502, + 0.014462342485785484, + 0.0028907384257763624, + 0.004542936105281115, + 0.003611920168623328, + 0.004786938428878784, + 0.0041803354397416115, + 0.0051679303869605064, + 0.003513649571686983, + 0.003359320340678096, + 0.0015998509479686618, + 0.0022761961445212364, + 0.0020409852731972933, + 0.002281089313328266, + 0.002539557171985507, + 0.001892059575766325, + 0.0021553742699325085, + 0.0019067187095060945, + 0.0044319541193544865, + 0.004035305697470903, + 0.00765569694340229, + 0.006368798669427633, + 0.004017035011202097, + 0.008164455182850361, + 0.0048921131528913975, + 0.006284208502620459, + 0.0032934516202658415, + 0.002705913269892335, + 0.0037787407636642456, + 0.005005687475204468, + 0.003146371804177761, + 0.004759975243359804, + 0.00657292827963829, + 0.003966955933719873, + 0.033235881477594376, + 0.03401470184326172, + 0.038563113659620285, + 0.05725782737135887, + 0.0773828998208046, + 0.14746521413326263, + 0.06542988121509552, + 0.0776364803314209, + 0.0038066119886934757, + 0.004169321618974209, + 0.004725676029920578, + 0.008210414089262486, + 0.004997790325433016, + 0.005446344614028931, + 0.009354191832244396, + 0.0067818728275597095, + 0.005391091573983431, + 0.006887507624924183, + 0.0073438649997115135, + 0.0066688512451946735, + 0.0060734874568879604, + 0.00880264863371849, + 0.005980686284601688, + 0.00744250975549221, + 0.004463793709874153, + 0.007071375381201506, + 0.002895601326599717, + 0.003861734177917242, + 0.005955291911959648, + 0.00540750240907073, + 0.00765138678252697, + 0.005648608785122633, + 0.013414040207862854, + 0.022938380017876625, + 0.009977484121918678, + 0.008833387866616249, + 0.012267546728253365, + 0.012629716657102108, + 0.012733510695397854, + 0.010093307122588158 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 9, + 0 + ], + "selected_reward": -0.0015200738562271 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" 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 shot descend slowly and steadily toward the center of Niagara Falls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the sky visible as bright and clear blue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the water depicted as deep turquoise that accelerates and churns into white rapids?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a vast torrent of water shown crashing over jagged, dark, and damp rock shelves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a vivid, shimmering rainbow visible arcing against the turbulent backdrop?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0021983047481626272 + ], + "opposite_logprob_tokens": [ + -6.127198219299316 + ], + "logprob_reward": -0.0021983047481626272, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.0021983047481626272, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13160979747772217, + 0.09817926585674286, + 0.023189159110188484, + 0.04745309427380562, + 0.1102670207619667, + 0.28014636039733887, + 0.07167303562164307, + 0.10458919405937195, + 0.053406089544296265, + 0.0794868990778923 + ], + "credit_map": [ + 0.012076791375875473, + 0.014043062925338745, + 0.009709370322525501, + 0.012923934496939182, + 0.013561118394136429, + 0.030120188370347023, + 0.013188636861741543, + 0.025986695662140846, + 0.006837492808699608, + 0.007700816728174686, + 0.007737466134130955, + 0.008308661170303822, + 0.012183768674731255, + 0.02985777147114277, + 0.010110193863511086, + 0.015443100593984127, + 0.002317064208909869, + 0.0017988740000873804, + 0.0026234646793454885, + 0.003075375221669674, + 0.003303148318082094, + 0.004579211585223675, + 0.0028151075821369886, + 0.002676914678886533, + 0.003614219604060054, + 0.0030928519554436207, + 0.0075154551304876804, + 0.005529389716684818, + 0.004932472016662359, + 0.011291418224573135, + 0.005064441356807947, + 0.006412849761545658, + 0.008867863565683365, + 0.008639910258352757, + 0.013774353079497814, + 0.011603610590100288, + 0.012565402314066887, + 0.02661946974694729, + 0.011432383209466934, + 0.016764037311077118, + 0.019764142110943794, + 0.016785617917776108, + 0.02120012231171131, + 0.028674237430095673, + 0.041324079036712646, + 0.078468918800354, + 0.031974174082279205, + 0.041955091059207916, + 0.007094878237694502, + 0.007005108054727316, + 0.006756778806447983, + 0.00840890221297741, + 0.006183238234370947, + 0.009878447279334068, + 0.008982807397842407, + 0.01736287958920002, + 0.013398460112512112, + 0.015630517154932022, + 0.011340536177158356, + 0.009869462810456753, + 0.011053540743887424, + 0.011553606018424034, + 0.014348858967423439, + 0.017394226044416428, + 0.007491726893931627, + 0.009644195437431335, + 0.005117138382047415, + 0.0042496128007769585, + 0.007630432955920696, + 0.00759044149890542, + 0.005619591102004051, + 0.006062955595552921, + 0.010532482527196407, + 0.018438469618558884, + 0.005188063718378544, + 0.006221522111445665, + 0.011798778548836708, + 0.01130819134414196, + 0.010182754136621952, + 0.0058166454546153545 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 4 + ], + "selected_reward": -0.0021983047481626272 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" 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 shot descend slowly and steadily toward the center of Niagara Falls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the sky visible as bright and clear blue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the water depicted as deep turquoise that accelerates and churns into white rapids?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a vast torrent of water shown crashing over jagged, dark, and damp rock shelves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a vivid, shimmering rainbow visible arcing against the turbulent backdrop?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"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": [ + 0.11621491611003876, + 0.11361284554004669, + 0.05151643231511116, + 0.10565836727619171, + 0.13952285051345825, + 0.3190487325191498, + 0.03934783861041069, + 0.03298890218138695, + 0.028799159452319145, + 0.053289979696273804 + ], + "credit_map": [ + 0.010092681273818016, + 0.012380193918943405, + 0.008580871857702732, + 0.011896226555109024, + 0.012462805956602097, + 0.026815365999937057, + 0.011999939568340778, + 0.021986840292811394, + 0.008564682677388191, + 0.00951191782951355, + 0.007513857912272215, + 0.008402573876082897, + 0.014642918482422829, + 0.03539058938622475, + 0.01011617947369814, + 0.019470131024718285, + 0.005807756446301937, + 0.004001003690063953, + 0.0053917597979307175, + 0.006250513251870871, + 0.006496620364487171, + 0.011319468729197979, + 0.006396128796041012, + 0.00585318123921752, + 0.009610376320779324, + 0.005626335274428129, + 0.02235073782503605, + 0.01377880945801735, + 0.009755249135196209, + 0.018022969365119934, + 0.010581547394394875, + 0.015932325273752213, + 0.008111597038805485, + 0.009710683487355709, + 0.014490896835923195, + 0.019130948930978775, + 0.015615093521773815, + 0.0276795681566, + 0.025850798934698105, + 0.01893327757716179, + 0.021702229976654053, + 0.02213365025818348, + 0.024448467418551445, + 0.03310689330101013, + 0.046231359243392944, + 0.08810345828533173, + 0.03836482763290405, + 0.04495782032608986, + 0.002967025153338909, + 0.006302007008343935, + 0.0031685172580182552, + 0.004667714703828096, + 0.004008965101093054, + 0.004837913904339075, + 0.007733072619885206, + 0.0056626200675964355, + 0.002994707552716136, + 0.004479054361581802, + 0.0030568670481443405, + 0.003807834582403302, + 0.004435111302882433, + 0.005207526497542858, + 0.0036386544816195965, + 0.00536914449185133, + 0.0035156579688191414, + 0.003942950628697872, + 0.002776671899482608, + 0.004380360245704651, + 0.002941496204584837, + 0.004677856341004372, + 0.0037386922631412745, + 0.0028254736680537462, + 0.00556907057762146, + 0.006853315979242325, + 0.004746587947010994, + 0.006235100328922272, + 0.003837950062006712, + 0.007154763676226139, + 0.013235734775662422, + 0.005657457280904055 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 4, + 0 + ], + "selected_reward": -0.014192597009241581 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An epic sweeping drone shot, slowly but steadily, descends toward the heart of Niagara Falls, capturing its immense scale against a bright, clear blue sky. The deep, turquoise water accelerates, its surface churning into a white rapid as it inexorably approaches the cliffs. Then, an unimaginably vast torrent of water crashes unstoppably over the jagged, dark, and damp rock shelves in a continuous, thunderous curtain, dropping with immense force into the chaotic cauldron below. The violent impact at the base sends a vast, billowing cloud of dense white mist into the air, drifting and swirling in the wind. Bright sunlight glints off the water's surface and refracts through the atomized spray, creating a vivid, shimmering rainbow that arcs gracefully against the turbulent backdrop, composing an awe-inspiring, primal, majestic atmosphere of natural power.", + "seed": 1104, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000121/group_3/index_2407_seed_1104_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": -0.040929918992333114, + "vlm_logprob": -0.040929918992333114, + "vlm_yes_no_margin": 4.15, + "vlm_token_credit": -0.040929918992333114, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" 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 shot descend slowly and steadily toward the center of Niagara Falls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the sky visible as bright and clear blue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the water depicted as deep turquoise that accelerates and churns into white rapids?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a vast torrent of water shown crashing over jagged, dark, and damp rock shelves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a vivid, shimmering rainbow visible arcing against the turbulent backdrop?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.12263371795415878 + ], + "opposite_logprob_tokens": [ + -2.247633695602417 + ], + "logprob_reward": -0.12263371795415878, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.12263371795415878, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08860942721366882, + 0.0892895981669426, + 0.07389658689498901, + 0.15710382163524628, + 0.13115817308425903, + 0.06654186546802521, + 0.09790610522031784, + 0.22684477269649506, + 0.034038711339235306, + 0.03461097925901413 + ], + "credit_map": [ + 0.010808156803250313, + 0.008914690464735031, + 0.0117372777312994, + 0.01508739311248064, + 0.007798903621733189, + 0.00883640069514513, + 0.015641335397958755, + 0.009785264730453491, + 0.0139042092487216, + 0.008055327460169792, + 0.01263532042503357, + 0.012064239010214806, + 0.009236571379005909, + 0.012605184689164162, + 0.012834039516746998, + 0.007954702712595463, + 0.009229375049471855, + 0.0064115021377801895, + 0.009029030799865723, + 0.01548703946173191, + 0.00637570908293128, + 0.006463023833930492, + 0.013279072940349579, + 0.007621824275702238, + 0.0189188364893198, + 0.02016664855182171, + 0.018698127940297127, + 0.03529469668865204, + 0.014723705127835274, + 0.0114939259365201, + 0.018414750695228577, + 0.01939312368631363, + 0.018342947587370872, + 0.027262626215815544, + 0.015237720683217049, + 0.015290764160454273, + 0.013001063838601112, + 0.009372545406222343, + 0.018038425594568253, + 0.014612059108912945, + 0.010141134262084961, + 0.007079283706843853, + 0.010416233912110329, + 0.01011117734014988, + 0.00878322497010231, + 0.003988755401223898, + 0.009873978793621063, + 0.006148071493953466, + 0.010275783017277718, + 0.008333101868629456, + 0.010572683997452259, + 0.015868820250034332, + 0.011993961408734322, + 0.007238718681037426, + 0.02209985814988613, + 0.011523166671395302, + 0.031449947506189346, + 0.037786200642585754, + 0.01931767351925373, + 0.015491570346057415, + 0.06382657587528229, + 0.016940012574195862, + 0.025653328746557236, + 0.016379447653889656, + 0.0042995368130505085, + 0.006869927980005741, + 0.004169512074440718, + 0.004834088031202555, + 0.004480691161006689, + 0.0030329010915011168, + 0.003354616230353713, + 0.00299743446521461, + 0.005240902304649353, + 0.003998737782239914, + 0.004677388351410627, + 0.003597790142521262, + 0.004661769140511751, + 0.004550594370812178, + 0.003168918425217271, + 0.004714878275990486 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 3, + 4 + ], + "selected_reward": -0.12263371795415878 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" 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 shot descend slowly and steadily toward the center of Niagara Falls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the sky visible as bright and clear blue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the water depicted as deep turquoise that accelerates and churns into white rapids?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a vast torrent of water shown crashing over jagged, dark, and damp rock shelves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a vivid, shimmering rainbow visible arcing against the turbulent backdrop?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009814932942390442 + ], + "opposite_logprob_tokens": [ + -4.634814739227295 + ], + "logprob_reward": -0.009814932942390442, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009814932942390442, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2560041844844818, + 0.2410339117050171, + 0.05218835175037384, + 0.038558438420295715, + 0.0627039298415184, + 0.053678449243307114, + 0.03187110647559166, + 0.2048116773366928, + 0.02964385785162449, + 0.029506102204322815 + ], + "credit_map": [ + 0.046994708478450775, + 0.02261020429432392, + 0.05048014968633652, + 0.02729238197207451, + 0.022131960839033127, + 0.031223919242620468, + 0.03242770954966545, + 0.022843137383461, + 0.040731001645326614, + 0.034504763782024384, + 0.028444578871130943, + 0.01967441663146019, + 0.03334162384271622, + 0.04320068284869194, + 0.0162309929728508, + 0.0249058548361063, + 0.006004874594509602, + 0.005296910181641579, + 0.005792509764432907, + 0.008463885635137558, + 0.007430991157889366, + 0.0049606915563344955, + 0.008807817474007607, + 0.005430667661130428, + 0.006589086260646582, + 0.0044998303055763245, + 0.004558699671179056, + 0.006070791278034449, + 0.003752607386559248, + 0.003692973405122757, + 0.005211484152823687, + 0.0041829668916761875, + 0.009683546610176563, + 0.008580436930060387, + 0.006989540997892618, + 0.006020853761583567, + 0.008233463391661644, + 0.005676650442183018, + 0.009494293481111526, + 0.008025141432881355, + 0.0053258370608091354, + 0.005664899945259094, + 0.00817269366234541, + 0.010345980525016785, + 0.0071408445946872234, + 0.0035002324730157852, + 0.008459262549877167, + 0.0050686923786997795, + 0.0041758096776902676, + 0.003844397608190775, + 0.003182802116498351, + 0.004446905571967363, + 0.004012727178633213, + 0.0026431374717503786, + 0.0059159197844564915, + 0.0036494059022516012, + 0.02878718450665474, + 0.03469458222389221, + 0.02025955356657505, + 0.012742148712277412, + 0.05932357534766197, + 0.01397024281322956, + 0.0206897035241127, + 0.014344654977321625, + 0.0037596491165459156, + 0.00434404332190752, + 0.0025481495540589094, + 0.002911124611273408, + 0.003876816714182496, + 0.0054586236365139484, + 0.004174189176410437, + 0.002571257995441556, + 0.003366922726854682, + 0.0024903982412070036, + 0.005837790202349424, + 0.002913023578003049, + 0.0033798443619161844, + 0.003925775643438101, + 0.003323381068184972, + 0.004268963821232319 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 7 + ], + "selected_reward": -0.009814932942390442 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" 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 shot descend slowly and steadily toward the center of Niagara Falls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the sky visible as bright and clear blue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the water depicted as deep turquoise that accelerates and churns into white rapids?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a vast torrent of water shown crashing over jagged, dark, and damp rock shelves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a vivid, shimmering rainbow visible arcing against the turbulent backdrop?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008649622090160847 + ], + "opposite_logprob_tokens": [ + -4.758649826049805 + ], + "logprob_reward": -0.008649622090160847, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008649622090160847, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10783974826335907, + 0.12808437645435333, + 0.17198055982589722, + 0.044093839824199677, + 0.056547220796346664, + 0.043803390115499496, + 0.023468149825930595, + 0.28245091438293457, + 0.10169296711683273, + 0.04003887623548508 + ], + "credit_map": [ + 0.012320750392973423, + 0.009598642587661743, + 0.015035753138363361, + 0.01048257201910019, + 0.010387438349425793, + 0.01760023832321167, + 0.019283698871731758, + 0.013130639679729939, + 0.018326397985219955, + 0.014640167355537415, + 0.016421392560005188, + 0.01694498211145401, + 0.014199157245457172, + 0.016888657584786415, + 0.017935195937752724, + 0.012728420086205006, + 0.019016128033399582, + 0.017278915271162987, + 0.018392186611890793, + 0.028412479907274246, + 0.03014376387000084, + 0.0164515171200037, + 0.027074623852968216, + 0.015210943296551704, + 0.005499345250427723, + 0.004277735482901335, + 0.005559451878070831, + 0.00617365213111043, + 0.005306069273501635, + 0.0034995595924556255, + 0.007206571288406849, + 0.006571450736373663, + 0.009251645766198635, + 0.009897702373564243, + 0.006044848822057247, + 0.0074859135784208775, + 0.005750260315835476, + 0.004389020148664713, + 0.006366860121488571, + 0.00736096641048789, + 0.005565529223531485, + 0.007947232574224472, + 0.005529035348445177, + 0.005118129774928093, + 0.006331632379442453, + 0.003354292828589678, + 0.006155984941869974, + 0.003801547223702073, + 0.003698024433106184, + 0.002650124253705144, + 0.0026904973201453686, + 0.00230662664398551, + 0.003915803506970406, + 0.001808341359719634, + 0.003690391546115279, + 0.0027083386667072773, + 0.039423342794179916, + 0.0452224463224411, + 0.02883470058441162, + 0.019004980102181435, + 0.08257120102643967, + 0.01945592276751995, + 0.02836153469979763, + 0.01957676373422146, + 0.013228045776486397, + 0.0161886066198349, + 0.009402425028383732, + 0.01184258796274662, + 0.01586936227977276, + 0.015778440982103348, + 0.010767806321382523, + 0.008615677244961262, + 0.005902295932173729, + 0.003929383587092161, + 0.003778979182243347, + 0.003681071335449815, + 0.005913779139518738, + 0.006608142983168364, + 0.00476544164121151, + 0.005459778476506472 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 2, + 1 + ], + "selected_reward": -0.008649622090160847 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" 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 shot descend slowly and steadily toward the center of Niagara Falls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the sky visible as bright and clear blue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the water depicted as deep turquoise that accelerates and churns into white rapids?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a vast torrent of water shown crashing over jagged, dark, and damp rock shelves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a vivid, shimmering rainbow visible arcing against the turbulent backdrop?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0015142414486035705 + ], + "opposite_logprob_tokens": [ + -6.501514434814453 + ], + "logprob_reward": -0.0015142414486035705, + "yes_no_margin_reward": 6.5, + "token_credit_reward": -0.0015142414486035705, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1276150941848755, + 0.07836932688951492, + 0.027626225724816322, + 0.04087270423769951, + 0.08853931725025177, + 0.05857390537858009, + 0.038387976586818695, + 0.35879960656166077, + 0.07855148613452911, + 0.10266447067260742 + ], + "credit_map": [ + 0.019582921639084816, + 0.00914051290601492, + 0.022524913772940636, + 0.014006680808961391, + 0.011453459039330482, + 0.015475817956030369, + 0.022966453805565834, + 0.012464320287108421, + 0.008935880847275257, + 0.0066075255163013935, + 0.011972693726420403, + 0.010650129057466984, + 0.011051034554839134, + 0.010959995910525322, + 0.012858983129262924, + 0.005333079490810633, + 0.0034874510020017624, + 0.003130712080746889, + 0.0031724979635328054, + 0.004369479604065418, + 0.002930235583335161, + 0.0025730489287525415, + 0.004554219078272581, + 0.003408578922972083, + 0.006301745772361755, + 0.0047560567036271095, + 0.005329651292413473, + 0.005594056099653244, + 0.004290065262466669, + 0.0034121673088520765, + 0.006139998324215412, + 0.005048963241279125, + 0.013926362618803978, + 0.015199407935142517, + 0.009181750938296318, + 0.013058476150035858, + 0.009260903112590313, + 0.00743538374081254, + 0.012482145801186562, + 0.007994873449206352, + 0.006049794144928455, + 0.006979060359299183, + 0.008359812200069427, + 0.011171968653798103, + 0.008413641713559628, + 0.003186185844242573, + 0.009104850701987743, + 0.005308583844453096, + 0.0049721840769052505, + 0.005143200512975454, + 0.004323933273553848, + 0.004171730484813452, + 0.004477974958717823, + 0.004398931283503771, + 0.006614357233047485, + 0.004285662900656462, + 0.050098054111003876, + 0.06251444667577744, + 0.033214207738637924, + 0.022012805566191673, + 0.10332189500331879, + 0.02468005008995533, + 0.037506140768527985, + 0.02545194886624813, + 0.009846939705312252, + 0.007947889156639576, + 0.010334794409573078, + 0.0174031313508749, + 0.009122947230935097, + 0.008068084716796875, + 0.009379938244819641, + 0.006447747349739075, + 0.013728589750826359, + 0.010889559984207153, + 0.01378662884235382, + 0.010589898563921452, + 0.015365874394774437, + 0.016941677778959274, + 0.009796608239412308, + 0.011565624736249447 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 9 + ], + "selected_reward": -0.0015142414486035705 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" 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 shot descend slowly and steadily toward the center of Niagara Falls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the sky visible as bright and clear blue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the water depicted as deep turquoise that accelerates and churns into white rapids?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a vast torrent of water shown crashing over jagged, dark, and damp rock shelves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a vivid, shimmering rainbow visible arcing against the turbulent backdrop?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"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": [ + 0.28831616044044495, + 0.0823296532034874, + 0.045394156128168106, + 0.06447460502386093, + 0.14359046518802643, + 0.04725717753171921, + 0.049201786518096924, + 0.19513265788555145, + 0.04330414533615112, + 0.0409991554915905 + ], + "credit_map": [ + 0.043533630669116974, + 0.020835427567362785, + 0.05145163834095001, + 0.03569205105304718, + 0.028003491461277008, + 0.04948798567056656, + 0.031873513013124466, + 0.02743840590119362, + 0.007754695136100054, + 0.008039338514208794, + 0.013658042065799236, + 0.009888340719044209, + 0.013492471538484097, + 0.012262828648090363, + 0.011628853157162666, + 0.0056050787679851055, + 0.004330137278884649, + 0.0041307322680950165, + 0.005055852234363556, + 0.008224308490753174, + 0.006403881590813398, + 0.00366512150503695, + 0.009022543206810951, + 0.004561576526612043, + 0.009045020677149296, + 0.007479485124349594, + 0.008449011482298374, + 0.012585833668708801, + 0.007612255867570639, + 0.004715694580227137, + 0.007794652134180069, + 0.006792647764086723, + 0.02162971906363964, + 0.023206200450658798, + 0.017454281449317932, + 0.016029508784413338, + 0.015856003388762474, + 0.012620329856872559, + 0.02090613543987274, + 0.015888284891843796, + 0.0058028558269143105, + 0.006534526124596596, + 0.006803096737712622, + 0.00931368488818407, + 0.005333614069968462, + 0.0024770379532128572, + 0.006635560188442469, + 0.004356799181550741, + 0.006055247969925404, + 0.005535563454031944, + 0.006512661464512348, + 0.0068657007068395615, + 0.004754941910505295, + 0.0036647513043135405, + 0.009216676466166973, + 0.006596241146326065, + 0.028000155463814735, + 0.0336025133728981, + 0.018686456605792046, + 0.013850701972842216, + 0.05383770167827606, + 0.012310424819588661, + 0.021285109221935272, + 0.013559588231146336, + 0.00610394636169076, + 0.007766821421682835, + 0.003433407051488757, + 0.005400711204856634, + 0.005849382374435663, + 0.006239712238311768, + 0.004803130868822336, + 0.0037070352118462324, + 0.004544205032289028, + 0.005271040368825197, + 0.0061812144704163074, + 0.0044792345724999905, + 0.00556519441306591, + 0.006000651512295008, + 0.0049868435598909855, + 0.003970773424953222 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 4 + ], + "selected_reward": -0.06203708052635193 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An epic sweeping drone shot, slowly but steadily, descends toward the heart of Niagara Falls, capturing its immense scale against a bright, clear blue sky. The deep, turquoise water accelerates, its surface churning into a white rapid as it inexorably approaches the cliffs. Then, an unimaginably vast torrent of water crashes unstoppably over the jagged, dark, and damp rock shelves in a continuous, thunderous curtain, dropping with immense force into the chaotic cauldron below. The violent impact at the base sends a vast, billowing cloud of dense white mist into the air, drifting and swirling in the wind. Bright sunlight glints off the water's surface and refracts through the atomized spray, creating a vivid, shimmering rainbow that arcs gracefully against the turbulent backdrop, composing an awe-inspiring, primal, majestic atmosphere of natural power.", + "seed": 1105, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000121/group_3/index_2407_seed_1105_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": -0.022403848962858318, + "vlm_logprob": -0.022403848962858318, + "vlm_yes_no_margin": 4.45, + "vlm_token_credit": -0.022403848962858318, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" 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 shot descend slowly and steadily toward the center of Niagara Falls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the sky visible as bright and clear blue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the water depicted as deep turquoise that accelerates and churns into white rapids?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a vast torrent of water shown crashing over jagged, dark, and damp rock shelves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a vivid, shimmering rainbow visible arcing against the turbulent backdrop?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.054111968725919724 + ], + "opposite_logprob_tokens": [ + -3.179111957550049 + ], + "logprob_reward": -0.054111968725919724, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.054111968725919724, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2469412386417389, + 0.07149184495210648, + 0.12298763543367386, + 0.06679260730743408, + 0.16063618659973145, + 0.03668832406401634, + 0.11293655633926392, + 0.06930772215127945, + 0.08257279545068741, + 0.02964509092271328 + ], + "credit_map": [ + 0.03717387467622757, + 0.03878253698348999, + 0.0297488234937191, + 0.034029800444841385, + 0.029971743002533913, + 0.02178558148443699, + 0.03124791756272316, + 0.024200938642024994, + 0.007161564193665981, + 0.010137535631656647, + 0.009962012991309166, + 0.00893143005669117, + 0.010164095088839531, + 0.006913991179317236, + 0.011838252656161785, + 0.006382959429174662, + 0.013049331493675709, + 0.020689707249403, + 0.011364474892616272, + 0.027832644060254097, + 0.010165928862988949, + 0.00884777307510376, + 0.022268524393439293, + 0.008769248612225056, + 0.007935176603496075, + 0.011927273124456406, + 0.006194970570504665, + 0.011901532299816608, + 0.005187016911804676, + 0.008076747879385948, + 0.008482951670885086, + 0.007086939178407192, + 0.0093515170738101, + 0.04232068359851837, + 0.013987541198730469, + 0.023882100358605385, + 0.017763540148735046, + 0.014947114512324333, + 0.026932241395115852, + 0.011451443657279015, + 0.003501882078126073, + 0.005586118903011084, + 0.002897560829296708, + 0.00429689371958375, + 0.006148508284240961, + 0.005305545870214701, + 0.004449631553143263, + 0.004502184689044952, + 0.010029145516455173, + 0.012033678591251373, + 0.01330264750868082, + 0.017877332866191864, + 0.01949361339211464, + 0.01500342134386301, + 0.011167820543050766, + 0.014028887264430523, + 0.0058888341300189495, + 0.00976258609443903, + 0.004771229345351458, + 0.00986036378890276, + 0.01475627813488245, + 0.008718777447938919, + 0.00886334665119648, + 0.006686311215162277, + 0.010773936286568642, + 0.010437723249197006, + 0.008868448436260223, + 0.014503711834549904, + 0.012266936711966991, + 0.009674109518527985, + 0.008238126523792744, + 0.007809807080775499, + 0.0020199422724545, + 0.004195416811853647, + 0.002812659367918968, + 0.006208632607012987, + 0.004494204185903072, + 0.0033356198109686375, + 0.003102315589785576, + 0.0034762986470013857 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 2 + ], + "selected_reward": -0.054111968725919724 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" 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 shot descend slowly and steadily toward the center of Niagara Falls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the sky visible as bright and clear blue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the water depicted as deep turquoise that accelerates and churns into white rapids?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a vast torrent of water shown crashing over jagged, dark, and damp rock shelves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a vivid, shimmering rainbow visible arcing against the turbulent backdrop?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0059780846349895 + ], + "opposite_logprob_tokens": [ + -5.130978107452393 + ], + "logprob_reward": -0.0059780846349895, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.0059780846349895, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22104786336421967, + 0.4480908513069153, + 0.03003372810781002, + 0.02344040758907795, + 0.050194256007671356, + 0.04433644935488701, + 0.0488990843296051, + 0.03196099400520325, + 0.053021129220724106, + 0.04897526279091835 + ], + "credit_map": [ + 0.04066679999232292, + 0.052404578775167465, + 0.023850638419389725, + 0.02660093456506729, + 0.0219646617770195, + 0.016462452709674835, + 0.020792094990611076, + 0.018305707722902298, + 0.056751567870378494, + 0.07115615159273148, + 0.05069057270884514, + 0.06319274753332138, + 0.06361539661884308, + 0.04084589704871178, + 0.06489090621471405, + 0.0369475893676281, + 0.003518969751894474, + 0.0044266097247600555, + 0.002440990647301078, + 0.006477010436356068, + 0.0031287716701626778, + 0.0023657307028770447, + 0.0052075618878006935, + 0.0024680804926902056, + 0.003008923726156354, + 0.003566864412277937, + 0.0021526894997805357, + 0.0039230454713106155, + 0.002444314304739237, + 0.0017260609893128276, + 0.004124602302908897, + 0.002493904670700431, + 0.003354999702423811, + 0.007983743213117123, + 0.004910093266516924, + 0.006202180869877338, + 0.008656006306409836, + 0.006849215365946293, + 0.0057626147754490376, + 0.006475402973592281, + 0.0024499802384525537, + 0.00651905033737421, + 0.0069592418149113655, + 0.007743213325738907, + 0.005318437237292528, + 0.004981872159987688, + 0.006625626236200333, + 0.0037390231154859066, + 0.004895997699350119, + 0.005439884029328823, + 0.006629119161516428, + 0.006144425831735134, + 0.008104143664240837, + 0.005723783280700445, + 0.005653133615851402, + 0.0063085914589464664, + 0.0029064142145216465, + 0.004892985802143812, + 0.0019217393128201365, + 0.004119525663554668, + 0.006424104329198599, + 0.003510934766381979, + 0.004434044938534498, + 0.003751241136342287, + 0.005690312013030052, + 0.008091910742223263, + 0.004170456435531378, + 0.004073213320225477, + 0.010537340305745602, + 0.006669002119451761, + 0.008034616708755493, + 0.005754274316132069, + 0.002125699305906892, + 0.008274157531559467, + 0.004312830977141857, + 0.007905728183686733, + 0.009239913895726204, + 0.004460718482732773, + 0.006258833222091198, + 0.006397378630936146 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 8 + ], + "selected_reward": -0.0059780846349895 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" 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 shot descend slowly and steadily toward the center of Niagara Falls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the sky visible as bright and clear blue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the water depicted as deep turquoise that accelerates and churns into white rapids?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a vast torrent of water shown crashing over jagged, dark, and damp rock shelves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a vivid, shimmering rainbow visible arcing against the turbulent backdrop?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003625488607212901 + ], + "opposite_logprob_tokens": [ + -5.628625392913818 + ], + "logprob_reward": -0.003625488607212901, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.003625488607212901, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10923276841640472, + 0.04020244628190994, + 0.02767898328602314, + 0.04777032136917114, + 0.07722584903240204, + 0.03866591677069664, + 0.06167902424931526, + 0.10129456967115402, + 0.11728201806545258, + 0.37896811962127686 + ], + "credit_map": [ + 0.017138130962848663, + 0.028224654495716095, + 0.008503238670527935, + 0.00987815298140049, + 0.013568672351539135, + 0.011025195010006428, + 0.01220779586583376, + 0.008686919696629047, + 0.0033331539016216993, + 0.005818624049425125, + 0.003424106864258647, + 0.004910397808998823, + 0.005345557350665331, + 0.004924198612570763, + 0.007913736626505852, + 0.004532668273895979, + 0.0031773787923157215, + 0.003976673353463411, + 0.0032186105381697416, + 0.004738639108836651, + 0.0027308908756822348, + 0.0026235857512801886, + 0.005150166340172291, + 0.0020630371291190386, + 0.004449702333658934, + 0.004318885505199432, + 0.001956833293661475, + 0.012730498798191547, + 0.002470232080668211, + 0.0030195333529263735, + 0.014637665823101997, + 0.00418696366250515, + 0.004772634711116552, + 0.01260143332183361, + 0.006119236815720797, + 0.008095135912299156, + 0.015307081863284111, + 0.011346619576215744, + 0.009947705082595348, + 0.00903598964214325, + 0.002380987862125039, + 0.007387606427073479, + 0.005715518724173307, + 0.005379807204008102, + 0.005618308670818806, + 0.0038029621355235577, + 0.005881146993488073, + 0.00249957712367177, + 0.004343350417912006, + 0.006397021934390068, + 0.005411429796367884, + 0.009319649077951908, + 0.01013953611254692, + 0.009959293529391289, + 0.007145876530557871, + 0.008962863124907017, + 0.009827527217566967, + 0.015831273049116135, + 0.0061117419973015785, + 0.00928635522723198, + 0.022308049723505974, + 0.012479646131396294, + 0.013922757469117641, + 0.011527212336659431, + 0.009886357001960278, + 0.01952839456498623, + 0.011803231202065945, + 0.009292848408222198, + 0.026589740067720413, + 0.016070656478405, + 0.012887954711914062, + 0.011222833767533302, + 0.016418030485510826, + 0.05909420922398567, + 0.023045014590024948, + 0.03690914809703827, + 0.062070269137620926, + 0.040143683552742004, + 0.06677361577749252, + 0.07451411336660385 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 0 + ], + "selected_reward": -0.003625488607212901 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" 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 shot descend slowly and steadily toward the center of Niagara Falls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the sky visible as bright and clear blue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the water depicted as deep turquoise that accelerates and churns into white rapids?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a vast torrent of water shown crashing over jagged, dark, and damp rock shelves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a vivid, shimmering rainbow visible arcing against the turbulent backdrop?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"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": [ + 0.06405757367610931, + 0.030793236568570137, + 0.04234856739640236, + 0.037852928042411804, + 0.2692203223705292, + 0.04320324584841728, + 0.1066225990653038, + 0.12266325205564499, + 0.16349168121814728, + 0.11974652111530304 + ], + "credit_map": [ + 0.007845249958336353, + 0.00963797140866518, + 0.007090126164257526, + 0.011155539192259312, + 0.008758575655519962, + 0.0057454900816082954, + 0.007895839400589466, + 0.005928782280534506, + 0.0033272516448050737, + 0.0036072994116693735, + 0.0036386114079505205, + 0.003498210571706295, + 0.0036107595078647137, + 0.0033965862821787596, + 0.0064749508164823055, + 0.0032395687885582447, + 0.004777080379426479, + 0.00747228367254138, + 0.0040734754875302315, + 0.008947945199906826, + 0.003644757205620408, + 0.0030317918863147497, + 0.006912941113114357, + 0.003488294780254364, + 0.005150679964572191, + 0.004990651737898588, + 0.003189821494743228, + 0.007514870725572109, + 0.0028924953658133745, + 0.003850302891805768, + 0.0068001155741512775, + 0.003463994013145566, + 0.015604531392455101, + 0.036160990595817566, + 0.020632466301321983, + 0.02352645993232727, + 0.059656303375959396, + 0.046319760382175446, + 0.031223567202687263, + 0.03609628602862358, + 0.0029843628872185946, + 0.004439395386725664, + 0.006651702802628279, + 0.006651440169662237, + 0.005956142209470272, + 0.00623454712331295, + 0.005362894851714373, + 0.004922761581838131, + 0.007910664193332195, + 0.011238932609558105, + 0.010700379498302937, + 0.01596967875957489, + 0.015830786898732185, + 0.015183668583631516, + 0.014208534732460976, + 0.015579956583678722, + 0.008871166966855526, + 0.01735459640622139, + 0.008952505886554718, + 0.014054137282073498, + 0.02328556217253208, + 0.018181083723902702, + 0.017664378508925438, + 0.01429982203990221, + 0.013897649943828583, + 0.025267411023378372, + 0.013573612086474895, + 0.014832265675067902, + 0.03302808478474617, + 0.022238437086343765, + 0.023343795910477638, + 0.01731044054031372, + 0.008742395788431168, + 0.01786256767809391, + 0.011877692304551601, + 0.010608116164803505, + 0.017658740282058716, + 0.01499788835644722, + 0.016659684479236603, + 0.021339448168873787 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 8, + 7 + ], + "selected_reward": -0.005253320559859276 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" 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 shot descend slowly and steadily toward the center of Niagara Falls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the sky visible as bright and clear blue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the water depicted as deep turquoise that accelerates and churns into white rapids?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a vast torrent of water shown crashing over jagged, dark, and damp rock shelves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a vivid, shimmering rainbow visible arcing against the turbulent backdrop?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.043050382286310196 + ], + "opposite_logprob_tokens": [ + -3.1680502891540527 + ], + "logprob_reward": -0.043050382286310196, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.043050382286310196, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07312937825918198, + 0.03405746445059776, + 0.02345466800034046, + 0.03173697739839554, + 0.05778127908706665, + 0.04496321454644203, + 0.06930205971002579, + 0.18499571084976196, + 0.15388517081737518, + 0.3266940712928772 + ], + "credit_map": [ + 0.007538356352597475, + 0.011071561835706234, + 0.007547958754003048, + 0.011737429536879063, + 0.010318362154066563, + 0.00854374561458826, + 0.009546286426484585, + 0.0068256789818406105, + 0.002896062098443508, + 0.005421603098511696, + 0.0023367127869278193, + 0.004809358157217503, + 0.003993234131485224, + 0.003380183130502701, + 0.007158709689974785, + 0.004061603453010321, + 0.00198718742467463, + 0.00258482969366014, + 0.001954493345692754, + 0.004832391161471605, + 0.002097895136103034, + 0.0020208631176501513, + 0.0061323256231844425, + 0.0018446823814883828, + 0.003951463382691145, + 0.003124756272882223, + 0.002599333645775914, + 0.00332680344581604, + 0.0033210802357643843, + 0.004326464142650366, + 0.004271011333912611, + 0.0068160658702254295, + 0.003340479452162981, + 0.006300441455096006, + 0.0035637293476611376, + 0.0042422618716955185, + 0.013568844646215439, + 0.01164210680872202, + 0.006298776250332594, + 0.008824638091027737, + 0.003584235906600952, + 0.006240753922611475, + 0.0064292773604393005, + 0.0061200810596346855, + 0.0061973026022315025, + 0.005943701136857271, + 0.005815263371914625, + 0.00463259918615222, + 0.005309220869094133, + 0.007129364181309938, + 0.006871039047837257, + 0.010321108624339104, + 0.012810926884412766, + 0.00938986986875534, + 0.008126672357320786, + 0.00934385322034359, + 0.01571512036025524, + 0.024685002863407135, + 0.011609384790062904, + 0.042303383350372314, + 0.030274203047156334, + 0.021812474355101585, + 0.023638106882572174, + 0.01495804451406002, + 0.015623373910784721, + 0.02149079367518425, + 0.01273761410266161, + 0.01573985442519188, + 0.03337439149618149, + 0.021030593663454056, + 0.01901196874678135, + 0.014876584522426128, + 0.020306268706917763, + 0.04203922674059868, + 0.03140358626842499, + 0.030633067712187767, + 0.041548870503902435, + 0.041434962302446365, + 0.05578123405575752, + 0.06354685872793198 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 7, + 8 + ], + "selected_reward": -0.043050382286310196 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An epic sweeping drone shot, slowly but steadily, descends toward the heart of Niagara Falls, capturing its immense scale against a bright, clear blue sky. The deep, turquoise water accelerates, its surface churning into a white rapid as it inexorably approaches the cliffs. Then, an unimaginably vast torrent of water crashes unstoppably over the jagged, dark, and damp rock shelves in a continuous, thunderous curtain, dropping with immense force into the chaotic cauldron below. The violent impact at the base sends a vast, billowing cloud of dense white mist into the air, drifting and swirling in the wind. Bright sunlight glints off the water's surface and refracts through the atomized spray, creating a vivid, shimmering rainbow that arcs gracefully against the turbulent backdrop, composing an awe-inspiring, primal, majestic atmosphere of natural power.", + "seed": 1106, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000121/group_3/index_2407_seed_1106_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": -0.025729131465777755, + "vlm_logprob": -0.025729131465777755, + "vlm_yes_no_margin": 4.45, + "vlm_token_credit": -0.025729131465777755, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" 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 shot descend slowly and steadily toward the center of Niagara Falls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the sky visible as bright and clear blue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the water depicted as deep turquoise that accelerates and churns into white rapids?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a vast torrent of water shown crashing over jagged, dark, and damp rock shelves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a vivid, shimmering rainbow visible arcing against the turbulent backdrop?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06468742340803146 + ], + "opposite_logprob_tokens": [ + -2.939687490463257 + ], + "logprob_reward": -0.06468742340803146, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.06468742340803146, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.27315792441368103, + 0.1136193498969078, + 0.07852691411972046, + 0.14399586617946625, + 0.09413626790046692, + 0.06346537172794342, + 0.039988454431295395, + 0.023565130308270454, + 0.09690763056278229, + 0.07263706624507904 + ], + "credit_map": [ + 0.01584627293050289, + 0.015232319943606853, + 0.02435830608010292, + 0.04195123910903931, + 0.056412842124700546, + 0.04004453495144844, + 0.05335044488310814, + 0.025961991399526596, + 0.011488113552331924, + 0.016637535765767097, + 0.013118553906679153, + 0.011474444530904293, + 0.02029237151145935, + 0.01459434162825346, + 0.013007285073399544, + 0.013006703928112984, + 0.007591282483190298, + 0.007619424723088741, + 0.009098745882511139, + 0.009779449552297592, + 0.009090793319046497, + 0.005731704644858837, + 0.018195820972323418, + 0.011419691145420074, + 0.01685524359345436, + 0.013025385327637196, + 0.016788767650723457, + 0.024934256449341774, + 0.01904277317225933, + 0.017723187804222107, + 0.0154443783685565, + 0.02018187753856182, + 0.011143871583044529, + 0.013974878005683422, + 0.01163448952138424, + 0.01627166196703911, + 0.015361540950834751, + 0.007080607116222382, + 0.010896327905356884, + 0.007772885728627443, + 0.006178246345371008, + 0.00796461757272482, + 0.011059788055717945, + 0.006872289348393679, + 0.012977964244782925, + 0.006998799741268158, + 0.005479054059833288, + 0.005934610031545162, + 0.004744348581880331, + 0.004859131760895252, + 0.004618078004568815, + 0.006206960417330265, + 0.005686586257070303, + 0.004389959387481213, + 0.005746456794440746, + 0.0037369343917816877, + 0.002507672179490328, + 0.003570390632376075, + 0.0037273275665938854, + 0.002592502860352397, + 0.004722913261502981, + 0.0019160809461027384, + 0.0018545369384810328, + 0.002673706505447626, + 0.00580692570656538, + 0.004153402056545019, + 0.006479629315435886, + 0.03178718313574791, + 0.02660038135945797, + 0.006438813637942076, + 0.010632498189806938, + 0.005008800886571407, + 0.00624182028695941, + 0.01043696142733097, + 0.008611640892922878, + 0.010616941377520561, + 0.009339362382888794, + 0.008686724118888378, + 0.012486637569963932, + 0.006216977722942829 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 1 + ], + "selected_reward": -0.06468742340803146 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" 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 shot descend slowly and steadily toward the center of Niagara Falls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the sky visible as bright and clear blue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the water depicted as deep turquoise that accelerates and churns into white rapids?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a vast torrent of water shown crashing over jagged, dark, and damp rock shelves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a vivid, shimmering rainbow visible arcing against the turbulent backdrop?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029833959415555 + ], + "opposite_logprob_tokens": [ + -3.529834032058716 + ], + "logprob_reward": -0.029833959415555, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029833959415555, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.685165524482727, + 0.19874940812587738, + 0.03716263920068741, + 0.00995177123695612, + 0.0151421083137393, + 0.010348611511290073, + 0.011370339430868626, + 0.01082673855125904, + 0.0084253940731287, + 0.012857465073466301 + ], + "credit_map": [ + 0.05697821453213692, + 0.05342823266983032, + 0.07953700423240662, + 0.14081694185733795, + 0.06984958797693253, + 0.06688190996646881, + 0.1353633552789688, + 0.0823102816939354, + 0.022052206099033356, + 0.022902878001332283, + 0.020986195653676987, + 0.020928483456373215, + 0.0442819707095623, + 0.02860320545732975, + 0.022146228700876236, + 0.016848230734467506, + 0.004400582052767277, + 0.005324934609234333, + 0.0042906408198177814, + 0.004035237245261669, + 0.004187515936791897, + 0.004364470019936562, + 0.006370710674673319, + 0.004188546445220709, + 0.0009160555200651288, + 0.001045833807438612, + 0.001210926566272974, + 0.0013804837362840772, + 0.001952179940417409, + 0.0008148932247422636, + 0.0016572249587625265, + 0.0009741734247654676, + 0.0016543358797207475, + 0.001399948145262897, + 0.0024184805806726217, + 0.0019188974983990192, + 0.002449429128319025, + 0.001713534235022962, + 0.0019786658231168985, + 0.001608817488886416, + 0.0009359169634990394, + 0.0008916828082874417, + 0.001177805825136602, + 0.0013349010841920972, + 0.0025562664959579706, + 0.001630341517738998, + 0.0009986577788367867, + 0.0008230391540564597, + 0.0014436972560361028, + 0.0010739152785390615, + 0.0010329148499295115, + 0.001808297005482018, + 0.0015496892156079412, + 0.0017740440089255571, + 0.0016883143689483404, + 0.0009994679130613804, + 0.0010528864804655313, + 0.0011743471259251237, + 0.0013916961615905166, + 0.001400106935761869, + 0.001617991365492344, + 0.001751056290231645, + 0.0014018993824720383, + 0.0010367551585659385, + 0.0005665732314810157, + 0.0009604239603504539, + 0.0008957714890129864, + 0.0020793729927390814, + 0.0014693704433739185, + 0.000654319825116545, + 0.0011075477814301848, + 0.0006920142332091928, + 0.0011250776005908847, + 0.0026418541092425585, + 0.0013795971171930432, + 0.0018983707996085286, + 0.0016861639451235533, + 0.0009838186670094728, + 0.001797060831449926, + 0.0013455224689096212 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.029833959415555 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" 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 shot descend slowly and steadily toward the center of Niagara Falls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the sky visible as bright and clear blue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the water depicted as deep turquoise that accelerates and churns into white rapids?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a vast torrent of water shown crashing over jagged, dark, and damp rock shelves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a vivid, shimmering rainbow visible arcing against the turbulent backdrop?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002824010793119669 + ], + "opposite_logprob_tokens": [ + -5.877823829650879 + ], + "logprob_reward": -0.002824010793119669, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.002824010793119669, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1932503879070282, + 0.19305142760276794, + 0.025904444977641106, + 0.033431652933359146, + 0.08664847165346146, + 0.034817468374967575, + 0.03463853895664215, + 0.043337177485227585, + 0.31529873609542847, + 0.03962162509560585 + ], + "credit_map": [ + 0.013407503254711628, + 0.009485960938036442, + 0.014603937044739723, + 0.01497782114893198, + 0.042468514293432236, + 0.0346207469701767, + 0.0401570126414299, + 0.023528898134827614, + 0.01673976331949234, + 0.016757849603891373, + 0.021454384550452232, + 0.0317096970975399, + 0.024953342974185944, + 0.027472183108329773, + 0.03141846880316734, + 0.022545749321579933, + 0.002753784414380789, + 0.0025636195205152035, + 0.0039056187961250544, + 0.0027323542162775993, + 0.003445412963628769, + 0.002280536573380232, + 0.0038596431259065866, + 0.004363476298749447, + 0.003262382699176669, + 0.003287960309535265, + 0.0041997735388576984, + 0.005240216385573149, + 0.0054809399880468845, + 0.0026140487752854824, + 0.0058583710342645645, + 0.0034879588056355715, + 0.01184718869626522, + 0.010099529288709164, + 0.011915910989046097, + 0.01256159134209156, + 0.015060664154589176, + 0.007227668073028326, + 0.01068253442645073, + 0.007253382820636034, + 0.00341969495639205, + 0.004704958759248257, + 0.005949296522885561, + 0.0040347035974264145, + 0.005013009998947382, + 0.0044363695196807384, + 0.002881306456401944, + 0.0043781292624771595, + 0.0035423252265900373, + 0.004045641049742699, + 0.0030802330002188683, + 0.004092515911906958, + 0.005012256558984518, + 0.007729748729616404, + 0.003637018147855997, + 0.0034987987019121647, + 0.004155884031206369, + 0.0032830180134624243, + 0.006316869053989649, + 0.004772906191647053, + 0.006017736624926329, + 0.008032895624637604, + 0.007039438467472792, + 0.0037184266839176416, + 0.018754327669739723, + 0.01209950353950262, + 0.019011573866009712, + 0.10749047249555588, + 0.09794417023658752, + 0.01700371503829956, + 0.03313200920820236, + 0.009862968698143959, + 0.0030990825034677982, + 0.005855489522218704, + 0.005751192104071379, + 0.004428043495863676, + 0.007067683152854443, + 0.0036177602596580982, + 0.0039716013707220554, + 0.005830772686749697 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 1 + ], + "selected_reward": -0.002824010793119669 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" 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 shot descend slowly and steadily toward the center of Niagara Falls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the sky visible as bright and clear blue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the water depicted as deep turquoise that accelerates and churns into white rapids?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a vast torrent of water shown crashing over jagged, dark, and damp rock shelves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a vivid, shimmering rainbow visible arcing against the turbulent backdrop?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00151305110193789 + ], + "opposite_logprob_tokens": [ + -6.5015130043029785 + ], + "logprob_reward": -0.00151305110193789, + "yes_no_margin_reward": 6.5, + "token_credit_reward": -0.00151305110193789, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15454111993312836, + 0.2899872958660126, + 0.06635049730539322, + 0.027973197400569916, + 0.06388509273529053, + 0.07103680819272995, + 0.018435612320899963, + 0.1742212027311325, + 0.07759527117013931, + 0.055973924696445465 + ], + "credit_map": [ + 0.008297467604279518, + 0.0075859385542571545, + 0.010312830097973347, + 0.017026197165250778, + 0.040014177560806274, + 0.02784121222794056, + 0.02581881731748581, + 0.0176444873213768, + 0.027492454275488853, + 0.02629248797893524, + 0.030299870297312737, + 0.040313471108675, + 0.050225261598825455, + 0.044071659445762634, + 0.04112165793776512, + 0.030170436948537827, + 0.008122250437736511, + 0.0076559968292713165, + 0.008597188629209995, + 0.0072530722245574, + 0.00834562350064516, + 0.00795319490134716, + 0.009505297057330608, + 0.008917875587940216, + 0.0028324590530246496, + 0.0033622158225625753, + 0.0034824542235583067, + 0.005703845992684364, + 0.003793137613683939, + 0.0019896298181265593, + 0.0035325041972100735, + 0.0032769523095339537, + 0.00715859467163682, + 0.00569434417411685, + 0.007868112996220589, + 0.011423546820878983, + 0.013806292787194252, + 0.006745520047843456, + 0.006439149845391512, + 0.004749529529362917, + 0.007242713589221239, + 0.008534668013453484, + 0.008692841976881027, + 0.009349553845822811, + 0.011736308224499226, + 0.01019923947751522, + 0.006954115815460682, + 0.008327380754053593, + 0.002029197756201029, + 0.0016263746656477451, + 0.0023765896912664175, + 0.002872498007491231, + 0.002404360333457589, + 0.0031981258653104305, + 0.00192835065536201, + 0.0020001158118247986, + 0.01778678223490715, + 0.013838440179824829, + 0.02412075735628605, + 0.0218117106705904, + 0.02145739085972309, + 0.03032572567462921, + 0.03121493011713028, + 0.013665463775396347, + 0.004710556007921696, + 0.004722906742244959, + 0.006535509135574102, + 0.020555254071950912, + 0.022879308089613914, + 0.007078288123011589, + 0.006097409874200821, + 0.005016037728637457, + 0.004467759281396866, + 0.007290937472134829, + 0.005864682141691446, + 0.008766568265855312, + 0.008582580834627151, + 0.005404770374298096, + 0.009954207576811314, + 0.005642421543598175 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 7, + 0 + ], + "selected_reward": -0.00151305110193789 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" 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 shot descend slowly and steadily toward the center of Niagara Falls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the sky visible as bright and clear blue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the water depicted as deep turquoise that accelerates and churns into white rapids?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a vast torrent of water shown crashing over jagged, dark, and damp rock shelves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a vivid, shimmering rainbow visible arcing against the turbulent backdrop?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02978721261024475 + ], + "opposite_logprob_tokens": [ + -3.529787302017212 + ], + "logprob_reward": -0.02978721261024475, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.02978721261024475, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3339290916919708, + 0.30292007327079773, + 0.07295674830675125, + 0.021321531385183334, + 0.05591895058751106, + 0.030986785888671875, + 0.01916060969233513, + 0.05018695443868637, + 0.05843492969870567, + 0.05418432131409645 + ], + "credit_map": [ + 0.02793089672923088, + 0.02626856230199337, + 0.035308003425598145, + 0.06945192068815231, + 0.03512420877814293, + 0.03222811594605446, + 0.062307097017765045, + 0.04531029239296913, + 0.029379060491919518, + 0.024834487587213516, + 0.031056605279445648, + 0.04297130927443504, + 0.053581420332193375, + 0.04245329275727272, + 0.04664323106408119, + 0.03200066462159157, + 0.007130291312932968, + 0.006743072997778654, + 0.009226936846971512, + 0.008213489316403866, + 0.010915035381913185, + 0.008268005214631557, + 0.011027072556316853, + 0.011432843282818794, + 0.0019111742731183767, + 0.0024172388948500156, + 0.0025950472336262465, + 0.00289242179132998, + 0.004393653012812138, + 0.002238408662378788, + 0.002991853980347514, + 0.0018817323725670576, + 0.006705352570861578, + 0.010304801166057587, + 0.007448000367730856, + 0.00909403059631586, + 0.005888583138585091, + 0.004667934030294418, + 0.00644603231921792, + 0.005364217329770327, + 0.003310099709779024, + 0.004181176424026489, + 0.00464176619425416, + 0.004037249367684126, + 0.0042281909845769405, + 0.004430556669831276, + 0.002781290328130126, + 0.003376457840204239, + 0.0021997035946696997, + 0.002207062439993024, + 0.0020546386949718, + 0.0018611837876960635, + 0.002825414529070258, + 0.00381411868147552, + 0.0018909586360678077, + 0.0023075281642377377, + 0.004229267127811909, + 0.003627491882070899, + 0.00602493342012167, + 0.006716698408126831, + 0.008921533823013306, + 0.010379440151154995, + 0.004919075407087803, + 0.005368510726839304, + 0.004049997311085463, + 0.005700426176190376, + 0.005560774356126785, + 0.012092462740838528, + 0.013344475999474525, + 0.005941677372902632, + 0.005379108712077141, + 0.006366003770381212, + 0.00577126070857048, + 0.007319385651499033, + 0.00673226872459054, + 0.005260871723294258, + 0.009180535562336445, + 0.00567544624209404, + 0.008777913637459278, + 0.00546663673594594 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.02978721261024475 + } + ], + "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/0000121/group_4/index_tensor([4694, 4694])_group_4_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000121/group_4/index_tensor([4694, 4694])_group_4_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..932d6acb20765805122f399c613d3ce87bd0b363 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000121/group_4/index_tensor([4694, 4694])_group_4_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "low-angle, medium shot first focuses on a pair of dark, muddy leather boots as a figure deliberately steps down the last plank of a weathered wooden porch, the old wood groaning softly under its weight. As the figure's form is obscured by a heavy, forest-green wool cloak, its hood pulled low, its feet firmly planted on a narrow dirt path, the camera begins a steady tracking motion backward and upward. The action follows the figures as they begin their journey, their backs to the unseen house, their postures resolute. The camera's retreat gradually widens the frame, revealing the full context: the path winds past the dilapidated porch and disappears into a vast, uncharted landscape, shrouded in a thick, swirling morning mist. Faint, pearlescent dawn light struggles to penetrate the mist, creating a soft, diffuse glow that captures the damp moss clinging to the gnarled roots of the ancient trees that flank the path. The scene is rendered in a desaturated, cool palette of earthy browns, deep greens, and ethereal grays, evoking the profound atmosphere of a solemn departure and the quiet beginning of an epic, uncertain quest.", + "seed": 4431, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000121/group_4/index_4694_seed_4431_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": -0.010969454608857631, + "vlm_logprob": -0.010969454608857631, + "vlm_yes_no_margin": 4.725, + "vlm_token_credit": -0.010969454608857631, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the initial shot a low-angle medium shot focusing on a pair of dark, muddy leather boots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does a figure wearing a heavy, forest-green wool cloak with a hood pulled low step down the last plank of a weathered wooden porch?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera execute a steady tracking motion backward and upward as the figure moves onto a narrow dirt path?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment revealed to include a path winding past a dilapidated porch into a vast landscape shrouded in thick, swirling morning mist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene rendered in a desaturated, cool palette of earthy browns, deep greens, and ethereal grays with faint, pearlescent dawn 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": [ + -0.01743115670979023 + ], + "opposite_logprob_tokens": [ + -4.642431259155273 + ], + "logprob_reward": -0.01743115670979023, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.01743115670979023, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20232708752155304, + 0.059863727539777756, + 0.41480374336242676, + 0.12182112038135529, + 0.030969955027103424, + 0.11296136677265167, + 0.014241186901926994, + 0.014071209356188774, + 0.014230262488126755, + 0.01471033040434122 + ], + "credit_map": [ + 0.043265484273433685, + 0.022903988137841225, + 0.06278769671916962, + 0.025104142725467682, + 0.011555821634829044, + 0.009995363652706146, + 0.015596725046634674, + 0.011117864400148392, + 0.0047285896725952625, + 0.012212879955768585, + 0.005575619637966156, + 0.007775033824145794, + 0.011544816195964813, + 0.008698421530425549, + 0.005170038435608149, + 0.00415832782164216, + 0.019727861508727074, + 0.15877768397331238, + 0.03852292150259018, + 0.0510125458240509, + 0.08844174444675446, + 0.0071246144361793995, + 0.036432649940252304, + 0.01476369984447956, + 0.0036977806594222784, + 0.017362080514431, + 0.009528002701699734, + 0.01500635128468275, + 0.028071707114577293, + 0.0022380203008651733, + 0.037282634526491165, + 0.008634541183710098, + 0.002149624517187476, + 0.002346013905480504, + 0.005717723164707422, + 0.004729083739221096, + 0.005371455568820238, + 0.002239981899037957, + 0.004765122663229704, + 0.0036509495694190264, + 0.016489215195178986, + 0.005547357257455587, + 0.009915824048221111, + 0.015418093651533127, + 0.01498891320079565, + 0.006528225261718035, + 0.03446226567029953, + 0.009611465036869049, + 0.0022020305041223764, + 0.001424264395609498, + 0.002190494444221258, + 0.00197219499386847, + 0.0021692838054150343, + 0.0009826291352510452, + 0.0020540086552500725, + 0.0012462818995118141, + 0.0021599200554192066, + 0.0015134427230805159, + 0.0029648602940142155, + 0.0013487170217558742, + 0.0012577128363773227, + 0.0011786507675424218, + 0.0020107338204979897, + 0.0016371719539165497, + 0.0021098495926707983, + 0.001990078017115593, + 0.0026960293762385845, + 0.0018131439574062824, + 0.0011203018948435783, + 0.001176274730823934, + 0.0018445926252752542, + 0.001479991478845477, + 0.0013355507981032133, + 0.0013479770859703422, + 0.0016353303799405694, + 0.0022708065807819366, + 0.002594838384538889, + 0.0018688739510253072, + 0.0018964362097904086, + 0.0017605177126824856 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 3 + ], + "selected_reward": -0.01743115670979023 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the initial shot a low-angle medium shot focusing on a pair of dark, muddy leather boots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does a figure wearing a heavy, forest-green wool cloak with a hood pulled low step down the last plank of a weathered wooden porch?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera execute a steady tracking motion backward and upward as the figure moves onto a narrow dirt path?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment revealed to include a path winding past a dilapidated porch into a vast landscape shrouded in thick, swirling morning mist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene rendered in a desaturated, cool palette of earthy browns, deep greens, and ethereal grays with faint, pearlescent dawn 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": [ + -0.012585620395839214 + ], + "opposite_logprob_tokens": [ + -4.387585639953613 + ], + "logprob_reward": -0.012585620395839214, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012585620395839214, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0911543071269989, + 0.07713358104228973, + 0.5530641674995422, + 0.07092545926570892, + 0.039291076362133026, + 0.07230469584465027, + 0.00831245630979538, + 0.024701692163944244, + 0.035040222108364105, + 0.02807234413921833 + ], + "credit_map": [ + 0.01338230911642313, + 0.014523993246257305, + 0.012309649027884007, + 0.019515354186296463, + 0.009398345835506916, + 0.005932587664574385, + 0.009806603193283081, + 0.006285469979047775, + 0.0066166408360004425, + 0.01703849621117115, + 0.006942570209503174, + 0.007585152518004179, + 0.015749013051390648, + 0.009133790619671345, + 0.008303618058562279, + 0.0057642958126962185, + 0.025910349562764168, + 0.2137662172317505, + 0.050608884543180466, + 0.06639407575130463, + 0.1192917451262474, + 0.00951207336038351, + 0.04796108603477478, + 0.01961972378194332, + 0.0030317557975649834, + 0.011492874473333359, + 0.00543831055983901, + 0.0069955079816281796, + 0.02007240429520607, + 0.0018340018577873707, + 0.017611201852560043, + 0.004449398256838322, + 0.00388154829852283, + 0.003626640187576413, + 0.006995527073740959, + 0.006930211558938026, + 0.005113726481795311, + 0.002121995436027646, + 0.005190939176827669, + 0.005430488381534815, + 0.01080436073243618, + 0.0038101859390735626, + 0.007082585711032152, + 0.010055468417704105, + 0.00970800593495369, + 0.0039686886593699455, + 0.02138392999768257, + 0.005491470452398062, + 0.0007904398953542113, + 0.0011562065919861197, + 0.0008477557566948235, + 0.0011977896792814136, + 0.0015435685636475682, + 0.000704143603798002, + 0.0011717100860551, + 0.0009008426568470895, + 0.0032199525739997625, + 0.002374898176640272, + 0.003061870578676462, + 0.0029675166588276625, + 0.00430225720629096, + 0.001933830208145082, + 0.004271426238119602, + 0.002569940872490406, + 0.003404358634725213, + 0.004496119916439056, + 0.0035873963497579098, + 0.004468227736651897, + 0.006556910928338766, + 0.004532218910753727, + 0.004834337159991264, + 0.003160656662657857, + 0.0018930751830339432, + 0.002961801365017891, + 0.00336070591583848, + 0.004729539155960083, + 0.004274028819054365, + 0.002834814367815852, + 0.003958335146307945, + 0.00406004348769784 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 1 + ], + "selected_reward": -0.012585620395839214 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the initial shot a low-angle medium shot focusing on a pair of dark, muddy leather boots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does a figure wearing a heavy, forest-green wool cloak with a hood pulled low step down the last plank of a weathered wooden porch?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera execute a steady tracking motion backward and upward as the figure moves onto a narrow dirt path?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment revealed to include a path winding past a dilapidated porch into a vast landscape shrouded in thick, swirling morning mist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene rendered in a desaturated, cool palette of earthy browns, deep greens, and ethereal grays with faint, pearlescent dawn 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": [ + -0.012558073736727238 + ], + "opposite_logprob_tokens": [ + -4.3875579833984375 + ], + "logprob_reward": -0.012558073736727238, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012558073736727238, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08571156114339828, + 0.044707272201776505, + 0.5392476916313171, + 0.11479856818914413, + 0.03402816876769066, + 0.09464391320943832, + 0.022506382316350937, + 0.014519916847348213, + 0.019513795152306557, + 0.030322788283228874 + ], + "credit_map": [ + 0.01634829491376877, + 0.00883907824754715, + 0.022721169516444206, + 0.013036070391535759, + 0.007981780916452408, + 0.00548179168254137, + 0.006158790551126003, + 0.005144587252289057, + 0.003880108240991831, + 0.008462448604404926, + 0.004750736523419619, + 0.004843286704272032, + 0.010211033746600151, + 0.004846255760639906, + 0.0038283877074718475, + 0.003885011887177825, + 0.026431340724229813, + 0.2071424275636673, + 0.05083546042442322, + 0.06514876335859299, + 0.114583820104599, + 0.009342228062450886, + 0.04634518548846245, + 0.019418440759181976, + 0.00336273736320436, + 0.021492987871170044, + 0.011043842881917953, + 0.023279380053281784, + 0.03179006651043892, + 0.0026856958866119385, + 0.016924798488616943, + 0.004219060763716698, + 0.0029704852495342493, + 0.0030853936914354563, + 0.005612625274807215, + 0.005440950393676758, + 0.004902096930891275, + 0.0023109884932637215, + 0.004888213239610195, + 0.004817413166165352, + 0.013830476440489292, + 0.003959920257329941, + 0.009018074721097946, + 0.013361546210944653, + 0.011961420997977257, + 0.005127773620188236, + 0.029098713770508766, + 0.00828598439693451, + 0.004361998289823532, + 0.001493664807640016, + 0.004215539433062077, + 0.002627447247505188, + 0.0025249274913221598, + 0.001447729766368866, + 0.003492083167657256, + 0.0023429919965565205, + 0.0010360564338043332, + 0.0019889087416231632, + 0.0022439993917942047, + 0.0017521069385111332, + 0.0024208752438426018, + 0.0011789104901254177, + 0.002133902395144105, + 0.0017651576781645417, + 0.0018047932535409927, + 0.0025592048186808825, + 0.002232164144515991, + 0.002255047904327512, + 0.0028932837303727865, + 0.002483383286744356, + 0.0030757130589336157, + 0.002210203791037202, + 0.0031157182529568672, + 0.0031809313222765923, + 0.004739614203572273, + 0.0038068562280386686, + 0.0041751740500330925, + 0.0036692386493086815, + 0.00462267454713583, + 0.003012582426890731 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 3, + 5 + ], + "selected_reward": -0.012558073736727238 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the initial shot a low-angle medium shot focusing on a pair of dark, muddy leather boots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does a figure wearing a heavy, forest-green wool cloak with a hood pulled low step down the last plank of a weathered wooden porch?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera execute a steady tracking motion backward and upward as the figure moves onto a narrow dirt path?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment revealed to include a path winding past a dilapidated porch into a vast landscape shrouded in thick, swirling morning mist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene rendered in a desaturated, cool palette of earthy browns, deep greens, and ethereal grays with faint, pearlescent dawn 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": [ + -0.004637557547539473 + ], + "opposite_logprob_tokens": [ + -5.379637718200684 + ], + "logprob_reward": -0.004637557547539473, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004637557547539473, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0752020999789238, + 0.037149567157030106, + 0.4995727837085724, + 0.05605554208159447, + 0.154013529419899, + 0.06339842081069946, + 0.01596996746957302, + 0.023628683760762215, + 0.034085195511579514, + 0.040924157947301865 + ], + "credit_map": [ + 0.011478519067168236, + 0.017261546105146408, + 0.008837699890136719, + 0.012098736129701138, + 0.008292521350085735, + 0.005228517111390829, + 0.006953858770430088, + 0.00505069550126791, + 0.0033674384467303753, + 0.00648973323404789, + 0.004662626422941685, + 0.004805822856724262, + 0.0062664831057190895, + 0.0053338962607085705, + 0.0029865228570997715, + 0.003237046068534255, + 0.023556100204586983, + 0.190450057387352, + 0.04593191295862198, + 0.06336545199155807, + 0.105881467461586, + 0.008754792623221874, + 0.044110555201768875, + 0.017522459849715233, + 0.002533717779442668, + 0.008280500769615173, + 0.0070900400169193745, + 0.015906017273664474, + 0.010213672183454037, + 0.0014715205179527402, + 0.007030966691672802, + 0.0035291051026433706, + 0.01630987972021103, + 0.015099997632205486, + 0.02648402936756611, + 0.02797418087720871, + 0.018326956778764725, + 0.010235446505248547, + 0.023167051374912262, + 0.016415990889072418, + 0.009283809922635555, + 0.003987184260040522, + 0.006217498332262039, + 0.008909748867154121, + 0.008245970122516155, + 0.004164563026279211, + 0.016795191913843155, + 0.005794459022581577, + 0.0017217735294252634, + 0.0015145378420129418, + 0.0022545871324837208, + 0.002240055473521352, + 0.002408036030828953, + 0.0013607227010652423, + 0.0024418088141828775, + 0.002028445713222027, + 0.0017679092707112432, + 0.0031820780131965876, + 0.0038068864960223436, + 0.003006520215421915, + 0.003163423389196396, + 0.002453276887536049, + 0.003444487461820245, + 0.002804102608934045, + 0.004655718803405762, + 0.004349114838987589, + 0.00560279143974185, + 0.004428200889378786, + 0.0034682373516261578, + 0.0033753158058971167, + 0.004281694535166025, + 0.003924119286239147, + 0.003501547034829855, + 0.0055524674244225025, + 0.0064388844184577465, + 0.006017275620251894, + 0.004299805499613285, + 0.004527985118329525, + 0.004889518488198519, + 0.005696677602827549 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 4, + 0 + ], + "selected_reward": -0.004637557547539473 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the initial shot a low-angle medium shot focusing on a pair of dark, muddy leather boots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does a figure wearing a heavy, forest-green wool cloak with a hood pulled low step down the last plank of a weathered wooden porch?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera execute a steady tracking motion backward and upward as the figure moves onto a narrow dirt path?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment revealed to include a path winding past a dilapidated porch into a vast landscape shrouded in thick, swirling morning mist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene rendered in a desaturated, cool palette of earthy browns, deep greens, and ethereal grays with faint, pearlescent dawn 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": [ + -0.007634864654392004 + ], + "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": [ + 0.14681588113307953, + 0.041405849158763885, + 0.4423373341560364, + 0.09127852320671082, + 0.1303364336490631, + 0.07555222511291504, + 0.019048336893320084, + 0.014123312197625637, + 0.01568976789712906, + 0.023412318900227547 + ], + "credit_map": [ + 0.022761914879083633, + 0.03203960880637169, + 0.021409282460808754, + 0.022556087002158165, + 0.014334909617900848, + 0.010476750321686268, + 0.011553527787327766, + 0.011683804914355278, + 0.003803675062954426, + 0.006090526934713125, + 0.005909423343837261, + 0.007213525474071503, + 0.0057260626927018166, + 0.0030259632039815187, + 0.005295733455568552, + 0.00434093875810504, + 0.020364362746477127, + 0.16827523708343506, + 0.04217962920665741, + 0.055804815143346786, + 0.09284316003322601, + 0.008258230984210968, + 0.03836236149072647, + 0.016249539330601692, + 0.001373556675389409, + 0.009033799171447754, + 0.0038661828730255365, + 0.009050773456692696, + 0.029578950256109238, + 0.001684393035247922, + 0.03183280676603317, + 0.004858054220676422, + 0.012956066057085991, + 0.01278891321271658, + 0.020613089203834534, + 0.021755021065473557, + 0.016333770006895065, + 0.007956563495099545, + 0.02378937415778637, + 0.014143641106784344, + 0.010696809738874435, + 0.004126410000026226, + 0.00684455456212163, + 0.010261344723403454, + 0.010118863545358181, + 0.004529021680355072, + 0.022223472595214844, + 0.006751754321157932, + 0.0023331185802817345, + 0.0018797683296725154, + 0.0027935346588492393, + 0.0035986530128866434, + 0.0015756961656734347, + 0.0019007088849321008, + 0.0026259543374180794, + 0.00234090443700552, + 0.0011324917431920767, + 0.0022111486177891493, + 0.0015882733277976513, + 0.001933528226800263, + 0.0014533204957842827, + 0.0016990280710160732, + 0.0018042379524558783, + 0.002301284344866872, + 0.0011094248620793223, + 0.0023278205189853907, + 0.00235788244754076, + 0.002609384013339877, + 0.001902659423649311, + 0.0015452317893505096, + 0.0020640443544834852, + 0.0017733215354382992, + 0.001968023832887411, + 0.0031115422025322914, + 0.004035646561533213, + 0.003074957523494959, + 0.002756711794063449, + 0.0025224019773304462, + 0.003117977874353528, + 0.002825055969879031 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 4 + ], + "selected_reward": -0.007634864654392004 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle, medium shot first focuses on a pair of dark, muddy leather boots as a figure deliberately steps down the last plank of a weathered wooden porch, the old wood groaning softly under its weight. As the figure's form is obscured by a heavy, forest-green wool cloak, its hood pulled low, its feet firmly planted on a narrow dirt path, the camera begins a steady tracking motion backward and upward. The action follows the figures as they begin their journey, their backs to the unseen house, their postures resolute. The camera's retreat gradually widens the frame, revealing the full context: the path winds past the dilapidated porch and disappears into a vast, uncharted landscape, shrouded in a thick, swirling morning mist. Faint, pearlescent dawn light struggles to penetrate the mist, creating a soft, diffuse glow that captures the damp moss clinging to the gnarled roots of the ancient trees that flank the path. The scene is rendered in a desaturated, cool palette of earthy browns, deep greens, and ethereal grays, evoking the profound atmosphere of a solemn departure and the quiet beginning of an epic, uncertain quest.", + "seed": 4432, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000121/group_4/index_4694_seed_4432_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": -0.00787739953957498, + "vlm_logprob": -0.00787739953957498, + "vlm_yes_no_margin": 5.225, + "vlm_token_credit": -0.00787739953957498, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the initial shot a low-angle medium shot focusing on a pair of dark, muddy leather boots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does a figure wearing a heavy, forest-green wool cloak with a hood pulled low step down the last plank of a weathered wooden porch?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera execute a steady tracking motion backward and upward as the figure moves onto a narrow dirt path?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment revealed to include a path winding past a dilapidated porch into a vast landscape shrouded in thick, swirling morning mist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene rendered in a desaturated, cool palette of earthy browns, deep greens, and ethereal grays with faint, pearlescent dawn 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": [ + -0.02132534049451351 + ], + "opposite_logprob_tokens": [ + -4.14632511138916 + ], + "logprob_reward": -0.02132534049451351, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.02132534049451351, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0783371552824974, + 0.0708843544125557, + 0.041791077703237534, + 0.026828959584236145, + 0.45639389753341675, + 0.05447380617260933, + 0.08307922631502151, + 0.16556791961193085, + 0.013978853821754456, + 0.00866478681564331 + ], + "credit_map": [ + 0.009167761541903019, + 0.009436579421162605, + 0.010723540559411049, + 0.01125377044081688, + 0.008806944824755192, + 0.010443825274705887, + 0.007357708644121885, + 0.011147030629217625, + 0.006770713720470667, + 0.008955061435699463, + 0.008106084540486336, + 0.007533927448093891, + 0.008633246645331383, + 0.008241388946771622, + 0.013773568905889988, + 0.008870366029441357, + 0.005678030662238598, + 0.005731622688472271, + 0.004194295033812523, + 0.0034073260612785816, + 0.00755241559818387, + 0.004124412313103676, + 0.005144721362739801, + 0.005958255380392075, + 0.004567113239318132, + 0.0039568375796079636, + 0.0032968639861792326, + 0.0029580697882920504, + 0.0028580366633832455, + 0.003930429462343454, + 0.002381605561822653, + 0.0028800040017813444, + 0.023159408941864967, + 0.18928954005241394, + 0.20186851918697357, + 0.01737452857196331, + 0.008137233555316925, + 0.005575574468821287, + 0.006038540042936802, + 0.004950566682964563, + 0.002394618233665824, + 0.01885257475078106, + 0.0029883228708058596, + 0.018624871969223022, + 0.0019094578456133604, + 0.002959497971460223, + 0.0028332911897450686, + 0.00391116738319397, + 0.007280962076038122, + 0.01926036924123764, + 0.007668785750865936, + 0.017391275614500046, + 0.004450655076652765, + 0.004311408847570419, + 0.011757775209844112, + 0.010958001017570496, + 0.022391045466065407, + 0.01958131417632103, + 0.009338824078440666, + 0.027760079130530357, + 0.009824676439166069, + 0.03427319601178169, + 0.018973561003804207, + 0.023425225168466568, + 0.001711413380689919, + 0.002666689921170473, + 0.002810273552313447, + 0.0018578837625682354, + 0.0008779082563705742, + 0.00164637656416744, + 0.001138298655860126, + 0.0012700099032372236, + 0.0014706969959661365, + 0.0012539481977000833, + 0.001448532217182219, + 0.0011217721039429307, + 0.0007729530916549265, + 0.0009309957385994494, + 0.0007005034713074565, + 0.0009653841843828559 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 7, + 6 + ], + "selected_reward": -0.02132534049451351 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the initial shot a low-angle medium shot focusing on a pair of dark, muddy leather boots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does a figure wearing a heavy, forest-green wool cloak with a hood pulled low step down the last plank of a weathered wooden porch?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera execute a steady tracking motion backward and upward as the figure moves onto a narrow dirt path?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment revealed to include a path winding past a dilapidated porch into a vast landscape shrouded in thick, swirling morning mist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene rendered in a desaturated, cool palette of earthy browns, deep greens, and ethereal grays with faint, pearlescent dawn 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": [ + -0.006762593984603882 + ], + "opposite_logprob_tokens": [ + -5.006762504577637 + ], + "logprob_reward": -0.006762593984603882, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006762593984603882, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05438746511936188, + 0.05133212357759476, + 0.0614289790391922, + 0.11598039418458939, + 0.2200792133808136, + 0.07936184108257294, + 0.10393135994672775, + 0.2669258713722229, + 0.03371667116880417, + 0.01285606063902378 + ], + "credit_map": [ + 0.006220813374966383, + 0.0061864098533988, + 0.007848222739994526, + 0.007930104620754719, + 0.006111699156463146, + 0.006433155853301287, + 0.005467192269861698, + 0.008189870975911617, + 0.0038367274682968855, + 0.005173517856746912, + 0.005926455836743116, + 0.005174499936401844, + 0.011228875257074833, + 0.005786991212517023, + 0.006557673681527376, + 0.007647388614714146, + 0.006179953459650278, + 0.005374039988964796, + 0.006541937589645386, + 0.004228489939123392, + 0.010661695152521133, + 0.008491682820022106, + 0.00937771424651146, + 0.01057347096502781, + 0.015813002362847328, + 0.015646064653992653, + 0.0113103948533535, + 0.007666691672056913, + 0.019462669268250465, + 0.01658700965344906, + 0.01027619931846857, + 0.019218361005187035, + 0.01056952029466629, + 0.09046291559934616, + 0.09419620037078857, + 0.00905281025916338, + 0.004972963128238916, + 0.00428885268047452, + 0.0034093589056283236, + 0.0031266005244106054, + 0.004295016638934612, + 0.026325389742851257, + 0.004583483096212149, + 0.026373019441962242, + 0.0029834136366844177, + 0.004259203094989061, + 0.0041221207939088345, + 0.006420197896659374, + 0.008445492014288902, + 0.024079887196421623, + 0.009363779798150063, + 0.02193763665854931, + 0.006205752957612276, + 0.006486917845904827, + 0.014436347410082817, + 0.012975550256669521, + 0.03686146438121796, + 0.03060407005250454, + 0.014506164938211441, + 0.04366964474320412, + 0.016199884936213493, + 0.05601327121257782, + 0.02916819602251053, + 0.0399031899869442, + 0.0054526855237782, + 0.005761709995567799, + 0.004770476371049881, + 0.003484695916995406, + 0.002457868540659547, + 0.004096189048141241, + 0.0037552639842033386, + 0.003937783185392618, + 0.001200379221700132, + 0.0014681803295388818, + 0.002169340383261442, + 0.002525669289752841, + 0.001751935575157404, + 0.0009825825691223145, + 0.0016294001834467053, + 0.0011285747168585658 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 4, + 3 + ], + "selected_reward": -0.006762593984603882 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the initial shot a low-angle medium shot focusing on a pair of dark, muddy leather boots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does a figure wearing a heavy, forest-green wool cloak with a hood pulled low step down the last plank of a weathered wooden porch?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera execute a steady tracking motion backward and upward as the figure moves onto a narrow dirt path?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment revealed to include a path winding past a dilapidated porch into a vast landscape shrouded in thick, swirling morning mist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene rendered in a desaturated, cool palette of earthy browns, deep greens, and ethereal grays with faint, pearlescent dawn 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": [ + -0.004105829633772373 + ], + "opposite_logprob_tokens": [ + -5.504106044769287 + ], + "logprob_reward": -0.004105829633772373, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004105829633772373, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08565573394298553, + 0.036048587411642075, + 0.07093025743961334, + 0.1315634548664093, + 0.4802474081516266, + 0.06707204878330231, + 0.01704070344567299, + 0.04512493312358856, + 0.02819753810763359, + 0.03811933100223541 + ], + "credit_map": [ + 0.008830446749925613, + 0.008883791044354439, + 0.01258364412933588, + 0.011679481714963913, + 0.00928344577550888, + 0.011068208143115044, + 0.008346907794475555, + 0.014979802072048187, + 0.0035966294817626476, + 0.003831852227449417, + 0.00428715581074357, + 0.004190700128674507, + 0.006151992361992598, + 0.003894090186804533, + 0.006393542978912592, + 0.003702624700963497, + 0.008199922740459442, + 0.006525808479636908, + 0.0069734444841742516, + 0.006038506515324116, + 0.010777207091450691, + 0.008029366843402386, + 0.012932470999658108, + 0.011453529819846153, + 0.028312668204307556, + 0.02109060063958168, + 0.016552956774830818, + 0.008146271109580994, + 0.022835133597254753, + 0.0180116668343544, + 0.008181300945580006, + 0.00843285396695137, + 0.024112720042467117, + 0.1935935914516449, + 0.20668798685073853, + 0.02164437249302864, + 0.009902405552566051, + 0.008768729865550995, + 0.007842167280614376, + 0.007695418316870928, + 0.0037601219955831766, + 0.023145973682403564, + 0.0046187397092580795, + 0.02128116972744465, + 0.002936571603640914, + 0.0036974730901420116, + 0.0039238352328538895, + 0.0037081667687743902, + 0.0020119722466915846, + 0.002447491744533181, + 0.0029577987734228373, + 0.0027842519339174032, + 0.0009783976711332798, + 0.0014319663168862462, + 0.0018685950199142098, + 0.0025602285750210285, + 0.005785290151834488, + 0.0074104792438447475, + 0.0028572555165737867, + 0.007890569977462292, + 0.0028738975524902344, + 0.008336828090250492, + 0.004982206970453262, + 0.00498840631917119, + 0.003192498115822673, + 0.0042592850513756275, + 0.003436187282204628, + 0.004721460863947868, + 0.0025451933033764362, + 0.003154823323711753, + 0.003977656364440918, + 0.002910435199737549, + 0.003564380342140794, + 0.005258318968117237, + 0.00539669394493103, + 0.005795183591544628, + 0.003938055131584406, + 0.003559594275429845, + 0.0036209342069923878, + 0.00698616961017251 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 3, + 0 + ], + "selected_reward": -0.004105829633772373 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the initial shot a low-angle medium shot focusing on a pair of dark, muddy leather boots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does a figure wearing a heavy, forest-green wool cloak with a hood pulled low step down the last plank of a weathered wooden porch?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera execute a steady tracking motion backward and upward as the figure moves onto a narrow dirt path?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment revealed to include a path winding past a dilapidated porch into a vast landscape shrouded in thick, swirling morning mist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene rendered in a desaturated, cool palette of earthy browns, deep greens, and ethereal grays with faint, pearlescent dawn 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": [ + -0.0019387274514883757 + ], + "opposite_logprob_tokens": [ + -6.251938819885254 + ], + "logprob_reward": -0.0019387274514883757, + "yes_no_margin_reward": 6.25, + "token_credit_reward": -0.0019387274514883757, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05269639194011688, + 0.03609582409262657, + 0.037826236337423325, + 0.044194526970386505, + 0.2458043098449707, + 0.26727089285850525, + 0.0842362642288208, + 0.14217527210712433, + 0.05745084211230278, + 0.03224942460656166 + ], + "credit_map": [ + 0.007436150219291449, + 0.005323460325598717, + 0.007619436364620924, + 0.006832449231296778, + 0.00508341146633029, + 0.007796728052198887, + 0.00403533224016428, + 0.008569426834583282, + 0.003156576305627823, + 0.004148681648075581, + 0.005049793049693108, + 0.003253246657550335, + 0.005508866161108017, + 0.004752982873469591, + 0.0045229471288621426, + 0.005702730268239975, + 0.004297684412449598, + 0.0037218783982098103, + 0.0033960766158998013, + 0.0033745411783456802, + 0.005494997370988131, + 0.003998556174337864, + 0.008353902027010918, + 0.005188599228858948, + 0.00823893491178751, + 0.006523012183606625, + 0.004346972331404686, + 0.0027049994096159935, + 0.007380821276456118, + 0.006136697717010975, + 0.003985356539487839, + 0.004877730738371611, + 0.011398223228752613, + 0.10003221035003662, + 0.104756198823452, + 0.010905430652201176, + 0.005099923815578222, + 0.004711325746029615, + 0.004407126922160387, + 0.004493862856179476, + 0.004037728067487478, + 0.11402434855699539, + 0.007416942156851292, + 0.11526156216859818, + 0.0052388203330338, + 0.005317308474332094, + 0.008253628388047218, + 0.0077205379493534565, + 0.00757528655230999, + 0.01950840651988983, + 0.007870164699852467, + 0.0164710134267807, + 0.004979623015969992, + 0.00518961763009429, + 0.011479935608804226, + 0.011162214912474155, + 0.017498880624771118, + 0.02117207832634449, + 0.008992951363325119, + 0.023652685806155205, + 0.009002527222037315, + 0.026042383164167404, + 0.015832750126719475, + 0.019981008023023605, + 0.008624411188066006, + 0.00870547629892826, + 0.009381457231938839, + 0.008202500641345978, + 0.0038427887484431267, + 0.006019988562911749, + 0.0069764768704771996, + 0.005697743967175484, + 0.0038433673325926065, + 0.004607131704688072, + 0.0062282634899020195, + 0.0060151685029268265, + 0.0026823950465768576, + 0.0032495877239853144, + 0.003025785321369767, + 0.0025977259501814842 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 4, + 7 + ], + "selected_reward": -0.0019387274514883757 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the initial shot a low-angle medium shot focusing on a pair of dark, muddy leather boots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does a figure wearing a heavy, forest-green wool cloak with a hood pulled low step down the last plank of a weathered wooden porch?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera execute a steady tracking motion backward and upward as the figure moves onto a narrow dirt path?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment revealed to include a path winding past a dilapidated porch into a vast landscape shrouded in thick, swirling morning mist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene rendered in a desaturated, cool palette of earthy browns, deep greens, and ethereal grays with faint, pearlescent dawn 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": [ + -0.005254506133496761 + ], + "opposite_logprob_tokens": [ + -5.25525426864624 + ], + "logprob_reward": -0.005254506133496761, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005254506133496761, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07597184926271439, + 0.09203799813985825, + 0.044764451682567596, + 0.037700701504945755, + 0.26710155606269836, + 0.27799564599990845, + 0.02696836367249489, + 0.0679137259721756, + 0.06390764564275742, + 0.0456380657851696 + ], + "credit_map": [ + 0.009580742567777634, + 0.006600996479392052, + 0.009899936616420746, + 0.012262812815606594, + 0.012062374502420425, + 0.007391962688416243, + 0.007003101985901594, + 0.011169922538101673, + 0.0077711851336061954, + 0.01124948263168335, + 0.010797925293445587, + 0.007772289682179689, + 0.01960679143667221, + 0.009603319689631462, + 0.013309369795024395, + 0.011927635408937931, + 0.005562797654420137, + 0.00506764929741621, + 0.003733886405825615, + 0.0036356074269860983, + 0.006832696031779051, + 0.005480409134179354, + 0.009114201180636883, + 0.005337206181138754, + 0.007379351649433374, + 0.00587458023801446, + 0.0032871789298951626, + 0.0038506092969328165, + 0.004855405073612928, + 0.004969046451151371, + 0.0034940813202410936, + 0.0039904480800032616, + 0.01332264393568039, + 0.1059088408946991, + 0.11259989440441132, + 0.012806648388504982, + 0.006960608996450901, + 0.005884815007448196, + 0.004839703440666199, + 0.004778383299708366, + 0.004689889494329691, + 0.12003559619188309, + 0.0072268652729690075, + 0.12148988246917725, + 0.005384456366300583, + 0.005254714284092188, + 0.008652936667203903, + 0.0052613005973398685, + 0.0032605098094791174, + 0.006982660386711359, + 0.0026220164727419615, + 0.004824579227715731, + 0.001674616476520896, + 0.0018937335116788745, + 0.003011872759088874, + 0.00269837467931211, + 0.00913916900753975, + 0.010594883933663368, + 0.006100211292505264, + 0.009920221753418446, + 0.004149461630731821, + 0.01369284838438034, + 0.006415871903300285, + 0.00790106039494276, + 0.00787215307354927, + 0.009319803677499294, + 0.013299687765538692, + 0.008074267767369747, + 0.0045367139391601086, + 0.006613471079617739, + 0.00685560517013073, + 0.007335945963859558, + 0.008061889559030533, + 0.007904774509370327, + 0.006425586994737387, + 0.005783292464911938, + 0.004179235082119703, + 0.005601037759333849, + 0.0033033103682100773, + 0.0043789385817945 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 4, + 1 + ], + "selected_reward": -0.005254506133496761 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle, medium shot first focuses on a pair of dark, muddy leather boots as a figure deliberately steps down the last plank of a weathered wooden porch, the old wood groaning softly under its weight. As the figure's form is obscured by a heavy, forest-green wool cloak, its hood pulled low, its feet firmly planted on a narrow dirt path, the camera begins a steady tracking motion backward and upward. The action follows the figures as they begin their journey, their backs to the unseen house, their postures resolute. The camera's retreat gradually widens the frame, revealing the full context: the path winds past the dilapidated porch and disappears into a vast, uncharted landscape, shrouded in a thick, swirling morning mist. Faint, pearlescent dawn light struggles to penetrate the mist, creating a soft, diffuse glow that captures the damp moss clinging to the gnarled roots of the ancient trees that flank the path. The scene is rendered in a desaturated, cool palette of earthy browns, deep greens, and ethereal grays, evoking the profound atmosphere of a solemn departure and the quiet beginning of an epic, uncertain quest.", + "seed": 4433, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000121/group_4/index_4694_seed_4433_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": -0.008224433357827365, + "vlm_logprob": -0.008224433357827365, + "vlm_yes_no_margin": 5.225, + "vlm_token_credit": -0.008224433357827365, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the initial shot a low-angle medium shot focusing on a pair of dark, muddy leather boots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does a figure wearing a heavy, forest-green wool cloak with a hood pulled low step down the last plank of a weathered wooden porch?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera execute a steady tracking motion backward and upward as the figure moves onto a narrow dirt path?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment revealed to include a path winding past a dilapidated porch into a vast landscape shrouded in thick, swirling morning mist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene rendered in a desaturated, cool palette of earthy browns, deep greens, and ethereal grays with faint, pearlescent dawn 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": [ + -0.020193351432681084 + ], + "opposite_logprob_tokens": [ + -4.395193576812744 + ], + "logprob_reward": -0.020193351432681084, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.020193351432681084, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2707947790622711, + 0.13808956742286682, + 0.06387142091989517, + 0.09726603329181671, + 0.04845517873764038, + 0.10117484629154205, + 0.05788207799196243, + 0.03497542440891266, + 0.03938380628824234, + 0.14810678362846375 + ], + "credit_map": [ + 0.031092410907149315, + 0.03687204793095589, + 0.0291514303535223, + 0.034642793238162994, + 0.05610405281186104, + 0.024919208139181137, + 0.03575432300567627, + 0.022258533164858818, + 0.009153255261480808, + 0.00789052527397871, + 0.02527773752808571, + 0.008167289197444916, + 0.023606445640325546, + 0.008331497199833393, + 0.04417463019490242, + 0.011488213203847408, + 0.006498757749795914, + 0.019404664635658264, + 0.006028853356838226, + 0.004148135427385569, + 0.015146289952099323, + 0.003514564363285899, + 0.004627050366252661, + 0.00450311740860343, + 0.00875693466514349, + 0.01249772310256958, + 0.01693207584321499, + 0.023161910474300385, + 0.004725371487438679, + 0.008254075422883034, + 0.007682051043957472, + 0.01525590755045414, + 0.004626681562513113, + 0.007403580937534571, + 0.004322932101786137, + 0.007888766005635262, + 0.006164311431348324, + 0.006357253063470125, + 0.0072257122956216335, + 0.004465948324650526, + 0.012019108980894089, + 0.0058218976482748985, + 0.019657662138342857, + 0.008455168455839157, + 0.01745770499110222, + 0.005495274905115366, + 0.023700939491391182, + 0.008567098528146744, + 0.006947435904294252, + 0.004885511938482523, + 0.0065894112922251225, + 0.006660806946456432, + 0.008729713968932629, + 0.004057278390973806, + 0.012222891673445702, + 0.007789035327732563, + 0.005642409436404705, + 0.002905638888478279, + 0.006097902078181505, + 0.004102712497115135, + 0.00411633774638176, + 0.0038976336363703012, + 0.003731379983946681, + 0.004481417126953602, + 0.0038159508258104324, + 0.005180394276976585, + 0.005242059472948313, + 0.007820980623364449, + 0.004146681632846594, + 0.004526244942098856, + 0.0032559894025325775, + 0.005395511165261269, + 0.00883942749351263, + 0.009802654385566711, + 0.03623967990279198, + 0.05315800756216049, + 0.019690778106451035, + 0.00479603186249733, + 0.006017952226102352, + 0.009562275372445583 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 1 + ], + "selected_reward": -0.020193351432681084 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the initial shot a low-angle medium shot focusing on a pair of dark, muddy leather boots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does a figure wearing a heavy, forest-green wool cloak with a hood pulled low step down the last plank of a weathered wooden porch?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera execute a steady tracking motion backward and upward as the figure moves onto a narrow dirt path?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment revealed to include a path winding past a dilapidated porch into a vast landscape shrouded in thick, swirling morning mist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene rendered in a desaturated, cool palette of earthy browns, deep greens, and ethereal grays with faint, pearlescent dawn 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": [ + -0.009823550470173359 + ], + "opposite_logprob_tokens": [ + -4.634823322296143 + ], + "logprob_reward": -0.009823550470173359, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009823550470173359, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07049915194511414, + 0.07117446511983871, + 0.33963653445243835, + 0.11986842006444931, + 0.06961844861507416, + 0.10093587636947632, + 0.08671076595783234, + 0.025965772569179535, + 0.05389023199677467, + 0.061700329184532166 + ], + "credit_map": [ + 0.007573978044092655, + 0.006616462487727404, + 0.007835735566914082, + 0.008471743203699589, + 0.015383180230855942, + 0.006385486572980881, + 0.011450386606156826, + 0.006782184820622206, + 0.005247997585684061, + 0.0044136447831988335, + 0.015039093792438507, + 0.0060754260048270226, + 0.00906443316489458, + 0.003823236096650362, + 0.022352734580636024, + 0.005157903768122196, + 0.024076733738183975, + 0.12018931657075882, + 0.02100072242319584, + 0.013805011287331581, + 0.10537046194076538, + 0.02106805518269539, + 0.018103793263435364, + 0.01602242887020111, + 0.008776983246207237, + 0.03211037069559097, + 0.012751004658639431, + 0.022060804069042206, + 0.006115783005952835, + 0.011074593290686607, + 0.009206011891365051, + 0.017772873863577843, + 0.0045912242494523525, + 0.013876702636480331, + 0.0063821058720350266, + 0.014784148894250393, + 0.01025449950248003, + 0.0054660034365952015, + 0.009587761014699936, + 0.004676004406064749, + 0.011002447456121445, + 0.0067457095719873905, + 0.018430031836032867, + 0.008232063613831997, + 0.01626075990498066, + 0.006298164837062359, + 0.02247442491352558, + 0.01149227935820818, + 0.009234296157956123, + 0.0056516616605222225, + 0.0075797573663294315, + 0.00571095198392868, + 0.011153459548950195, + 0.0060910796746611595, + 0.026749122887849808, + 0.01454043947160244, + 0.002658837242051959, + 0.0033025131560862064, + 0.0030744357500225306, + 0.004470278043299913, + 0.0028703943826258183, + 0.0036362712271511555, + 0.0022948761470615864, + 0.0036581656895577908, + 0.005866652820259333, + 0.010947556234896183, + 0.00629290658980608, + 0.007986240088939667, + 0.006839174777269363, + 0.004529834724962711, + 0.004897968377918005, + 0.006529902573674917, + 0.007421172223985195, + 0.0065244692377746105, + 0.01067817211151123, + 0.012731516733765602, + 0.00786537118256092, + 0.0047382996417582035, + 0.0052547273226082325, + 0.0064866021275520325 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 3, + 5 + ], + "selected_reward": -0.009823550470173359 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the initial shot a low-angle medium shot focusing on a pair of dark, muddy leather boots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does a figure wearing a heavy, forest-green wool cloak with a hood pulled low step down the last plank of a weathered wooden porch?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera execute a steady tracking motion backward and upward as the figure moves onto a narrow dirt path?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment revealed to include a path winding past a dilapidated porch into a vast landscape shrouded in thick, swirling morning mist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene rendered in a desaturated, cool palette of earthy browns, deep greens, and ethereal grays with faint, pearlescent dawn 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": [ + -0.00596374599263072 + ], + "opposite_logprob_tokens": [ + -5.1309638023376465 + ], + "logprob_reward": -0.00596374599263072, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.00596374599263072, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.25236383080482483, + 0.020109860226511955, + 0.20091769099235535, + 0.1492975950241089, + 0.03299440070986748, + 0.10023445636034012, + 0.051012154668569565, + 0.09554693102836609, + 0.0437316857278347, + 0.05379146337509155 + ], + "credit_map": [ + 0.020553262904286385, + 0.026186538860201836, + 0.016831593587994576, + 0.021717043593525887, + 0.07040687650442123, + 0.021305890753865242, + 0.05281203240156174, + 0.022550564259290695, + 0.0024005556479096413, + 0.0018916652770712972, + 0.00216131960041821, + 0.00263915560208261, + 0.00299321417696774, + 0.0016805344494059682, + 0.0037217310164123774, + 0.0026216828264296055, + 0.01125524751842022, + 0.0767526775598526, + 0.00843759160488844, + 0.007604003883898258, + 0.06685242801904678, + 0.012287279590964317, + 0.009907650761306286, + 0.007820786908268929, + 0.012711303308606148, + 0.011577566154301167, + 0.040991757065057755, + 0.03857723996043205, + 0.0062393671832978725, + 0.010674245655536652, + 0.01010105386376381, + 0.01842505671083927, + 0.0033065497409552336, + 0.006681057158857584, + 0.0025447632651776075, + 0.006232197862118483, + 0.003331528278067708, + 0.0037898330483585596, + 0.004511198960244656, + 0.002597266109660268, + 0.011937347240746021, + 0.005327292717993259, + 0.017869995906949043, + 0.007957501336932182, + 0.018751444295048714, + 0.004885587841272354, + 0.02550460770726204, + 0.00800066627562046, + 0.006978423800319433, + 0.004463266581296921, + 0.0059435609728097916, + 0.005618925206363201, + 0.0058152019046247005, + 0.0034942461643368006, + 0.010711722075939178, + 0.007986803539097309, + 0.01393645815551281, + 0.010124363005161285, + 0.02310035191476345, + 0.016183877363801003, + 0.007250267080962658, + 0.00576513959094882, + 0.007033881731331348, + 0.012152575887739658, + 0.0045249550603330135, + 0.006495182402431965, + 0.005527016241103411, + 0.006687892600893974, + 0.005894719623029232, + 0.004337192513048649, + 0.0036898418329656124, + 0.006574878469109535, + 0.005595546681433916, + 0.00874232780188322, + 0.009535173885524273, + 0.011960508301854134, + 0.005315856542438269, + 0.003808064851909876, + 0.004179169423878193, + 0.004654809832572937 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 3 + ], + "selected_reward": -0.00596374599263072 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the initial shot a low-angle medium shot focusing on a pair of dark, muddy leather boots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does a figure wearing a heavy, forest-green wool cloak with a hood pulled low step down the last plank of a weathered wooden porch?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera execute a steady tracking motion backward and upward as the figure moves onto a narrow dirt path?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment revealed to include a path winding past a dilapidated porch into a vast landscape shrouded in thick, swirling morning mist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene rendered in a desaturated, cool palette of earthy browns, deep greens, and ethereal grays with faint, pearlescent dawn 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": [ + -0.0019418209558352828 + ], + "opposite_logprob_tokens": [ + -6.251941680908203 + ], + "logprob_reward": -0.0019418209558352828, + "yes_no_margin_reward": 6.25, + "token_credit_reward": -0.0019418209558352828, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.143304705619812, + 0.024734681472182274, + 0.039386432617902756, + 0.08898323774337769, + 0.023310301825404167, + 0.08927521109580994, + 0.16406090557575226, + 0.053242746740579605, + 0.13526341319084167, + 0.23843839764595032 + ], + "credit_map": [ + 0.012746436521410942, + 0.014562349766492844, + 0.008980289101600647, + 0.011302826926112175, + 0.04053093492984772, + 0.012260877527296543, + 0.030013902112841606, + 0.01290707103908062, + 0.002061741892248392, + 0.0022703539580106735, + 0.005087758414447308, + 0.0029312053229659796, + 0.004336402285844088, + 0.0015522799221798778, + 0.0032024523243308067, + 0.003292484674602747, + 0.0038074790500104427, + 0.012581033632159233, + 0.0032074800692498684, + 0.0024645961821079254, + 0.009060646407306194, + 0.0027984564658254385, + 0.00310504250228405, + 0.002361695049330592, + 0.007891430519521236, + 0.007000330835580826, + 0.02254723757505417, + 0.02151135355234146, + 0.003990457858890295, + 0.00609192531555891, + 0.008970529772341251, + 0.010979972779750824, + 0.0025698132812976837, + 0.00375084625557065, + 0.002075629774481058, + 0.00425179535523057, + 0.0027964257169514894, + 0.002772596897557378, + 0.0032760249450802803, + 0.0018171672709286213, + 0.010678695514798164, + 0.005192829295992851, + 0.01590975560247898, + 0.00753048388287425, + 0.01618793047964573, + 0.004050122108310461, + 0.022484734654426575, + 0.007240652106702328, + 0.019756929948925972, + 0.008773899637162685, + 0.014325774274766445, + 0.010812666267156601, + 0.020068375393748283, + 0.01005130261182785, + 0.05247112363576889, + 0.027800820767879486, + 0.0070071471855044365, + 0.00839979201555252, + 0.006666201166808605, + 0.007318882271647453, + 0.004417950287461281, + 0.005233496427536011, + 0.006110697984695435, + 0.008088573813438416, + 0.01405138149857521, + 0.016649603843688965, + 0.015190927311778069, + 0.028966257348656654, + 0.013060570694506168, + 0.01216138619929552, + 0.00964225921779871, + 0.025541020557284355, + 0.012468208558857441, + 0.012794018723070621, + 0.05825258046388626, + 0.08911380171775818, + 0.03429245203733444, + 0.0061939372681081295, + 0.010513076558709145, + 0.014810298569500446 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 6, + 0 + ], + "selected_reward": -0.0019418209558352828 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the initial shot a low-angle medium shot focusing on a pair of dark, muddy leather boots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does a figure wearing a heavy, forest-green wool cloak with a hood pulled low step down the last plank of a weathered wooden porch?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera execute a steady tracking motion backward and upward as the figure moves onto a narrow dirt path?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment revealed to include a path winding past a dilapidated porch into a vast landscape shrouded in thick, swirling morning mist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene rendered in a desaturated, cool palette of earthy browns, deep greens, and ethereal grays with faint, pearlescent dawn 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": [ + -0.0031996979378163815 + ], + "opposite_logprob_tokens": [ + -5.753199577331543 + ], + "logprob_reward": -0.0031996979378163815, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.0031996979378163815, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11973779648542404, + 0.08375707268714905, + 0.02141496166586876, + 0.04678771272301674, + 0.02252974733710289, + 0.08113548159599304, + 0.18782031536102295, + 0.021767815575003624, + 0.0563306026160717, + 0.35871848464012146 + ], + "credit_map": [ + 0.009122023358941078, + 0.011506101116538048, + 0.008910827338695526, + 0.01129602175205946, + 0.031995490193367004, + 0.011347435414791107, + 0.024481317028403282, + 0.011078571900725365, + 0.004532746504992247, + 0.004051226191222668, + 0.01788584142923355, + 0.005054190289229155, + 0.01223240327090025, + 0.004413966555148363, + 0.028646722435951233, + 0.006939976941794157, + 0.003464136738330126, + 0.0037853007670491934, + 0.003147741314023733, + 0.0019451328553259373, + 0.002746966667473316, + 0.0018327276920899749, + 0.002447029110044241, + 0.0020459271036088467, + 0.003908884711563587, + 0.004715989343822002, + 0.007667881436645985, + 0.008039549924433231, + 0.003098714165389538, + 0.006205669604241848, + 0.0069061401300132275, + 0.006244883872568607, + 0.0022654160857200623, + 0.0040621026419103146, + 0.001936614979058504, + 0.0038557974621653557, + 0.002157981274649501, + 0.002999256132170558, + 0.003427890595048666, + 0.0018246895633637905, + 0.009558971039950848, + 0.005200732033699751, + 0.014728783629834652, + 0.0067306444980204105, + 0.014417766593396664, + 0.003849519183859229, + 0.019778825342655182, + 0.0068702357821166515, + 0.023053523153066635, + 0.009772504679858685, + 0.016579771414399147, + 0.010708346031606197, + 0.02266676351428032, + 0.011240805499255657, + 0.06062118709087372, + 0.03317740932106972, + 0.0038427882827818394, + 0.001435266574844718, + 0.003486748319119215, + 0.002984510501846671, + 0.002310423878952861, + 0.002096801996231079, + 0.00281908200122416, + 0.0027921954169869423, + 0.006813493091613054, + 0.00564311258494854, + 0.008190402761101723, + 0.01175299659371376, + 0.004275349900126457, + 0.004903176799416542, + 0.0037722759880125523, + 0.01097980048507452, + 0.018077125772833824, + 0.02105109766125679, + 0.08115456998348236, + 0.1336704045534134, + 0.05760210007429123, + 0.008958583697676659, + 0.015865737572312355, + 0.022338880226016045 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 6, + 0 + ], + "selected_reward": -0.0031996979378163815 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle, medium shot first focuses on a pair of dark, muddy leather boots as a figure deliberately steps down the last plank of a weathered wooden porch, the old wood groaning softly under its weight. As the figure's form is obscured by a heavy, forest-green wool cloak, its hood pulled low, its feet firmly planted on a narrow dirt path, the camera begins a steady tracking motion backward and upward. The action follows the figures as they begin their journey, their backs to the unseen house, their postures resolute. The camera's retreat gradually widens the frame, revealing the full context: the path winds past the dilapidated porch and disappears into a vast, uncharted landscape, shrouded in a thick, swirling morning mist. Faint, pearlescent dawn light struggles to penetrate the mist, creating a soft, diffuse glow that captures the damp moss clinging to the gnarled roots of the ancient trees that flank the path. The scene is rendered in a desaturated, cool palette of earthy browns, deep greens, and ethereal grays, evoking the profound atmosphere of a solemn departure and the quiet beginning of an epic, uncertain quest.", + "seed": 4434, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000121/group_4/index_4694_seed_4434_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": -0.007442139461636544, + "vlm_logprob": -0.007442139461636544, + "vlm_yes_no_margin": 5.125, + "vlm_token_credit": -0.007442139461636544, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the initial shot a low-angle medium shot focusing on a pair of dark, muddy leather boots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does a figure wearing a heavy, forest-green wool cloak with a hood pulled low step down the last plank of a weathered wooden porch?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera execute a steady tracking motion backward and upward as the figure moves onto a narrow dirt path?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment revealed to include a path winding past a dilapidated porch into a vast landscape shrouded in thick, swirling morning mist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene rendered in a desaturated, cool palette of earthy browns, deep greens, and ethereal grays with faint, pearlescent dawn 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": [ + -0.012290103361010551 + ], + "opposite_logprob_tokens": [ + -4.887290000915527 + ], + "logprob_reward": -0.012290103361010551, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.012290103361010551, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13713234663009644, + 0.0914262905716896, + 0.3904390335083008, + 0.06887529790401459, + 0.051082443445920944, + 0.08418729156255722, + 0.092805877327919, + 0.022395877167582512, + 0.021533936262130737, + 0.0401216521859169 + ], + "credit_map": [ + 0.020162511616945267, + 0.026369178667664528, + 0.01441668625921011, + 0.016535909846425056, + 0.01290903054177761, + 0.011712574400007725, + 0.014366721734404564, + 0.020659735426306725, + 0.00937176588922739, + 0.013933178037405014, + 0.01503010094165802, + 0.011667810380458832, + 0.01285721454769373, + 0.012001781724393368, + 0.00998502317816019, + 0.006579416338354349, + 0.02551831491291523, + 0.06250600516796112, + 0.021422380581498146, + 0.16364240646362305, + 0.04544738307595253, + 0.012506846338510513, + 0.04464671015739441, + 0.014748993329703808, + 0.006696362514048815, + 0.008995907381176949, + 0.010828118771314621, + 0.010335855185985565, + 0.007277403026819229, + 0.007989258505403996, + 0.005423463881015778, + 0.011328932829201221, + 0.0038127764128148556, + 0.010859436355531216, + 0.0055445898324251175, + 0.008646451868116856, + 0.004555806517601013, + 0.00402371771633625, + 0.0073933894746005535, + 0.006246278528124094, + 0.009115422144532204, + 0.014761305414140224, + 0.00884758960455656, + 0.00864466279745102, + 0.0066937669180333614, + 0.009706418961286545, + 0.012016743421554565, + 0.014401384629309177, + 0.010552136227488518, + 0.004598991479724646, + 0.02719002030789852, + 0.006319292355328798, + 0.00671019870787859, + 0.00538188312202692, + 0.0228602085262537, + 0.00919315218925476, + 0.001908651553094387, + 0.0030339781660586596, + 0.0025835218839347363, + 0.0020127969328314066, + 0.002968286629766226, + 0.002832242986187339, + 0.0034019711893051863, + 0.0036544271279126406, + 0.0033256064634770155, + 0.002891330514103174, + 0.002996811643242836, + 0.002702069701626897, + 0.0023913912009447813, + 0.00189510069321841, + 0.002479806775227189, + 0.0028518212493509054, + 0.009239915758371353, + 0.007057354785501957, + 0.005707497708499432, + 0.004396588075906038, + 0.004048564471304417, + 0.0032904569525271654, + 0.0029443451203405857, + 0.0034369281493127346 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 6 + ], + "selected_reward": -0.012290103361010551 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the initial shot a low-angle medium shot focusing on a pair of dark, muddy leather boots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does a figure wearing a heavy, forest-green wool cloak with a hood pulled low step down the last plank of a weathered wooden porch?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera execute a steady tracking motion backward and upward as the figure moves onto a narrow dirt path?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment revealed to include a path winding past a dilapidated porch into a vast landscape shrouded in thick, swirling morning mist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene rendered in a desaturated, cool palette of earthy browns, deep greens, and ethereal grays with faint, pearlescent dawn 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": [ + -0.009819064289331436 + ], + "opposite_logprob_tokens": [ + -4.634819030761719 + ], + "logprob_reward": -0.009819064289331436, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009819064289331436, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1571187525987625, + 0.0585404634475708, + 0.27638864517211914, + 0.05014485865831375, + 0.09446653723716736, + 0.041862424463033676, + 0.02782571129500866, + 0.048448771238327026, + 0.03596861660480499, + 0.20923517644405365 + ], + "credit_map": [ + 0.012412453070282936, + 0.022484414279460907, + 0.02119433507323265, + 0.02057846635580063, + 0.018337305635213852, + 0.01374487392604351, + 0.018989592790603638, + 0.029377326369285583, + 0.0039621260948479176, + 0.007829797454178333, + 0.008779279887676239, + 0.007820308208465576, + 0.008104478009045124, + 0.008453402668237686, + 0.008539627306163311, + 0.005051448475569487, + 0.01738261990249157, + 0.04239703342318535, + 0.014697183854877949, + 0.11257025599479675, + 0.03276530280709267, + 0.012725531123578548, + 0.03125884383916855, + 0.012591884471476078, + 0.003038728376850486, + 0.011784547008574009, + 0.007852962240576744, + 0.009051061235368252, + 0.003481650957837701, + 0.005921205505728722, + 0.004563533701002598, + 0.004451173823326826, + 0.006468862295150757, + 0.021080046892166138, + 0.006819442845880985, + 0.024139275774359703, + 0.006998662371188402, + 0.009799656458199024, + 0.011433359235525131, + 0.007727230899035931, + 0.00258525088429451, + 0.005635117180645466, + 0.0035091196186840534, + 0.004803204908967018, + 0.0032573905773460865, + 0.006303778383880854, + 0.008482160046696663, + 0.007286404725164175, + 0.0027919821441173553, + 0.0020784337539225817, + 0.006689154542982578, + 0.0023818975314497948, + 0.003242045408114791, + 0.0025749467313289642, + 0.0036617040168493986, + 0.004405550193041563, + 0.005524607840925455, + 0.006688761990517378, + 0.007223841268569231, + 0.0034381647128611803, + 0.0034361248835921288, + 0.005841776728630066, + 0.00808449275791645, + 0.008211005479097366, + 0.00389997661113739, + 0.00615677097812295, + 0.006337949074804783, + 0.004622291307896376, + 0.0031725529115647078, + 0.0032521781977266073, + 0.004044832196086645, + 0.004482070449739695, + 0.0503116212785244, + 0.026490986347198486, + 0.06439486891031265, + 0.026368027552962303, + 0.0063996040262281895, + 0.007375330198556185, + 0.013085213489830494, + 0.014809530228376389 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 9, + 0 + ], + "selected_reward": -0.009819064289331436 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the initial shot a low-angle medium shot focusing on a pair of dark, muddy leather boots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does a figure wearing a heavy, forest-green wool cloak with a hood pulled low step down the last plank of a weathered wooden porch?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera execute a steady tracking motion backward and upward as the figure moves onto a narrow dirt path?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment revealed to include a path winding past a dilapidated porch into a vast landscape shrouded in thick, swirling morning mist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene rendered in a desaturated, cool palette of earthy browns, deep greens, and ethereal grays with faint, pearlescent dawn 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": [ + -0.0046503725461661816 + ], + "opposite_logprob_tokens": [ + -5.379650592803955 + ], + "logprob_reward": -0.0046503725461661816, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.0046503725461661816, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07543017715215683, + 0.08258894830942154, + 0.3990756571292877, + 0.05812152475118637, + 0.05084864795207977, + 0.02775539644062519, + 0.0385584831237793, + 0.05886199697852135, + 0.020165784284472466, + 0.18859341740608215 + ], + "credit_map": [ + 0.010809524916112423, + 0.011704613454639912, + 0.012554152868688107, + 0.008180622942745686, + 0.007160557899624109, + 0.007271712180227041, + 0.007241065613925457, + 0.010507929138839245, + 0.008287047035992146, + 0.014486156404018402, + 0.01329098828136921, + 0.00930880755186081, + 0.009207268245518208, + 0.009118900634348392, + 0.01342710666358471, + 0.0054626790806651115, + 0.02949349768459797, + 0.06319285184144974, + 0.022937268018722534, + 0.1654394268989563, + 0.046230100095272064, + 0.0123133584856987, + 0.044729363173246384, + 0.014739783480763435, + 0.004226852208375931, + 0.00918495561927557, + 0.008436226285994053, + 0.008143095299601555, + 0.003959891851991415, + 0.008267425000667572, + 0.004631184507161379, + 0.01127189677208662, + 0.004283200949430466, + 0.009727304801344872, + 0.0032760847825556993, + 0.014765326865017414, + 0.0030626594088971615, + 0.003233806462958455, + 0.007564560044556856, + 0.0049357027746737, + 0.0031491671688854694, + 0.003651274833828211, + 0.002989057684317231, + 0.0026351436972618103, + 0.002101973630487919, + 0.004089540336281061, + 0.004137955605983734, + 0.005001282785087824, + 0.0035956883803009987, + 0.005443071015179157, + 0.007439525797963142, + 0.0029208301566541195, + 0.003290597116574645, + 0.002307055750861764, + 0.008833085186779499, + 0.004728631116449833, + 0.005082913674414158, + 0.005811057984828949, + 0.00797725934535265, + 0.003631498431786895, + 0.004218964371830225, + 0.004862927366048098, + 0.012995127588510513, + 0.01428224891424179, + 0.002147821942344308, + 0.002564777387306094, + 0.003271574154496193, + 0.0023806910030543804, + 0.0013855606084689498, + 0.0025175062473863363, + 0.0029189721681177616, + 0.002978881821036339, + 0.028052864596247673, + 0.00889759510755539, + 0.0730394572019577, + 0.02537640556693077, + 0.008941574953496456, + 0.007085566874593496, + 0.01317406166344881, + 0.02402588538825512 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 9, + 1 + ], + "selected_reward": -0.0046503725461661816 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the initial shot a low-angle medium shot focusing on a pair of dark, muddy leather boots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does a figure wearing a heavy, forest-green wool cloak with a hood pulled low step down the last plank of a weathered wooden porch?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera execute a steady tracking motion backward and upward as the figure moves onto a narrow dirt path?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment revealed to include a path winding past a dilapidated porch into a vast landscape shrouded in thick, swirling morning mist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene rendered in a desaturated, cool palette of earthy browns, deep greens, and ethereal grays with faint, pearlescent dawn 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": [ + -0.002818067092448473 + ], + "opposite_logprob_tokens": [ + -5.8778181076049805 + ], + "logprob_reward": -0.002818067092448473, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.002818067092448473, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11773601919412613, + 0.05806555598974228, + 0.0903191864490509, + 0.06868501007556915, + 0.03631262853741646, + 0.05068881809711456, + 0.06242217496037483, + 0.133137047290802, + 0.12890006601810455, + 0.2537335455417633 + ], + "credit_map": [ + 0.014864569529891014, + 0.01776108890771866, + 0.014068922027945518, + 0.013222075067460537, + 0.012349709868431091, + 0.013122687116265297, + 0.014494111761450768, + 0.017852842807769775, + 0.006201134994626045, + 0.010381325148046017, + 0.010383239947259426, + 0.006639045663177967, + 0.005688267294317484, + 0.006827293895184994, + 0.007944716140627861, + 0.004000527318567038, + 0.007061407435685396, + 0.012743431143462658, + 0.009671897627413273, + 0.017788760364055634, + 0.008863809518516064, + 0.012449347414076328, + 0.012302962131798267, + 0.009437565691769123, + 0.0053725880570709705, + 0.009946594946086407, + 0.012855521403253078, + 0.009686806239187717, + 0.004522472620010376, + 0.008527434431016445, + 0.005723417736589909, + 0.012050166726112366, + 0.0023853445891290903, + 0.006195664405822754, + 0.0036449688486754894, + 0.008995341137051582, + 0.002816242165863514, + 0.003247134853154421, + 0.004794110078364611, + 0.004233820363879204, + 0.004071491304785013, + 0.005946297198534012, + 0.004195780027657747, + 0.005818542558699846, + 0.0044286069460213184, + 0.009575746022164822, + 0.006934203673154116, + 0.009718148969113827, + 0.006390407681465149, + 0.004078851081430912, + 0.005720148794353008, + 0.004636013880372047, + 0.00715645169839263, + 0.0049176812171936035, + 0.01739179529249668, + 0.01213082205504179, + 0.013933387584984303, + 0.01934071071445942, + 0.019898338243365288, + 0.00975647009909153, + 0.008737822994589806, + 0.013263879343867302, + 0.02182629145681858, + 0.026380129158496857, + 0.021109625697135925, + 0.028860801830887794, + 0.01987234689295292, + 0.012216172181069851, + 0.009426274336874485, + 0.00820721685886383, + 0.016724543645977974, + 0.012483065947890282, + 0.03747250512242317, + 0.010350115597248077, + 0.10441212356090546, + 0.03709584102034569, + 0.007495839148759842, + 0.006890356540679932, + 0.019711781293153763, + 0.030304960906505585 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 7, + 8 + ], + "selected_reward": -0.002818067092448473 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the initial shot a low-angle medium shot focusing on a pair of dark, muddy leather boots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does a figure wearing a heavy, forest-green wool cloak with a hood pulled low step down the last plank of a weathered wooden porch?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera execute a steady tracking motion backward and upward as the figure moves onto a narrow dirt path?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment revealed to include a path winding past a dilapidated porch into a vast landscape shrouded in thick, swirling morning mist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene rendered in a desaturated, cool palette of earthy browns, deep greens, and ethereal grays with faint, pearlescent dawn 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": [ + -0.007633090019226074 + ], + "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": [ + 0.11619850993156433, + 0.049850672483444214, + 0.3680669963359833, + 0.044017206877470016, + 0.07954361289739609, + 0.04177503660321236, + 0.05348868668079376, + 0.09057340770959854, + 0.041951145976781845, + 0.11453469842672348 + ], + "credit_map": [ + 0.014375539496541023, + 0.0170073751360178, + 0.015495168045163155, + 0.013090666383504868, + 0.011586179956793785, + 0.010864905081689358, + 0.015575158409774303, + 0.01820351742208004, + 0.0038357560988515615, + 0.007611976936459541, + 0.009083091281354427, + 0.009424008429050446, + 0.005781728308647871, + 0.0055853198282420635, + 0.005749173928052187, + 0.0027796190697699785, + 0.023211581632494926, + 0.05851379409432411, + 0.020384173840284348, + 0.15347714722156525, + 0.042085129767656326, + 0.014444909058511257, + 0.041295237839221954, + 0.014655051752924919, + 0.003405890893191099, + 0.009481167420744896, + 0.007371863350272179, + 0.0066462308168411255, + 0.0033310146536678076, + 0.004939627833664417, + 0.0035144148860126734, + 0.005326997488737106, + 0.004989952314645052, + 0.018096379935741425, + 0.00589882442727685, + 0.024139881134033203, + 0.004547274671494961, + 0.0055466219782829285, + 0.010575855150818825, + 0.005748825613409281, + 0.005130813457071781, + 0.0058366782031953335, + 0.005565233062952757, + 0.0036881673149764538, + 0.003638356225565076, + 0.004507311619818211, + 0.006936768535524607, + 0.006471707485616207, + 0.0056079369969666, + 0.003423244459554553, + 0.007529766298830509, + 0.0036342418752610683, + 0.005118987523019314, + 0.0038911052979528904, + 0.01550668478012085, + 0.008776718750596046, + 0.009055178612470627, + 0.01412448100745678, + 0.01061324030160904, + 0.0061196014285087585, + 0.0070763505063951015, + 0.008256718516349792, + 0.01626606285572052, + 0.019061774015426636, + 0.006042472552508116, + 0.009178116917610168, + 0.0063118101097643375, + 0.00413654837757349, + 0.0030858551617711782, + 0.003411484882235527, + 0.005717615131288767, + 0.004067244473844767, + 0.014633606187999249, + 0.005822605919092894, + 0.0361420102417469, + 0.016595439985394478, + 0.009644757024943829, + 0.00809410959482193, + 0.008203179575502872, + 0.015398996882140636 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 9 + ], + "selected_reward": -0.007633090019226074 + } + ], + "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/0000121/group_5/index_tensor([945, 945])_group_5_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000121/group_5/index_tensor([945, 945])_group_5_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..2f116558a6ea62ecaa16450702d129fb26636cb3 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000121/group_5/index_tensor([945, 945])_group_5_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "static, slightly high-angled medium shot frames a scene steeped in the minimalist elegance of traditional Japanese ink painting. A delicate watering can, crafted from antique bronze, displaying a soft patina and subtle emerald green, rests atop a meticulously cultivated potted pine. The tree, with its gnarled, ancient-looking trunk and dense, deep-green needles, sits within a simple, unadorned ceramic pot. Against a dark wooden surface, a soft, out-of-focus gray hue recalls the empty space of an ink scroll. In mesmerizing slow motion, a perfectly spherical drop of clear water deliberately oozes from the pot's slender spout, suspended for a gestational moment, reflecting diffuse light, before beginning its graceful descent. As the drop contacts the dark, damp soil, it erupts with a silent, beautiful splash, instantly transforming into a shimmering, liquid golden ink that radiates outward in a breathtaking corona. Its brilliant metallic sheen contrasts starkly and beautifully with the scene's otherwise muted, earthy palette, infusing it with a moment of profound, contemplative magic.", + "seed": 24623, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000121/group_5/index_945_seed_24623_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": -0.037772034853696825, + "vlm_logprob": -0.037772034853696825, + "vlm_yes_no_margin": 4.525, + "vlm_token_credit": -0.037772034853696825, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene presented in a static, slightly high-angled medium shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style consistent with traditional Japanese ink painting (minimalist elegance)?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a watering can made of antique bronze with a soft patina and subtle emerald green visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a perfectly spherical drop of clear water ooze from the pot's spout in slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the water drop transform into shimmering liquid golden ink upon hitting the soil?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02725399285554886 + ], + "opposite_logprob_tokens": [ + -7.027254104614258 + ], + "logprob_reward": -0.02725399285554886, + "yes_no_margin_reward": 7.0, + "token_credit_reward": -0.02725399285554886, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19454297423362732, + 0.059865739196538925, + 0.06087573990225792, + 0.21067200601100922, + 0.02763892523944378, + 0.06794866174459457, + 0.0757322907447815, + 0.07048802077770233, + 0.15168043971061707, + 0.08055519312620163 + ], + "credit_map": [ + 0.016898972913622856, + 0.024478457868099213, + 0.02605106122791767, + 0.029609736055135727, + 0.02502388320863247, + 0.021122362464666367, + 0.02365422621369362, + 0.0277042705565691, + 0.006664094515144825, + 0.008597450330853462, + 0.00821742694824934, + 0.009219007566571236, + 0.006215526722371578, + 0.006895532365888357, + 0.00717881927266717, + 0.006877881940454245, + 0.007760460022836924, + 0.011326445266604424, + 0.009906640276312828, + 0.005437307059764862, + 0.006588061805814505, + 0.008021092973649502, + 0.004880085121840239, + 0.006955641787499189, + 0.021906021982431412, + 0.037367191165685654, + 0.02523125521838665, + 0.03294616937637329, + 0.01914931647479534, + 0.020786894485354424, + 0.022152187302708626, + 0.031132984906435013, + 0.0030048158951103687, + 0.004271760117262602, + 0.0034824355971068144, + 0.0030231447890400887, + 0.003083539195358753, + 0.0036921098362654448, + 0.00343863177113235, + 0.0036424894351512194, + 0.007379861082881689, + 0.009150844067335129, + 0.01123007945716381, + 0.007853548973798752, + 0.006754586007446051, + 0.008997062221169472, + 0.006420916877686977, + 0.01016176212579012, + 0.007697275374084711, + 0.011131621897220612, + 0.00954437255859375, + 0.007031934801489115, + 0.008902129717171192, + 0.014009104110300541, + 0.006218770518898964, + 0.01119707990437746, + 0.0071406555362045765, + 0.008512424305081367, + 0.009671548381447792, + 0.007647079415619373, + 0.012047357857227325, + 0.005840474274009466, + 0.012227254919707775, + 0.007401225157082081, + 0.016746578738093376, + 0.021055111661553383, + 0.01929304748773575, + 0.016324488446116447, + 0.022478358820080757, + 0.023045698180794716, + 0.016667405143380165, + 0.016069753095507622, + 0.009849241003394127, + 0.013609785586595535, + 0.00959050003439188, + 0.009153027087450027, + 0.01208460796624422, + 0.008487265557050705, + 0.009873876348137856, + 0.007906895130872726 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 8 + ], + "selected_reward": -0.02725399285554886 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene presented in a static, slightly high-angled medium shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style consistent with traditional Japanese ink painting (minimalist elegance)?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a watering can made of antique bronze with a soft patina and subtle emerald green visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a perfectly spherical drop of clear water ooze from the pot's spout in slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the water drop transform into shimmering liquid golden ink upon hitting the soil?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08904905617237091 + ], + "opposite_logprob_tokens": [ + -2.4640491008758545 + ], + "logprob_reward": -0.08904905617237091, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08904905617237091, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13533736765384674, + 0.22707116603851318, + 0.07783839106559753, + 0.03759613633155823, + 0.05170370265841484, + 0.2603139877319336, + 0.0927468091249466, + 0.03945727273821831, + 0.03282909467816353, + 0.04510604962706566 + ], + "credit_map": [ + 0.012979308143258095, + 0.019564691931009293, + 0.018550017848610878, + 0.01791011355817318, + 0.01452952716499567, + 0.014498005621135235, + 0.018075821921229362, + 0.019229881465435028, + 0.022521628066897392, + 0.027336647734045982, + 0.027863087132573128, + 0.03700939193367958, + 0.026214681565761566, + 0.0334014929831028, + 0.024300936609506607, + 0.028423292562365532, + 0.008052737452089787, + 0.016181131824851036, + 0.010135386139154434, + 0.010908715426921844, + 0.008636251091957092, + 0.005981774069368839, + 0.00903937965631485, + 0.00890301913022995, + 0.003751646261662245, + 0.00706152617931366, + 0.006607289891690016, + 0.005040884483605623, + 0.003919337876141071, + 0.003544914536178112, + 0.003614655928686261, + 0.004055881407111883, + 0.0053040143102407455, + 0.007833365350961685, + 0.0078040314838290215, + 0.007009418681263924, + 0.005057152826339006, + 0.004470015875995159, + 0.006871568970382214, + 0.007354133762419224, + 0.02889774926006794, + 0.04399017617106438, + 0.039098966866731644, + 0.03224971890449524, + 0.026141390204429626, + 0.021300317719578743, + 0.026964163407683372, + 0.041671495884656906, + 0.010841351933777332, + 0.01234386395663023, + 0.011316891759634018, + 0.009722677059471607, + 0.008915870450437069, + 0.01237002294510603, + 0.00926780141890049, + 0.017968328669667244, + 0.004949428141117096, + 0.004198204260319471, + 0.006205142475664616, + 0.005006425082683563, + 0.004094429314136505, + 0.0031750230118632317, + 0.004327268339693546, + 0.007501353044062853, + 0.0034993102308362722, + 0.004441703669726849, + 0.00409616157412529, + 0.004964744206517935, + 0.0044659581035375595, + 0.004396907053887844, + 0.0034350447822362185, + 0.0035292631946504116, + 0.00588432839140296, + 0.00648265378549695, + 0.00754664046689868, + 0.00673568993806839, + 0.0058545381762087345, + 0.002585380570963025, + 0.005776192992925644, + 0.004240625072270632 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 1, + 0 + ], + "selected_reward": -0.08904905617237091 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene presented in a static, slightly high-angled medium shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style consistent with traditional Japanese ink painting (minimalist elegance)?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a watering can made of antique bronze with a soft patina and subtle emerald green visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a perfectly spherical drop of clear water ooze from the pot's spout in slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the water drop transform into shimmering liquid golden ink upon hitting the soil?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06204918026924133 + ], + "opposite_logprob_tokens": [ + -2.812049150466919 + ], + "logprob_reward": -0.06204918026924133, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06204918026924133, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3060462772846222, + 0.15930062532424927, + 0.08491375297307968, + 0.07912170886993408, + 0.1082296371459961, + 0.02880117855966091, + 0.05321024730801582, + 0.034798916429281235, + 0.06996091455221176, + 0.075616754591465 + ], + "credit_map": [ + 0.031578898429870605, + 0.04246543347835541, + 0.03854698687791824, + 0.044088054448366165, + 0.03420712798833847, + 0.037730053067207336, + 0.03820480778813362, + 0.03922489657998085, + 0.01433597318828106, + 0.01956922560930252, + 0.020063772797584534, + 0.026382673531770706, + 0.018482685089111328, + 0.026866428554058075, + 0.014537427574396133, + 0.019062433391809464, + 0.008690912276506424, + 0.010755822993814945, + 0.011289033107459545, + 0.011983506381511688, + 0.010272061452269554, + 0.014801761135458946, + 0.008217066526412964, + 0.008903589099645615, + 0.00675311079248786, + 0.01156479213386774, + 0.010566356591880322, + 0.012244077399373055, + 0.007880480028688908, + 0.011428826488554478, + 0.008942699991166592, + 0.00974136684089899, + 0.011454490944743156, + 0.01300333347171545, + 0.016734713688492775, + 0.014781957492232323, + 0.010627069510519505, + 0.013513053767383099, + 0.014299995265901089, + 0.013815023936331272, + 0.003385808551684022, + 0.0029885435942560434, + 0.005063876509666443, + 0.0036907843314111233, + 0.0027783012483268976, + 0.0030722764786332846, + 0.003995863255113363, + 0.003825725754722953, + 0.006497771013528109, + 0.007133715786039829, + 0.006699505262076855, + 0.005045752041041851, + 0.004867104347795248, + 0.006431027315557003, + 0.005090735387057066, + 0.011444637551903725, + 0.004454717971384525, + 0.0035776416771113873, + 0.005120737478137016, + 0.00510374503210187, + 0.0029386812821030617, + 0.002232542261481285, + 0.004865995608270168, + 0.006504856515675783, + 0.00739676458761096, + 0.009625595062971115, + 0.008518104441463947, + 0.006140751764178276, + 0.010531951673328876, + 0.01148220244795084, + 0.008215127512812614, + 0.008050411939620972, + 0.00816467497497797, + 0.011334876529872417, + 0.008236735127866268, + 0.005905545316636562, + 0.011325879022479057, + 0.013700447976589203, + 0.007116363849490881, + 0.009832235053181648 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 4 + ], + "selected_reward": -0.06204918026924133 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene presented in a static, slightly high-angled medium shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style consistent with traditional Japanese ink painting (minimalist elegance)?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a watering can made of antique bronze with a soft patina and subtle emerald green visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a perfectly spherical drop of clear water ooze from the pot's spout in slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the water drop transform into shimmering liquid golden ink upon hitting the soil?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005251660011708736 + ], + "opposite_logprob_tokens": [ + -5.255251884460449 + ], + "logprob_reward": -0.005251660011708736, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005251660011708736, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18221110105514526, + 0.1815817654132843, + 0.08447060734033585, + 0.10085006803274155, + 0.026037907227873802, + 0.03731587156653404, + 0.014158469624817371, + 0.08872522413730621, + 0.10859563946723938, + 0.17605330049991608 + ], + "credit_map": [ + 0.019082607701420784, + 0.026630938053131104, + 0.022223135456442833, + 0.028314242139458656, + 0.0199385117739439, + 0.019547853618860245, + 0.018618278205394745, + 0.027855543419718742, + 0.01743260584771633, + 0.019369369372725487, + 0.022208351641893387, + 0.029154932126402855, + 0.021320944651961327, + 0.032267287373542786, + 0.018015457317233086, + 0.021812837570905685, + 0.008668520487844944, + 0.012094122357666492, + 0.009627440012991428, + 0.013004058040678501, + 0.009582294151186943, + 0.015027298592031002, + 0.007360934745520353, + 0.009105936624109745, + 0.009380333125591278, + 0.012956924736499786, + 0.01459027174860239, + 0.0169976819306612, + 0.00970740057528019, + 0.013245327398180962, + 0.01071932353079319, + 0.013252799399197102, + 0.00230701407417655, + 0.0031247849110513926, + 0.0032771583646535873, + 0.0028113177977502346, + 0.002860414795577526, + 0.002365077380090952, + 0.005171461030840874, + 0.00412067910656333, + 0.0035067738499492407, + 0.0060049486346542835, + 0.0050680385902523994, + 0.006244150456041098, + 0.0038395607843995094, + 0.003289460903033614, + 0.003746120259165764, + 0.005616816692054272, + 0.0014582729199901223, + 0.0021214133594185114, + 0.0023635304532945156, + 0.0016141722444444895, + 0.0017458873335272074, + 0.001117610721848905, + 0.002038500504568219, + 0.0016990824369713664, + 0.009753088466823101, + 0.011097426526248455, + 0.009691828861832619, + 0.007388348691165447, + 0.013438543304800987, + 0.01609763689339161, + 0.00949897337704897, + 0.0117593789473176, + 0.010831957682967186, + 0.015400640666484833, + 0.015666048973798752, + 0.016864681616425514, + 0.015575706958770752, + 0.007573601324111223, + 0.015315220691263676, + 0.011367774568498135, + 0.021106118336319923, + 0.026161493733525276, + 0.027019692584872246, + 0.02436252497136593, + 0.026236049830913544, + 0.01288262102752924, + 0.022251620888710022, + 0.016033189371228218 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 9 + ], + "selected_reward": -0.005251660011708736 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene presented in a static, slightly high-angled medium shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style consistent with traditional Japanese ink painting (minimalist elegance)?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a watering can made of antique bronze with a soft patina and subtle emerald green visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a perfectly spherical drop of clear water ooze from the pot's spout in slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the water drop transform into shimmering liquid golden ink upon hitting the soil?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005256284959614277 + ], + "opposite_logprob_tokens": [ + -5.255256175994873 + ], + "logprob_reward": -0.005256284959614277, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005256284959614277, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19991013407707214, + 0.1456853300333023, + 0.04389328509569168, + 0.022672539576888084, + 0.07832624018192291, + 0.0713520497083664, + 0.05049321800470352, + 0.06545915454626083, + 0.1821311116218567, + 0.14007693529129028 + ], + "credit_map": [ + 0.023529568687081337, + 0.032134801149368286, + 0.023028453812003136, + 0.025989968329668045, + 0.018005074933171272, + 0.01629958674311638, + 0.026352759450674057, + 0.03456993028521538, + 0.014743153937160969, + 0.02108052186667919, + 0.01821296103298664, + 0.025018922984600067, + 0.01484856940805912, + 0.016362743452191353, + 0.017316356301307678, + 0.018102087080478668, + 0.005023716948926449, + 0.007112936116755009, + 0.006184940226376057, + 0.004603108856827021, + 0.003845225553959608, + 0.0057805320248007774, + 0.006059730425477028, + 0.005283096805214882, + 0.002110767876729369, + 0.0033764056861400604, + 0.003972744103521109, + 0.0026381155475974083, + 0.0025638004299253225, + 0.0025836622808128595, + 0.0031120956409722567, + 0.0023149491753429174, + 0.0070236981846392155, + 0.010541774332523346, + 0.010868394747376442, + 0.010187216103076935, + 0.008116527460515499, + 0.010334914550185204, + 0.011321991682052612, + 0.009931721724569798, + 0.007438036613166332, + 0.008575046434998512, + 0.010963141918182373, + 0.008567714132368565, + 0.006601841654628515, + 0.007483833469450474, + 0.00900418870151043, + 0.012718243524432182, + 0.005676920060068369, + 0.006825434044003487, + 0.0064006345346570015, + 0.005052628926932812, + 0.00540734501555562, + 0.006757283117622137, + 0.005034296773374081, + 0.009338678792119026, + 0.007742172572761774, + 0.007177487947046757, + 0.008948961272835732, + 0.00509507954120636, + 0.008759930729866028, + 0.007837187498807907, + 0.01192407961934805, + 0.007974251173436642, + 0.02053057961165905, + 0.029330911114811897, + 0.020566392689943314, + 0.013879704289138317, + 0.028799742460250854, + 0.030854498967528343, + 0.018258560448884964, + 0.019910713657736778, + 0.01996241882443428, + 0.01374009158462286, + 0.028279192745685577, + 0.020209603011608124, + 0.015104098245501518, + 0.006856762338429689, + 0.022688576951622963, + 0.013236192055046558 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 1 + ], + "selected_reward": -0.005256284959614277 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, slightly high-angled medium shot frames a scene steeped in the minimalist elegance of traditional Japanese ink painting. A delicate watering can, crafted from antique bronze, displaying a soft patina and subtle emerald green, rests atop a meticulously cultivated potted pine. The tree, with its gnarled, ancient-looking trunk and dense, deep-green needles, sits within a simple, unadorned ceramic pot. Against a dark wooden surface, a soft, out-of-focus gray hue recalls the empty space of an ink scroll. In mesmerizing slow motion, a perfectly spherical drop of clear water deliberately oozes from the pot's slender spout, suspended for a gestational moment, reflecting diffuse light, before beginning its graceful descent. As the drop contacts the dark, damp soil, it erupts with a silent, beautiful splash, instantly transforming into a shimmering, liquid golden ink that radiates outward in a breathtaking corona. Its brilliant metallic sheen contrasts starkly and beautifully with the scene's otherwise muted, earthy palette, infusing it with a moment of profound, contemplative magic.", + "seed": 24624, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000121/group_5/index_945_seed_24624_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": -0.041404615715146066, + "vlm_logprob": -0.041404615715146066, + "vlm_yes_no_margin": 3.9, + "vlm_token_credit": -0.041404615715146066, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene presented in a static, slightly high-angled medium shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style consistent with traditional Japanese ink painting (minimalist elegance)?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a watering can made of antique bronze with a soft patina and subtle emerald green visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a perfectly spherical drop of clear water ooze from the pot's spout in slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the water drop transform into shimmering liquid golden ink upon hitting the soil?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.022779574617743492 + ], + "opposite_logprob_tokens": [ + -6.14777946472168 + ], + "logprob_reward": -0.022779574617743492, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.022779574617743492, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1078634187579155, + 0.24457883834838867, + 0.018894195556640625, + 0.02907012775540352, + 0.030529823154211044, + 0.027147138491272926, + 0.37335464358329773, + 0.05756388232111931, + 0.060045406222343445, + 0.05095251649618149 + ], + "credit_map": [ + 0.03451094031333923, + 0.007100921124219894, + 0.03726617619395256, + 0.008333008736371994, + 0.003771968651562929, + 0.005208497866988182, + 0.005332431755959988, + 0.006339481566101313, + 0.08814536780118942, + 0.015533712692558765, + 0.08818325400352478, + 0.01680377684533596, + 0.006425526458770037, + 0.009369020350277424, + 0.0071031092666089535, + 0.013015102595090866, + 0.0027466369792819023, + 0.0029986214358359575, + 0.002728551160544157, + 0.0019388985820114613, + 0.001962933922186494, + 0.002025413326919079, + 0.0019944028463214636, + 0.0024987405631691217, + 0.00857915822416544, + 0.0023700387682765722, + 0.009328573942184448, + 0.002093593357130885, + 0.001111527904868126, + 0.0011827800190076232, + 0.0018847811734303832, + 0.002519675763323903, + 0.004145693499594927, + 0.007770376279950142, + 0.0033302148804068565, + 0.004336275160312653, + 0.003160210559144616, + 0.0029894241597503424, + 0.002370716305449605, + 0.002426913008093834, + 0.008421743288636208, + 0.0021006627939641476, + 0.008873384445905685, + 0.0024385100696235895, + 0.0009540045284666121, + 0.0011023130500689149, + 0.0014414830366149545, + 0.001815039082430303, + 0.1354689747095108, + 0.02405771054327488, + 0.13716700673103333, + 0.024242300540208817, + 0.008241694420576096, + 0.014513382688164711, + 0.006518011447042227, + 0.023145588114857674, + 0.01939326710999012, + 0.0036537342239171267, + 0.01990179903805256, + 0.003782460233196616, + 0.0017268545925617218, + 0.0025738703552633524, + 0.0022064002696424723, + 0.004325500689446926, + 0.021712342277169228, + 0.004507731180638075, + 0.019737977534532547, + 0.0038274317048490047, + 0.001897229696623981, + 0.002478240057826042, + 0.0023469801526516676, + 0.003537481650710106, + 0.008360086008906364, + 0.009325411170721054, + 0.006724328268319368, + 0.007396230474114418, + 0.005755083169788122, + 0.0033037776593118906, + 0.006376969162374735, + 0.003710636403411627 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 1, + 0 + ], + "selected_reward": -0.022779574617743492 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene presented in a static, slightly high-angled medium shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style consistent with traditional Japanese ink painting (minimalist elegance)?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a watering can made of antique bronze with a soft patina and subtle emerald green visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a perfectly spherical drop of clear water ooze from the pot's spout in slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the water drop transform into shimmering liquid golden ink upon hitting the soil?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08905243873596191 + ], + "opposite_logprob_tokens": [ + -2.464052438735962 + ], + "logprob_reward": -0.08905243873596191, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08905243873596191, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06779736280441284, + 0.06656629592180252, + 0.02686312422156334, + 0.4537813067436218, + 0.038870129734277725, + 0.05273715406656265, + 0.09714896231889725, + 0.09896549582481384, + 0.045873984694480896, + 0.05139622092247009 + ], + "credit_map": [ + 0.015777254477143288, + 0.007377756759524345, + 0.01964608021080494, + 0.0089723477140069, + 0.0033267096150666475, + 0.0035761960316449404, + 0.004931084346026182, + 0.0041899289935827255, + 0.02235718071460724, + 0.0038796612061560154, + 0.02324563078582287, + 0.0054889339953660965, + 0.0019570505246520042, + 0.0028516289312392473, + 0.002797180786728859, + 0.003989025019109249, + 0.00572749599814415, + 0.0026592374779284, + 0.007224116008728743, + 0.002194303320720792, + 0.0019485942320898175, + 0.0025040467735379934, + 0.0016837927978485823, + 0.0029215365648269653, + 0.16538375616073608, + 0.0273220706731081, + 0.1686108410358429, + 0.03014315478503704, + 0.009038706310093403, + 0.016673019155859947, + 0.01134483888745308, + 0.02526485174894333, + 0.009407462552189827, + 0.007617333438247442, + 0.008157645352184772, + 0.0037669993471354246, + 0.003110130550339818, + 0.0022084233351051807, + 0.002413827693089843, + 0.0021883053705096245, + 0.01743290200829506, + 0.0044972836039960384, + 0.019062872976064682, + 0.0035327738150954247, + 0.0012720946688205004, + 0.0020419179927557707, + 0.0016852733679115772, + 0.0032120291143655777, + 0.032153401523828506, + 0.008697587065398693, + 0.032412316650152206, + 0.007675490342080593, + 0.003061344614252448, + 0.0037641068920493126, + 0.002631158335134387, + 0.006753548979759216, + 0.03295071795582771, + 0.007881427183747292, + 0.03354867547750473, + 0.0061841364949941635, + 0.003517392324283719, + 0.00437477370724082, + 0.002918286481872201, + 0.007590084802359343, + 0.016665220260620117, + 0.003585874568670988, + 0.01569126918911934, + 0.0024716921616345644, + 0.001288259169086814, + 0.002025400986894965, + 0.0012214690214022994, + 0.002924792468547821, + 0.018082918599247932, + 0.004607155919075012, + 0.016314014792442322, + 0.003034790512174368, + 0.0020548852626234293, + 0.0020561814308166504, + 0.002271408913657069, + 0.002974856412038207 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 7, + 6 + ], + "selected_reward": -0.08905243873596191 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene presented in a static, slightly high-angled medium shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style consistent with traditional Japanese ink painting (minimalist elegance)?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a watering can made of antique bronze with a soft patina and subtle emerald green visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a perfectly spherical drop of clear water ooze from the pot's spout in slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the water drop transform into shimmering liquid golden ink upon hitting the soil?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06205982342362404 + ], + "opposite_logprob_tokens": [ + -2.8120598793029785 + ], + "logprob_reward": -0.06205982342362404, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06205982342362404, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.056218333542346954, + 0.0290791317820549, + 0.0169425830245018, + 0.6480700373649597, + 0.019200708717107773, + 0.029981018975377083, + 0.024427590891718864, + 0.09297864139080048, + 0.01661752350628376, + 0.06648439913988113 + ], + "credit_map": [ + 0.008284465409815311, + 0.008414043113589287, + 0.01100965030491352, + 0.010303094051778316, + 0.0029878588393330574, + 0.003637814661487937, + 0.006220133975148201, + 0.005361276678740978, + 0.00985939521342516, + 0.0021649436093866825, + 0.00927968043833971, + 0.002108084736391902, + 0.001316144596785307, + 0.0011714727152138948, + 0.0015921343583613634, + 0.0015872756484895945, + 0.00446196086704731, + 0.0011428389698266983, + 0.004783871583640575, + 0.0013442935887724161, + 0.0012314830673858523, + 0.0008847739663906395, + 0.0010869435500353575, + 0.002006416441872716, + 0.23562175035476685, + 0.039266958832740784, + 0.24092669785022736, + 0.043212514370679855, + 0.012972105294466019, + 0.023849288001656532, + 0.01628468558192253, + 0.03593601658940315, + 0.0025764715392142534, + 0.004822106566280127, + 0.002770158927887678, + 0.001942535862326622, + 0.002571079647168517, + 0.0017077227821573615, + 0.0015848991461098194, + 0.0012257361086085439, + 0.009952818043529987, + 0.002412824658676982, + 0.010822257958352566, + 0.002009135438129306, + 0.0006858599954284728, + 0.001179850660264492, + 0.0009377326932735741, + 0.0019805384799838066, + 0.008322642184793949, + 0.0017406464321538806, + 0.008444705046713352, + 0.0016179317608475685, + 0.0006979804602451622, + 0.0009705395204946399, + 0.001072933548130095, + 0.0015602123457938433, + 0.03265867754817009, + 0.006478773429989815, + 0.033015258610248566, + 0.0054280138574540615, + 0.002217636676505208, + 0.004102928563952446, + 0.002165845362469554, + 0.006911503616720438, + 0.0034054864663630724, + 0.0033694638404995203, + 0.0038312918040901423, + 0.001438742969185114, + 0.0013654828071594238, + 0.0011794341262429953, + 0.0009387290338054299, + 0.0010888922261074185, + 0.01639709435403347, + 0.011703316122293472, + 0.013346978463232517, + 0.006144045386463404, + 0.005769461393356323, + 0.003460707375779748, + 0.005952144041657448, + 0.0037106513045728207 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 7, + 9 + ], + "selected_reward": -0.06205982342362404 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene presented in a static, slightly high-angled medium shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style consistent with traditional Japanese ink painting (minimalist elegance)?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a watering can made of antique bronze with a soft patina and subtle emerald green visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a perfectly spherical drop of clear water ooze from the pot's spout in slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the water drop transform into shimmering liquid golden ink upon hitting the soil?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012539708986878395 + ], + "opposite_logprob_tokens": [ + -4.387539863586426 + ], + "logprob_reward": -0.012539708986878395, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012539708986878395, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07451321929693222, + 0.21046049892902374, + 0.021403800696134567, + 0.13362477719783783, + 0.0356317013502121, + 0.03069395199418068, + 0.13581940531730652, + 0.0618339478969574, + 0.02536013163626194, + 0.27065858244895935 + ], + "credit_map": [ + 0.015241983346641064, + 0.010596465319395065, + 0.020414860919117928, + 0.00992116890847683, + 0.0038274433463811874, + 0.004323325119912624, + 0.005160465836524963, + 0.005027507897466421, + 0.07605431973934174, + 0.012860246933996677, + 0.0771176815032959, + 0.014097343198955059, + 0.0047453916631639, + 0.007749889045953751, + 0.0058515495620667934, + 0.011984076350927353, + 0.004840850364416838, + 0.002135469578206539, + 0.006171360146254301, + 0.00214927620254457, + 0.0011099271941930056, + 0.001371968537569046, + 0.0017760144546627998, + 0.001848933519795537, + 0.04848496615886688, + 0.008324950933456421, + 0.04927287623286247, + 0.008996703661978245, + 0.0027264272794127464, + 0.004851756151765585, + 0.0034840861335396767, + 0.007483022753149271, + 0.011900100857019424, + 0.0024145508650690317, + 0.012198539450764656, + 0.0024257078766822815, + 0.0017395875183865428, + 0.001753867487423122, + 0.0013156788190826774, + 0.0018836662638932467, + 0.009772722609341145, + 0.0023587567266076803, + 0.010340452194213867, + 0.00288121635094285, + 0.0009922570316120982, + 0.0011372157605364919, + 0.001377892796881497, + 0.0018334370106458664, + 0.046347614377737045, + 0.008600838482379913, + 0.04980650916695595, + 0.010079286061227322, + 0.003149050986394286, + 0.005464759189635515, + 0.0031886256765574217, + 0.009182724170386791, + 0.021952280774712563, + 0.003992903046309948, + 0.020665807649493217, + 0.0038470495492219925, + 0.0018631468992680311, + 0.002453405875712633, + 0.0028381324373185635, + 0.004221226088702679, + 0.006800848059356213, + 0.0043258643709123135, + 0.005017428193241358, + 0.0023700736928731203, + 0.0020655596163123846, + 0.0015479165595024824, + 0.001758819678798318, + 0.001473622745834291, + 0.08999404311180115, + 0.026972971856594086, + 0.08140599727630615, + 0.018879232928156853, + 0.013167232275009155, + 0.011176981031894684, + 0.01413164846599102, + 0.014930474571883678 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 1, + 6 + ], + "selected_reward": -0.012539708986878395 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene presented in a static, slightly high-angled medium shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style consistent with traditional Japanese ink painting (minimalist elegance)?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a watering can made of antique bronze with a soft patina and subtle emerald green visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a perfectly spherical drop of clear water ooze from the pot's spout in slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the water drop transform into shimmering liquid golden ink upon hitting the soil?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020591532811522484 + ], + "opposite_logprob_tokens": [ + -3.8955914974212646 + ], + "logprob_reward": -0.020591532811522484, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020591532811522484, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06891535222530365, + 0.026917781680822372, + 0.016202429309487343, + 0.06394414603710175, + 0.02386157587170601, + 0.023143745958805084, + 0.05121198669075966, + 0.05231967195868492, + 0.03210887685418129, + 0.641374409198761 + ], + "credit_map": [ + 0.019210029393434525, + 0.006673940923064947, + 0.023086117580533028, + 0.005527340807020664, + 0.002725901547819376, + 0.003463942790403962, + 0.004159728996455669, + 0.004068349488079548, + 0.006013907492160797, + 0.0026437919586896896, + 0.006419653072953224, + 0.002402448793873191, + 0.002065777312964201, + 0.002920586848631501, + 0.002260429784655571, + 0.0021911868825554848, + 0.0044230748899281025, + 0.0012491535162553191, + 0.004950432106852531, + 0.0015652672154828906, + 0.0008895283099263906, + 0.0008935460937209427, + 0.0011488981544971466, + 0.0010825295466929674, + 0.021794233471155167, + 0.003924969583749771, + 0.022055139765143394, + 0.0047997282817959785, + 0.0025920795742422342, + 0.0027156744617968798, + 0.002302501816302538, + 0.003759816288948059, + 0.007842019200325012, + 0.0018705110996961594, + 0.007872534915804863, + 0.0018634641310200095, + 0.0007798201404511929, + 0.001027242629788816, + 0.0014141013380140066, + 0.0011918836971744895, + 0.007447212003171444, + 0.0015316790668293834, + 0.007814137265086174, + 0.0020857425406575203, + 0.0006766283768229187, + 0.0009604755323380232, + 0.0011082511628046632, + 0.0015196185559034348, + 0.015907879918813705, + 0.004429080057889223, + 0.01733716018497944, + 0.004243177827447653, + 0.0018030557548627257, + 0.0021736917551606894, + 0.0016393413534387946, + 0.003678601700812578, + 0.017436832189559937, + 0.004226361867040396, + 0.01591200940310955, + 0.0034601360093802214, + 0.002115411451086402, + 0.002307467395439744, + 0.002928414847701788, + 0.003933041822165251, + 0.009742402471601963, + 0.004562026355415583, + 0.007675597909837961, + 0.00253181136213243, + 0.0021747169084846973, + 0.001671128673478961, + 0.0019402733305469155, + 0.0018109193770214915, + 0.2193658947944641, + 0.04907967895269394, + 0.20825710892677307, + 0.04506419971585274, + 0.02408641204237938, + 0.026591073721647263, + 0.03298158198595047, + 0.035948481410741806 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 3 + ], + "selected_reward": -0.020591532811522484 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, slightly high-angled medium shot frames a scene steeped in the minimalist elegance of traditional Japanese ink painting. A delicate watering can, crafted from antique bronze, displaying a soft patina and subtle emerald green, rests atop a meticulously cultivated potted pine. The tree, with its gnarled, ancient-looking trunk and dense, deep-green needles, sits within a simple, unadorned ceramic pot. Against a dark wooden surface, a soft, out-of-focus gray hue recalls the empty space of an ink scroll. In mesmerizing slow motion, a perfectly spherical drop of clear water deliberately oozes from the pot's slender spout, suspended for a gestational moment, reflecting diffuse light, before beginning its graceful descent. As the drop contacts the dark, damp soil, it erupts with a silent, beautiful splash, instantly transforming into a shimmering, liquid golden ink that radiates outward in a breathtaking corona. Its brilliant metallic sheen contrasts starkly and beautifully with the scene's otherwise muted, earthy palette, infusing it with a moment of profound, contemplative magic.", + "seed": 24625, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000121/group_5/index_945_seed_24625_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": -0.0885607697069645, + "vlm_logprob": -0.0885607697069645, + "vlm_yes_no_margin": 2.9749999761581423, + "vlm_token_credit": -0.0885607697069645, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene presented in a static, slightly high-angled medium shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style consistent with traditional Japanese ink painting (minimalist elegance)?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a watering can made of antique bronze with a soft patina and subtle emerald green visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a perfectly spherical drop of clear water ooze from the pot's spout in slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the water drop transform into shimmering liquid golden ink upon hitting the soil?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.034401971846818924 + ], + "opposite_logprob_tokens": [ + -5.409401893615723 + ], + "logprob_reward": -0.034401971846818924, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.034401971846818924, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3462507128715515, + 0.0655490979552269, + 0.03290111571550369, + 0.05568215250968933, + 0.06308849900960922, + 0.0886777937412262, + 0.11240284144878387, + 0.11288043856620789, + 0.041738253086805344, + 0.08082916587591171 + ], + "credit_map": [ + 0.04357899725437164, + 0.045014843344688416, + 0.04157580807805061, + 0.04703288897871971, + 0.05006789788603783, + 0.04755336791276932, + 0.04051961004734039, + 0.030907263979315758, + 0.0069546992890536785, + 0.008626038208603859, + 0.01036662794649601, + 0.009824097156524658, + 0.008232412859797478, + 0.009306934662163258, + 0.006216667592525482, + 0.0060216207057237625, + 0.0024306837003678083, + 0.0036452016793191433, + 0.0037477905862033367, + 0.0047839037142694, + 0.004099439363926649, + 0.0044797915033996105, + 0.004241117741912603, + 0.005473187658935785, + 0.005556109361350536, + 0.00715499185025692, + 0.00944098737090826, + 0.008075281977653503, + 0.007272371090948582, + 0.00696109514683485, + 0.004941244143992662, + 0.006280071102082729, + 0.012625695206224918, + 0.010171784088015556, + 0.00944228284060955, + 0.004533464554697275, + 0.006211344618350267, + 0.010945053771138191, + 0.004293280187994242, + 0.004865589551627636, + 0.009447473101317883, + 0.010631128214299679, + 0.011307599022984505, + 0.010654703713953495, + 0.012480063363909721, + 0.01928633451461792, + 0.007448517717421055, + 0.0074219778180122375, + 0.012952997349202633, + 0.02432049997150898, + 0.015018039382994175, + 0.014854376204311848, + 0.010109351947903633, + 0.018426669761538506, + 0.008543942123651505, + 0.008176970295608044, + 0.010823667049407959, + 0.022723421454429626, + 0.015447312965989113, + 0.01583932712674141, + 0.01156723964959383, + 0.016291102394461632, + 0.009877050295472145, + 0.010311314836144447, + 0.00458401907235384, + 0.0050667654722929, + 0.005861676763743162, + 0.004156031180173159, + 0.0035465231630951166, + 0.010278087109327316, + 0.004228607285767794, + 0.004016542341560125, + 0.006820210721343756, + 0.010302177630364895, + 0.0078044612891972065, + 0.015462975949048996, + 0.008014055900275707, + 0.012405361980199814, + 0.011879785917699337, + 0.008140130899846554 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 6 + ], + "selected_reward": -0.034401971846818924 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene presented in a static, slightly high-angled medium shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style consistent with traditional Japanese ink painting (minimalist elegance)?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a watering can made of antique bronze with a soft patina and subtle emerald green visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a perfectly spherical drop of clear water ooze from the pot's spout in slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the water drop transform into shimmering liquid golden ink upon hitting the soil?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"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": [ + 0.3412683308124542, + 0.0504143089056015, + 0.03873544931411743, + 0.04079164192080498, + 0.20012718439102173, + 0.0717366561293602, + 0.11949082463979721, + 0.07759515941143036, + 0.03549414128065109, + 0.024346239864826202 + ], + "credit_map": [ + 0.040771499276161194, + 0.03252052143216133, + 0.053523000329732895, + 0.04851626604795456, + 0.044237490743398666, + 0.04371442645788193, + 0.03905192390084267, + 0.03893322870135307, + 0.0046735187061131, + 0.00636273855343461, + 0.006034019868820906, + 0.005399317014962435, + 0.007527659647166729, + 0.006920264568179846, + 0.007271953392773867, + 0.0062248436734080315, + 0.004252893850207329, + 0.004715478513389826, + 0.00560815678909421, + 0.005020801909267902, + 0.004611512646079063, + 0.005842791870236397, + 0.004355229437351227, + 0.004328589420765638, + 0.004959230776876211, + 0.0053261336870491505, + 0.005809021648019552, + 0.004090915434062481, + 0.00641285628080368, + 0.005755401682108641, + 0.003383366856724024, + 0.005054718814790249, + 0.02267875149846077, + 0.03321735933423042, + 0.023971693590283394, + 0.023869890719652176, + 0.03582354262471199, + 0.026761699467897415, + 0.018007678911089897, + 0.015796585008502007, + 0.007583863567560911, + 0.014497941359877586, + 0.010625057853758335, + 0.007103465497493744, + 0.00946056004613638, + 0.00817681010812521, + 0.007479079533368349, + 0.006809882819652557, + 0.016855284571647644, + 0.028738562017679214, + 0.015965262427926064, + 0.0073315296322107315, + 0.010225328616797924, + 0.023639405146241188, + 0.007167493458837271, + 0.009567970409989357, + 0.010126672685146332, + 0.013197189196944237, + 0.01024987455457449, + 0.008832897990942001, + 0.007907237857580185, + 0.010082907043397427, + 0.00847640074789524, + 0.008721988648176193, + 0.005610412918031216, + 0.005544675979763269, + 0.003934512380510569, + 0.002914808690547943, + 0.003331428859382868, + 0.007428452838212252, + 0.003707251278683543, + 0.003022602992132306, + 0.002433182904496789, + 0.004852261394262314, + 0.0036865812726318836, + 0.001855557202361524, + 0.0020676522981375456, + 0.0029780662152916193, + 0.0031381419394165277, + 0.0033347993157804012 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 6 + ], + "selected_reward": -0.08904688060283661 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene presented in a static, slightly high-angled medium shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style consistent with traditional Japanese ink painting (minimalist elegance)?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a watering can made of antique bronze with a soft patina and subtle emerald green visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a perfectly spherical drop of clear water ooze from the pot's spout in slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the water drop transform into shimmering liquid golden ink upon hitting the soil?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.14281399548053741 + ], + "opposite_logprob_tokens": [ + -2.0178139209747314 + ], + "logprob_reward": -0.14281399548053741, + "yes_no_margin_reward": 1.8749998807907104, + "token_credit_reward": -0.14281399548053741, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3729485869407654, + 0.07535108178853989, + 0.07215020060539246, + 0.13269884884357452, + 0.04200582206249237, + 0.05141730234026909, + 0.060906507074832916, + 0.07845015078783035, + 0.03244553133845329, + 0.08162591606378555 + ], + "credit_map": [ + 0.04709785804152489, + 0.039099957793951035, + 0.06834954768419266, + 0.04304174333810806, + 0.048823583871126175, + 0.04950980842113495, + 0.04501982405781746, + 0.032006263732910156, + 0.007272458169609308, + 0.007754203863441944, + 0.011698856949806213, + 0.009419582784175873, + 0.012804246507585049, + 0.010538688860833645, + 0.009610473178327084, + 0.006252567283809185, + 0.006822189316153526, + 0.008338609710335732, + 0.010157006792724133, + 0.012075122445821762, + 0.008502991870045662, + 0.010591551661491394, + 0.00893145240843296, + 0.006731288507580757, + 0.013871765695512295, + 0.019680500030517578, + 0.01688901148736477, + 0.01645561307668686, + 0.023657873272895813, + 0.018719753250479698, + 0.012916427105665207, + 0.010507921688258648, + 0.005457399878650904, + 0.007670893799513578, + 0.006548221223056316, + 0.0034654971677809954, + 0.005979730281978846, + 0.0066699194721877575, + 0.0030989826191216707, + 0.0031151832081377506, + 0.009822488762438297, + 0.005714914761483669, + 0.00921957939863205, + 0.004278988111764193, + 0.006674290634691715, + 0.007015479728579521, + 0.004749753978103399, + 0.003941809292882681, + 0.006791506893932819, + 0.012863288633525372, + 0.0071958391927182674, + 0.008258731104433537, + 0.006657008547335863, + 0.008138985373079777, + 0.005458971951156855, + 0.005542182829231024, + 0.007781079970300198, + 0.011359010823071003, + 0.009808183647692204, + 0.009666561149060726, + 0.008864990435540676, + 0.013490856625139713, + 0.008445416577160358, + 0.009034055285155773, + 0.0032838021870702505, + 0.005073792766779661, + 0.0038887697737663984, + 0.00398539612069726, + 0.003636663779616356, + 0.005720502231270075, + 0.0034379891585558653, + 0.0034186160191893578, + 0.009648120030760765, + 0.010306209325790405, + 0.009582617320120335, + 0.01162132527679205, + 0.010153600014746189, + 0.009579949080944061, + 0.010563281364738941, + 0.01017081830650568 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 9 + ], + "selected_reward": -0.14281399548053741 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene presented in a static, slightly high-angled medium shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style consistent with traditional Japanese ink painting (minimalist elegance)?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a watering can made of antique bronze with a soft patina and subtle emerald green visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a perfectly spherical drop of clear water ooze from the pot's spout in slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the water drop transform into shimmering liquid golden ink upon hitting the soil?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03370778262615204 + ], + "opposite_logprob_tokens": [ + -3.408707857131958 + ], + "logprob_reward": -0.03370778262615204, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03370778262615204, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3595997095108032, + 0.08107462525367737, + 0.059795405715703964, + 0.07689066231250763, + 0.08334365487098694, + 0.0527566634118557, + 0.07055838406085968, + 0.08138362318277359, + 0.0821860283613205, + 0.052411239594221115 + ], + "credit_map": [ + 0.0415460579097271, + 0.03952373191714287, + 0.044075798243284225, + 0.05106440931558609, + 0.05188009515404701, + 0.045644938945770264, + 0.04939183592796326, + 0.03647281602025032, + 0.009687955491244793, + 0.012025906704366207, + 0.009788946248590946, + 0.01065654307603836, + 0.010963904671370983, + 0.010565460659563541, + 0.008222105912864208, + 0.009163806214928627, + 0.006208166014403105, + 0.006771143060177565, + 0.008855964057147503, + 0.008970524184405804, + 0.007972917519509792, + 0.008272739127278328, + 0.00662292679771781, + 0.006121025420725346, + 0.00873236358165741, + 0.013617374934256077, + 0.01021605171263218, + 0.0053813280537724495, + 0.01383861992508173, + 0.01007173303514719, + 0.008109943941235542, + 0.00692323874682188, + 0.008526426739990711, + 0.012684714049100876, + 0.013638640753924847, + 0.01168185006827116, + 0.008663511835038662, + 0.01690826751291752, + 0.0063664899207651615, + 0.004873753525316715, + 0.008852959610521793, + 0.005733297672122717, + 0.012422316707670689, + 0.005677121691405773, + 0.0056171175092458725, + 0.0069326795637607574, + 0.0036660348996520042, + 0.0038551348261535168, + 0.008009872399270535, + 0.011445817537605762, + 0.009018671698868275, + 0.008062141947448254, + 0.00983875896781683, + 0.010680856183171272, + 0.007191927172243595, + 0.006310335826128721, + 0.00651724124327302, + 0.00912025012075901, + 0.011153330095112324, + 0.011214135214686394, + 0.009689176455140114, + 0.015034257434308529, + 0.008520511910319328, + 0.010134723037481308, + 0.007096794433891773, + 0.013333400711417198, + 0.010777047835290432, + 0.012627930380403996, + 0.00883612409234047, + 0.010912995785474777, + 0.009922838769853115, + 0.00867889728397131, + 0.006127092521637678, + 0.006365950219333172, + 0.00734527874737978, + 0.006499860435724258, + 0.005435366649180651, + 0.008030024357140064, + 0.00604912219569087, + 0.006558545865118504 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 8 + ], + "selected_reward": -0.03370778262615204 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene presented in a static, slightly high-angled medium shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style consistent with traditional Japanese ink painting (minimalist elegance)?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a watering can made of antique bronze with a soft patina and subtle emerald green visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a perfectly spherical drop of clear water ooze from the pot's spout in slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the water drop transform into shimmering liquid golden ink upon hitting the soil?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.14283321797847748 + ], + "opposite_logprob_tokens": [ + -2.0178332328796387 + ], + "logprob_reward": -0.14283321797847748, + "yes_no_margin_reward": 1.875, + "token_credit_reward": -0.14283321797847748, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3013475835323334, + 0.13792684674263, + 0.08448686450719833, + 0.06351830810308456, + 0.07996492087841034, + 0.07909262925386429, + 0.0666390061378479, + 0.08028895407915115, + 0.04911605641245842, + 0.057618919759988785 + ], + "credit_map": [ + 0.034555304795503616, + 0.040318649262189865, + 0.05152879282832146, + 0.03269541263580322, + 0.03998608514666557, + 0.04570465534925461, + 0.025502976030111313, + 0.03105570189654827, + 0.016723209992051125, + 0.01727956160902977, + 0.02222558669745922, + 0.019836649298667908, + 0.016301410272717476, + 0.019473521038889885, + 0.013174638152122498, + 0.012912261299788952, + 0.008784192614257336, + 0.008442476391792297, + 0.010092437267303467, + 0.015642808750271797, + 0.010946525260806084, + 0.013636244460940361, + 0.009831667877733707, + 0.007110514212399721, + 0.006177523173391819, + 0.00627133110538125, + 0.009113765321671963, + 0.012390093877911568, + 0.0070198094472289085, + 0.00848308578133583, + 0.00699700228869915, + 0.007065690588206053, + 0.008711325004696846, + 0.010982642881572247, + 0.01310779806226492, + 0.005873735062777996, + 0.010373832657933235, + 0.018201395869255066, + 0.006544293370097876, + 0.006169890519231558, + 0.01140561606734991, + 0.009746324270963669, + 0.009764578193426132, + 0.0071539548225700855, + 0.010645974427461624, + 0.01768234744668007, + 0.006246452685445547, + 0.006447367835789919, + 0.007540373131632805, + 0.009993743151426315, + 0.010004019364714622, + 0.008354933001101017, + 0.007608420215547085, + 0.008060107938945293, + 0.006864678580313921, + 0.008212722837924957, + 0.0076841507107019424, + 0.011011818423867226, + 0.010111753828823566, + 0.011699865572154522, + 0.00933997891843319, + 0.012091767974197865, + 0.00866603571921587, + 0.009683574549853802, + 0.005336469039320946, + 0.006817953661084175, + 0.007459522224962711, + 0.0046171327121555805, + 0.005662394221872091, + 0.005868701729923487, + 0.006188654340803623, + 0.007165227085351944, + 0.00543648935854435, + 0.00711797783151269, + 0.007335296832025051, + 0.005236811935901642, + 0.004787679761648178, + 0.008544368669390678, + 0.008570054545998573, + 0.010590237565338612 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.14283321797847748 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, slightly high-angled medium shot frames a scene steeped in the minimalist elegance of traditional Japanese ink painting. A delicate watering can, crafted from antique bronze, displaying a soft patina and subtle emerald green, rests atop a meticulously cultivated potted pine. The tree, with its gnarled, ancient-looking trunk and dense, deep-green needles, sits within a simple, unadorned ceramic pot. Against a dark wooden surface, a soft, out-of-focus gray hue recalls the empty space of an ink scroll. In mesmerizing slow motion, a perfectly spherical drop of clear water deliberately oozes from the pot's slender spout, suspended for a gestational moment, reflecting diffuse light, before beginning its graceful descent. As the drop contacts the dark, damp soil, it erupts with a silent, beautiful splash, instantly transforming into a shimmering, liquid golden ink that radiates outward in a breathtaking corona. Its brilliant metallic sheen contrasts starkly and beautifully with the scene's otherwise muted, earthy palette, infusing it with a moment of profound, contemplative magic.", + "seed": 24626, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000121/group_5/index_945_seed_24626_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": -0.11431298982352019, + "vlm_logprob": -0.11431298982352019, + "vlm_yes_no_margin": 3.3999999880790712, + "vlm_token_credit": -0.11431298982352019, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene presented in a static, slightly high-angled medium shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style consistent with traditional Japanese ink painting (minimalist elegance)?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a watering can made of antique bronze with a soft patina and subtle emerald green visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a perfectly spherical drop of clear water ooze from the pot's spout in slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the water drop transform into shimmering liquid golden ink upon hitting the soil?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.031122779473662376 + ], + "opposite_logprob_tokens": [ + -6.65612268447876 + ], + "logprob_reward": -0.031122779473662376, + "yes_no_margin_reward": 6.625, + "token_credit_reward": -0.031122779473662376, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14138028025627136, + 0.04890687018632889, + 0.16600514948368073, + 0.10497923195362091, + 0.07516549527645111, + 0.09401700645685196, + 0.09894801676273346, + 0.10919945687055588, + 0.10367611795663834, + 0.05772232636809349 + ], + "credit_map": [ + 0.02090913988649845, + 0.020494753494858742, + 0.022256722673773766, + 0.0109154237434268, + 0.02067975513637066, + 0.014843114651739597, + 0.012629218399524689, + 0.018652157858014107, + 0.004135968629270792, + 0.007357263937592506, + 0.005814767442643642, + 0.004396543372422457, + 0.007125317119061947, + 0.007055880036205053, + 0.006881849840283394, + 0.006139282137155533, + 0.011928257532417774, + 0.009327196516096592, + 0.009454888291656971, + 0.004550031386315823, + 0.010905592702329159, + 0.05675283074378967, + 0.009016048163175583, + 0.05407030135393143, + 0.007459732703864574, + 0.008620267733931541, + 0.013584031723439693, + 0.01442574430257082, + 0.00790083035826683, + 0.017327409237623215, + 0.01738894358277321, + 0.01827227510511875, + 0.009027295745909214, + 0.0075284321792423725, + 0.008285842835903168, + 0.004978502169251442, + 0.013959405943751335, + 0.015612873248755932, + 0.0063652233220636845, + 0.009407921694219112, + 0.015807904303073883, + 0.011826513335108757, + 0.009829331189393997, + 0.008833675645291805, + 0.010274634696543217, + 0.01223843079060316, + 0.012486879713833332, + 0.012719634920358658, + 0.008990730158984661, + 0.010188866406679153, + 0.012003444135189056, + 0.012355107814073563, + 0.012833517976105213, + 0.01189751923084259, + 0.01721668802201748, + 0.013462141156196594, + 0.009926908649504185, + 0.009309888817369938, + 0.010156272910535336, + 0.01208310853689909, + 0.010829154402017593, + 0.016357628628611565, + 0.02306908369064331, + 0.01746741123497486, + 0.00745333731174469, + 0.007574601098895073, + 0.010809154249727726, + 0.009616991505026817, + 0.008533063344657421, + 0.020286960527300835, + 0.02045583352446556, + 0.018946180120110512, + 0.006720909848809242, + 0.007490025367587805, + 0.006750856060534716, + 0.005956741515547037, + 0.00484371418133378, + 0.0061080544255673885, + 0.011710869148373604, + 0.00814115535467863 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 7 + ], + "selected_reward": -0.031122779473662376 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene presented in a static, slightly high-angled medium shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style consistent with traditional Japanese ink painting (minimalist elegance)?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a watering can made of antique bronze with a soft patina and subtle emerald green visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a perfectly spherical drop of clear water ooze from the pot's spout in slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the water drop transform into shimmering liquid golden ink upon hitting the soil?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.12706106901168823 + ], + "opposite_logprob_tokens": [ + -2.127061128616333 + ], + "logprob_reward": -0.12706106901168823, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.12706106901168823, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09523748606443405, + 0.05939400941133499, + 0.2515673041343689, + 0.03383314982056618, + 0.06573372334241867, + 0.11682868003845215, + 0.11939047276973724, + 0.09198711812496185, + 0.12015436589717865, + 0.04587369039654732 + ], + "credit_map": [ + 0.008360622450709343, + 0.010638896375894547, + 0.010151272639632225, + 0.008185158483684063, + 0.01772867515683174, + 0.015623727813363075, + 0.012462719343602657, + 0.012086396105587482, + 0.006874491926282644, + 0.0077063883654773235, + 0.005593252833932638, + 0.005618907045572996, + 0.006717219017446041, + 0.009056665934622288, + 0.006948657799512148, + 0.010878418572247028, + 0.01431906595826149, + 0.013503435999155045, + 0.00925842858850956, + 0.00482280133292079, + 0.018548468127846718, + 0.09170126169919968, + 0.009707769379019737, + 0.08970604836940765, + 0.0034689391031861305, + 0.003339156974107027, + 0.004609137773513794, + 0.00379434903152287, + 0.0034038154408335686, + 0.0054393731988966465, + 0.005068371072411537, + 0.004710004664957523, + 0.009988714940845966, + 0.006645296700298786, + 0.008207358419895172, + 0.003757710102945566, + 0.011772731319069862, + 0.011617019772529602, + 0.005137939006090164, + 0.008606944233179092, + 0.0177691001445055, + 0.01245448924601078, + 0.011235128156840801, + 0.012606698088347912, + 0.011562655679881573, + 0.017952436581254005, + 0.017108768224716187, + 0.01613938808441162, + 0.014895926229655743, + 0.01354891899973154, + 0.015553714707493782, + 0.01244234386831522, + 0.012658920139074326, + 0.017805375158786774, + 0.017117315903306007, + 0.015367946587502956, + 0.006605279166251421, + 0.006451127119362354, + 0.007147527765482664, + 0.007966580800712109, + 0.007530533242970705, + 0.016531789675354958, + 0.01951303333044052, + 0.020241236314177513, + 0.007316755596548319, + 0.007534725591540337, + 0.006760769989341497, + 0.006274751387536526, + 0.00792487058788538, + 0.0378265343606472, + 0.01166179496794939, + 0.034854158759117126, + 0.004168847110122442, + 0.0040925489738583565, + 0.005053543485701084, + 0.004028282128274441, + 0.009107863530516624, + 0.006555633619427681, + 0.007461157627403736, + 0.005405808333307505 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 8, + 6 + ], + "selected_reward": -0.12706106901168823 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene presented in a static, slightly high-angled medium shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style consistent with traditional Japanese ink painting (minimalist elegance)?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a watering can made of antique bronze with a soft patina and subtle emerald green visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a perfectly spherical drop of clear water ooze from the pot's spout in slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the water drop transform into shimmering liquid golden ink upon hitting the soil?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.348633348941803 + ], + "opposite_logprob_tokens": [ + -1.2236332893371582 + ], + "logprob_reward": -0.348633348941803, + "yes_no_margin_reward": 0.8749999403953552, + "token_credit_reward": -0.348633348941803, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09938310831785202, + 0.07393240183591843, + 0.09162008762359619, + 0.0338907390832901, + 0.13438309729099274, + 0.11653929948806763, + 0.07689881324768066, + 0.16945821046829224, + 0.12513892352581024, + 0.07875535637140274 + ], + "credit_map": [ + 0.0063732280395925045, + 0.00942004844546318, + 0.02045227587223053, + 0.012187646701931953, + 0.008003021590411663, + 0.011409112252295017, + 0.019669227302074432, + 0.011868542060256004, + 0.008615578524768353, + 0.010540910996496677, + 0.014014896005392075, + 0.006549668963998556, + 0.005178616847842932, + 0.009075785987079144, + 0.009897231124341488, + 0.010059713385999203, + 0.0065762498416006565, + 0.009254416450858116, + 0.006970553193241358, + 0.003822133643552661, + 0.0054978178814053535, + 0.02644106186926365, + 0.005069874692708254, + 0.027987971901893616, + 0.004143579863011837, + 0.005965772084891796, + 0.0043303570710122585, + 0.0024376404471695423, + 0.003053493332117796, + 0.004251998383551836, + 0.0046672639437019825, + 0.0050406320951879025, + 0.020998597145080566, + 0.010966220870614052, + 0.018249478191137314, + 0.006882444489747286, + 0.022670768201351166, + 0.024713946506381035, + 0.008862103335559368, + 0.021039534360170364, + 0.018011562526226044, + 0.012371785007417202, + 0.012030127458274364, + 0.011676017194986343, + 0.015325188636779785, + 0.017258241772651672, + 0.015183237381279469, + 0.01468313205987215, + 0.006294569931924343, + 0.007721923291683197, + 0.007705275900661945, + 0.008959440514445305, + 0.009879708290100098, + 0.011652347631752491, + 0.014315317384898663, + 0.010370222851634026, + 0.015937117859721184, + 0.015675678849220276, + 0.0112337376922369, + 0.009861310012638569, + 0.014069274999201298, + 0.04461880400776863, + 0.013163389638066292, + 0.04489889740943909, + 0.00909219216555357, + 0.007219852413982153, + 0.006384388078004122, + 0.004942566156387329, + 0.008332853205502033, + 0.037253618240356445, + 0.013327449560165405, + 0.038586001843214035, + 0.016176767647266388, + 0.011696051806211472, + 0.00802149809896946, + 0.005163438618183136, + 0.005555595271289349, + 0.009719779714941978, + 0.007873477414250374, + 0.01454874500632286 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 4, + 8 + ], + "selected_reward": -0.348633348941803 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene presented in a static, slightly high-angled medium shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style consistent with traditional Japanese ink painting (minimalist elegance)?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a watering can made of antique bronze with a soft patina and subtle emerald green visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a perfectly spherical drop of clear water ooze from the pot's spout in slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the water drop transform into shimmering liquid golden ink upon hitting the soil?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009780817665159702 + ], + "opposite_logprob_tokens": [ + -4.6347808837890625 + ], + "logprob_reward": -0.009780817665159702, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009780817665159702, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.027276020497083664, + 0.0248576607555151, + 0.058498553931713104, + 0.031261999160051346, + 0.01726449653506279, + 0.0104981055483222, + 0.02052580565214157, + 0.0990351215004921, + 0.525284469127655, + 0.18549782037734985 + ], + "credit_map": [ + 0.002585491631180048, + 0.004121428821235895, + 0.0025554499588906765, + 0.0033361848909407854, + 0.0031780130229890347, + 0.0042226193472743034, + 0.0033932658843696117, + 0.003883566940203309, + 0.00196343963034451, + 0.0029976314399391413, + 0.0026311641559004784, + 0.0016954921884462237, + 0.0017195127438753843, + 0.005808206740766764, + 0.002064094413071871, + 0.0059781200252473354, + 0.0038705188781023026, + 0.00439028674736619, + 0.0035157082602381706, + 0.0018447206821292639, + 0.004180427640676498, + 0.01937018893659115, + 0.002514471998438239, + 0.018812231719493866, + 0.0031506200321018696, + 0.00392568577080965, + 0.0037541775964200497, + 0.0032878059428185225, + 0.002619638340547681, + 0.005439938046038151, + 0.00399745861068368, + 0.005086672957986593, + 0.0016300957649946213, + 0.002780631883069873, + 0.0023389921989291906, + 0.0019948892295360565, + 0.0016491444548591971, + 0.0023354177828878164, + 0.002425653161481023, + 0.0021096726413816214, + 0.0011410453589633107, + 0.0015200948109850287, + 0.0010451306588947773, + 0.0009453778620809317, + 0.0013556093908846378, + 0.0015534890117123723, + 0.00143842538818717, + 0.001498933183029294, + 0.0025488934479653835, + 0.003705250099301338, + 0.0018785466672852635, + 0.0013146710116416216, + 0.0036526801995933056, + 0.0023758511524647474, + 0.002125880215317011, + 0.0029240334406495094, + 0.010007491335272789, + 0.006158802658319473, + 0.0043639144860208035, + 0.003981439862400293, + 0.008164765313267708, + 0.029245931655168533, + 0.009423641487956047, + 0.0276891328394413, + 0.028413204476237297, + 0.029663210734725, + 0.02110450156033039, + 0.01838594116270542, + 0.0350206159055233, + 0.18647269904613495, + 0.03101835586130619, + 0.17520596086978912, + 0.011477508582174778, + 0.015239725820720196, + 0.017085738480091095, + 0.013996869325637817, + 0.03140146657824516, + 0.034489572048187256, + 0.028726790100336075, + 0.03308013826608658 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 7 + ], + "selected_reward": -0.009780817665159702 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene presented in a static, slightly high-angled medium shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style consistent with traditional Japanese ink painting (minimalist elegance)?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a watering can made of antique bronze with a soft patina and subtle emerald green visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a perfectly spherical drop of clear water ooze from the pot's spout in slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the water drop transform into shimmering liquid golden ink upon hitting the soil?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05496693402528763 + ], + "opposite_logprob_tokens": [ + -2.929966926574707 + ], + "logprob_reward": -0.05496693402528763, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05496693402528763, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.015161947347223759, + 0.02329975739121437, + 0.01932930015027523, + 0.033159345388412476, + 0.01228374894708395, + 0.013341734185814857, + 0.010321715846657753, + 0.10053618252277374, + 0.6146527528762817, + 0.15791356563568115 + ], + "credit_map": [ + 0.0015202605864033103, + 0.0020401307847350836, + 0.00226715044118464, + 0.001475732889957726, + 0.0016922492068260908, + 0.0021215889137238264, + 0.0021293077152222395, + 0.0019155265763401985, + 0.0022894146386533976, + 0.0031708853784948587, + 0.0023698722943663597, + 0.0016977416817098856, + 0.001541370409540832, + 0.00466258404776454, + 0.0022360640577971935, + 0.005331826861947775, + 0.0020072441548109055, + 0.0023529333993792534, + 0.0033722405787557364, + 0.0015783210983499885, + 0.001242112834006548, + 0.003932259976863861, + 0.0016312566585838795, + 0.0032129313331097364, + 0.002677914686501026, + 0.003801344893872738, + 0.0037734268698841333, + 0.003569894703105092, + 0.0028141289949417114, + 0.006056996528059244, + 0.004713514354079962, + 0.005752123426645994, + 0.001443470478989184, + 0.0017355331219732761, + 0.0014773388393223286, + 0.001135153230279684, + 0.0012228635605424643, + 0.0020793406292796135, + 0.001538605778478086, + 0.001651443657465279, + 0.0023205436300486326, + 0.0016904189251363277, + 0.0016135892365127802, + 0.0010532890446484089, + 0.0017017970094457269, + 0.0015593133866786957, + 0.0018346975557506084, + 0.001568086096085608, + 0.0013942266814410686, + 0.001647247583605349, + 0.001092637307010591, + 0.0009120253962464631, + 0.0015429117484018207, + 0.0010782768949866295, + 0.0010010674595832825, + 0.0016533235320821404, + 0.009132607840001583, + 0.006390783935785294, + 0.004309358075261116, + 0.0036459744442254305, + 0.008469327352941036, + 0.030139563605189323, + 0.009952479042112827, + 0.028496084734797478, + 0.03380320221185684, + 0.03609105199575424, + 0.02420666813850403, + 0.015249358490109444, + 0.04214046895503998, + 0.22220344841480255, + 0.03155750781297684, + 0.20940101146697998, + 0.009304017759859562, + 0.011940174736082554, + 0.017408622428774834, + 0.0128614641726017, + 0.02685258537530899, + 0.02779575064778328, + 0.02744351513683796, + 0.02430744655430317 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 7 + ], + "selected_reward": -0.05496693402528763 + } + ], + "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/0000121/group_6/index_tensor([5563, 5563])_group_6_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000121/group_6/index_tensor([5563, 5563])_group_6_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..653a5c87fea6a88ccb4da91692b80de62f2ea85a --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000121/group_6/index_tensor([5563, 5563])_group_6_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "Beginning with a medium shot from a slightly high angle, the camera begins a slow, steady push-in, immersing the viewer in the understated, intimate, and focused scene. A mother cat, a domestic shorthair with a beautiful calico coat marked with distinct white, orange, and black spots, curls protectively around her litter in a tight \"C\" formation. They nestle in a soft, dark gray kennel lined with plush cream-colored faux fur. The central subject is a group of squirming newborn kittens, their eyes still closed, their sparse fur barely covering their pink skin. They huddle in their mother's warmth, some instinctively nuzzling for milk. The only illumination in the otherwise dark room comes from a nearby smart pet device on the floor, casting a warm, gentle, golden glow. This single light source frames the subject, capturing the soft texture of the mother's fur and creating soft highlights on the plush kennel lining, while casting long, deep shadows that stretch across the room, obscuring the background in darkness. The mother cat remains completely still, her expression one of serene alertness, creating a powerful atmosphere of tranquility, safety, and focused maternal love, while the deliberate movement of the camera draws the viewer close to this private, tender moment.", + "seed": 24022, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000121/group_6/index_5563_seed_24022_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": -0.049532475508749486, + "vlm_logprob": -0.049532475508749486, + "vlm_yes_no_margin": 3.5, + "vlm_token_credit": -0.049532475508749486, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a mother cat with a calico coat featuring white, orange, and black spots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are newborn kittens with closed eyes and pink skin visible huddled in the mother cat's arms?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene set inside a dark gray kennel lined with plush cream-colored faux fur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by a single warm golden light source from a smart pet device on the floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera begin with a medium shot from a slightly high angle and perform a slow, steady push-in?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.16182605922222137 + ], + "opposite_logprob_tokens": [ + -1.9118260145187378 + ], + "logprob_reward": -0.16182605922222137, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.16182605922222137, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.39844319224357605, + 0.35584312677383423, + 0.04817212373018265, + 0.0501517578959465, + 0.03668384253978729, + 0.018157681450247765, + 0.026198117062449455, + 0.020571265369653702, + 0.02067004144191742, + 0.02510886639356613 + ], + "credit_map": [ + 0.1326489895582199, + 0.021802229806780815, + 0.02597593516111374, + 0.016305452212691307, + 0.02979912795126438, + 0.140045627951622, + 0.018840868026018143, + 0.013024971820414066, + 0.07919159531593323, + 0.028454909101128578, + 0.017703836783766747, + 0.06093265116214752, + 0.018697869032621384, + 0.08986248821020126, + 0.04645906016230583, + 0.01454070582985878, + 0.0064618936739861965, + 0.004440261051058769, + 0.005394441075623035, + 0.010708729736506939, + 0.004947464447468519, + 0.005576173309236765, + 0.006626514717936516, + 0.004016648977994919, + 0.006172364577651024, + 0.005390503443777561, + 0.004524543881416321, + 0.008390314877033234, + 0.008780728094279766, + 0.006342705804854631, + 0.005341839510947466, + 0.0052087558433413506, + 0.003718300722539425, + 0.0032302141189575195, + 0.0030584384221583605, + 0.008780538104474545, + 0.0032197425607591867, + 0.005827025510370731, + 0.005831859540194273, + 0.0030177216976881027, + 0.0018192697316408157, + 0.0024239865597337484, + 0.0015925696352496743, + 0.0036400402896106243, + 0.00212057470344007, + 0.001757331076078117, + 0.0026789039839059114, + 0.0021250054705888033, + 0.0022316675167530775, + 0.0025722719728946686, + 0.0024402246344834566, + 0.004270141012966633, + 0.004906700924038887, + 0.0029408591799438, + 0.004286335315555334, + 0.002549916272982955, + 0.0016516539035364985, + 0.002682246034964919, + 0.0019049295224249363, + 0.0034387661144137383, + 0.002809763653203845, + 0.00231350539252162, + 0.0031197103671729565, + 0.002650690730661154, + 0.0021264278329908848, + 0.002210992621257901, + 0.002021457301452756, + 0.00316416728310287, + 0.0023824095260351896, + 0.0028084837831556797, + 0.0041969106532633305, + 0.0017591931391507387, + 0.001891454914584756, + 0.004112336318939924, + 0.0019090065034106374, + 0.003738693194463849, + 0.004250549245625734, + 0.0036125986371189356, + 0.0028681079857051373, + 0.002726119477301836 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 3 + ], + "selected_reward": -0.16182605922222137 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a mother cat with a calico coat featuring white, orange, and black spots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are newborn kittens with closed eyes and pink skin visible huddled in the mother cat's arms?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene set inside a dark gray kennel lined with plush cream-colored faux fur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by a single warm golden light source from a smart pet device on the floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera begin with a medium shot from a slightly high angle and perform a slow, steady push-in?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0160874892026186 + ], + "opposite_logprob_tokens": [ + -4.141087532043457 + ], + "logprob_reward": -0.0160874892026186, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.0160874892026186, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.581346333026886, + 0.11608292162418365, + 0.053052835166454315, + 0.04022025689482689, + 0.026744825765490532, + 0.027696378529071808, + 0.03702932968735695, + 0.018152007833123207, + 0.0337706133723259, + 0.06590443849563599 + ], + "credit_map": [ + 0.22971117496490479, + 0.020841922610998154, + 0.020217912271618843, + 0.013620973564684391, + 0.017361275851726532, + 0.249694362282753, + 0.01629280298948288, + 0.013605913147330284, + 0.02634655125439167, + 0.008663115091621876, + 0.005957822315394878, + 0.015305224806070328, + 0.009888031519949436, + 0.03110801987349987, + 0.012154550291597843, + 0.006659608334302902, + 0.00689316401258111, + 0.005555187352001667, + 0.007479977793991566, + 0.008473150432109833, + 0.006598757579922676, + 0.006643041968345642, + 0.006286276504397392, + 0.005123277194797993, + 0.004838716238737106, + 0.003734840312972665, + 0.00409944960847497, + 0.005515007302165031, + 0.0068228961899876595, + 0.004878029227256775, + 0.004631497897207737, + 0.005699818953871727, + 0.0026750206016004086, + 0.0029642689041793346, + 0.0025351594667881727, + 0.005828056018799543, + 0.002974008908495307, + 0.0033782750833779573, + 0.003761101281270385, + 0.002628935966640711, + 0.0029042549431324005, + 0.003201127517968416, + 0.0019462855998426676, + 0.006791905965656042, + 0.003275969298556447, + 0.0028398120775818825, + 0.003355505643412471, + 0.0033815153874456882, + 0.003256796160712838, + 0.0038550137542188168, + 0.0032921128440648317, + 0.006398153491318226, + 0.0061578829772770405, + 0.00510295107960701, + 0.005549614783376455, + 0.003416802268475294, + 0.0016422998160123825, + 0.0024252363946288824, + 0.002039010636508465, + 0.0029679297003895044, + 0.0025799807626754045, + 0.002173062413930893, + 0.002251317258924246, + 0.0020731710828840733, + 0.003205992514267564, + 0.004150078631937504, + 0.0033246329985558987, + 0.00731061864644289, + 0.004730098880827427, + 0.003962419461458921, + 0.004385278094559908, + 0.0027014962397515774, + 0.004995531868189573, + 0.005596659146249294, + 0.0066571105271577835, + 0.01580747775733471, + 0.00849471427500248, + 0.007889061234891415, + 0.010581397451460361, + 0.005882485304027796 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 9 + ], + "selected_reward": -0.0160874892026186 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a mother cat with a calico coat featuring white, orange, and black spots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are newborn kittens with closed eyes and pink skin visible huddled in the mother cat's arms?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene set inside a dark gray kennel lined with plush cream-colored faux fur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by a single warm golden light source from a smart pet device on the floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera begin with a medium shot from a slightly high angle and perform a slow, steady push-in?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0380902998149395 + ], + "opposite_logprob_tokens": [ + -3.288090229034424 + ], + "logprob_reward": -0.0380902998149395, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.0380902998149395, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3739529252052307, + 0.40093228220939636, + 0.037028077989816666, + 0.021507341414690018, + 0.038803935050964355, + 0.01625591330230236, + 0.039515867829322815, + 0.03647753596305847, + 0.018437251448631287, + 0.017088884487748146 + ], + "credit_map": [ + 0.14673765003681183, + 0.01403591874986887, + 0.012515151873230934, + 0.009849222376942635, + 0.012389909476041794, + 0.15818345546722412, + 0.01171667966991663, + 0.008524879813194275, + 0.09886132180690765, + 0.028809454292058945, + 0.018407924100756645, + 0.05888384208083153, + 0.022848371416330338, + 0.11145125329494476, + 0.04388676956295967, + 0.01778331771492958, + 0.005386815872043371, + 0.003487390000373125, + 0.00344307953491807, + 0.005340555217117071, + 0.004232069943100214, + 0.005169072654098272, + 0.005457230843603611, + 0.0045118615962564945, + 0.00288359634578228, + 0.002040407620370388, + 0.0026497808285057545, + 0.0026435761246830225, + 0.0027382660191506147, + 0.003229673020541668, + 0.0024229257833212614, + 0.002899112645536661, + 0.004062655847519636, + 0.004991040099412203, + 0.0036052113864570856, + 0.007739415392279625, + 0.00429107528179884, + 0.005978700704872608, + 0.004801260773092508, + 0.0033345725387334824, + 0.001903231837786734, + 0.0018509384244680405, + 0.001146945171058178, + 0.004376453813165426, + 0.0017410635482519865, + 0.001716539729386568, + 0.0017733298009261489, + 0.0017474093474447727, + 0.003841179423034191, + 0.005563267506659031, + 0.004505004268139601, + 0.004934885539114475, + 0.005929615348577499, + 0.004540836438536644, + 0.00550316646695137, + 0.004697911441326141, + 0.0034639027435332537, + 0.0037435193080455065, + 0.003814330790191889, + 0.006285117939114571, + 0.005046967417001724, + 0.003995951730757952, + 0.0062248315662145615, + 0.0039029137697070837, + 0.0016032173298299313, + 0.0020497904624789953, + 0.0018624881049618125, + 0.002786110620945692, + 0.0020877020433545113, + 0.002203136682510376, + 0.003917835187166929, + 0.0019269688054919243, + 0.0012214697198942304, + 0.002834274433553219, + 0.001330420607700944, + 0.001943604671396315, + 0.003631619503721595, + 0.0027840258553624153, + 0.0020574049558490515, + 0.0012860631104558706 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 6 + ], + "selected_reward": -0.0380902998149395 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a mother cat with a calico coat featuring white, orange, and black spots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are newborn kittens with closed eyes and pink skin visible huddled in the mother cat's arms?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene set inside a dark gray kennel lined with plush cream-colored faux fur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by a single warm golden light source from a smart pet device on the floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera begin with a medium shot from a slightly high angle and perform a slow, steady push-in?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020587796345353127 + ], + "opposite_logprob_tokens": [ + -3.895587682723999 + ], + "logprob_reward": -0.020587796345353127, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020587796345353127, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.46396785974502563, + 0.3137153089046478, + 0.03934253379702568, + 0.02719767950475216, + 0.03903944045305252, + 0.017008386552333832, + 0.021828951314091682, + 0.036329105496406555, + 0.013913264498114586, + 0.027657510712742805 + ], + "credit_map": [ + 0.17395371198654175, + 0.02075449749827385, + 0.014732194133102894, + 0.015463920310139656, + 0.019677473232150078, + 0.19090420007705688, + 0.01603725738823414, + 0.012444637715816498, + 0.07640203088521957, + 0.023486098274588585, + 0.01426143292337656, + 0.04608415439724922, + 0.018820971250534058, + 0.08749094605445862, + 0.03436803072690964, + 0.012801655568182468, + 0.005414106883108616, + 0.003967147320508957, + 0.003957224544137716, + 0.006705653388053179, + 0.004410253372043371, + 0.006464814767241478, + 0.005322285462170839, + 0.003101048292592168, + 0.0033842814154922962, + 0.0029871584847569466, + 0.0023497019428759813, + 0.004517924040555954, + 0.004987113643437624, + 0.0035495625343173742, + 0.0029050910379737616, + 0.0025168450083583593, + 0.0045059132389724255, + 0.004445309285074472, + 0.003751274896785617, + 0.005179266911000013, + 0.005483994260430336, + 0.0058019403368234634, + 0.005333125591278076, + 0.004538617562502623, + 0.0018482455052435398, + 0.00225376570597291, + 0.00259581720456481, + 0.002327547874301672, + 0.00251912884414196, + 0.001989418873563409, + 0.0014266804791986942, + 0.0020477839279919863, + 0.0021215949673205614, + 0.0025610816664993763, + 0.002533608814701438, + 0.0033728277776390314, + 0.002712275367230177, + 0.002605923218652606, + 0.003999400418251753, + 0.0019222383853048086, + 0.00349525292403996, + 0.004001906607300043, + 0.004428975284099579, + 0.004514731932431459, + 0.004927051719278097, + 0.0047909836284816265, + 0.005907666869461536, + 0.0042625353671610355, + 0.001633798354305327, + 0.0015444244490936399, + 0.0015407797181978822, + 0.002316152211278677, + 0.0016347079072147608, + 0.0018052321393042803, + 0.002017824212089181, + 0.0014203459722921252, + 0.0021136135328561068, + 0.005692463368177414, + 0.0019338997080922127, + 0.002925089793279767, + 0.00510270893573761, + 0.005297033116221428, + 0.0025799176655709743, + 0.0020127836614847183 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.020587796345353127 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a mother cat with a calico coat featuring white, orange, and black spots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are newborn kittens with closed eyes and pink skin visible huddled in the mother cat's arms?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene set inside a dark gray kennel lined with plush cream-colored faux fur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by a single warm golden light source from a smart pet device on the floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera begin with a medium shot from a slightly high angle and perform a slow, steady push-in?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011070732958614826 + ], + "opposite_logprob_tokens": [ + -4.511070728302002 + ], + "logprob_reward": -0.011070732958614826, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011070732958614826, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5003882646560669, + 0.08089224994182587, + 0.07358100265264511, + 0.1282072812318802, + 0.02800009958446026, + 0.017900671809911728, + 0.0332513190805912, + 0.039415959268808365, + 0.04049745574593544, + 0.05786572024226189 + ], + "credit_map": [ + 0.1382167786359787, + 0.03563240170478821, + 0.062752366065979, + 0.02673008106648922, + 0.04145752638578415, + 0.14708085358142853, + 0.02605244517326355, + 0.022465800866484642, + 0.012092509306967258, + 0.009208681993186474, + 0.00566124590113759, + 0.015074392780661583, + 0.007168630138039589, + 0.011912248097360134, + 0.012552253901958466, + 0.007222285959869623, + 0.008862845599651337, + 0.007816614583134651, + 0.009145556949079037, + 0.015695925801992416, + 0.009170247241854668, + 0.007514419034123421, + 0.008755375631153584, + 0.006620013155043125, + 0.016091860830783844, + 0.008574453182518482, + 0.011976977810263634, + 0.03879060223698616, + 0.010021730326116085, + 0.014534765854477882, + 0.017535455524921417, + 0.01068144477903843, + 0.0031656641513109207, + 0.0034524197690188885, + 0.0029755565337836742, + 0.004757206421345472, + 0.003796161152422428, + 0.0033406002912670374, + 0.003353620646521449, + 0.0031588696874678135, + 0.0019255831139162183, + 0.0017168294871225953, + 0.0020388963166624308, + 0.004112069960683584, + 0.0020359382033348083, + 0.001888548955321312, + 0.002362699480727315, + 0.0018201059428974986, + 0.003132236422970891, + 0.003808199195191264, + 0.003672626568004489, + 0.00453775143250823, + 0.005183996632695198, + 0.004374406300485134, + 0.0048503512516617775, + 0.003691752441227436, + 0.004141589626669884, + 0.00481427414342761, + 0.003979465924203396, + 0.007432680111378431, + 0.004985653329640627, + 0.004352424293756485, + 0.0052946703508496284, + 0.004415201023221016, + 0.004381646402180195, + 0.004578860010951757, + 0.00433075986802578, + 0.00716426270082593, + 0.004166373983025551, + 0.004917412996292114, + 0.006414329633116722, + 0.004543812945485115, + 0.004434951115399599, + 0.009044518694281578, + 0.006486131809651852, + 0.009215768426656723, + 0.009445971809327602, + 0.008851923979818821, + 0.005208376329392195, + 0.005178078077733517 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 1 + ], + "selected_reward": -0.011070732958614826 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "Beginning with a medium shot from a slightly high angle, the camera begins a slow, steady push-in, immersing the viewer in the understated, intimate, and focused scene. A mother cat, a domestic shorthair with a beautiful calico coat marked with distinct white, orange, and black spots, curls protectively around her litter in a tight \"C\" formation. They nestle in a soft, dark gray kennel lined with plush cream-colored faux fur. The central subject is a group of squirming newborn kittens, their eyes still closed, their sparse fur barely covering their pink skin. They huddle in their mother's warmth, some instinctively nuzzling for milk. The only illumination in the otherwise dark room comes from a nearby smart pet device on the floor, casting a warm, gentle, golden glow. This single light source frames the subject, capturing the soft texture of the mother's fur and creating soft highlights on the plush kennel lining, while casting long, deep shadows that stretch across the room, obscuring the background in darkness. The mother cat remains completely still, her expression one of serene alertness, creating a powerful atmosphere of tranquility, safety, and focused maternal love, while the deliberate movement of the camera draws the viewer close to this private, tender moment.", + "seed": 24023, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000121/group_6/index_5563_seed_24023_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": -0.026461805496364833, + "vlm_logprob": -0.026461805496364833, + "vlm_yes_no_margin": 3.825, + "vlm_token_credit": -0.026461805496364833, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a mother cat with a calico coat featuring white, orange, and black spots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are newborn kittens with closed eyes and pink skin visible huddled in the mother cat's arms?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene set inside a dark gray kennel lined with plush cream-colored faux fur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by a single warm golden light source from a smart pet device on the floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera begin with a medium shot from a slightly high angle and perform a slow, steady push-in?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05039123445749283 + ], + "opposite_logprob_tokens": [ + -3.05039119720459 + ], + "logprob_reward": -0.05039123445749283, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.05039123445749283, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2631914019584656, + 0.07115647941827774, + 0.041798703372478485, + 0.05871923640370369, + 0.13656145334243774, + 0.09575162082910538, + 0.037270963191986084, + 0.10521020740270615, + 0.11326759308576584, + 0.07707231491804123 + ], + "credit_map": [ + 0.021773559972643852, + 0.01780259795486927, + 0.05171974375844002, + 0.032107897102832794, + 0.02739468216896057, + 0.04275284707546234, + 0.04820510372519493, + 0.021434970200061798, + 0.0056660110130906105, + 0.00789661705493927, + 0.010253135114908218, + 0.009587160311639309, + 0.008148958906531334, + 0.012120045721530914, + 0.013651528395712376, + 0.0038330235984176397, + 0.0034325970336794853, + 0.0041162436828017235, + 0.006399692036211491, + 0.0060797459445893764, + 0.005610058084130287, + 0.006990913301706314, + 0.006401146296411753, + 0.0027683046646416187, + 0.004763071890920401, + 0.005209794733673334, + 0.006953502539545298, + 0.007474177982658148, + 0.007355276495218277, + 0.009814275428652763, + 0.009228145703673363, + 0.007920990698039532, + 0.020343296229839325, + 0.010356909595429897, + 0.016600163653492928, + 0.02282370813190937, + 0.0148340268060565, + 0.023143520578742027, + 0.01578872837126255, + 0.012671095319092274, + 0.009078415110707283, + 0.007735507562756538, + 0.010408615693449974, + 0.012333045713603497, + 0.009999619796872139, + 0.017572518438100815, + 0.018160004168748856, + 0.010463896207511425, + 0.003929083235561848, + 0.0030513203237205744, + 0.004146651364862919, + 0.004396991338580847, + 0.0034068727400153875, + 0.006213533692061901, + 0.007008193992078304, + 0.005118315573781729, + 0.007152183912694454, + 0.009713935665786266, + 0.008867128752171993, + 0.010186326690018177, + 0.01277621928602457, + 0.01789841428399086, + 0.014500085264444351, + 0.0241159126162529, + 0.011366002261638641, + 0.009270697832107544, + 0.014470010995864868, + 0.015188234858214855, + 0.011422951705753803, + 0.01798781007528305, + 0.012903155758976936, + 0.020658716559410095, + 0.006228918209671974, + 0.011657199822366238, + 0.00732471002265811, + 0.008017217740416527, + 0.008403848856687546, + 0.011546286754310131, + 0.011103495955467224, + 0.012790641747415066 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 8 + ], + "selected_reward": -0.05039123445749283 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a mother cat with a calico coat featuring white, orange, and black spots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are newborn kittens with closed eyes and pink skin visible huddled in the mother cat's arms?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene set inside a dark gray kennel lined with plush cream-colored faux fur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by a single warm golden light source from a smart pet device on the floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera begin with a medium shot from a slightly high angle and perform a slow, steady push-in?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0181954987347126 + ], + "opposite_logprob_tokens": [ + -4.018195629119873 + ], + "logprob_reward": -0.0181954987347126, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.0181954987347126, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3284938633441925, + 0.10705850273370743, + 0.038997989147901535, + 0.09076548367738724, + 0.06392987817525864, + 0.0481150783598423, + 0.033230531960725784, + 0.09090233594179153, + 0.1330406665802002, + 0.06546564400196075 + ], + "credit_map": [ + 0.03281985968351364, + 0.02468562312424183, + 0.05800096318125725, + 0.04339848458766937, + 0.0343097560107708, + 0.054240476340055466, + 0.062258731573820114, + 0.018779966980218887, + 0.009066936559975147, + 0.011863325722515583, + 0.018354391679167747, + 0.014322474598884583, + 0.013303184881806374, + 0.0188575591892004, + 0.014650203287601471, + 0.006640419363975525, + 0.0042301444336771965, + 0.00381515733897686, + 0.006134075112640858, + 0.004222263116389513, + 0.00494851591065526, + 0.006648745853453875, + 0.004623074550181627, + 0.004376012831926346, + 0.009815208613872528, + 0.004878072068095207, + 0.013143809512257576, + 0.012512647546827793, + 0.01293755043298006, + 0.014475006610155106, + 0.015244982205331326, + 0.007758213672786951, + 0.009845998138189316, + 0.005700356792658567, + 0.007979150861501694, + 0.008382156491279602, + 0.006528869271278381, + 0.010220094583928585, + 0.009027289226651192, + 0.006245957687497139, + 0.004771780222654343, + 0.004883166868239641, + 0.0051299999468028545, + 0.0061014494858682156, + 0.005300022196024656, + 0.008448599837720394, + 0.00911638978868723, + 0.004363670013844967, + 0.0030383821576833725, + 0.003426316659897566, + 0.0038882605731487274, + 0.004769925959408283, + 0.003663316834717989, + 0.006005255039781332, + 0.0053741345182061195, + 0.003064938588067889, + 0.0064652240835130215, + 0.006775398273020983, + 0.007035733200609684, + 0.008001599460840225, + 0.008704598993062973, + 0.016711400821805, + 0.013920266181230545, + 0.023288113996386528, + 0.015584612265229225, + 0.01723651960492134, + 0.012847281992435455, + 0.014406169764697552, + 0.018003147095441818, + 0.020114032551646233, + 0.020044054836034775, + 0.014804843813180923, + 0.005884143523871899, + 0.008658063597977161, + 0.007570814341306686, + 0.009672548621892929, + 0.010288811288774014, + 0.009414264000952244, + 0.007203569170087576, + 0.006773429922759533 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 1 + ], + "selected_reward": -0.0181954987347126 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a mother cat with a calico coat featuring white, orange, and black spots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are newborn kittens with closed eyes and pink skin visible huddled in the mother cat's arms?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene set inside a dark gray kennel lined with plush cream-colored faux fur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by a single warm golden light source from a smart pet device on the floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera begin with a medium shot from a slightly high angle and perform a slow, steady push-in?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"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": [ + 0.31138062477111816, + 0.08641793578863144, + 0.09443490207195282, + 0.057124435901641846, + 0.10922688245773315, + 0.0612572506070137, + 0.05391587316989899, + 0.04295869916677475, + 0.11017527431249619, + 0.07310814410448074 + ], + "credit_map": [ + 0.028200561180710793, + 0.02693256549537182, + 0.03596142679452896, + 0.03141295909881592, + 0.053226273506879807, + 0.05107253044843674, + 0.06170995533466339, + 0.02286435477435589, + 0.008311094716191292, + 0.008917974308133125, + 0.013429699465632439, + 0.01008023414760828, + 0.010674127377569675, + 0.015353384427726269, + 0.012463380582630634, + 0.007188047748059034, + 0.006644327659159899, + 0.008795736357569695, + 0.014980933628976345, + 0.012233611196279526, + 0.013369081541895866, + 0.017123492434620857, + 0.01359782088547945, + 0.007689899764955044, + 0.005188079550862312, + 0.0036574359983205795, + 0.007730937097221613, + 0.009325768798589706, + 0.00618382403627038, + 0.009684297256171703, + 0.011481530033051968, + 0.003872564760968089, + 0.014721703715622425, + 0.00878725852817297, + 0.01297503337264061, + 0.01381892990320921, + 0.010355653241276741, + 0.01737116277217865, + 0.014557897113263607, + 0.016639241948723793, + 0.006615773309022188, + 0.004666615277528763, + 0.008081554435193539, + 0.007359076291322708, + 0.006723718252032995, + 0.01137506403028965, + 0.011582248844206333, + 0.0048532020300626755, + 0.004650502000004053, + 0.004613874945789576, + 0.004995862953364849, + 0.0074595073238015175, + 0.00440458208322525, + 0.007931340485811234, + 0.008047028444707394, + 0.01181317400187254, + 0.004141570534557104, + 0.00472938222810626, + 0.0038502218667417765, + 0.0051810527220368385, + 0.004946235567331314, + 0.006894232239574194, + 0.006172323599457741, + 0.007043677847832441, + 0.010478489100933075, + 0.01457177009433508, + 0.012364927679300308, + 0.011897340416908264, + 0.017042694613337517, + 0.019001325592398643, + 0.01048043929040432, + 0.014338284730911255, + 0.007868196815252304, + 0.011142170056700706, + 0.007207822520285845, + 0.007750202436000109, + 0.008528958074748516, + 0.010620185174047947, + 0.007207846734672785, + 0.012782759964466095 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 4 + ], + "selected_reward": -0.03369637206196785 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a mother cat with a calico coat featuring white, orange, and black spots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are newborn kittens with closed eyes and pink skin visible huddled in the mother cat's arms?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene set inside a dark gray kennel lined with plush cream-colored faux fur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by a single warm golden light source from a smart pet device on the floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera begin with a medium shot from a slightly high angle and perform a slow, steady push-in?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02329506166279316 + ], + "opposite_logprob_tokens": [ + -3.7732951641082764 + ], + "logprob_reward": -0.02329506166279316, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.02329506166279316, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23835240304470062, + 0.10962865501642227, + 0.07313597947359085, + 0.10842418670654297, + 0.08275134116411209, + 0.07670256495475769, + 0.10076010227203369, + 0.058637846261262894, + 0.08402663469314575, + 0.06758034974336624 + ], + "credit_map": [ + 0.025855308398604393, + 0.019702782854437828, + 0.03845025599002838, + 0.025270679965615273, + 0.03199384734034538, + 0.036716945469379425, + 0.03404619172215462, + 0.026316380128264427, + 0.00898424070328474, + 0.007259204983711243, + 0.016020694747567177, + 0.013234657235443592, + 0.015809131786227226, + 0.018109722062945366, + 0.01776752807199955, + 0.012443477287888527, + 0.007469289004802704, + 0.005536308512091637, + 0.014523587189614773, + 0.006816093809902668, + 0.009135019965469837, + 0.012868900783360004, + 0.010930152609944344, + 0.005856623407453299, + 0.009084410965442657, + 0.007305988110601902, + 0.016172705218195915, + 0.015402844175696373, + 0.010696791112422943, + 0.017863180488348007, + 0.0220131054520607, + 0.009885158389806747, + 0.012452730908989906, + 0.006258574780076742, + 0.009088896214962006, + 0.012131002731621265, + 0.010591749101877213, + 0.013428127393126488, + 0.009944196790456772, + 0.008856057189404964, + 0.006848304532468319, + 0.004747831728309393, + 0.009767349809408188, + 0.010917038656771183, + 0.007547279354184866, + 0.01409265585243702, + 0.01506973709911108, + 0.007712358143180609, + 0.009042342193424702, + 0.007620346266776323, + 0.010343731380999088, + 0.016327867284417152, + 0.009508046321570873, + 0.016108093783259392, + 0.019846946001052856, + 0.011962724849581718, + 0.004267358686774969, + 0.003989006858319044, + 0.0036251505371183157, + 0.005331122782081366, + 0.00567272026091814, + 0.010615820065140724, + 0.00862084235996008, + 0.01651582308113575, + 0.013319556601345539, + 0.009474054910242558, + 0.008728988468647003, + 0.00913243368268013, + 0.010009770281612873, + 0.012575316242873669, + 0.01061009056866169, + 0.010176421143114567, + 0.009760795161128044, + 0.007980777882039547, + 0.008046789094805717, + 0.007570602465420961, + 0.006841483525931835, + 0.01029796339571476, + 0.0064725554548203945, + 0.010609383694827557 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 3 + ], + "selected_reward": -0.02329506166279316 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a mother cat with a calico coat featuring white, orange, and black spots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are newborn kittens with closed eyes and pink skin visible huddled in the mother cat's arms?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene set inside a dark gray kennel lined with plush cream-colored faux fur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by a single warm golden light source from a smart pet device on the floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera begin with a medium shot from a slightly high angle and perform a slow, steady push-in?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006730860564857721 + ], + "opposite_logprob_tokens": [ + -5.006731033325195 + ], + "logprob_reward": -0.006730860564857721, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006730860564857721, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19186151027679443, + 0.058185454457998276, + 0.047288306057453156, + 0.04747534543275833, + 0.06694963574409485, + 0.055176135152578354, + 0.07570862770080566, + 0.08418016880750656, + 0.3128538131713867, + 0.06032105162739754 + ], + "credit_map": [ + 0.017381908372044563, + 0.014568603597581387, + 0.026836352422833443, + 0.02512984536588192, + 0.026028485968708992, + 0.031124599277973175, + 0.03720070794224739, + 0.013591005466878414, + 0.003934981767088175, + 0.003976395353674889, + 0.009984704665839672, + 0.006203516386449337, + 0.008888257667422295, + 0.009822391904890537, + 0.011465541087090969, + 0.00390966609120369, + 0.00411071814596653, + 0.004341285675764084, + 0.006282387301325798, + 0.007391613442450762, + 0.005876863840967417, + 0.008283626288175583, + 0.00770998140797019, + 0.003291826695203781, + 0.003504385007545352, + 0.004144751932471991, + 0.005513378418982029, + 0.006351838354021311, + 0.005076924338936806, + 0.008365187793970108, + 0.008574677631258965, + 0.005944200791418552, + 0.008998876437544823, + 0.004659563302993774, + 0.009437164291739464, + 0.009423233568668365, + 0.007063782308250666, + 0.010530599392950535, + 0.010875449515879154, + 0.005960965063422918, + 0.004472864791750908, + 0.005128325428813696, + 0.005839601159095764, + 0.008739490061998367, + 0.005736092105507851, + 0.008586042560636997, + 0.007049208506941795, + 0.009624509140849113, + 0.007121074944734573, + 0.007968706078827381, + 0.007257685996592045, + 0.008665822446346283, + 0.008446984924376011, + 0.010842723771929741, + 0.009793448261916637, + 0.01561217661947012, + 0.005973941646516323, + 0.008719053119421005, + 0.005998832639306784, + 0.01070066262036562, + 0.008857382461428642, + 0.01423656102269888, + 0.010761303827166557, + 0.01893243007361889, + 0.02952698804438114, + 0.03329441323876381, + 0.029305262491106987, + 0.03814880922436714, + 0.044222235679626465, + 0.05120918154716492, + 0.03643928840756416, + 0.05070764571428299, + 0.006343451794236898, + 0.006215888075530529, + 0.0063480488024652, + 0.006639262195676565, + 0.0068489802069962025, + 0.008124181069433689, + 0.007720643654465675, + 0.012080592103302479 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 7 + ], + "selected_reward": -0.006730860564857721 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "Beginning with a medium shot from a slightly high angle, the camera begins a slow, steady push-in, immersing the viewer in the understated, intimate, and focused scene. A mother cat, a domestic shorthair with a beautiful calico coat marked with distinct white, orange, and black spots, curls protectively around her litter in a tight \"C\" formation. They nestle in a soft, dark gray kennel lined with plush cream-colored faux fur. The central subject is a group of squirming newborn kittens, their eyes still closed, their sparse fur barely covering their pink skin. They huddle in their mother's warmth, some instinctively nuzzling for milk. The only illumination in the otherwise dark room comes from a nearby smart pet device on the floor, casting a warm, gentle, golden glow. This single light source frames the subject, capturing the soft texture of the mother's fur and creating soft highlights on the plush kennel lining, while casting long, deep shadows that stretch across the room, obscuring the background in darkness. The mother cat remains completely still, her expression one of serene alertness, creating a powerful atmosphere of tranquility, safety, and focused maternal love, while the deliberate movement of the camera draws the viewer close to this private, tender moment.", + "seed": 24024, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000121/group_6/index_5563_seed_24024_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": -0.036924075521528724, + "vlm_logprob": -0.036924075521528724, + "vlm_yes_no_margin": 3.5, + "vlm_token_credit": -0.036924075521528724, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a mother cat with a calico coat featuring white, orange, and black spots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are newborn kittens with closed eyes and pink skin visible huddled in the mother cat's arms?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene set inside a dark gray kennel lined with plush cream-colored faux fur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by a single warm golden light source from a smart pet device on the floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera begin with a medium shot from a slightly high angle and perform a slow, steady push-in?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07123947888612747 + ], + "opposite_logprob_tokens": [ + -2.696239471435547 + ], + "logprob_reward": -0.07123947888612747, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07123947888612747, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21633656322956085, + 0.04030891880393028, + 0.1043238416314125, + 0.08462116122245789, + 0.04269656538963318, + 0.14785626530647278, + 0.2238207757472992, + 0.047349825501441956, + 0.06224808096885681, + 0.030438022688031197 + ], + "credit_map": [ + 0.024898631498217583, + 0.037195947021245956, + 0.025578642264008522, + 0.03520849719643593, + 0.02669525519013405, + 0.030590971931815147, + 0.021857470273971558, + 0.014311150647699833, + 0.007997683249413967, + 0.00394271919503808, + 0.007778619881719351, + 0.0036121502052992582, + 0.0032746775541454554, + 0.005696401931345463, + 0.004102238453924656, + 0.0039044281002134085, + 0.015445293858647346, + 0.015108779072761536, + 0.01580379530787468, + 0.012638920918107033, + 0.012635010294616222, + 0.013593972660601139, + 0.010791141539812088, + 0.008306928910315037, + 0.009169016033411026, + 0.011225220747292042, + 0.009562528692185879, + 0.018756719306111336, + 0.009972459636628628, + 0.010031312704086304, + 0.004801720380783081, + 0.011102181859314442, + 0.005062719341367483, + 0.008496117778122425, + 0.004673640709370375, + 0.005560433492064476, + 0.004550304263830185, + 0.005706330761313438, + 0.004278471693396568, + 0.004368546884506941, + 0.015854597091674805, + 0.013797319494187832, + 0.019762858748435974, + 0.014754658564925194, + 0.013403579592704773, + 0.024812141433358192, + 0.021266698837280273, + 0.024204416200518608, + 0.026773156598210335, + 0.02691512741148472, + 0.03064052388072014, + 0.02273027040064335, + 0.0348348543047905, + 0.024238571524620056, + 0.02780074253678322, + 0.029887530952692032, + 0.005551355890929699, + 0.005852177739143372, + 0.005282173864543438, + 0.011284814216196537, + 0.0041340715251863, + 0.004948768299072981, + 0.005030380096286535, + 0.005266085267066956, + 0.0074566020630300045, + 0.009886577725410461, + 0.00641826493665576, + 0.007463068701326847, + 0.005702096503227949, + 0.010408241301774979, + 0.006310419179499149, + 0.008602811954915524, + 0.002621588297188282, + 0.00342898559756577, + 0.003887048689648509, + 0.0038881609216332436, + 0.0035143750719726086, + 0.004838587716221809, + 0.002916487865149975, + 0.005342790856957436 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 5 + ], + "selected_reward": -0.07123947888612747 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a mother cat with a calico coat featuring white, orange, and black spots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are newborn kittens with closed eyes and pink skin visible huddled in the mother cat's arms?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene set inside a dark gray kennel lined with plush cream-colored faux fur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by a single warm golden light source from a smart pet device on the floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera begin with a medium shot from a slightly high angle and perform a slow, steady push-in?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016085142269730568 + ], + "opposite_logprob_tokens": [ + -4.141085147857666 + ], + "logprob_reward": -0.016085142269730568, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016085142269730568, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17733877897262573, + 0.08629539608955383, + 0.061853863298892975, + 0.06099002808332443, + 0.026961171999573708, + 0.07382465898990631, + 0.3430337607860565, + 0.05092231556773186, + 0.07514391839504242, + 0.04363609105348587 + ], + "credit_map": [ + 0.015007027424871922, + 0.0371243879199028, + 0.02044975757598877, + 0.025173025205731392, + 0.01953861117362976, + 0.02292671985924244, + 0.02138807252049446, + 0.015731163322925568, + 0.015172973275184631, + 0.008711051195859909, + 0.01834779419004917, + 0.008763167075812817, + 0.009133411571383476, + 0.01256486400961876, + 0.0069039505906403065, + 0.006698186043649912, + 0.008864477276802063, + 0.009199243038892746, + 0.008839897811412811, + 0.007577182725071907, + 0.00742258969694376, + 0.008296250365674496, + 0.0058484929613769054, + 0.00580572709441185, + 0.006932420656085014, + 0.008918545208871365, + 0.007139996625483036, + 0.009376181289553642, + 0.007240639068186283, + 0.007810860872268677, + 0.004471674095839262, + 0.00909971259534359, + 0.0032340711914002895, + 0.004238670226186514, + 0.0030241194181144238, + 0.003054013941437006, + 0.003947877790778875, + 0.003172749886289239, + 0.0029982137493789196, + 0.003291455563157797, + 0.00862074177712202, + 0.009015093557536602, + 0.008907400071620941, + 0.006066542584449053, + 0.00811578705906868, + 0.012285834178328514, + 0.009836883284151554, + 0.010976382531225681, + 0.030953213572502136, + 0.029622111469507217, + 0.0324249342083931, + 0.03074244223535061, + 0.04687715321779251, + 0.06370605528354645, + 0.04533885791897774, + 0.06336899101734161, + 0.00662460969761014, + 0.005306295119225979, + 0.004998762626200914, + 0.0047857328318059444, + 0.005882828030735254, + 0.007706677541136742, + 0.007846198976039886, + 0.007771210744976997, + 0.009122965857386589, + 0.011478903703391552, + 0.009596279822289944, + 0.008473655208945274, + 0.00854862667620182, + 0.0100975027307868, + 0.008704248815774918, + 0.00912174116820097, + 0.004988283384591341, + 0.005341588519513607, + 0.004727749153971672, + 0.005938875023275614, + 0.00628645159304142, + 0.006110533606261015, + 0.005230581387877464, + 0.0050120302475988865 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 1 + ], + "selected_reward": -0.016085142269730568 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a mother cat with a calico coat featuring white, orange, and black spots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are newborn kittens with closed eyes and pink skin visible huddled in the mother cat's arms?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene set inside a dark gray kennel lined with plush cream-colored faux fur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by a single warm golden light source from a smart pet device on the floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera begin with a medium shot from a slightly high angle and perform a slow, steady push-in?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05494651198387146 + ], + "opposite_logprob_tokens": [ + -2.9299464225769043 + ], + "logprob_reward": -0.05494651198387146, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05494651198387146, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1174040138721466, + 0.049648646265268326, + 0.03667069226503372, + 0.04910387843847275, + 0.04696670547127724, + 0.06301770359277725, + 0.526557981967926, + 0.06539115309715271, + 0.028789937496185303, + 0.016449306160211563 + ], + "credit_map": [ + 0.010724926367402077, + 0.023196201771497726, + 0.011674742214381695, + 0.019370298832654953, + 0.015510371886193752, + 0.016681604087352753, + 0.011839737184345722, + 0.008406130596995354, + 0.008779922500252724, + 0.0052388752810657024, + 0.010666906833648682, + 0.00518891541287303, + 0.004848232958465815, + 0.007106298115104437, + 0.003929973114281893, + 0.0038895215839147568, + 0.004389383364468813, + 0.005737170577049255, + 0.005417352542281151, + 0.004842952359467745, + 0.004560907371342182, + 0.004966570530086756, + 0.002806619508191943, + 0.003949734847992659, + 0.003595947753638029, + 0.0067367106676101685, + 0.004696644376963377, + 0.01524319127202034, + 0.003907274920493364, + 0.006540076341480017, + 0.0031710562761873007, + 0.005212979391217232, + 0.0051187993958592415, + 0.006081698462367058, + 0.006251721177250147, + 0.005357499700039625, + 0.00509506044909358, + 0.006397060118615627, + 0.004636900499463081, + 0.008027961477637291, + 0.007545184809714556, + 0.00714879110455513, + 0.008034033700823784, + 0.004939528182148933, + 0.00734159629791975, + 0.009013592265546322, + 0.007966194301843643, + 0.011028779670596123, + 0.05813099816441536, + 0.05897848680615425, + 0.06509135663509369, + 0.05256211757659912, + 0.0825977697968483, + 0.05817287787795067, + 0.0789380893111229, + 0.07208627462387085, + 0.007509591057896614, + 0.0070700799115002155, + 0.008805525489151478, + 0.008948305621743202, + 0.007204381283372641, + 0.00854068249464035, + 0.00794647540897131, + 0.009366116486489773, + 0.0029524676501750946, + 0.0031765266321599483, + 0.0030543962493538857, + 0.002430969150736928, + 0.002829714212566614, + 0.00528269587084651, + 0.004115161951631308, + 0.004948006477206945, + 0.002151592168956995, + 0.001731737982481718, + 0.0019643004052340984, + 0.0017907802248373628, + 0.0016620007809251547, + 0.0020909367594867945, + 0.002680588513612747, + 0.002377368276938796 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 7 + ], + "selected_reward": -0.05494651198387146 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a mother cat with a calico coat featuring white, orange, and black spots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are newborn kittens with closed eyes and pink skin visible huddled in the mother cat's arms?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene set inside a dark gray kennel lined with plush cream-colored faux fur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by a single warm golden light source from a smart pet device on the floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera begin with a medium shot from a slightly high angle and perform a slow, steady push-in?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029816951602697372 + ], + "opposite_logprob_tokens": [ + -3.5298168659210205 + ], + "logprob_reward": -0.029816951602697372, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029816951602697372, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11859744042158127, + 0.10586372017860413, + 0.10828720778226852, + 0.0790613442659378, + 0.14434000849723816, + 0.20615693926811218, + 0.12019667029380798, + 0.06033502519130707, + 0.031281083822250366, + 0.02588057331740856 + ], + "credit_map": [ + 0.011344153434038162, + 0.020821981132030487, + 0.012137400917708874, + 0.016008194535970688, + 0.01313608605414629, + 0.01701684109866619, + 0.012219438329339027, + 0.015913350507616997, + 0.022296423092484474, + 0.010793491266667843, + 0.01968713477253914, + 0.0083463778719306, + 0.007471035234630108, + 0.01669245958328247, + 0.008041725493967533, + 0.01253507100045681, + 0.01706649363040924, + 0.013858291320502758, + 0.016507089138031006, + 0.012824846431612968, + 0.012344570830464363, + 0.01470927894115448, + 0.010231629945337772, + 0.010745014995336533, + 0.011820086278021336, + 0.00884850975126028, + 0.010495174676179886, + 0.007394591812044382, + 0.006096397526562214, + 0.012651990167796612, + 0.007235999219119549, + 0.014518593437969685, + 0.01596701517701149, + 0.015150823630392551, + 0.012944488786160946, + 0.01594008319079876, + 0.01435462012887001, + 0.02248077653348446, + 0.011799175292253494, + 0.03570302575826645, + 0.02222810499370098, + 0.016723215579986572, + 0.026542218402028084, + 0.016308628022670746, + 0.0183656495064497, + 0.03510604798793793, + 0.030366435647010803, + 0.04051664471626282, + 0.014448238536715508, + 0.014274956658482552, + 0.01584751531481743, + 0.013463162817060947, + 0.015759848058223724, + 0.013846911489963531, + 0.0163438618183136, + 0.016212182119488716, + 0.00761388149112463, + 0.0058311284519732, + 0.007434878032654524, + 0.010941307060420513, + 0.00575271574780345, + 0.008193129673600197, + 0.005847549065947533, + 0.008720438927412033, + 0.004024916794151068, + 0.003713341662660241, + 0.003415122162550688, + 0.0036870455369353294, + 0.0023830088321119547, + 0.004969542380422354, + 0.004413823131471872, + 0.004674283787608147, + 0.0027843709103763103, + 0.0023937581572681665, + 0.00298320478759706, + 0.006162737961858511, + 0.002517433138564229, + 0.003106585703790188, + 0.0027760223019868135, + 0.003156461985781789 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 4, + 6 + ], + "selected_reward": -0.029816951602697372 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a mother cat with a calico coat featuring white, orange, and black spots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are newborn kittens with closed eyes and pink skin visible huddled in the mother cat's arms?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene set inside a dark gray kennel lined with plush cream-colored faux fur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by a single warm golden light source from a smart pet device on the floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera begin with a medium shot from a slightly high angle and perform a slow, steady push-in?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"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": [ + 0.12375867366790771, + 0.23148968815803528, + 0.18691164255142212, + 0.12591788172721863, + 0.019920699298381805, + 0.09314313530921936, + 0.1163896769285202, + 0.029920512810349464, + 0.028005125001072884, + 0.044542886316776276 + ], + "credit_map": [ + 0.01258176751434803, + 0.02453947439789772, + 0.014091728255152702, + 0.015847399830818176, + 0.012866524048149586, + 0.01668720506131649, + 0.016175242140889168, + 0.010969343595206738, + 0.030737027525901794, + 0.03247690573334694, + 0.040850717574357986, + 0.032442349940538406, + 0.025412267073988914, + 0.03298323228955269, + 0.02016034908592701, + 0.01642683893442154, + 0.03000161238014698, + 0.02928799018263817, + 0.02889701910316944, + 0.022258298471570015, + 0.021590793505311012, + 0.024367347359657288, + 0.01704220660030842, + 0.013466399163007736, + 0.016351277008652687, + 0.019999422132968903, + 0.015310692600905895, + 0.020085135474801064, + 0.01486233714967966, + 0.01608426496386528, + 0.008525337092578411, + 0.014699436724185944, + 0.0020356702152639627, + 0.002507288008928299, + 0.0021650572307407856, + 0.0018887656042352319, + 0.001942453091032803, + 0.003065733704715967, + 0.0019783591851592064, + 0.0043373736552894115, + 0.0107890535145998, + 0.008430983871221542, + 0.012407990172505379, + 0.007852553389966488, + 0.008519771508872509, + 0.015103496611118317, + 0.014629517681896687, + 0.015409781597554684, + 0.011912766844034195, + 0.014028451405465603, + 0.011118317022919655, + 0.01025253813713789, + 0.016216138377785683, + 0.01968751847743988, + 0.012776251882314682, + 0.02039770409464836, + 0.0042120288126170635, + 0.003344061318784952, + 0.0037673618644475937, + 0.0038500528316944838, + 0.0034052219707518816, + 0.0040493318811059, + 0.003253004513680935, + 0.004039451479911804, + 0.005348104517906904, + 0.0034993356093764305, + 0.0025864194612950087, + 0.0036934639792889357, + 0.0024477625265717506, + 0.0036754142493009567, + 0.003116610459983349, + 0.003638015827164054, + 0.0039376127533614635, + 0.004369055852293968, + 0.004854577127844095, + 0.011907116509974003, + 0.005384199321269989, + 0.004609423223882914, + 0.004582277499139309, + 0.004898629151284695 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 2, + 3 + ], + "selected_reward": -0.012532292865216732 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "Beginning with a medium shot from a slightly high angle, the camera begins a slow, steady push-in, immersing the viewer in the understated, intimate, and focused scene. A mother cat, a domestic shorthair with a beautiful calico coat marked with distinct white, orange, and black spots, curls protectively around her litter in a tight \"C\" formation. They nestle in a soft, dark gray kennel lined with plush cream-colored faux fur. The central subject is a group of squirming newborn kittens, their eyes still closed, their sparse fur barely covering their pink skin. They huddle in their mother's warmth, some instinctively nuzzling for milk. The only illumination in the otherwise dark room comes from a nearby smart pet device on the floor, casting a warm, gentle, golden glow. This single light source frames the subject, capturing the soft texture of the mother's fur and creating soft highlights on the plush kennel lining, while casting long, deep shadows that stretch across the room, obscuring the background in darkness. The mother cat remains completely still, her expression one of serene alertness, creating a powerful atmosphere of tranquility, safety, and focused maternal love, while the deliberate movement of the camera draws the viewer close to this private, tender moment.", + "seed": 24025, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000121/group_6/index_5563_seed_24025_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": -0.04214998837560415, + "vlm_logprob": -0.04214998837560415, + "vlm_yes_no_margin": 3.425, + "vlm_token_credit": -0.04214998837560415, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a mother cat with a calico coat featuring white, orange, and black spots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are newborn kittens with closed eyes and pink skin visible huddled in the mother cat's arms?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene set inside a dark gray kennel lined with plush cream-colored faux fur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by a single warm golden light source from a smart pet device on the floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera begin with a medium shot from a slightly high angle and perform a slow, steady push-in?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0803455263376236 + ], + "opposite_logprob_tokens": [ + -2.580345630645752 + ], + "logprob_reward": -0.0803455263376236, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.0803455263376236, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18393352627754211, + 0.1297151744365692, + 0.06657425314188004, + 0.06560533493757248, + 0.16650798916816711, + 0.05726176127791405, + 0.17197874188423157, + 0.029034147039055824, + 0.05959632992744446, + 0.06979277729988098 + ], + "credit_map": [ + 0.02076561190187931, + 0.036001820117235184, + 0.02499835565686226, + 0.02632324770092964, + 0.02048545703291893, + 0.014153936877846718, + 0.01793556846678257, + 0.0232695285230875, + 0.015664495527744293, + 0.016608212143182755, + 0.021017413586378098, + 0.01596355251967907, + 0.015602023340761662, + 0.009186450392007828, + 0.017070665955543518, + 0.018602365627884865, + 0.011158059351146221, + 0.0070100086741149426, + 0.010318492539227009, + 0.007055747322738171, + 0.0067500886507332325, + 0.006509415339678526, + 0.007548419758677483, + 0.01022401824593544, + 0.010720495134592056, + 0.0072396136820316315, + 0.011902682483196259, + 0.00840874295681715, + 0.005637198220938444, + 0.006678602658212185, + 0.005241124890744686, + 0.00977687444537878, + 0.027754275128245354, + 0.021043725311756134, + 0.02716945670545101, + 0.020648520439863205, + 0.017555806785821915, + 0.012114329263567924, + 0.019854949787259102, + 0.020366940647363663, + 0.010032890364527702, + 0.008240501396358013, + 0.010562931187450886, + 0.005339560564607382, + 0.004041391424834728, + 0.005526326596736908, + 0.004666147753596306, + 0.008852011524140835, + 0.029106639325618744, + 0.02452869713306427, + 0.028437301516532898, + 0.023013802245259285, + 0.013050765730440617, + 0.018000677227973938, + 0.019077233970165253, + 0.016763608902692795, + 0.00517336605116725, + 0.004915241152048111, + 0.004039151128381491, + 0.002306886250153184, + 0.002816751366481185, + 0.0035413126461207867, + 0.0032471981830894947, + 0.0029942402616143227, + 0.008955645374953747, + 0.006839117966592312, + 0.008728251792490482, + 0.010072965174913406, + 0.004746058024466038, + 0.00520089128986001, + 0.010254749096930027, + 0.00479865213856101, + 0.009898596443235874, + 0.01233382523059845, + 0.01398110669106245, + 0.006515766493976116, + 0.0064139035530388355, + 0.00732190627604723, + 0.00713870394974947, + 0.006188967730849981 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 4 + ], + "selected_reward": -0.0803455263376236 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a mother cat with a calico coat featuring white, orange, and black spots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are newborn kittens with closed eyes and pink skin visible huddled in the mother cat's arms?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene set inside a dark gray kennel lined with plush cream-colored faux fur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by a single warm golden light source from a smart pet device on the floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera begin with a medium shot from a slightly high angle and perform a slow, steady push-in?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"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": [ + 0.09036389738321304, + 0.05068061873316765, + 0.0863502249121666, + 0.10065455734729767, + 0.11824766546487808, + 0.06340830773115158, + 0.1434134989976883, + 0.08060234785079956, + 0.050248973071575165, + 0.21602995693683624 + ], + "credit_map": [ + 0.012028208933770657, + 0.016554761677980423, + 0.01353397499769926, + 0.011499724350869656, + 0.012097084894776344, + 0.007926864549517632, + 0.00829700194299221, + 0.008426274172961712, + 0.006325419060885906, + 0.009142919443547726, + 0.007572265341877937, + 0.005040373187512159, + 0.0058525800704956055, + 0.006321764085441828, + 0.005613744258880615, + 0.004811553284525871, + 0.014043056406080723, + 0.0090408343821764, + 0.013946185819804668, + 0.011902796104550362, + 0.006862055975943804, + 0.008500990457832813, + 0.008101961575448513, + 0.013952340930700302, + 0.015518373809754848, + 0.01672760397195816, + 0.015060324221849442, + 0.011546099558472633, + 0.009294312447309494, + 0.009873072616755962, + 0.010081229731440544, + 0.01255354005843401, + 0.028413088992238045, + 0.009948886930942535, + 0.03876842185854912, + 0.010038003325462341, + 0.005718579515814781, + 0.007207774091511965, + 0.006331842392683029, + 0.011821076273918152, + 0.010071189142763615, + 0.005614405497908592, + 0.013327962718904018, + 0.007766827940940857, + 0.005236477125436068, + 0.005795696284621954, + 0.00908233318477869, + 0.0065134186297655106, + 0.02344578132033348, + 0.019475601613521576, + 0.022526947781443596, + 0.016726357862353325, + 0.011794372461736202, + 0.013180574402213097, + 0.018825743347406387, + 0.017438121140003204, + 0.01205459889024496, + 0.010820052586495876, + 0.010995431803166866, + 0.007608909625560045, + 0.007483732886612415, + 0.0069727408699691296, + 0.01523695420473814, + 0.009429931640625, + 0.007132512517273426, + 0.006101140286773443, + 0.011034279130399227, + 0.006500755436718464, + 0.00416563032194972, + 0.004361884202808142, + 0.007042108569294214, + 0.0039106630720198154, + 0.030270595103502274, + 0.03906578570604324, + 0.025575796142220497, + 0.029491839930415154, + 0.031069805845618248, + 0.020553244277834892, + 0.023580553010106087, + 0.0164223350584507 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 6, + 4 + ], + "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 mother cat with a calico coat featuring white, orange, and black spots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are newborn kittens with closed eyes and pink skin visible huddled in the mother cat's arms?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene set inside a dark gray kennel lined with plush cream-colored faux fur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by a single warm golden light source from a smart pet device on the floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera begin with a medium shot from a slightly high angle and perform a slow, steady push-in?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.062045034021139145 + ], + "opposite_logprob_tokens": [ + -2.812045097351074 + ], + "logprob_reward": -0.062045034021139145, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.062045034021139145, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04818826541304588, + 0.03933678939938545, + 0.05756732448935509, + 0.08979751914739609, + 0.19052955508232117, + 0.16555899381637573, + 0.2545289695262909, + 0.07820827513933182, + 0.010504485107958317, + 0.06577978283166885 + ], + "credit_map": [ + 0.006703156046569347, + 0.005710506346076727, + 0.007310754153877497, + 0.007098275236785412, + 0.004763192031532526, + 0.004406497813761234, + 0.004329074639827013, + 0.007866813801229, + 0.006198555696755648, + 0.003727972973138094, + 0.006399557925760746, + 0.004368534777313471, + 0.0032213896047323942, + 0.0030904291197657585, + 0.004223574418574572, + 0.008106771856546402, + 0.011971812695264816, + 0.004635324236005545, + 0.01096320804208517, + 0.0064138928428292274, + 0.005549625493586063, + 0.005541067570447922, + 0.006440992001444101, + 0.0060514044016599655, + 0.014192048460245132, + 0.008712978102266788, + 0.013449510559439659, + 0.01088931504637003, + 0.006687157787382603, + 0.00839291699230671, + 0.010881217196583748, + 0.016592375934123993, + 0.030522765591740608, + 0.02986125834286213, + 0.03627822920680046, + 0.02214973047375679, + 0.019751066341996193, + 0.013169101439416409, + 0.0214848592877388, + 0.01731254532933235, + 0.025581466034054756, + 0.024939564988017082, + 0.027375170961022377, + 0.01634984277188778, + 0.012971503660082817, + 0.014152214862406254, + 0.022340379655361176, + 0.021848853677511215, + 0.0443679541349411, + 0.03625140339136124, + 0.04413815215229988, + 0.021232301369309425, + 0.030886122956871986, + 0.023032519966363907, + 0.02744133211672306, + 0.027179179713129997, + 0.013926813378930092, + 0.013619338162243366, + 0.014191930182278156, + 0.00701812794432044, + 0.006095645017921925, + 0.005835721269249916, + 0.00851548369973898, + 0.009005211293697357, + 0.0017655567498877645, + 0.0014756092568859458, + 0.0019852304831147194, + 0.0007816728902980685, + 0.0009123042691498995, + 0.001287389430217445, + 0.0012184601509943604, + 0.001078262459486723, + 0.01056599523872137, + 0.015249725431203842, + 0.009094726294279099, + 0.006440786179155111, + 0.006359057035297155, + 0.006930618546903133, + 0.006028904113918543, + 0.005109966266900301 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 4, + 5 + ], + "selected_reward": -0.062045034021139145 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a mother cat with a calico coat featuring white, orange, and black spots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are newborn kittens with closed eyes and pink skin visible huddled in the mother cat's arms?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene set inside a dark gray kennel lined with plush cream-colored faux fur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by a single warm golden light source from a smart pet device on the floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera begin with a medium shot from a slightly high angle and perform a slow, steady push-in?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"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": [ + 0.10028732568025589, + 0.06435059756040573, + 0.10734958201646805, + 0.10102929919958115, + 0.2185373455286026, + 0.16715764999389648, + 0.11747545748949051, + 0.05398623272776604, + 0.028994930908083916, + 0.040831539779901505 + ], + "credit_map": [ + 0.013577191159129143, + 0.012143483385443687, + 0.013549532741308212, + 0.012642394751310349, + 0.014308042824268341, + 0.008911298587918282, + 0.012126365676522255, + 0.013029023073613644, + 0.008742243982851505, + 0.006041625514626503, + 0.010975565761327744, + 0.007797546684741974, + 0.006206826772540808, + 0.0065032439306378365, + 0.008394730277359486, + 0.009688818827271461, + 0.019258886575698853, + 0.010517638176679611, + 0.015255594626069069, + 0.01620873250067234, + 0.013990153558552265, + 0.00789942778646946, + 0.014598257839679718, + 0.009620890952646732, + 0.01701827347278595, + 0.011696457862854004, + 0.019188929349184036, + 0.013138676062226295, + 0.010589828714728355, + 0.007186272647231817, + 0.00900867860764265, + 0.013202185742557049, + 0.03966381773352623, + 0.038141582161188126, + 0.04575882852077484, + 0.02047470398247242, + 0.016475338488817215, + 0.017170891165733337, + 0.016141634434461594, + 0.024710573256015778, + 0.031607735902071, + 0.01942364126443863, + 0.03329024463891983, + 0.013547717593610287, + 0.008751888759434223, + 0.0208723284304142, + 0.014509745873510838, + 0.025154368951916695, + 0.019492460414767265, + 0.019408082589507103, + 0.019823163747787476, + 0.010723093524575233, + 0.009644027799367905, + 0.010564817115664482, + 0.013043462298810482, + 0.014776352792978287, + 0.009968151338398457, + 0.007437553256750107, + 0.011253674514591694, + 0.003701288253068924, + 0.0056842933408916, + 0.005245306994765997, + 0.005796367302536964, + 0.004899602383375168, + 0.004813590552657843, + 0.00445374334231019, + 0.00610268022865057, + 0.0023420490324497223, + 0.002474214881658554, + 0.003994176164269447, + 0.0027124162297695875, + 0.0021020632702857256, + 0.005314302165061235, + 0.006863588001579046, + 0.00448976643383503, + 0.005468135233968496, + 0.004723984748125076, + 0.004354291129857302, + 0.005388779100030661, + 0.0042286948300898075 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 5, + 6 + ], + "selected_reward": -0.04307607561349869 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a mother cat with a calico coat featuring white, orange, and black spots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are newborn kittens with closed eyes and pink skin visible huddled in the mother cat's arms?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene set inside a dark gray kennel lined with plush cream-colored faux fur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by a single warm golden light source from a smart pet device on the floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera begin with a medium shot from a slightly high angle and perform a slow, steady push-in?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011069553904235363 + ], + "opposite_logprob_tokens": [ + -4.5110697746276855 + ], + "logprob_reward": -0.011069553904235363, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011069553904235363, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08622895181179047, + 0.02763691172003746, + 0.05490723252296448, + 0.08489461243152618, + 0.25984251499176025, + 0.09745980054140091, + 0.13330279290676117, + 0.0558026060461998, + 0.15441976487636566, + 0.04550488293170929 + ], + "credit_map": [ + 0.009355826303362846, + 0.016215724870562553, + 0.010272081010043621, + 0.013222930952906609, + 0.008763255551457405, + 0.007417130284011364, + 0.008252475410699844, + 0.012729532085359097, + 0.0032249558717012405, + 0.0036680237390100956, + 0.004127228632569313, + 0.004001673310995102, + 0.0029048779979348183, + 0.0029489987064152956, + 0.003591142361983657, + 0.0031700083054602146, + 0.010721642524003983, + 0.0057615675032138824, + 0.008980564773082733, + 0.00584301445633173, + 0.009598552249372005, + 0.003889941843226552, + 0.006342209409922361, + 0.0037697399966418743, + 0.013873957097530365, + 0.018625350669026375, + 0.011854414828121662, + 0.008008135482668877, + 0.009524385444819927, + 0.008880849927663803, + 0.0065430449321866035, + 0.007584478240460157, + 0.057887002825737, + 0.01708674430847168, + 0.0783621221780777, + 0.025330103933811188, + 0.015972618013620377, + 0.013809473253786564, + 0.022551730275154114, + 0.028842708095908165, + 0.015862876549363136, + 0.010908917523920536, + 0.017520736902952194, + 0.01018951740115881, + 0.008055290207266808, + 0.009268363006412983, + 0.010759733617305756, + 0.014894366264343262, + 0.02413298934698105, + 0.01854761876165867, + 0.021749241277575493, + 0.011789257638156414, + 0.011242729611694813, + 0.016153380274772644, + 0.012348202988505363, + 0.017339378595352173, + 0.008747927844524384, + 0.007405691314488649, + 0.009989793412387371, + 0.003693401115015149, + 0.0043546827509999275, + 0.005577012896537781, + 0.008721367456018925, + 0.007312730886042118, + 0.024357963353395462, + 0.029160531237721443, + 0.02991531416773796, + 0.01622910611331463, + 0.011050185188651085, + 0.01253199391067028, + 0.017979370430111885, + 0.013195301406085491, + 0.006929166615009308, + 0.010807343758642673, + 0.00806410051882267, + 0.004171257372945547, + 0.004795162472873926, + 0.0036233013961464167, + 0.003544227685779333, + 0.0035703275352716446 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 8, + 6 + ], + "selected_reward": -0.011069553904235363 + } + ], + "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/0000121/group_7/index_tensor([6007, 6007])_group_7_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000121/group_7/index_tensor([6007, 6007])_group_7_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..151d2bf17b4a37108d105fbbd2c2f66fd12bd60c --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000121/group_7/index_tensor([6007, 6007])_group_7_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "This medium-wide shot, initially static and humorously stylized, captures the portly samurai Azumi Jigorou attempting to gracefully cross a vast, shimmering frozen pond beneath a crisp winter sky. He wears a traditional dark blue robe and vest, his enormous belly straining against the fabric, and a katana neatly tucked into his belt. He comically tiptoes across the slick ice with delicate, tentative steps, his face a mask of stoic concentration. The massive tsunami wave looms impossibly behind him, its massive crest a crystalline sculpture of emerald green ice and still white foam, dwarfing the surrounding snow-dusted pine trees. As he turns his head slightly, his stoic expression instantly melts into pure slapstick terror; the camera zooms in quickly and harshly, tightly framing his wide, panicked eyes as they first register the absurd, frozen catastrophe. In this extreme close-up, a perfectly rendered bead of sweat flies from his forehead in exaggerated slow motion, arcing through the air and capturing the bright, clear light that punctuates his moment of cartoonish terror.", + "seed": 32259, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000121/group_7/index_6007_seed_32259_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": -0.035805993364192545, + "vlm_logprob": -0.035805993364192545, + "vlm_yes_no_margin": 4.45, + "vlm_token_credit": -0.035805993364192545, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main character a portly samurai named Azumi Jigorou wearing a traditional dark blue robe and vest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character comically tiptoeing across a vast, shimmering frozen pond with a massive tsunami wave of emerald green ice looming behind him?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the character's expression change from stoic concentration to slapstick terror as he turns his head?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera zoom in quickly and harshly to an extreme close-up of the character's wide, panicked eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a perfectly rendered bead of sweat shown flying from the character's forehead in exaggerated 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": [ + -0.10680640488862991 + ], + "opposite_logprob_tokens": [ + -2.356806516647339 + ], + "logprob_reward": -0.10680640488862991, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10680640488862991, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4347890019416809, + 0.06504781544208527, + 0.026034660637378693, + 0.07960064709186554, + 0.07684078812599182, + 0.024467650800943375, + 0.0501965656876564, + 0.1386057585477829, + 0.038366228342056274, + 0.06605078279972076 + ], + "credit_map": [ + 0.08422814309597015, + 0.05261343717575073, + 0.02998504415154457, + 0.1680103987455368, + 0.03353465721011162, + 0.017840508371591568, + 0.03598843142390251, + 0.012588446028530598, + 0.009038659743964672, + 0.011796637438237667, + 0.008120809681713581, + 0.007807914633303881, + 0.00943544041365385, + 0.009307186119258404, + 0.005830319132655859, + 0.0037108552642166615, + 0.003909226972609758, + 0.002077830955386162, + 0.002781486604362726, + 0.00370756140910089, + 0.003805624321103096, + 0.00433653499931097, + 0.0032444088719785213, + 0.002171990694478154, + 0.005847979336977005, + 0.004358949605375528, + 0.021061362698674202, + 0.026733793318271637, + 0.0038252107333391905, + 0.00551019050180912, + 0.007067137397825718, + 0.005196033976972103, + 0.012500930577516556, + 0.011791963130235672, + 0.011717959307134151, + 0.010402543470263481, + 0.0077549791894853115, + 0.006551424041390419, + 0.01080285757780075, + 0.005318147595971823, + 0.00268630962818861, + 0.0032519851811230183, + 0.004927400965243578, + 0.002848957199603319, + 0.0030628975946456194, + 0.002141005825251341, + 0.004012280143797398, + 0.0015368192689493299, + 0.0056449188850820065, + 0.006071218755096197, + 0.006836879067122936, + 0.006213354412466288, + 0.0036364609841257334, + 0.006437415257096291, + 0.007632439490407705, + 0.007723886985331774, + 0.02010822854936123, + 0.019358443096280098, + 0.016124093905091286, + 0.016973720863461494, + 0.012098988518118858, + 0.018429359421133995, + 0.020160416141152382, + 0.015352526679635048, + 0.009452815167605877, + 0.004147327970713377, + 0.0025905794464051723, + 0.003714915132150054, + 0.004194600507616997, + 0.005765464622527361, + 0.004515858832746744, + 0.003984672948718071, + 0.011498253792524338, + 0.006984712556004524, + 0.00239339261315763, + 0.009583061560988426, + 0.007471802178770304, + 0.01656179130077362, + 0.008021657355129719, + 0.003536124946549535 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 3 + ], + "selected_reward": -0.10680640488862991 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main character a portly samurai named Azumi Jigorou wearing a traditional dark blue robe and vest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character comically tiptoeing across a vast, shimmering frozen pond with a massive tsunami wave of emerald green ice looming behind him?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the character's expression change from stoic concentration to slapstick terror as he turns his head?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera zoom in quickly and harshly to an extreme close-up of the character's wide, panicked eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a perfectly rendered bead of sweat shown flying from the character's forehead in exaggerated 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": [ + -0.0036272702272981405 + ], + "opposite_logprob_tokens": [ + -5.628627300262451 + ], + "logprob_reward": -0.0036272702272981405, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.0036272702272981405, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4190414249897003, + 0.20657551288604736, + 0.03976908698678017, + 0.04065583273768425, + 0.07661235332489014, + 0.03416068106889725, + 0.04328005015850067, + 0.0654965415596962, + 0.036787260323762894, + 0.03762117400765419 + ], + "credit_map": [ + 0.09102562069892883, + 0.04109162092208862, + 0.02269565314054489, + 0.171016126871109, + 0.026902157813310623, + 0.018969345837831497, + 0.03745662793517113, + 0.009884342551231384, + 0.025647616013884544, + 0.04213202744722366, + 0.025736838579177856, + 0.017748195677995682, + 0.04632207006216049, + 0.021121269091963768, + 0.016560645774006844, + 0.011306866072118282, + 0.005924595985561609, + 0.004091761540621519, + 0.005512069445103407, + 0.005259376019239426, + 0.007178097497671843, + 0.004788239020854235, + 0.003513609990477562, + 0.0035013400483876467, + 0.006109293550252914, + 0.003059378359466791, + 0.011526023969054222, + 0.006507159676402807, + 0.0033241072669625282, + 0.004015488084405661, + 0.003333357162773609, + 0.0027810288593173027, + 0.010369714349508286, + 0.010601174086332321, + 0.015742897987365723, + 0.010675597004592419, + 0.006049558520317078, + 0.009309432469308376, + 0.00825431663542986, + 0.005609669256955385, + 0.0038419526536017656, + 0.004609725438058376, + 0.0058593242429196835, + 0.0037170497234910727, + 0.005130062811076641, + 0.0026950084138661623, + 0.006071896757930517, + 0.0022356645204126835, + 0.008163769729435444, + 0.005816866643726826, + 0.0048628575168550014, + 0.005436449311673641, + 0.0030812732875347137, + 0.008224650286138058, + 0.0046859560534358025, + 0.003008228028193116, + 0.009875844232738018, + 0.007424868177622557, + 0.009002753533422947, + 0.008191880770027637, + 0.004876984748989344, + 0.008908507414162159, + 0.009610333479940891, + 0.007605371996760368, + 0.007501781452447176, + 0.006250497419387102, + 0.0020921737886965275, + 0.0037740010302513838, + 0.005214828997850418, + 0.005775717552751303, + 0.003714128863066435, + 0.0024641333147883415, + 0.006510548293590546, + 0.0038027195259928703, + 0.0020590582862496376, + 0.005871635861694813, + 0.00409326795488596, + 0.008377209305763245, + 0.004001120571047068, + 0.0029056179337203503 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 4 + ], + "selected_reward": -0.0036272702272981405 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main character a portly samurai named Azumi Jigorou wearing a traditional dark blue robe and vest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character comically tiptoeing across a vast, shimmering frozen pond with a massive tsunami wave of emerald green ice looming behind him?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the character's expression change from stoic concentration to slapstick terror as he turns his head?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera zoom in quickly and harshly to an extreme close-up of the character's wide, panicked eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a perfectly rendered bead of sweat shown flying from the character's forehead in exaggerated 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": [ + -0.0019422968616709113 + ], + "opposite_logprob_tokens": [ + -6.251942157745361 + ], + "logprob_reward": -0.0019422968616709113, + "yes_no_margin_reward": 6.25, + "token_credit_reward": -0.0019422968616709113, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5734855532646179, + 0.06197851151227951, + 0.020170897245407104, + 0.03143100067973137, + 0.019629018381237984, + 0.024707172065973282, + 0.036416053771972656, + 0.040609460324048996, + 0.11496169120073318, + 0.07661066949367523 + ], + "credit_map": [ + 0.13055966794490814, + 0.058647431433200836, + 0.03521931543946266, + 0.2520085275173187, + 0.019242331385612488, + 0.014989927411079407, + 0.0495799221098423, + 0.013238430954515934, + 0.007638594601303339, + 0.012604190967977047, + 0.007247913163155317, + 0.005860333330929279, + 0.01348677184432745, + 0.007215694058686495, + 0.00438039842993021, + 0.003544613951817155, + 0.003940184134989977, + 0.0018308153375983238, + 0.0021396318916231394, + 0.0035879267379641533, + 0.002852512989193201, + 0.002540215151384473, + 0.0018180517945438623, + 0.0014615588588640094, + 0.002879827283322811, + 0.0018459181301295757, + 0.004490130115300417, + 0.007053124252706766, + 0.002835745457559824, + 0.004997132811695337, + 0.004422962199896574, + 0.002906161593273282, + 0.0033034253865480423, + 0.002837661886587739, + 0.001956506399437785, + 0.002463816897943616, + 0.0019954585004597902, + 0.0032549814786762, + 0.0019729810301214457, + 0.0018441876163706183, + 0.003402445465326309, + 0.0037978270556777716, + 0.003983887378126383, + 0.0024618085008114576, + 0.0030738513451069593, + 0.002145644510164857, + 0.003941396716982126, + 0.0019003101624548435, + 0.005071881227195263, + 0.006191962864249945, + 0.005555588286370039, + 0.004406003747135401, + 0.003551153466105461, + 0.005037572234869003, + 0.0039541213773190975, + 0.0026477694045752287, + 0.00868558045476675, + 0.005365883000195026, + 0.003809780813753605, + 0.004570512566715479, + 0.0034943940117955208, + 0.006248227786272764, + 0.004951725713908672, + 0.0034833571407943964, + 0.022445859387516975, + 0.014096717350184917, + 0.007633657660335302, + 0.017183413729071617, + 0.012020534835755825, + 0.014748243615031242, + 0.016166741028428078, + 0.01066652312874794, + 0.014341619797050953, + 0.0069761802442371845, + 0.0031827203929424286, + 0.011248789727687836, + 0.007103012874722481, + 0.021669644862413406, + 0.0075197056867182255, + 0.0045689987018704414 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 9 + ], + "selected_reward": -0.0019422968616709113 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main character a portly samurai named Azumi Jigorou wearing a traditional dark blue robe and vest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character comically tiptoeing across a vast, shimmering frozen pond with a massive tsunami wave of emerald green ice looming behind him?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the character's expression change from stoic concentration to slapstick terror as he turns his head?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera zoom in quickly and harshly to an extreme close-up of the character's wide, panicked eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a perfectly rendered bead of sweat shown flying from the character's forehead in exaggerated 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": [ + -0.004633048549294472 + ], + "opposite_logprob_tokens": [ + -5.379632949829102 + ], + "logprob_reward": -0.004633048549294472, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004633048549294472, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16344520449638367, + 0.08817349374294281, + 0.03261032700538635, + 0.12612031400203705, + 0.06019528955221176, + 0.07266949117183685, + 0.11485828459262848, + 0.06440036743879318, + 0.11331018060445786, + 0.1642169952392578 + ], + "credit_map": [ + 0.03277158364653587, + 0.026142871007323265, + 0.012278852052986622, + 0.043017081916332245, + 0.021547306329011917, + 0.01106538437306881, + 0.010233267210423946, + 0.006388859823346138, + 0.011625991202890873, + 0.015025285072624683, + 0.011756060644984245, + 0.00885094702243805, + 0.01688242517411709, + 0.011625233106315136, + 0.007772098761051893, + 0.0046354555524885654, + 0.005615259986370802, + 0.0036112850066274405, + 0.003900517476722598, + 0.004703067243099213, + 0.004018977750092745, + 0.004219597205519676, + 0.0035707936622202396, + 0.0029708310030400753, + 0.00816999189555645, + 0.007096339017152786, + 0.04314453899860382, + 0.0463322289288044, + 0.006274177692830563, + 0.00608578696846962, + 0.005128357093781233, + 0.0038888880517333746, + 0.008770221844315529, + 0.007096129469573498, + 0.009898045100271702, + 0.007397200912237167, + 0.006740344222635031, + 0.004825873300433159, + 0.006665787659585476, + 0.008801690302789211, + 0.009272039867937565, + 0.009674402885138988, + 0.009507996030151844, + 0.008650208823382854, + 0.009334082715213299, + 0.007704385556280613, + 0.012085183523595333, + 0.006441192235797644, + 0.015764405950903893, + 0.017272094264626503, + 0.010492892004549503, + 0.019580921158194542, + 0.011574040167033672, + 0.015235482715070248, + 0.015087147243320942, + 0.009851298294961452, + 0.009754092432558537, + 0.007821260020136833, + 0.008398571982979774, + 0.007071909960359335, + 0.00568184582516551, + 0.008833931758999825, + 0.009620008990168571, + 0.007218748331069946, + 0.018076548352837563, + 0.018249113112688065, + 0.007336885202676058, + 0.011796033009886742, + 0.02050403691828251, + 0.015034371055662632, + 0.013235969468951225, + 0.00907721184194088, + 0.03213758021593094, + 0.01706191897392273, + 0.007539394311606884, + 0.023345354944467545, + 0.011803768575191498, + 0.0477711483836174, + 0.01589382253587246, + 0.008663996122777462 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 3 + ], + "selected_reward": -0.004633048549294472 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main character a portly samurai named Azumi Jigorou wearing a traditional dark blue robe and vest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character comically tiptoeing across a vast, shimmering frozen pond with a massive tsunami wave of emerald green ice looming behind him?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the character's expression change from stoic concentration to slapstick terror as he turns his head?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera zoom in quickly and harshly to an extreme close-up of the character's wide, panicked eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a perfectly rendered bead of sweat shown flying from the character's forehead in exaggerated 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": [ + -0.06202094629406929 + ], + "opposite_logprob_tokens": [ + -2.812021017074585 + ], + "logprob_reward": -0.06202094629406929, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06202094629406929, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.59279465675354, + 0.12209004908800125, + 0.016467919573187828, + 0.028065016493201256, + 0.021786389872431755, + 0.03127114474773407, + 0.040823422372341156, + 0.0435485914349556, + 0.046431202441453934, + 0.05672162398695946 + ], + "credit_map": [ + 0.14092501997947693, + 0.0594666562974453, + 0.03221973404288292, + 0.2583380937576294, + 0.020806951448321342, + 0.016328515484929085, + 0.05032291263341904, + 0.014386787079274654, + 0.01999986730515957, + 0.015299666672945023, + 0.015505589544773102, + 0.01366580743342638, + 0.02410900592803955, + 0.013962688855826855, + 0.012784963473677635, + 0.006762458942830563, + 0.0017637369455769658, + 0.0019853347912430763, + 0.0012918125139549375, + 0.002395352115854621, + 0.003353578969836235, + 0.0026463414542376995, + 0.0013695016968995333, + 0.0016622617840766907, + 0.00306920288130641, + 0.0029040956869721413, + 0.004164404701441526, + 0.003493346506729722, + 0.0034097093157470226, + 0.003781696781516075, + 0.004245433956384659, + 0.002997125731781125, + 0.0024937065318226814, + 0.003528804751113057, + 0.0028570385184139013, + 0.002334363292902708, + 0.0025314129889011383, + 0.003217945108190179, + 0.0021533386316150427, + 0.0026697805151343346, + 0.003743389155715704, + 0.003617370268329978, + 0.0037617264315485954, + 0.003105516778305173, + 0.005469752941280603, + 0.0037894032429903746, + 0.004506370052695274, + 0.0032776189036667347, + 0.004968300927430391, + 0.006797707639634609, + 0.003792364848777652, + 0.009324155747890472, + 0.0034666634164750576, + 0.004763209726661444, + 0.004958839155733585, + 0.0027521797455847263, + 0.008058993145823479, + 0.005328852217644453, + 0.005985247902572155, + 0.004835621453821659, + 0.0033998957369476557, + 0.005700747016817331, + 0.006370616145431995, + 0.0038686171174049377, + 0.005981347057968378, + 0.008895764127373695, + 0.002627864247187972, + 0.0067638009786605835, + 0.007096653338521719, + 0.005211622454226017, + 0.006101307459175587, + 0.0037528423126786947, + 0.01046090479940176, + 0.005448011681437492, + 0.002186601748690009, + 0.008281232789158821, + 0.005835691932588816, + 0.016219817101955414, + 0.005469960160553455, + 0.002819401677697897 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 9 + ], + "selected_reward": -0.06202094629406929 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "This medium-wide shot, initially static and humorously stylized, captures the portly samurai Azumi Jigorou attempting to gracefully cross a vast, shimmering frozen pond beneath a crisp winter sky. He wears a traditional dark blue robe and vest, his enormous belly straining against the fabric, and a katana neatly tucked into his belt. He comically tiptoes across the slick ice with delicate, tentative steps, his face a mask of stoic concentration. The massive tsunami wave looms impossibly behind him, its massive crest a crystalline sculpture of emerald green ice and still white foam, dwarfing the surrounding snow-dusted pine trees. As he turns his head slightly, his stoic expression instantly melts into pure slapstick terror; the camera zooms in quickly and harshly, tightly framing his wide, panicked eyes as they first register the absurd, frozen catastrophe. In this extreme close-up, a perfectly rendered bead of sweat flies from his forehead in exaggerated slow motion, arcing through the air and capturing the bright, clear light that punctuates his moment of cartoonish terror.", + "seed": 32260, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000121/group_7/index_6007_seed_32260_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": -0.0426481700502336, + "vlm_logprob": -0.0426481700502336, + "vlm_yes_no_margin": 3.85, + "vlm_token_credit": -0.0426481700502336, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main character a portly samurai named Azumi Jigorou wearing a traditional dark blue robe and vest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character comically tiptoeing across a vast, shimmering frozen pond with a massive tsunami wave of emerald green ice looming behind him?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the character's expression change from stoic concentration to slapstick terror as he turns his head?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera zoom in quickly and harshly to an extreme close-up of the character's wide, panicked eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a perfectly rendered bead of sweat shown flying from the character's forehead in exaggerated 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": [ + -0.07667648047208786 + ], + "opposite_logprob_tokens": [ + -2.701676368713379 + ], + "logprob_reward": -0.07667648047208786, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07667648047208786, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08168676495552063, + 0.18547672033309937, + 0.03344417363405228, + 0.03148926421999931, + 0.23858441412448883, + 0.11988168209791183, + 0.09928754717111588, + 0.06229778751730919, + 0.08755340427160263, + 0.060298167169094086 + ], + "credit_map": [ + 0.007528812624514103, + 0.00923913810402155, + 0.007926199585199356, + 0.0062458934262394905, + 0.010056179016828537, + 0.01548725739121437, + 0.010642469860613346, + 0.014560816809535027, + 0.055904682725667953, + 0.012671391479671001, + 0.0608435682952404, + 0.012828128412365913, + 0.009395319037139416, + 0.012221029959619045, + 0.013023106381297112, + 0.008589502424001694, + 0.004983812570571899, + 0.003722032532095909, + 0.00912279449403286, + 0.0033659276086837053, + 0.0027815760113298893, + 0.0037484781350940466, + 0.0028940029442310333, + 0.002825549803674221, + 0.00414134981110692, + 0.0022421167232096195, + 0.007591244298964739, + 0.00297733792103827, + 0.0019757847767323256, + 0.005025217775255442, + 0.00355293950997293, + 0.003983271773904562, + 0.0263033676892519, + 0.030942237004637718, + 0.039283692836761475, + 0.039289750158786774, + 0.025499677285552025, + 0.024646878242492676, + 0.02892978861927986, + 0.023689033463597298, + 0.019291561096906662, + 0.020649336278438568, + 0.010536490008234978, + 0.015008503571152687, + 0.013577837496995926, + 0.015521072782576084, + 0.013638301752507687, + 0.011658580042421818, + 0.012389348819851875, + 0.010495644994080067, + 0.01022286806255579, + 0.010664938017725945, + 0.013451674021780491, + 0.011350645683705807, + 0.013124670833349228, + 0.01758776232600212, + 0.009219652973115444, + 0.00391471479088068, + 0.005375554319471121, + 0.005064164288341999, + 0.007863548584282398, + 0.016904711723327637, + 0.0044786278158426285, + 0.009476812556385994, + 0.008138809353113174, + 0.007736317813396454, + 0.013695355504751205, + 0.0100374361500144, + 0.01619425043463707, + 0.010426236316561699, + 0.010492497123777866, + 0.010832499712705612, + 0.006960439030081034, + 0.0060186320915818214, + 0.006416254676878452, + 0.008345277048647404, + 0.006614003796130419, + 0.008701865561306477, + 0.010781901888549328, + 0.0064597963355481625 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 1, + 5 + ], + "selected_reward": -0.07667648047208786 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main character a portly samurai named Azumi Jigorou wearing a traditional dark blue robe and vest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character comically tiptoeing across a vast, shimmering frozen pond with a massive tsunami wave of emerald green ice looming behind him?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the character's expression change from stoic concentration to slapstick terror as he turns his head?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera zoom in quickly and harshly to an extreme close-up of the character's wide, panicked eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a perfectly rendered bead of sweat shown flying from the character's forehead in exaggerated 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": [ + -0.0067533585242927074 + ], + "opposite_logprob_tokens": [ + -5.006753444671631 + ], + "logprob_reward": -0.0067533585242927074, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.0067533585242927074, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11598749458789825, + 0.14831319451332092, + 0.07456806302070618, + 0.057918038219213486, + 0.19106455147266388, + 0.0845428928732872, + 0.09209324419498444, + 0.09654778242111206, + 0.05178869888186455, + 0.08717604726552963 + ], + "credit_map": [ + 0.01307748444378376, + 0.010168914683163166, + 0.014743172563612461, + 0.007183095905929804, + 0.014113659039139748, + 0.016457224264740944, + 0.01781780645251274, + 0.02242613583803177, + 0.020519014447927475, + 0.019016319885849953, + 0.013435380533337593, + 0.012872115708887577, + 0.02219850942492485, + 0.01893710158765316, + 0.022617477923631668, + 0.01871727779507637, + 0.009610814973711967, + 0.0057924590073525906, + 0.008996238000690937, + 0.005036638118326664, + 0.007900362834334373, + 0.009120323695242405, + 0.013111520558595657, + 0.014999707229435444, + 0.0067728785797953606, + 0.0037500180769711733, + 0.011289071291685104, + 0.005929505918174982, + 0.0038476826157420874, + 0.008207232691347599, + 0.007215054240077734, + 0.010906594805419445, + 0.022014496847987175, + 0.013247174210846424, + 0.028443612158298492, + 0.012434154748916626, + 0.015225168317556381, + 0.019856614992022514, + 0.023666469380259514, + 0.05617685988545418, + 0.01113143190741539, + 0.01566913351416588, + 0.0062771933153271675, + 0.010128996334969997, + 0.011335990391671658, + 0.01043492928147316, + 0.010281665250658989, + 0.009283561259508133, + 0.010730022564530373, + 0.012447281740605831, + 0.007682796102017164, + 0.013751396909356117, + 0.00878399983048439, + 0.007754061371088028, + 0.013301716186106205, + 0.017641963437199593, + 0.015857137739658356, + 0.005680474452674389, + 0.007996915839612484, + 0.007113541942089796, + 0.008771030232310295, + 0.031129663810133934, + 0.006170476321130991, + 0.013828546740114689, + 0.005760009400546551, + 0.003935143351554871, + 0.00820770114660263, + 0.005581257399171591, + 0.006595856510102749, + 0.0070559000596404076, + 0.005466710310429335, + 0.009186120703816414, + 0.009603877551853657, + 0.009052586741745472, + 0.010244381614029408, + 0.012175975367426872, + 0.009813380427658558, + 0.008320851251482964, + 0.015555987134575844, + 0.012409011833369732 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 1, + 0 + ], + "selected_reward": -0.0067533585242927074 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main character a portly samurai named Azumi Jigorou wearing a traditional dark blue robe and vest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character comically tiptoeing across a vast, shimmering frozen pond with a massive tsunami wave of emerald green ice looming behind him?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the character's expression change from stoic concentration to slapstick terror as he turns his head?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera zoom in quickly and harshly to an extreme close-up of the character's wide, panicked eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a perfectly rendered bead of sweat shown flying from the character's forehead in exaggerated 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": [ + -0.003193756565451622 + ], + "opposite_logprob_tokens": [ + -5.7531938552856445 + ], + "logprob_reward": -0.003193756565451622, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.003193756565451622, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10840358585119247, + 0.09580223262310028, + 0.06520965695381165, + 0.05232135206460953, + 0.03622984513640404, + 0.06035308167338371, + 0.12943685054779053, + 0.18213512003421783, + 0.07487323135137558, + 0.19523502886295319 + ], + "credit_map": [ + 0.01137989480048418, + 0.014516154304146767, + 0.011238053441047668, + 0.009212189354002476, + 0.012538979761302471, + 0.021385416388511658, + 0.013680766336619854, + 0.014452127739787102, + 0.014273491688072681, + 0.012005697004497051, + 0.015107092447578907, + 0.010080278851091862, + 0.013246226124465466, + 0.010733972303569317, + 0.01004348136484623, + 0.010311986319720745, + 0.008302652277052402, + 0.006314811296761036, + 0.005796598270535469, + 0.005847661755979061, + 0.008520673029124737, + 0.007925924845039845, + 0.011187909170985222, + 0.01131343375891447, + 0.007280105724930763, + 0.007092246785759926, + 0.0074122315272688866, + 0.008037290535867214, + 0.00457308953627944, + 0.0057249292731285095, + 0.005934616085141897, + 0.006266843993216753, + 0.004420917481184006, + 0.004093982744961977, + 0.005252243019640446, + 0.005202003754675388, + 0.004184477496892214, + 0.003927378915250301, + 0.004431483801454306, + 0.004717355594038963, + 0.008451854810118675, + 0.010873572900891304, + 0.006469403859227896, + 0.00647760508581996, + 0.0074299974367022514, + 0.007209542207419872, + 0.007220957428216934, + 0.0062201484106481075, + 0.01595223881304264, + 0.019795823842287064, + 0.011727728880941868, + 0.02578771486878395, + 0.012382812798023224, + 0.01013085339218378, + 0.014258969575166702, + 0.019400710240006447, + 0.030531570315361023, + 0.019084500148892403, + 0.019636696204543114, + 0.014784363098442554, + 0.018732944503426552, + 0.049172356724739075, + 0.010595318861305714, + 0.0195973739027977, + 0.010662119835615158, + 0.005713192280381918, + 0.006139086093753576, + 0.005523946136236191, + 0.010993091389536858, + 0.014159688726067543, + 0.009029394946992397, + 0.012652707286179066, + 0.015262527391314507, + 0.02521033026278019, + 0.016454361379146576, + 0.03368952125310898, + 0.02705942839384079, + 0.019868316128849983, + 0.03502220660448074, + 0.02266833744943142 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 7, + 6 + ], + "selected_reward": -0.003193756565451622 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main character a portly samurai named Azumi Jigorou wearing a traditional dark blue robe and vest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character comically tiptoeing across a vast, shimmering frozen pond with a massive tsunami wave of emerald green ice looming behind him?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the character's expression change from stoic concentration to slapstick terror as he turns his head?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera zoom in quickly and harshly to an extreme close-up of the character's wide, panicked eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a perfectly rendered bead of sweat shown flying from the character's forehead in exaggerated 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": [ + -0.026335397735238075 + ], + "opposite_logprob_tokens": [ + -3.6513354778289795 + ], + "logprob_reward": -0.026335397735238075, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026335397735238075, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06243184208869934, + 0.050578318536281586, + 0.07849353551864624, + 0.04694847762584686, + 0.07313402742147446, + 0.032299868762493134, + 0.11655280739068985, + 0.16227009892463684, + 0.0927530825138092, + 0.2845379710197449 + ], + "credit_map": [ + 0.00660566333681345, + 0.00897129625082016, + 0.008295315317809582, + 0.0056227329187095165, + 0.00801593717187643, + 0.009187459945678711, + 0.009605667553842068, + 0.006127769127488136, + 0.013338078744709492, + 0.0035666432231664658, + 0.012047793716192245, + 0.003056054003536701, + 0.004808368161320686, + 0.0036347373388707638, + 0.005805606953799725, + 0.004321039188653231, + 0.009408153593540192, + 0.006314971949905157, + 0.010770825669169426, + 0.006588059943169355, + 0.010521785356104374, + 0.009561521001160145, + 0.012803108431398869, + 0.012525113299489021, + 0.005870429798960686, + 0.004033701028674841, + 0.005976931657642126, + 0.0038226228207349777, + 0.0028857968281954527, + 0.006367027759552002, + 0.006925799418240786, + 0.011066164821386337, + 0.01183150615543127, + 0.007654903456568718, + 0.015445219352841377, + 0.010150153189897537, + 0.005559877958148718, + 0.008790413849055767, + 0.006410769186913967, + 0.007291179616004229, + 0.003928621299564838, + 0.0050987317226827145, + 0.00285753863863647, + 0.00494479900225997, + 0.004684274084866047, + 0.003524461295455694, + 0.004347353707998991, + 0.0029140887781977654, + 0.019051726907491684, + 0.010174437426030636, + 0.013735847547650337, + 0.013217155821621418, + 0.014936668798327446, + 0.014980879612267017, + 0.015674224123358727, + 0.01478186622262001, + 0.021591918542981148, + 0.014642054215073586, + 0.022768910974264145, + 0.011083897203207016, + 0.019132176414132118, + 0.03495927155017853, + 0.018151162192225456, + 0.019940704107284546, + 0.014780920930206776, + 0.007624673191457987, + 0.0074579743668437, + 0.008418027311563492, + 0.010833016596734524, + 0.018131254240870476, + 0.010644447058439255, + 0.014862771145999432, + 0.021865341812372208, + 0.03814014047384262, + 0.025704974308609962, + 0.04289798438549042, + 0.03878166899085045, + 0.02646736614406109, + 0.05143185332417488, + 0.03924862667918205 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 7, + 6 + ], + "selected_reward": -0.026335397735238075 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main character a portly samurai named Azumi Jigorou wearing a traditional dark blue robe and vest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character comically tiptoeing across a vast, shimmering frozen pond with a massive tsunami wave of emerald green ice looming behind him?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the character's expression change from stoic concentration to slapstick terror as he turns his head?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera zoom in quickly and harshly to an extreme close-up of the character's wide, panicked eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a perfectly rendered bead of sweat shown flying from the character's forehead in exaggerated 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": [ + -0.10028185695409775 + ], + "opposite_logprob_tokens": [ + -2.3502819538116455 + ], + "logprob_reward": -0.10028185695409775, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10028185695409775, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10463497787714005, + 0.029102999716997147, + 0.09252426773309708, + 0.06537822633981705, + 0.05901413783431053, + 0.05147172138094902, + 0.07130887359380722, + 0.2680273652076721, + 0.10363732278347015, + 0.15490013360977173 + ], + "credit_map": [ + 0.012735925614833832, + 0.007013911847025156, + 0.016311608254909515, + 0.00723373144865036, + 0.010294786654412746, + 0.015111049637198448, + 0.015724340453743935, + 0.020209623500704765, + 0.004279823042452335, + 0.0030765384435653687, + 0.0029738713055849075, + 0.002049277536571026, + 0.0035625125747174025, + 0.003794843563809991, + 0.004887785762548447, + 0.004478347487747669, + 0.010911716148257256, + 0.005110237281769514, + 0.01102577243000269, + 0.006942232605069876, + 0.011471432633697987, + 0.011945214122533798, + 0.018620561808347702, + 0.01649709790945053, + 0.008279694244265556, + 0.004231211729347706, + 0.017249206081032753, + 0.006482359021902084, + 0.005528851877897978, + 0.0075675412081182, + 0.006828009616583586, + 0.009211353957653046, + 0.006320602260529995, + 0.007467626128345728, + 0.008395718410611153, + 0.006854468956589699, + 0.006355235353112221, + 0.005383532494306564, + 0.007873019203543663, + 0.010363934561610222, + 0.008867081254720688, + 0.008677151054143906, + 0.004738344810903072, + 0.006381968501955271, + 0.006132757291197777, + 0.0072212074883282185, + 0.005243881139904261, + 0.004209327511489391, + 0.008671742863953114, + 0.008328457362949848, + 0.0078020584769546986, + 0.010628260672092438, + 0.007818718440830708, + 0.0074546365067362785, + 0.009156277403235435, + 0.011448724195361137, + 0.04680351912975311, + 0.010450458154082298, + 0.016938170418143272, + 0.014541905373334885, + 0.01728077419102192, + 0.10195492208003998, + 0.011644269339740276, + 0.04841333255171776, + 0.012550745159387589, + 0.006707340478897095, + 0.014259249903261662, + 0.009376171045005322, + 0.012705039232969284, + 0.010487133637070656, + 0.011761842295527458, + 0.025789804756641388, + 0.022144049406051636, + 0.0166923888027668, + 0.016657834872603416, + 0.02637307345867157, + 0.01610933057963848, + 0.021554192528128624, + 0.01811162568628788, + 0.01725764386355877 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 9, + 0 + ], + "selected_reward": -0.10028185695409775 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "This medium-wide shot, initially static and humorously stylized, captures the portly samurai Azumi Jigorou attempting to gracefully cross a vast, shimmering frozen pond beneath a crisp winter sky. He wears a traditional dark blue robe and vest, his enormous belly straining against the fabric, and a katana neatly tucked into his belt. He comically tiptoes across the slick ice with delicate, tentative steps, his face a mask of stoic concentration. The massive tsunami wave looms impossibly behind him, its massive crest a crystalline sculpture of emerald green ice and still white foam, dwarfing the surrounding snow-dusted pine trees. As he turns his head slightly, his stoic expression instantly melts into pure slapstick terror; the camera zooms in quickly and harshly, tightly framing his wide, panicked eyes as they first register the absurd, frozen catastrophe. In this extreme close-up, a perfectly rendered bead of sweat flies from his forehead in exaggerated slow motion, arcing through the air and capturing the bright, clear light that punctuates his moment of cartoonish terror.", + "seed": 32261, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000121/group_7/index_6007_seed_32261_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": -0.05069414728786796, + "vlm_logprob": -0.05069414728786796, + "vlm_yes_no_margin": 4.15, + "vlm_token_credit": -0.05069414728786796, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main character a portly samurai named Azumi Jigorou wearing a traditional dark blue robe and vest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character comically tiptoeing across a vast, shimmering frozen pond with a massive tsunami wave of emerald green ice looming behind him?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the character's expression change from stoic concentration to slapstick terror as he turns his head?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera zoom in quickly and harshly to an extreme close-up of the character's wide, panicked eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a perfectly rendered bead of sweat shown flying from the character's forehead in exaggerated 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": [ + -0.1377791166305542 + ], + "opposite_logprob_tokens": [ + -2.1377792358398438 + ], + "logprob_reward": -0.1377791166305542, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.1377791166305542, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09035991132259369, + 0.02956426702439785, + 0.06799808889627457, + 0.2681368589401245, + 0.09909647703170776, + 0.026572616770863533, + 0.07084175944328308, + 0.23707319796085358, + 0.046506620943546295, + 0.06385015696287155 + ], + "credit_map": [ + 0.011433730833232403, + 0.01413390226662159, + 0.020551906898617744, + 0.01041621994227171, + 0.009842324070632458, + 0.010276325047016144, + 0.006749886088073254, + 0.0069556157104671, + 0.00562302628532052, + 0.006007364951074123, + 0.005076370667666197, + 0.0030980578158050776, + 0.0018602003110572696, + 0.004335127305239439, + 0.0017253548139706254, + 0.0018387663876637816, + 0.008511781692504883, + 0.009673584252595901, + 0.011950766667723656, + 0.01637127250432968, + 0.004206106066703796, + 0.006460106931626797, + 0.005235457327216864, + 0.00558901485055685, + 0.013996714726090431, + 0.009233690798282623, + 0.10069648176431656, + 0.10817975550889969, + 0.005773778073489666, + 0.008151107467710972, + 0.012870515696704388, + 0.00923481397330761, + 0.011366059072315693, + 0.00947414617985487, + 0.018798889592289925, + 0.01303805224597454, + 0.017444003373384476, + 0.008474433794617653, + 0.010845686309039593, + 0.009655208326876163, + 0.0031784570310264826, + 0.0032271486707031727, + 0.002864900743588805, + 0.0034494006540626287, + 0.00504795229062438, + 0.0031587034463882446, + 0.0024593831039965153, + 0.0031866682693362236, + 0.006465632934123278, + 0.006187559105455875, + 0.00728799682110548, + 0.0045486255548894405, + 0.011050187982618809, + 0.01595970056951046, + 0.008682197891175747, + 0.010659860447049141, + 0.012895474210381508, + 0.019274014979600906, + 0.020143315196037292, + 0.04140942171216011, + 0.018877148628234863, + 0.02571031264960766, + 0.0724787563085556, + 0.026284748688340187, + 0.005054114386439323, + 0.005360233597457409, + 0.006300867069512606, + 0.007152578793466091, + 0.006880602799355984, + 0.005226736888289452, + 0.0057385992258787155, + 0.004792890045791864, + 0.006447826512157917, + 0.006087239366024733, + 0.0047980076633393764, + 0.015162822790443897, + 0.007268896326422691, + 0.005936931353062391, + 0.013367271050810814, + 0.004781166557222605 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 7, + 4 + ], + "selected_reward": -0.1377791166305542 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main character a portly samurai named Azumi Jigorou wearing a traditional dark blue robe and vest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character comically tiptoeing across a vast, shimmering frozen pond with a massive tsunami wave of emerald green ice looming behind him?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the character's expression change from stoic concentration to slapstick terror as he turns his head?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera zoom in quickly and harshly to an extreme close-up of the character's wide, panicked eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a perfectly rendered bead of sweat shown flying from the character's forehead in exaggerated 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": [ + -0.005269329063594341 + ], + "opposite_logprob_tokens": [ + -5.255269527435303 + ], + "logprob_reward": -0.005269329063594341, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005269329063594341, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14102783799171448, + 0.09948442876338959, + 0.08226761966943741, + 0.13849975168704987, + 0.03851089999079704, + 0.047748640179634094, + 0.12513627111911774, + 0.2588088810443878, + 0.04399691894650459, + 0.024518776684999466 + ], + "credit_map": [ + 0.029532531276345253, + 0.013506544753909111, + 0.033907677978277206, + 0.017838845029473305, + 0.00918961688876152, + 0.014861834235489368, + 0.010043537244200706, + 0.012147241272032261, + 0.03387080505490303, + 0.009134641848504543, + 0.028425194323062897, + 0.00909686554223299, + 0.0033941930159926414, + 0.006094950716942549, + 0.004099155776202679, + 0.005368625279515982, + 0.009372475557029247, + 0.016213223338127136, + 0.016880212351679802, + 0.016247648745775223, + 0.006707062013447285, + 0.004366857465356588, + 0.007355165667831898, + 0.005124965682625771, + 0.008317852392792702, + 0.006947880145162344, + 0.04920641705393791, + 0.05757652595639229, + 0.003574503120034933, + 0.0031221515964716673, + 0.006054540164768696, + 0.0036998819559812546, + 0.006940087769180536, + 0.005386917851865292, + 0.007733031641691923, + 0.0058965785428881645, + 0.0036879563704133034, + 0.0033106852788478136, + 0.0028709303587675095, + 0.00268471403978765, + 0.008579992689192295, + 0.006806297693401575, + 0.009497690945863724, + 0.005467270966619253, + 0.004823237657546997, + 0.004625524394214153, + 0.0035224799066781998, + 0.004426144529134035, + 0.007458406034857035, + 0.007947985082864761, + 0.015105345286428928, + 0.008839010260999203, + 0.018467361107468605, + 0.032167691737413406, + 0.01650097966194153, + 0.018649498000741005, + 0.012776092626154423, + 0.018052205443382263, + 0.026993034407496452, + 0.049445316195487976, + 0.015142202377319336, + 0.01420943159610033, + 0.09812188893556595, + 0.024068690836429596, + 0.0051124002784490585, + 0.004934131633490324, + 0.005234532989561558, + 0.0033538537099957466, + 0.007949569262564182, + 0.0073190126568078995, + 0.0035867784172296524, + 0.006506639998406172, + 0.00408794404938817, + 0.002317323349416256, + 0.0024816645309329033, + 0.003109552199020982, + 0.002307671355083585, + 0.0036595722194761038, + 0.003032462205737829, + 0.003522587241604924 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 3 + ], + "selected_reward": -0.005269329063594341 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main character a portly samurai named Azumi Jigorou wearing a traditional dark blue robe and vest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character comically tiptoeing across a vast, shimmering frozen pond with a massive tsunami wave of emerald green ice looming behind him?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the character's expression change from stoic concentration to slapstick terror as he turns his head?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera zoom in quickly and harshly to an extreme close-up of the character's wide, panicked eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a perfectly rendered bead of sweat shown flying from the character's forehead in exaggerated 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": [ + -0.001515788841061294 + ], + "opposite_logprob_tokens": [ + -6.501515865325928 + ], + "logprob_reward": -0.001515788841061294, + "yes_no_margin_reward": 6.5, + "token_credit_reward": -0.001515788841061294, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.034086037427186966, + 0.05764567479491234, + 0.04703286662697792, + 0.35281333327293396, + 0.03506665304303169, + 0.020576685667037964, + 0.09414830803871155, + 0.19849233329296112, + 0.1004256010055542, + 0.05971251428127289 + ], + "credit_map": [ + 0.0036494897212833166, + 0.003282075747847557, + 0.003973633982241154, + 0.005143062677234411, + 0.004335497505962849, + 0.007667765486985445, + 0.0030252388678491116, + 0.0030092725064605474, + 0.019490521401166916, + 0.003965860232710838, + 0.020091509446501732, + 0.004941004328429699, + 0.0020917165093123913, + 0.002434835536405444, + 0.0018022325821220875, + 0.00282799219712615, + 0.00674054678529501, + 0.012195256538689137, + 0.004708967171609402, + 0.006581021938472986, + 0.006234373431652784, + 0.003626546123996377, + 0.004144663456827402, + 0.0028014928102493286, + 0.017854293808341026, + 0.01303680520504713, + 0.1369897574186325, + 0.1523894965648651, + 0.005417569074779749, + 0.005277674179524183, + 0.015905115753412247, + 0.005942641291767359, + 0.004082368686795235, + 0.00395117187872529, + 0.005761649925261736, + 0.003644393989816308, + 0.005945208016782999, + 0.003523102728649974, + 0.004163491073995829, + 0.003995264880359173, + 0.0027174437418580055, + 0.002927405061200261, + 0.002227296819910407, + 0.0025540306232869625, + 0.003684058552607894, + 0.0019521609647199512, + 0.0023862402886152267, + 0.002128050895407796, + 0.004935466218739748, + 0.005359261762350798, + 0.013991036452353, + 0.007254055235534906, + 0.015000843442976475, + 0.016695193946361542, + 0.016069380566477776, + 0.014843079261481762, + 0.011886602267622948, + 0.01493725087493658, + 0.01724342443048954, + 0.03490126505494118, + 0.014148849993944168, + 0.017995547503232956, + 0.06622115522623062, + 0.02115824446082115, + 0.014863526448607445, + 0.009134151041507721, + 0.009661002084612846, + 0.016034690663218498, + 0.011542207561433315, + 0.01541272085160017, + 0.013928533531725407, + 0.009848772548139095, + 0.0060482146218419075, + 0.003952478524297476, + 0.00339208054356277, + 0.016196759417653084, + 0.005988570395857096, + 0.005384285002946854, + 0.013845102861523628, + 0.0049050250090658665 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 7, + 8 + ], + "selected_reward": -0.001515788841061294 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main character a portly samurai named Azumi Jigorou wearing a traditional dark blue robe and vest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character comically tiptoeing across a vast, shimmering frozen pond with a massive tsunami wave of emerald green ice looming behind him?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the character's expression change from stoic concentration to slapstick terror as he turns his head?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera zoom in quickly and harshly to an extreme close-up of the character's wide, panicked eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a perfectly rendered bead of sweat shown flying from the character's forehead in exaggerated 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": [ + -0.00863284058868885 + ], + "opposite_logprob_tokens": [ + -4.758632659912109 + ], + "logprob_reward": -0.00863284058868885, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.00863284058868885, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.02112417109310627, + 0.01945512928068638, + 0.04994652047753334, + 0.5265815258026123, + 0.023179952055215836, + 0.04588949307799339, + 0.08795055747032166, + 0.08170755952596664, + 0.08952932059764862, + 0.054635751992464066 + ], + "credit_map": [ + 0.002733757020905614, + 0.0034497007727622986, + 0.0025598916690796614, + 0.0029231267981231213, + 0.0016804379411041737, + 0.0021593186538666487, + 0.003225007327273488, + 0.0023929309099912643, + 0.004357133060693741, + 0.0032625123858451843, + 0.004557301290333271, + 0.0017694245325401425, + 0.0012596829328686, + 0.0020357936155050993, + 0.001106667099520564, + 0.0011066137813031673, + 0.007518143858760595, + 0.011878988705575466, + 0.007757209241390228, + 0.007697057444602251, + 0.004845927003771067, + 0.00329581112600863, + 0.004332251381129026, + 0.0026211340446025133, + 0.027420205995440483, + 0.018939033150672913, + 0.2070581167936325, + 0.2276240885257721, + 0.006722734775394201, + 0.008003977127373219, + 0.023456469178199768, + 0.00735690351575613, + 0.0038813662249594927, + 0.0034949504770338535, + 0.0040621329098939896, + 0.0033644218929111958, + 0.0028473574202507734, + 0.0019605373963713646, + 0.0017923875711858273, + 0.0017767988611012697, + 0.0059208679012954235, + 0.006624035537242889, + 0.00651968689635396, + 0.0043165613897144794, + 0.00755064794793725, + 0.005549453664571047, + 0.004415122792124748, + 0.00499311788007617, + 0.006374918390065432, + 0.007578763645142317, + 0.010307577438652515, + 0.006121825892478228, + 0.016859980300068855, + 0.012624207884073257, + 0.016039235517382622, + 0.012044055387377739, + 0.006678484380245209, + 0.007525444496423006, + 0.01410779170691967, + 0.010000599548220634, + 0.0065587661229074, + 0.015590164810419083, + 0.011704685166478157, + 0.009541617706418037, + 0.010848955251276493, + 0.010178487747907639, + 0.009557841345667839, + 0.007095730397850275, + 0.016818711534142494, + 0.0151216434314847, + 0.0077269175089895725, + 0.012181036174297333, + 0.008568615652620792, + 0.006336571183055639, + 0.0054432968609035015, + 0.007782179396599531, + 0.006644464563578367, + 0.007181046064943075, + 0.007058910559862852, + 0.005620666779577732 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 8, + 6 + ], + "selected_reward": -0.00863284058868885 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main character a portly samurai named Azumi Jigorou wearing a traditional dark blue robe and vest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character comically tiptoeing across a vast, shimmering frozen pond with a massive tsunami wave of emerald green ice looming behind him?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the character's expression change from stoic concentration to slapstick terror as he turns his head?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera zoom in quickly and harshly to an extreme close-up of the character's wide, panicked eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a perfectly rendered bead of sweat shown flying from the character's forehead in exaggerated 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": [ + -0.10027366131544113 + ], + "opposite_logprob_tokens": [ + -2.350273609161377 + ], + "logprob_reward": -0.10027366131544113, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10027366131544113, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05313156172633171, + 0.06767811626195908, + 0.13753214478492737, + 0.12409897893667221, + 0.07936180382966995, + 0.06629630923271179, + 0.14058174192905426, + 0.12993474304676056, + 0.07870808243751526, + 0.12267650663852692 + ], + "credit_map": [ + 0.0064704460091888905, + 0.007112196180969477, + 0.00864956621080637, + 0.0069113438948988914, + 0.004583293106406927, + 0.006734361872076988, + 0.006925798021256924, + 0.0057445610873401165, + 0.017107399180531502, + 0.013007860630750656, + 0.01447226945310831, + 0.005817935802042484, + 0.00482727587223053, + 0.003912081941962242, + 0.004694882780313492, + 0.003838419681414962, + 0.026008598506450653, + 0.03250829130411148, + 0.019989073276519775, + 0.017019977793097496, + 0.0107932910323143, + 0.010475289076566696, + 0.010403935797512531, + 0.01033368706703186, + 0.007169805932790041, + 0.006370197515934706, + 0.04479464888572693, + 0.04929136112332344, + 0.0037593096494674683, + 0.0027739100623875856, + 0.0058098747394979, + 0.0041298759169876575, + 0.014749444089829922, + 0.011221419088542461, + 0.01609741523861885, + 0.012155482545495033, + 0.00730146886780858, + 0.007211620919406414, + 0.005592316389083862, + 0.005032640881836414, + 0.01157443132251501, + 0.00991598330438137, + 0.008007023483514786, + 0.0068355645053088665, + 0.011216799728572369, + 0.006109851412475109, + 0.006192914675921202, + 0.0064437510445714, + 0.012982259504497051, + 0.011539027094841003, + 0.01685260608792305, + 0.010480774566531181, + 0.024127230048179626, + 0.02191641367971897, + 0.022312801331281662, + 0.02037065662443638, + 0.008708776906132698, + 0.011806774884462357, + 0.009500568732619286, + 0.022296829149127007, + 0.011651553213596344, + 0.012606830336153507, + 0.03919222578406334, + 0.01417118962854147, + 0.01869354583323002, + 0.0056753396056592464, + 0.007353209424763918, + 0.005989638157188892, + 0.008186853490769863, + 0.015450401231646538, + 0.00698336586356163, + 0.010375741869211197, + 0.015393943525850773, + 0.008747887797653675, + 0.010393931530416012, + 0.024595431983470917, + 0.012457845732569695, + 0.013689558953046799, + 0.02396102249622345, + 0.013436894863843918 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 2, + 7 + ], + "selected_reward": -0.10027366131544113 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "This medium-wide shot, initially static and humorously stylized, captures the portly samurai Azumi Jigorou attempting to gracefully cross a vast, shimmering frozen pond beneath a crisp winter sky. He wears a traditional dark blue robe and vest, his enormous belly straining against the fabric, and a katana neatly tucked into his belt. He comically tiptoes across the slick ice with delicate, tentative steps, his face a mask of stoic concentration. The massive tsunami wave looms impossibly behind him, its massive crest a crystalline sculpture of emerald green ice and still white foam, dwarfing the surrounding snow-dusted pine trees. As he turns his head slightly, his stoic expression instantly melts into pure slapstick terror; the camera zooms in quickly and harshly, tightly framing his wide, panicked eyes as they first register the absurd, frozen catastrophe. In this extreme close-up, a perfectly rendered bead of sweat flies from his forehead in exaggerated slow motion, arcing through the air and capturing the bright, clear light that punctuates his moment of cartoonish terror.", + "seed": 32262, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000121/group_7/index_6007_seed_32262_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": -0.0457801359705627, + "vlm_logprob": -0.0457801359705627, + "vlm_yes_no_margin": 3.9, + "vlm_token_credit": -0.0457801359705627, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main character a portly samurai named Azumi Jigorou wearing a traditional dark blue robe and vest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character comically tiptoeing across a vast, shimmering frozen pond with a massive tsunami wave of emerald green ice looming behind him?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the character's expression change from stoic concentration to slapstick terror as he turns his head?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera zoom in quickly and harshly to an extreme close-up of the character's wide, panicked eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a perfectly rendered bead of sweat shown flying from the character's forehead in exaggerated 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": [ + -0.091835618019104 + ], + "opposite_logprob_tokens": [ + -2.4668354988098145 + ], + "logprob_reward": -0.091835618019104, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.091835618019104, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13529248535633087, + 0.07258018106222153, + 0.05988384410738945, + 0.07002641260623932, + 0.09827879071235657, + 0.06837809830904007, + 0.1282501369714737, + 0.21344004571437836, + 0.08069335669279099, + 0.07317660003900528 + ], + "credit_map": [ + 0.020896362140774727, + 0.0157975684851408, + 0.017132673412561417, + 0.0194474495947361, + 0.01282472163438797, + 0.019000839442014694, + 0.014411035925149918, + 0.015781840309500694, + 0.011689296923577785, + 0.008079598657786846, + 0.007342259865254164, + 0.006950233597308397, + 0.0070876795798540115, + 0.008601929992437363, + 0.014839861541986465, + 0.00798933207988739, + 0.00865980889648199, + 0.005350104067474604, + 0.006298774387687445, + 0.00623674038797617, + 0.008001502603292465, + 0.009842666797339916, + 0.007508464157581329, + 0.007985788397490978, + 0.01168378908187151, + 0.008370448835194111, + 0.008113916032016277, + 0.009370923042297363, + 0.006767793092876673, + 0.009203078225255013, + 0.010269694030284882, + 0.006246786564588547, + 0.02106550522148609, + 0.009119762107729912, + 0.011752422899007797, + 0.008307977579534054, + 0.013480939902365208, + 0.008131244219839573, + 0.01299335341900587, + 0.013427598401904106, + 0.012812329456210136, + 0.006764418911188841, + 0.006656197365373373, + 0.007781879510730505, + 0.00964885763823986, + 0.005423156078904867, + 0.011993179097771645, + 0.007298090495169163, + 0.02048446610569954, + 0.01435160357505083, + 0.01943078078329563, + 0.014652684330940247, + 0.018525399267673492, + 0.01026151329278946, + 0.01770291104912758, + 0.012840780429542065, + 0.015351534821093082, + 0.020700758323073387, + 0.013202453963458538, + 0.0808703601360321, + 0.009919052943587303, + 0.010099519975483418, + 0.026178766041994095, + 0.0371176078915596, + 0.011973030865192413, + 0.006443849299103022, + 0.012849526479840279, + 0.0060819475911557674, + 0.008080041967332363, + 0.012536775320768356, + 0.006968989036977291, + 0.01575920358300209, + 0.006983794271945953, + 0.008737201802432537, + 0.006483334116637707, + 0.008586319163441658, + 0.013034775853157043, + 0.008609765209257603, + 0.010950321331620216, + 0.009791095741093159 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 6 + ], + "selected_reward": -0.091835618019104 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main character a portly samurai named Azumi Jigorou wearing a traditional dark blue robe and vest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character comically tiptoeing across a vast, shimmering frozen pond with a massive tsunami wave of emerald green ice looming behind him?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the character's expression change from stoic concentration to slapstick terror as he turns his head?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera zoom in quickly and harshly to an extreme close-up of the character's wide, panicked eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a perfectly rendered bead of sweat shown flying from the character's forehead in exaggerated 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": [ + -0.014220568351447582 + ], + "opposite_logprob_tokens": [ + -4.264220714569092 + ], + "logprob_reward": -0.014220568351447582, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014220568351447582, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20482409000396729, + 0.08452984690666199, + 0.06299956887960434, + 0.08585584908723831, + 0.11407271772623062, + 0.0922994613647461, + 0.04883129894733429, + 0.15967808663845062, + 0.09466733783483505, + 0.0522417314350605 + ], + "credit_map": [ + 0.0423566997051239, + 0.013531542383134365, + 0.027643539011478424, + 0.018191467970609665, + 0.01563512161374092, + 0.04508230462670326, + 0.017526954412460327, + 0.024856463074684143, + 0.013033200055360794, + 0.011531571857631207, + 0.007422488182783127, + 0.01068789977580309, + 0.014117005281150341, + 0.0077700368128716946, + 0.0115165114402771, + 0.008451140485703945, + 0.009261732921004295, + 0.004832920618355274, + 0.006737005431205034, + 0.007399615366011858, + 0.007740480359643698, + 0.012059226632118225, + 0.006625381298363209, + 0.008343209512531757, + 0.013242630288004875, + 0.010027516633272171, + 0.009253829717636108, + 0.011407346464693546, + 0.008265233598649502, + 0.010708112269639969, + 0.01341817807406187, + 0.009533003903925419, + 0.019627107307314873, + 0.017825135961174965, + 0.012290987186133862, + 0.013778685592114925, + 0.01679706946015358, + 0.00964799989014864, + 0.01366044394671917, + 0.010445299558341503, + 0.013618063181638718, + 0.011472160927951336, + 0.014936450868844986, + 0.01960998959839344, + 0.00921402033418417, + 0.004818108398467302, + 0.01060551032423973, + 0.008025158196687698, + 0.007479333318769932, + 0.0061273640021681786, + 0.006528148893266916, + 0.007482272107154131, + 0.005622653756290674, + 0.0033774576149880886, + 0.007171630393713713, + 0.005042439326643944, + 0.012196642346680164, + 0.007674805354326963, + 0.011381897144019604, + 0.0561220645904541, + 0.0057282643392682076, + 0.008672453463077545, + 0.012704201973974705, + 0.045197777450084686, + 0.012463890016078949, + 0.009793177247047424, + 0.019533295184373856, + 0.007697323802858591, + 0.009951906278729439, + 0.00719077792018652, + 0.008875702507793903, + 0.019161274656653404, + 0.006188106723129749, + 0.004381598439067602, + 0.007321301382035017, + 0.005604717414826155, + 0.007482850458472967, + 0.005788835696876049, + 0.006176045164465904, + 0.009298284538090229 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 4 + ], + "selected_reward": -0.014220568351447582 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main character a portly samurai named Azumi Jigorou wearing a traditional dark blue robe and vest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character comically tiptoeing across a vast, shimmering frozen pond with a massive tsunami wave of emerald green ice looming behind him?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the character's expression change from stoic concentration to slapstick terror as he turns his head?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera zoom in quickly and harshly to an extreme close-up of the character's wide, panicked eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a perfectly rendered bead of sweat shown flying from the character's forehead in exaggerated 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": [ + -0.0031956578604876995 + ], + "opposite_logprob_tokens": [ + -5.753195762634277 + ], + "logprob_reward": -0.0031956578604876995, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.0031956578604876995, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08218700438737869, + 0.07045898586511612, + 0.06259467452764511, + 0.0174237173050642, + 0.03277963399887085, + 0.014503862708806992, + 0.03712991997599602, + 0.3554416000843048, + 0.2099965214729309, + 0.11748409271240234 + ], + "credit_map": [ + 0.015301092527806759, + 0.011754955165088177, + 0.007391588296741247, + 0.008166812360286713, + 0.008774257265031338, + 0.014436504803597927, + 0.007756202481687069, + 0.00860558357089758, + 0.011868548579514027, + 0.010800842195749283, + 0.0071008289232850075, + 0.005897682625800371, + 0.008080887608230114, + 0.01035287044942379, + 0.010689178481698036, + 0.005668139085173607, + 0.010648536495864391, + 0.009194708429276943, + 0.007781929802149534, + 0.0050736465491354465, + 0.008743670769035816, + 0.008825021795928478, + 0.005957028362900019, + 0.006370130926370621, + 0.0029485286213457584, + 0.0022607569117099047, + 0.001652029692195356, + 0.0015026063192635775, + 0.003408986609429121, + 0.0017566254828125238, + 0.0021538357250392437, + 0.0017403482925146818, + 0.007759610190987587, + 0.00334316142834723, + 0.003405855968594551, + 0.00398819986730814, + 0.0038209278136491776, + 0.002574353013187647, + 0.003882765769958496, + 0.00400475924834609, + 0.0018435061210766435, + 0.0014109647599980235, + 0.001571092288941145, + 0.001891423831693828, + 0.002585135167464614, + 0.001309187849983573, + 0.0017930768663063645, + 0.0020994755905121565, + 0.009698085486888885, + 0.0028190326411277056, + 0.0045377276837825775, + 0.004382371436804533, + 0.0030220618937164545, + 0.005828080698847771, + 0.0035957100335508585, + 0.0032468491699546576, + 0.022302819415926933, + 0.014330361969769001, + 0.021069129928946495, + 0.15596899390220642, + 0.011919070966541767, + 0.01019539125263691, + 0.04398709535598755, + 0.07566870748996735, + 0.023075643926858902, + 0.010891863144934177, + 0.05311640352010727, + 0.01238850224763155, + 0.01711871847510338, + 0.01829160936176777, + 0.019448116421699524, + 0.05566566064953804, + 0.011835585348308086, + 0.011104646138846874, + 0.013293152675032616, + 0.013906640000641346, + 0.019567102193832397, + 0.014119796454906464, + 0.01758408173918724, + 0.016073079779744148 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 8, + 9 + ], + "selected_reward": -0.0031956578604876995 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main character a portly samurai named Azumi Jigorou wearing a traditional dark blue robe and vest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character comically tiptoeing across a vast, shimmering frozen pond with a massive tsunami wave of emerald green ice looming behind him?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the character's expression change from stoic concentration to slapstick terror as he turns his head?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera zoom in quickly and harshly to an extreme close-up of the character's wide, panicked eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a perfectly rendered bead of sweat shown flying from the character's forehead in exaggerated 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": [ + -0.006732991896569729 + ], + "opposite_logprob_tokens": [ + -5.006732940673828 + ], + "logprob_reward": -0.006732991896569729, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006732991896569729, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09719692915678024, + 0.05064840242266655, + 0.032241351902484894, + 0.0716869980096817, + 0.045172907412052155, + 0.02154899574816227, + 0.05276240035891533, + 0.2369484305381775, + 0.23652145266532898, + 0.15527214109897614 + ], + "credit_map": [ + 0.017338337376713753, + 0.009996018372476101, + 0.0094712283462286, + 0.013032698072493076, + 0.008229145780205727, + 0.016863800585269928, + 0.011121295392513275, + 0.011144410818815231, + 0.008432717062532902, + 0.005808474496006966, + 0.0045566251501441, + 0.005056720692664385, + 0.006342936772853136, + 0.0063987006433308125, + 0.008998644538223743, + 0.005053583532571793, + 0.005263142287731171, + 0.004325227811932564, + 0.0047780489549040794, + 0.003691920544952154, + 0.003952844999730587, + 0.002224559197202325, + 0.005003849510103464, + 0.0030017609242349863, + 0.01297107245773077, + 0.008708484470844269, + 0.009382265619933605, + 0.009076923131942749, + 0.006647005677223206, + 0.009596003219485283, + 0.008192035369575024, + 0.007113209925591946, + 0.008861572481691837, + 0.004892221186310053, + 0.007179768290370703, + 0.0044196066446602345, + 0.004586486611515284, + 0.003395366482436657, + 0.006137266289442778, + 0.005700620356947184, + 0.0036093515809625387, + 0.0024375906214118004, + 0.0029568427707999945, + 0.0021272730082273483, + 0.003851582994684577, + 0.0017674288246780634, + 0.002943129977211356, + 0.0018557955045253038, + 0.007208848837763071, + 0.006835370324552059, + 0.006546610966324806, + 0.007408709265291691, + 0.004999080207198858, + 0.006023309659212828, + 0.008628851734101772, + 0.005111621227115393, + 0.027538582682609558, + 0.014706762507557869, + 0.01653756946325302, + 0.08993867039680481, + 0.012145040556788445, + 0.008979388512670994, + 0.03639911115169525, + 0.03070330061018467, + 0.024683887138962746, + 0.021602187305688858, + 0.05096884071826935, + 0.015496090985834599, + 0.02769334800541401, + 0.02489691600203514, + 0.022034915164113045, + 0.04914525896310806, + 0.016519375145435333, + 0.016250455752015114, + 0.025532325729727745, + 0.021859372034668922, + 0.014416278339922428, + 0.018577074632048607, + 0.018114138394594193, + 0.02400312013924122 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 8, + 9 + ], + "selected_reward": -0.006732991896569729 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main character a portly samurai named Azumi Jigorou wearing a traditional dark blue robe and vest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character comically tiptoeing across a vast, shimmering frozen pond with a massive tsunami wave of emerald green ice looming behind him?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the character's expression change from stoic concentration to slapstick terror as he turns his head?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera zoom in quickly and harshly to an extreme close-up of the character's wide, panicked eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a perfectly rendered bead of sweat shown flying from the character's forehead in exaggerated 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": [ + -0.11291584372520447 + ], + "opposite_logprob_tokens": [ + -2.2379157543182373 + ], + "logprob_reward": -0.11291584372520447, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11291584372520447, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08994951099157333, + 0.09155961871147156, + 0.04884377121925354, + 0.03737877309322357, + 0.023364447057247162, + 0.06916197389364243, + 0.06979665160179138, + 0.33819958567619324, + 0.1396753191947937, + 0.09207037091255188 + ], + "credit_map": [ + 0.014561976306140423, + 0.011692746542394161, + 0.008408564142882824, + 0.013329212553799152, + 0.009106414392590523, + 0.00974029116332531, + 0.013493926264345646, + 0.009616382420063019, + 0.015379908494651318, + 0.011736384592950344, + 0.007736437954008579, + 0.007441400550305843, + 0.012090153992176056, + 0.012743178755044937, + 0.014475027099251747, + 0.009957123547792435, + 0.008093737997114658, + 0.004021633416414261, + 0.006210668012499809, + 0.005250596906989813, + 0.004071696195751429, + 0.009385619312524796, + 0.004615378566086292, + 0.00719444127753377, + 0.006892697885632515, + 0.0059178969822824, + 0.0035656311083585024, + 0.005216005723923445, + 0.004105555359274149, + 0.004685672000050545, + 0.0040835607796907425, + 0.002911752089858055, + 0.0037535340525209904, + 0.0020184635650366545, + 0.003174317069351673, + 0.002887494396418333, + 0.0024328031577169895, + 0.0025397276040166616, + 0.003729988355189562, + 0.002828118158504367, + 0.013068919070065022, + 0.006516595371067524, + 0.007221473380923271, + 0.009441325441002846, + 0.008438284508883953, + 0.003629185725003481, + 0.014736929908394814, + 0.00610926141962409, + 0.013040739111602306, + 0.006940181367099285, + 0.012828213162720203, + 0.00815695058554411, + 0.005440854467451572, + 0.009146453812718391, + 0.008172700181603432, + 0.006070553790777922, + 0.018601879477500916, + 0.014440659433603287, + 0.021604109555482864, + 0.1443411260843277, + 0.010917472653090954, + 0.011539801955223083, + 0.037527553737163544, + 0.07922697812318802, + 0.01747553050518036, + 0.011731602251529694, + 0.028638917952775955, + 0.012758663855493069, + 0.012971080839633942, + 0.011612439528107643, + 0.01966811902821064, + 0.024818958714604378, + 0.009865039959549904, + 0.009506271220743656, + 0.01053512655198574, + 0.010682602412998676, + 0.015048307366669178, + 0.010115321725606918, + 0.013737639412283897, + 0.012580069713294506 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 8, + 9 + ], + "selected_reward": -0.11291584372520447 + } + ], + "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/0000121/group_8/index_tensor([1633, 1633])_group_8_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000121/group_8/index_tensor([1633, 1633])_group_8_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..b3f853a7dc2ef4533455f65ed53aedf401e8b555 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000121/group_8/index_tensor([1633, 1633])_group_8_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "stunning macro still shot, captured from a slightly overhead perspective, focuses the viewer's attention on the claw of a young purple dragon, carefully pressing down on the crown of a jeweled egg. The claw, though small, exudes power, covered in densely packed, fine scales that shimmer with the iridescent sheen of amethyst. Under steady pressure, a hairline crack appears on the surface of the eggshell, resembling raw, milky quartz. Then, with a slow yet steady rhythm, the crack begins to branch and spread outward, like a spiderweb. Truly breathtaking, a warm, luminous golden light pulsates from within the eggshell, filtering through each new crack and casting sharp beams of light against the nearly black, out-of-focus background. This holy light, originating from within, illuminates the dragon's claws from below in an almost magical way, outlining the edges of each purple scale with a dazzling golden edge. Its translucent texture and complex grain are displayed with unprecedented clarity, creating an atmosphere of anticipation, tranquility, and powerful vitality at the moment of magic's birth. The picture is extremely rich in details and full of epic fantasy.", + "seed": 27990, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000121/group_8/index_1633_seed_27990_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": -0.030431907810270786, + "vlm_logprob": -0.030431907810270786, + "vlm_yes_no_margin": 3.65, + "vlm_token_credit": -0.030431907810270786, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return 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 purple dragon's claw pressing down on a jeweled egg?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the dragon's claw scales depicted as densely packed, fine, and shimmering with an iridescent amethyst sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a hairline crack appear on the eggshell and branch outward like a spiderweb?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a warm, luminous golden light pulsating from within the egg and casting sharp beams against a nearly black, out-of-focus background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera positioned at a slightly overhead perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018912116065621376 + ], + "opposite_logprob_tokens": [ + -4.143912315368652 + ], + "logprob_reward": -0.018912116065621376, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.018912116065621376, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24919110536575317, + 0.1418471783399582, + 0.06967109441757202, + 0.06327039748430252, + 0.05597848445177078, + 0.07026765495538712, + 0.0401344820857048, + 0.07353099435567856, + 0.19486749172210693, + 0.041241127997636795 + ], + "credit_map": [ + 0.012284066528081894, + 0.08815136551856995, + 0.011763227172195911, + 0.0839531421661377, + 0.015259208157658577, + 0.009966445155441761, + 0.016117533668875694, + 0.01169611420482397, + 0.010372452437877655, + 0.013954664580523968, + 0.03644465655088425, + 0.018486227840185165, + 0.013658160343766212, + 0.013274074532091618, + 0.012156345881521702, + 0.023500604555010796, + 0.008470865897834301, + 0.008237907662987709, + 0.01129288598895073, + 0.00900140218436718, + 0.00807779747992754, + 0.005846727639436722, + 0.012005183845758438, + 0.006738319527357817, + 0.0094999885186553, + 0.009630277752876282, + 0.007900308817625046, + 0.007534521631896496, + 0.0070858001708984375, + 0.007356688380241394, + 0.007215477991849184, + 0.007047331891953945, + 0.0076808026060462, + 0.006262975744903088, + 0.00948732066899538, + 0.011129789985716343, + 0.006155023816972971, + 0.0050714705139398575, + 0.005944970063865185, + 0.004246133379638195, + 0.010820873081684113, + 0.009842048399150372, + 0.005838995799422264, + 0.010307195596396923, + 0.011620420031249523, + 0.0062127853743731976, + 0.010140920989215374, + 0.005484410561621189, + 0.007183391600847244, + 0.005306928418576717, + 0.0034936468582600355, + 0.004852082580327988, + 0.004562846384942532, + 0.0057364231906831264, + 0.004926267080008984, + 0.004072893410921097, + 0.010963467881083488, + 0.006770876236259937, + 0.006689072586596012, + 0.015819111838936806, + 0.006134634837508202, + 0.004705419763922691, + 0.013908113352954388, + 0.008540298789739609, + 0.027798889204859734, + 0.01153203658759594, + 0.01233805250376463, + 0.07616511732339859, + 0.006454478017985821, + 0.010644515044987202, + 0.041418492794036865, + 0.008515892550349236, + 0.003326058853417635, + 0.004183935932815075, + 0.006280958652496338, + 0.010543384589254856, + 0.003719739615917206, + 0.001920518116094172, + 0.007343600504100323, + 0.0039229323156178 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 1 + ], + "selected_reward": -0.018912116065621376 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return 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 purple dragon's claw pressing down on a jeweled egg?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the dragon's claw scales depicted as densely packed, fine, and shimmering with an iridescent amethyst sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a hairline crack appear on the eggshell and branch outward like a spiderweb?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a warm, luminous golden light pulsating from within the egg and casting sharp beams against a nearly black, out-of-focus background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera positioned at a slightly overhead perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03372587636113167 + ], + "opposite_logprob_tokens": [ + -3.4087259769439697 + ], + "logprob_reward": -0.03372587636113167, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03372587636113167, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.25040945410728455, + 0.1288227140903473, + 0.11642441898584366, + 0.10385360568761826, + 0.04449973255395889, + 0.059494324028491974, + 0.047022778540849686, + 0.08416631817817688, + 0.1331443190574646, + 0.032162345945835114 + ], + "credit_map": [ + 0.01783466711640358, + 0.08472084254026413, + 0.016840703785419464, + 0.06114012375473976, + 0.020053057000041008, + 0.011503523215651512, + 0.019744159653782845, + 0.018572349101305008, + 0.011345723643898964, + 0.015993859618902206, + 0.03154188022017479, + 0.014265559613704681, + 0.017094284296035767, + 0.012424184940755367, + 0.012373301200568676, + 0.01378391869366169, + 0.014641879126429558, + 0.020719271153211594, + 0.015206428244709969, + 0.012514302507042885, + 0.013229225762188435, + 0.014574757777154446, + 0.01371045969426632, + 0.01182809192687273, + 0.01158205047249794, + 0.017787592485547066, + 0.020645057782530785, + 0.01023005973547697, + 0.011860090307891369, + 0.007663398049771786, + 0.013173663057386875, + 0.010911686345934868, + 0.006340998224914074, + 0.005153680220246315, + 0.0063629006035625935, + 0.006056478712707758, + 0.00590177346020937, + 0.004241893067955971, + 0.005549459718167782, + 0.004892548080533743, + 0.009759076870977879, + 0.010261453688144684, + 0.0048070792108774185, + 0.006905560381710529, + 0.012409734539687634, + 0.005300516728311777, + 0.006842124275863171, + 0.0032087743747979403, + 0.0062658642418682575, + 0.00518202455714345, + 0.0036062858998775482, + 0.006595016922801733, + 0.0046294392086565495, + 0.003189753508195281, + 0.011891265399754047, + 0.005663127638399601, + 0.01067668478935957, + 0.008111607283353806, + 0.009079771116375923, + 0.01563083566725254, + 0.007047633174806833, + 0.007185556925833225, + 0.014429610222578049, + 0.0120046092197299, + 0.018813040107488632, + 0.007514896336942911, + 0.008702058345079422, + 0.052780598402023315, + 0.004102461040019989, + 0.007378268521279097, + 0.028432929888367653, + 0.005420058034360409, + 0.002206274541094899, + 0.0034791119396686554, + 0.002324008150026202, + 0.005048372782766819, + 0.004287495743483305, + 0.0018436021637171507, + 0.008415896445512772, + 0.00455758161842823 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 1 + ], + "selected_reward": -0.03372587636113167 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return 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 purple dragon's claw pressing down on a jeweled egg?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the dragon's claw scales depicted as densely packed, fine, and shimmering with an iridescent amethyst sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a hairline crack appear on the eggshell and branch outward like a spiderweb?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a warm, luminous golden light pulsating from within the egg and casting sharp beams against a nearly black, out-of-focus background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera positioned at a slightly overhead perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.062042009085416794 + ], + "opposite_logprob_tokens": [ + -2.812041997909546 + ], + "logprob_reward": -0.062042009085416794, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.062042009085416794, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2942780554294586, + 0.042196422815322876, + 0.021801361814141273, + 0.027160769328475, + 0.055185213685035706, + 0.03579125553369522, + 0.03022422082722187, + 0.03408690169453621, + 0.25885462760925293, + 0.20042113959789276 + ], + "credit_map": [ + 0.0098763108253479, + 0.10391651839017868, + 0.007746649440377951, + 0.12068916857242584, + 0.015051157213747501, + 0.010789039544761181, + 0.016646837815642357, + 0.009562370367348194, + 0.00421682745218277, + 0.005340827628970146, + 0.010391135700047016, + 0.004759920295327902, + 0.004263701383024454, + 0.0028958541806787252, + 0.004973898641765118, + 0.005354259163141251, + 0.0021598280873149633, + 0.003823030972853303, + 0.0034234742633998394, + 0.0027391775511205196, + 0.0022904069628566504, + 0.0027793189510703087, + 0.002141533885151148, + 0.0024445909075438976, + 0.004041143227368593, + 0.003613371169194579, + 0.0033271091524511576, + 0.00481913797557354, + 0.0031946836970746517, + 0.002036584075540304, + 0.0034926310181617737, + 0.002636109245941043, + 0.008117902092635632, + 0.006913752760738134, + 0.005541068036109209, + 0.009587869048118591, + 0.00619765417650342, + 0.004995920229703188, + 0.007310859393328428, + 0.006520187947899103, + 0.006071614101529121, + 0.003946002572774887, + 0.003285468090325594, + 0.006213600281625986, + 0.003599771298468113, + 0.003549499437212944, + 0.005526548717170954, + 0.003598753595724702, + 0.004342161118984222, + 0.0031432679388672113, + 0.0026879231445491314, + 0.005123366601765156, + 0.002803865587338805, + 0.005140690132975578, + 0.004616322927176952, + 0.0023666219785809517, + 0.004143537953495979, + 0.003468069713562727, + 0.003649515565484762, + 0.006254028994590044, + 0.0037795365788042545, + 0.0027211816050112247, + 0.0057062930427491665, + 0.004364741034805775, + 0.036180395632982254, + 0.014722725376486778, + 0.01705286279320717, + 0.10299063473939896, + 0.00753640104085207, + 0.013772457838058472, + 0.05494173616170883, + 0.011657431721687317, + 0.01580638624727726, + 0.029377669095993042, + 0.02479943074285984, + 0.030008260160684586, + 0.020890722051262856, + 0.011319897137582302, + 0.033439043909311295, + 0.03477972373366356 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 9 + ], + "selected_reward": -0.062042009085416794 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return 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 purple dragon's claw pressing down on a jeweled egg?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the dragon's claw scales depicted as densely packed, fine, and shimmering with an iridescent amethyst sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a hairline crack appear on the eggshell and branch outward like a spiderweb?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a warm, luminous golden light pulsating from within the egg and casting sharp beams against a nearly black, out-of-focus background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera positioned at a slightly overhead perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014196122996509075 + ], + "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": [ + 0.2058386504650116, + 0.08403898030519485, + 0.02647610381245613, + 0.14236366748809814, + 0.09617749601602554, + 0.054894354194402695, + 0.08317775279283524, + 0.10622615367174149, + 0.16648390889167786, + 0.034322936087846756 + ], + "credit_map": [ + 0.009227036498486996, + 0.066224604845047, + 0.010205181315541267, + 0.07888969779014587, + 0.011878956109285355, + 0.00823263917118311, + 0.012485173530876637, + 0.00869536679238081, + 0.006620473228394985, + 0.007113626226782799, + 0.025768162682652473, + 0.007257353980094194, + 0.008568383753299713, + 0.007544939871877432, + 0.005668180994689465, + 0.015497860498726368, + 0.003055343870073557, + 0.00299057480879128, + 0.0030942761804908514, + 0.004101119004189968, + 0.0024381866678595543, + 0.0031395412515848875, + 0.004172610118985176, + 0.003484449815005064, + 0.017196865752339363, + 0.022864826023578644, + 0.027390768751502037, + 0.015639612451195717, + 0.014329278841614723, + 0.010729719884693623, + 0.017877589911222458, + 0.016335001215338707, + 0.014916742220520973, + 0.011007474735379219, + 0.01590719446539879, + 0.013832293450832367, + 0.01104216743260622, + 0.00795914325863123, + 0.011634581722319126, + 0.009877897799015045, + 0.009762204252183437, + 0.005973716266453266, + 0.005185538437217474, + 0.008573496714234352, + 0.006192171014845371, + 0.005409647710621357, + 0.008433319628238678, + 0.005364257842302322, + 0.012100824154913425, + 0.009989888407289982, + 0.005563757847994566, + 0.010781963355839252, + 0.009105413220822811, + 0.009386626072227955, + 0.016554154455661774, + 0.009695124812424183, + 0.01287574227899313, + 0.009520557709038258, + 0.010765230283141136, + 0.016144616529345512, + 0.009820032864809036, + 0.009959985502064228, + 0.022308479994535446, + 0.014831510372459888, + 0.023357991129159927, + 0.009633546695113182, + 0.010486508719623089, + 0.06464789062738419, + 0.006248032674193382, + 0.008581753820180893, + 0.03599674999713898, + 0.007531440816819668, + 0.0029732927214354277, + 0.003810533322393894, + 0.003424200229346752, + 0.005294215399771929, + 0.0037648885045200586, + 0.002064622240141034, + 0.007560769096016884, + 0.0054304152727127075 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 3 + ], + "selected_reward": -0.014196122996509075 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return 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 purple dragon's claw pressing down on a jeweled egg?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the dragon's claw scales depicted as densely packed, fine, and shimmering with an iridescent amethyst sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a hairline crack appear on the eggshell and branch outward like a spiderweb?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a warm, luminous golden light pulsating from within the egg and casting sharp beams against a nearly black, out-of-focus background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera positioned at a slightly overhead perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02328341454267502 + ], + "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": [ + 0.1582263708114624, + 0.06648174673318863, + 0.047378044575452805, + 0.08385777473449707, + 0.03975820168852806, + 0.056372225284576416, + 0.052704937756061554, + 0.027441086247563362, + 0.44193947315216064, + 0.025840120390057564 + ], + "credit_map": [ + 0.008966050110757351, + 0.04569308087229729, + 0.008670762181282043, + 0.05554474890232086, + 0.010608506388962269, + 0.008398007601499557, + 0.012018891982734203, + 0.008326321840286255, + 0.006014341488480568, + 0.004176671151071787, + 0.016813671216368675, + 0.006677098106592894, + 0.005329350009560585, + 0.005541130900382996, + 0.006606638431549072, + 0.01532285287976265, + 0.004922062158584595, + 0.0085127679631114, + 0.005676585715264082, + 0.005784695502370596, + 0.004189719446003437, + 0.005235872231423855, + 0.006135770119726658, + 0.006920570973306894, + 0.010091032832860947, + 0.01110796257853508, + 0.015297843143343925, + 0.010007384233176708, + 0.008729619905352592, + 0.006341591477394104, + 0.011414486914873123, + 0.010867850854992867, + 0.005217206664383411, + 0.0044080778025090694, + 0.0058991145342588425, + 0.005880743730813265, + 0.005947411991655827, + 0.005164917558431625, + 0.003911342937499285, + 0.0033293848391622305, + 0.008009210228919983, + 0.010542291216552258, + 0.003501674858853221, + 0.004248275887221098, + 0.016114285215735435, + 0.004724147729575634, + 0.005404192488640547, + 0.0038281488232314587, + 0.008855784311890602, + 0.00948945339769125, + 0.0050523835234344006, + 0.006287097465246916, + 0.005578530486673117, + 0.005116085056215525, + 0.007814385928213596, + 0.004511214327067137, + 0.0047474876046180725, + 0.002160064410418272, + 0.0023135270457714796, + 0.006840934511274099, + 0.0020842363592237234, + 0.002553810365498066, + 0.00454646535217762, + 0.002194562228396535, + 0.06271839141845703, + 0.025279782712459564, + 0.02776050567626953, + 0.1760045886039734, + 0.013011213392019272, + 0.024050317704677582, + 0.09526470303535461, + 0.017849985510110855, + 0.003106503514572978, + 0.0022351343650370836, + 0.0032210545614361763, + 0.005905506666749716, + 0.002214570064097643, + 0.0012931666569784284, + 0.005033970344811678, + 0.0028302152641117573 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 3 + ], + "selected_reward": -0.02328341454267502 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "stunning macro still shot, captured from a slightly overhead perspective, focuses the viewer's attention on the claw of a young purple dragon, carefully pressing down on the crown of a jeweled egg. The claw, though small, exudes power, covered in densely packed, fine scales that shimmer with the iridescent sheen of amethyst. Under steady pressure, a hairline crack appears on the surface of the eggshell, resembling raw, milky quartz. Then, with a slow yet steady rhythm, the crack begins to branch and spread outward, like a spiderweb. Truly breathtaking, a warm, luminous golden light pulsates from within the eggshell, filtering through each new crack and casting sharp beams of light against the nearly black, out-of-focus background. This holy light, originating from within, illuminates the dragon's claws from below in an almost magical way, outlining the edges of each purple scale with a dazzling golden edge. Its translucent texture and complex grain are displayed with unprecedented clarity, creating an atmosphere of anticipation, tranquility, and powerful vitality at the moment of magic's birth. The picture is extremely rich in details and full of epic fantasy.", + "seed": 27991, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000121/group_8/index_1633_seed_27991_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": -0.015688705164939164, + "vlm_logprob": -0.015688705164939164, + "vlm_yes_no_margin": 4.625, + "vlm_token_credit": -0.015688705164939164, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return 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 purple dragon's claw pressing down on a jeweled egg?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the dragon's claw scales depicted as densely packed, fine, and shimmering with an iridescent amethyst sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a hairline crack appear on the eggshell and branch outward like a spiderweb?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a warm, luminous golden light pulsating from within the egg and casting sharp beams against a nearly black, out-of-focus background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera positioned at a slightly overhead perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00775339687243104 + ], + "opposite_logprob_tokens": [ + -5.257753372192383 + ], + "logprob_reward": -0.00775339687243104, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.00775339687243104, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10582317411899567, + 0.06784448772668839, + 0.09379642456769943, + 0.11231568455696106, + 0.03923496603965759, + 0.02233169786632061, + 0.05193307250738144, + 0.3328830897808075, + 0.08792468905448914, + 0.08591273427009583 + ], + "credit_map": [ + 0.01931013911962509, + 0.005904133897274733, + 0.02037719078361988, + 0.00980863906443119, + 0.01210083719342947, + 0.01177248451858759, + 0.018996521830558777, + 0.007553231902420521, + 0.0149392606690526, + 0.003852250287309289, + 0.016446078196167946, + 0.009003777988255024, + 0.005174299702048302, + 0.006510361563414335, + 0.007773163262754679, + 0.004145297221839428, + 0.013728913851082325, + 0.011507143266499043, + 0.01626427285373211, + 0.0074172550812363625, + 0.01288729440420866, + 0.00903704296797514, + 0.00885023083537817, + 0.014104272238910198, + 0.026114221662282944, + 0.015638001263141632, + 0.03139210864901543, + 0.009202823974192142, + 0.007621456403285265, + 0.007701146882027388, + 0.0064214798621833324, + 0.008224439807236195, + 0.005900110118091106, + 0.0034856584388762712, + 0.006312632001936436, + 0.00483498303219676, + 0.005048165563493967, + 0.005535654723644257, + 0.0038695731200277805, + 0.004248191602528095, + 0.004880047403275967, + 0.002222595037892461, + 0.005017567425966263, + 0.0015366974985226989, + 0.002248668810352683, + 0.00244202115572989, + 0.0018642895156517625, + 0.002119810087606311, + 0.015958057716488838, + 0.002577026141807437, + 0.017546385526657104, + 0.0025541812647134066, + 0.00328235374763608, + 0.0036393250338733196, + 0.003274408634752035, + 0.0031013311818242073, + 0.13380058109760284, + 0.013633654452860355, + 0.14150281250476837, + 0.010492788627743721, + 0.009863861836493015, + 0.004748926032334566, + 0.010944307781755924, + 0.007896166294813156, + 0.023707427084445953, + 0.0076461536809802055, + 0.025770042091608047, + 0.007169515825808048, + 0.005734436213970184, + 0.004619520157575607, + 0.008255726657807827, + 0.0050218673422932625, + 0.01668485626578331, + 0.007295117247849703, + 0.017038706690073013, + 0.014806813560426235, + 0.005587355233728886, + 0.004213636741042137, + 0.01310359500348568, + 0.007182649802416563 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 3, + 0 + ], + "selected_reward": -0.00775339687243104 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return 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 purple dragon's claw pressing down on a jeweled egg?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the dragon's claw scales depicted as densely packed, fine, and shimmering with an iridescent amethyst sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a hairline crack appear on the eggshell and branch outward like a spiderweb?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a warm, luminous golden light pulsating from within the egg and casting sharp beams against a nearly black, out-of-focus background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera positioned at a slightly overhead perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016085024923086166 + ], + "opposite_logprob_tokens": [ + -4.141085147857666 + ], + "logprob_reward": -0.016085024923086166, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016085024923086166, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12904563546180725, + 0.06726954132318497, + 0.10458679497241974, + 0.05735805630683899, + 0.030171139165759087, + 0.030349699780344963, + 0.033006519079208374, + 0.36458638310432434, + 0.13449017703533173, + 0.04913606867194176 + ], + "credit_map": [ + 0.03364188224077225, + 0.005749811884015799, + 0.03814589977264404, + 0.010200929827988148, + 0.010189351625740528, + 0.009386652149260044, + 0.014532068744301796, + 0.007199029438197613, + 0.010706687346100807, + 0.004701222293078899, + 0.010300626046955585, + 0.009555746801197529, + 0.007689381949603558, + 0.0066448659636080265, + 0.009696139954030514, + 0.007974870502948761, + 0.04127546399831772, + 0.003227998735383153, + 0.043100178241729736, + 0.004630979150533676, + 0.00328645552508533, + 0.0027958855498582125, + 0.0035114260390400887, + 0.0027584051713347435, + 0.010258599184453487, + 0.009862943552434444, + 0.012307467870414257, + 0.006196978967636824, + 0.0050871470011770725, + 0.0031701787374913692, + 0.005031289998441935, + 0.00544344587251544, + 0.004816840402781963, + 0.002819639630615711, + 0.005805963650345802, + 0.003013961249962449, + 0.0031218952499330044, + 0.004423904232680798, + 0.00313011696562171, + 0.0030388159211724997, + 0.005894731264561415, + 0.003718409687280655, + 0.007271819282323122, + 0.003644261509180069, + 0.00267927092500031, + 0.0028863619081676006, + 0.002104836981743574, + 0.002150005893781781, + 0.010054796002805233, + 0.00168887700419873, + 0.010924304835498333, + 0.0020504777785390615, + 0.001787817687727511, + 0.002251695143058896, + 0.00233454629778862, + 0.00191399862524122, + 0.1489948034286499, + 0.014829745516180992, + 0.1572425812482834, + 0.011565431021153927, + 0.009656731970608234, + 0.004973588977009058, + 0.011852472089231014, + 0.005471002776175737, + 0.051748815923929214, + 0.006032031029462814, + 0.05516006797552109, + 0.00540938600897789, + 0.004451577086001635, + 0.002991801593452692, + 0.005274338647723198, + 0.003422148758545518, + 0.011261086910963058, + 0.004263955168426037, + 0.011942971497774124, + 0.005206677597016096, + 0.0038814577274024487, + 0.004270180594176054, + 0.004772885702550411, + 0.0035368509124964476 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 8, + 0 + ], + "selected_reward": -0.016085024923086166 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return 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 purple dragon's claw pressing down on a jeweled egg?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the dragon's claw scales depicted as densely packed, fine, and shimmering with an iridescent amethyst sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a hairline crack appear on the eggshell and branch outward like a spiderweb?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a warm, luminous golden light pulsating from within the egg and casting sharp beams against a nearly black, out-of-focus background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera positioned at a slightly overhead perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023297857493162155 + ], + "opposite_logprob_tokens": [ + -3.7732977867126465 + ], + "logprob_reward": -0.023297857493162155, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023297857493162155, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1396954357624054, + 0.08784161508083344, + 0.06119014322757721, + 0.07151596993207932, + 0.03316289559006691, + 0.015699712559580803, + 0.031983014196157455, + 0.4085069000720978, + 0.09959003329277039, + 0.050814319401979446 + ], + "credit_map": [ + 0.036129336804151535, + 0.008161570876836777, + 0.03965485841035843, + 0.008749844506382942, + 0.01401176955550909, + 0.011220602318644524, + 0.014494793489575386, + 0.007272640243172646, + 0.030600756406784058, + 0.002951037371531129, + 0.03210537135601044, + 0.003730680327862501, + 0.005487511865794659, + 0.00541571993380785, + 0.0038979218807071447, + 0.0036526068579405546, + 0.022340204566717148, + 0.0018878815462812781, + 0.023420389741659164, + 0.003061053343117237, + 0.0032762009650468826, + 0.002446291735395789, + 0.0025159174110740423, + 0.002242203103378415, + 0.016950614750385284, + 0.008749374188482761, + 0.019396962597966194, + 0.006056374404579401, + 0.005151952616870403, + 0.0051340642385184765, + 0.005587887018918991, + 0.004488733131438494, + 0.004956706427037716, + 0.004609758500009775, + 0.006141247693449259, + 0.004436818417161703, + 0.004219491500407457, + 0.0029211388900876045, + 0.003172825090587139, + 0.002704903017729521, + 0.0029919175431132317, + 0.0014521231641992927, + 0.0034935956355184317, + 0.0016038534231483936, + 0.001778866397216916, + 0.001799795194528997, + 0.0013642238918691874, + 0.0012153356801718473, + 0.00897711981087923, + 0.0017856336198747158, + 0.00874410942196846, + 0.002341903280466795, + 0.0019492694409564137, + 0.0026841757353395224, + 0.00276197399944067, + 0.0027388266753405333, + 0.16213878989219666, + 0.01781122013926506, + 0.1744554042816162, + 0.013078683987259865, + 0.011605867184698582, + 0.0069214659743011, + 0.014694667421281338, + 0.007800797000527382, + 0.030663419514894485, + 0.00572088360786438, + 0.0323898084461689, + 0.009662920609116554, + 0.004591077566146851, + 0.0032279137521982193, + 0.007035934831947088, + 0.006298068445175886, + 0.009290790185332298, + 0.004193686880171299, + 0.00983813963830471, + 0.008188539184629917, + 0.004102636594325304, + 0.003482598112896085, + 0.007237113546580076, + 0.004480809438973665 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 8 + ], + "selected_reward": -0.023297857493162155 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return 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 purple dragon's claw pressing down on a jeweled egg?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the dragon's claw scales depicted as densely packed, fine, and shimmering with an iridescent amethyst sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a hairline crack appear on the eggshell and branch outward like a spiderweb?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a warm, luminous golden light pulsating from within the egg and casting sharp beams against a nearly black, out-of-focus background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera positioned at a slightly overhead perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0015144795179367065 + ], + "opposite_logprob_tokens": [ + -6.501514434814453 + ], + "logprob_reward": -0.0015144795179367065, + "yes_no_margin_reward": 6.5, + "token_credit_reward": -0.0015144795179367065, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08834633976221085, + 0.034939687699079514, + 0.03720628470182419, + 0.016514142975211143, + 0.034065648913383484, + 0.03481617569923401, + 0.07795809209346771, + 0.43831223249435425, + 0.05011763423681259, + 0.1877238005399704 + ], + "credit_map": [ + 0.024052025750279427, + 0.004030945245176554, + 0.02673337236046791, + 0.006360697094351053, + 0.00681455060839653, + 0.0060951560735702515, + 0.009881707839667797, + 0.004377881996333599, + 0.0101948706433177, + 0.002027902752161026, + 0.010328994132578373, + 0.0026981758419424295, + 0.0016491363057866693, + 0.0029936626087874174, + 0.0022264178842306137, + 0.0028205234557390213, + 0.013676105998456478, + 0.0017626839689910412, + 0.014212590642273426, + 0.0015997987939044833, + 0.0017545603914186358, + 0.001451384392566979, + 0.001287573017179966, + 0.0014615883119404316, + 0.0037355117965489626, + 0.001594957197085023, + 0.0040534911677241325, + 0.0013502738438546658, + 0.001425029942765832, + 0.0012253373861312866, + 0.0015977309085428715, + 0.0015318120131269097, + 0.006955733057111502, + 0.0034402625169605017, + 0.006457423325628042, + 0.004011351615190506, + 0.0033357355277985334, + 0.0037159097846597433, + 0.002844270085915923, + 0.0033049595076590776, + 0.011903847567737103, + 0.001675813109613955, + 0.01219253521412611, + 0.0015926988562569022, + 0.001635111984796822, + 0.0020047537982463837, + 0.0018279835348948836, + 0.0019834288395941257, + 0.029027298092842102, + 0.0033037918619811535, + 0.031531162559986115, + 0.0030064971651881933, + 0.0024776472710072994, + 0.0031016699504107237, + 0.0031767941545695066, + 0.0023332275450229645, + 0.18021811544895172, + 0.01817985810339451, + 0.19078749418258667, + 0.01284609641879797, + 0.011502275243401527, + 0.005150691606104374, + 0.014287720434367657, + 0.005340024828910828, + 0.008602909743785858, + 0.005744025111198425, + 0.00964987650513649, + 0.007112542167305946, + 0.005531232804059982, + 0.0033265629317611456, + 0.005521009210497141, + 0.004629473201930523, + 0.051606711000204086, + 0.01550264097750187, + 0.0552559420466423, + 0.017038071528077126, + 0.011491436511278152, + 0.010834832675755024, + 0.014783341437578201, + 0.011210811324417591 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 9, + 0 + ], + "selected_reward": -0.0015144795179367065 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return 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 purple dragon's claw pressing down on a jeweled egg?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the dragon's claw scales depicted as densely packed, fine, and shimmering with an iridescent amethyst sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a hairline crack appear on the eggshell and branch outward like a spiderweb?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a warm, luminous golden light pulsating from within the egg and casting sharp beams against a nearly black, out-of-focus background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera positioned at a slightly overhead perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029792767018079758 + ], + "opposite_logprob_tokens": [ + -3.5297927856445312 + ], + "logprob_reward": -0.029792767018079758, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029792767018079758, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07016021013259888, + 0.057199250906705856, + 0.04224410653114319, + 0.05336613580584526, + 0.03435000404715538, + 0.019093463197350502, + 0.07730065286159515, + 0.5683730840682983, + 0.024330882355570793, + 0.053582172840833664 + ], + "credit_map": [ + 0.011237086728215218, + 0.006627175956964493, + 0.012001121416687965, + 0.006500471383333206, + 0.01229699794203043, + 0.0073664868250489235, + 0.0072951954789459705, + 0.006835676729679108, + 0.015081607736647129, + 0.00319259543903172, + 0.015691595152020454, + 0.004455452784895897, + 0.004843881819397211, + 0.0034132522996515036, + 0.005041477736085653, + 0.0054793888702988625, + 0.008614842779934406, + 0.0041733235120773315, + 0.009242087602615356, + 0.003527835477143526, + 0.0037975928280502558, + 0.003284511622041464, + 0.0033554427791386843, + 0.0062484717927873135, + 0.009697236120700836, + 0.004924904089421034, + 0.010911842808127403, + 0.005291698966175318, + 0.0059506455436348915, + 0.005538644269108772, + 0.006646629422903061, + 0.00440453365445137, + 0.006267063319683075, + 0.0039431508630514145, + 0.006967134308069944, + 0.0030756548512727022, + 0.003874698653817177, + 0.003653699764981866, + 0.003326543839648366, + 0.0032420572824776173, + 0.0031529641710221767, + 0.002802414121106267, + 0.0033087001647800207, + 0.0025898567400872707, + 0.002208984922617674, + 0.0018970740493386984, + 0.0016415754798799753, + 0.0014918934321030974, + 0.028268633410334587, + 0.003746797563508153, + 0.02941747196018696, + 0.0035979440435767174, + 0.0020355014130473137, + 0.0036040551494807005, + 0.003621258307248354, + 0.003008990315720439, + 0.23371455073356628, + 0.023094456642866135, + 0.24770519137382507, + 0.01756283827126026, + 0.013614698313176632, + 0.006240615155547857, + 0.018691400066018105, + 0.007749300450086594, + 0.003919200971722603, + 0.002296366496011615, + 0.004528735764324665, + 0.0030064494349062443, + 0.0030210905242711306, + 0.0020665053743869066, + 0.0029515509959310293, + 0.002540983958169818, + 0.013822727836668491, + 0.0047843377105891705, + 0.01439651008695364, + 0.005672001279890537, + 0.003511956427246332, + 0.0032157343812286854, + 0.0047849868424236774, + 0.003393918974325061 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 6, + 0 + ], + "selected_reward": -0.029792767018079758 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "stunning macro still shot, captured from a slightly overhead perspective, focuses the viewer's attention on the claw of a young purple dragon, carefully pressing down on the crown of a jeweled egg. The claw, though small, exudes power, covered in densely packed, fine scales that shimmer with the iridescent sheen of amethyst. Under steady pressure, a hairline crack appears on the surface of the eggshell, resembling raw, milky quartz. Then, with a slow yet steady rhythm, the crack begins to branch and spread outward, like a spiderweb. Truly breathtaking, a warm, luminous golden light pulsates from within the eggshell, filtering through each new crack and casting sharp beams of light against the nearly black, out-of-focus background. This holy light, originating from within, illuminates the dragon's claws from below in an almost magical way, outlining the edges of each purple scale with a dazzling golden edge. Its translucent texture and complex grain are displayed with unprecedented clarity, creating an atmosphere of anticipation, tranquility, and powerful vitality at the moment of magic's birth. The picture is extremely rich in details and full of epic fantasy.", + "seed": 27992, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000121/group_8/index_1633_seed_27992_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": -0.022925542341545223, + "vlm_logprob": -0.022925542341545223, + "vlm_yes_no_margin": 4.275, + "vlm_token_credit": -0.022925542341545223, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return 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 purple dragon's claw pressing down on a jeweled egg?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the dragon's claw scales depicted as densely packed, fine, and shimmering with an iridescent amethyst sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a hairline crack appear on the eggshell and branch outward like a spiderweb?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a warm, luminous golden light pulsating from within the egg and casting sharp beams against a nearly black, out-of-focus background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera positioned at a slightly overhead perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011477158404886723 + ], + "opposite_logprob_tokens": [ + -4.761476993560791 + ], + "logprob_reward": -0.011477158404886723, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.011477158404886723, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17817270755767822, + 0.1391637623310089, + 0.06944488734006882, + 0.0797402635216713, + 0.15810996294021606, + 0.06789602339267731, + 0.0721818059682846, + 0.04055102914571762, + 0.1110084131360054, + 0.08373111486434937 + ], + "credit_map": [ + 0.037831373512744904, + 0.017997821792960167, + 0.031295791268348694, + 0.012097659520804882, + 0.01874617487192154, + 0.017473196610808372, + 0.030143387615680695, + 0.012587301433086395, + 0.026239557191729546, + 0.009196260012686253, + 0.026035621762275696, + 0.013988508842885494, + 0.022295523434877396, + 0.015097502619028091, + 0.015959717333316803, + 0.010351053439080715, + 0.011962332762777805, + 0.0039264787919819355, + 0.013554448261857033, + 0.005796000361442566, + 0.009084673598408699, + 0.007196743041276932, + 0.012218274176120758, + 0.0057059344835579395, + 0.011632137931883335, + 0.007049201987683773, + 0.012531750835478306, + 0.007365294732153416, + 0.010961969383060932, + 0.008027543313801289, + 0.015681466087698936, + 0.00649089552462101, + 0.02235231176018715, + 0.012718317098915577, + 0.019856352359056473, + 0.016946597024798393, + 0.016500089317560196, + 0.02038058079779148, + 0.019725117832422256, + 0.029630597680807114, + 0.013305970467627048, + 0.0041392226703464985, + 0.01089337095618248, + 0.006742451339960098, + 0.006426322739571333, + 0.007793438620865345, + 0.010255894623696804, + 0.0083393519744277, + 0.015562423504889011, + 0.004132863134145737, + 0.015142107382416725, + 0.007176058366894722, + 0.007663330063223839, + 0.00763336569070816, + 0.006721725221723318, + 0.008149928413331509, + 0.005277588032186031, + 0.003877925220876932, + 0.006552782841026783, + 0.0049883658066391945, + 0.004288654308766127, + 0.004529121331870556, + 0.007076393812894821, + 0.003960201051086187, + 0.020663246512413025, + 0.010134679265320301, + 0.021356286481022835, + 0.012662135064601898, + 0.010251735337078571, + 0.011331569403409958, + 0.014184719882905483, + 0.010424036532640457, + 0.010964908637106419, + 0.007415466010570526, + 0.011854161508381367, + 0.007563911844044924, + 0.015020851045846939, + 0.008351312950253487, + 0.01311421301215887, + 0.009446288459002972 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 1 + ], + "selected_reward": -0.011477158404886723 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return 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 purple dragon's claw pressing down on a jeweled egg?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the dragon's claw scales depicted as densely packed, fine, and shimmering with an iridescent amethyst sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a hairline crack appear on the eggshell and branch outward like a spiderweb?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a warm, luminous golden light pulsating from within the egg and casting sharp beams against a nearly black, out-of-focus background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera positioned at a slightly overhead perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029833264648914337 + ], + "opposite_logprob_tokens": [ + -3.5298333168029785 + ], + "logprob_reward": -0.029833264648914337, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029833264648914337, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19970844686031342, + 0.07909073680639267, + 0.16674567759037018, + 0.1457321047782898, + 0.11857213824987411, + 0.043029092252254486, + 0.04538227617740631, + 0.12322266399860382, + 0.035751823335886, + 0.0427650585770607 + ], + "credit_map": [ + 0.032611098140478134, + 0.027924926951527596, + 0.03246296942234039, + 0.011938227340579033, + 0.025503193959593773, + 0.0310746431350708, + 0.02246277593076229, + 0.0157306045293808, + 0.015232755802571774, + 0.005725470371544361, + 0.013422713615000248, + 0.008242175914347172, + 0.012220636010169983, + 0.01029662974178791, + 0.009394513443112373, + 0.004555845633149147, + 0.033334098756313324, + 0.01526882778853178, + 0.029729964211583138, + 0.014142464846372604, + 0.022766979411244392, + 0.016103651374578476, + 0.025017503648996353, + 0.010382190346717834, + 0.029366012662649155, + 0.009039712138473988, + 0.02626175992190838, + 0.010347082279622555, + 0.01520216092467308, + 0.010923628695309162, + 0.032709576189517975, + 0.011882172897458076, + 0.01860116980969906, + 0.010332951322197914, + 0.015192566439509392, + 0.011326831765472889, + 0.012896578758955002, + 0.01464944425970316, + 0.016402488574385643, + 0.01917010359466076, + 0.007012428715825081, + 0.003370867809280753, + 0.006976953241974115, + 0.005439225118607283, + 0.0042622946202754974, + 0.005241053644567728, + 0.005793911404907703, + 0.00493235420435667, + 0.00874357856810093, + 0.0030299208592623472, + 0.008059077896177769, + 0.0037782860454171896, + 0.004715811461210251, + 0.005642326083034277, + 0.006439140997827053, + 0.004974135663360357, + 0.019895881414413452, + 0.0076698847115039825, + 0.01972251385450363, + 0.014262079261243343, + 0.011511238291859627, + 0.009949587285518646, + 0.02742602862417698, + 0.012785445898771286, + 0.005100179463624954, + 0.0026176157407462597, + 0.005415911786258221, + 0.004881583619862795, + 0.0038620184641331434, + 0.0033451071940362453, + 0.005893077235668898, + 0.004636332392692566, + 0.006208340637385845, + 0.003481578314676881, + 0.006452653557062149, + 0.006329548545181751, + 0.004010444972664118, + 0.004844037350267172, + 0.005906722974032164, + 0.005531732924282551 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 3 + ], + "selected_reward": -0.029833264648914337 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return 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 purple dragon's claw pressing down on a jeweled egg?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the dragon's claw scales depicted as densely packed, fine, and shimmering with an iridescent amethyst sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a hairline crack appear on the eggshell and branch outward like a spiderweb?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a warm, luminous golden light pulsating from within the egg and casting sharp beams against a nearly black, out-of-focus background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera positioned at a slightly overhead perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03370881825685501 + ], + "opposite_logprob_tokens": [ + -3.4087088108062744 + ], + "logprob_reward": -0.03370881825685501, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03370881825685501, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1320287436246872, + 0.08309537172317505, + 0.06292560696601868, + 0.09765129536390305, + 0.09101679921150208, + 0.11071642488241196, + 0.18097996711730957, + 0.11875788122415543, + 0.056423742324113846, + 0.06640414893627167 + ], + "credit_map": [ + 0.019620146602392197, + 0.018221385776996613, + 0.02162203937768936, + 0.013088036328554153, + 0.0158431064337492, + 0.010203051380813122, + 0.025152679532766342, + 0.008278309367597103, + 0.013502007350325584, + 0.004349196329712868, + 0.01588764786720276, + 0.00675842585042119, + 0.01398314256221056, + 0.009892318397760391, + 0.011743979528546333, + 0.006978664547204971, + 0.009307654574513435, + 0.004821040201932192, + 0.011519298888742924, + 0.005117885768413544, + 0.008604886010289192, + 0.006333418656140566, + 0.011865214444696903, + 0.0053562140092253685, + 0.018150288611650467, + 0.007665077690035105, + 0.01799554005265236, + 0.01071971096098423, + 0.009357104077935219, + 0.008246704936027527, + 0.016168974339962006, + 0.009347901679575443, + 0.012074378319084644, + 0.010287242010235786, + 0.014133620075881481, + 0.007980456575751305, + 0.011346397921442986, + 0.009372013621032238, + 0.018505899235606194, + 0.007316803093999624, + 0.016017479822039604, + 0.008642144501209259, + 0.017455261200666428, + 0.013156109489500523, + 0.012444681487977505, + 0.011286354623734951, + 0.020747171714901924, + 0.010967228561639786, + 0.029673809185624123, + 0.014879635535180569, + 0.027810191735625267, + 0.024525444954633713, + 0.021506208926439285, + 0.01935671456158161, + 0.024766117334365845, + 0.01846187189221382, + 0.016331138089299202, + 0.006309750955551863, + 0.02029474452137947, + 0.015493259765207767, + 0.013101101852953434, + 0.011459680274128914, + 0.025282179936766624, + 0.010486037470400333, + 0.011351210996508598, + 0.004309435375034809, + 0.010836882516741753, + 0.005766274407505989, + 0.005271218251436949, + 0.005796727258712053, + 0.00794255081564188, + 0.00514944875612855, + 0.009622691199183464, + 0.005041958764195442, + 0.010793255642056465, + 0.008039223030209541, + 0.006379291880875826, + 0.0067578088492155075, + 0.010105744935572147, + 0.00966417696326971 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 7 + ], + "selected_reward": -0.03370881825685501 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return 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 purple dragon's claw pressing down on a jeweled egg?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the dragon's claw scales depicted as densely packed, fine, and shimmering with an iridescent amethyst sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a hairline crack appear on the eggshell and branch outward like a spiderweb?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a warm, luminous golden light pulsating from within the egg and casting sharp beams against a nearly black, out-of-focus background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera positioned at a slightly overhead perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0015149556566029787 + ], + "opposite_logprob_tokens": [ + -6.501514911651611 + ], + "logprob_reward": -0.0015149556566029787, + "yes_no_margin_reward": 6.5, + "token_credit_reward": -0.0015149556566029787, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13583646714687347, + 0.07460582256317139, + 0.05467734485864639, + 0.06536581367254257, + 0.2978217303752899, + 0.09115483611822128, + 0.04775632545351982, + 0.09530714899301529, + 0.0640067458152771, + 0.07346773892641068 + ], + "credit_map": [ + 0.02942301146686077, + 0.013360440731048584, + 0.02355693094432354, + 0.010646271519362926, + 0.015446703881025314, + 0.01430165022611618, + 0.017384584993124008, + 0.011716878041625023, + 0.015094541013240814, + 0.006457340903580189, + 0.011799409985542297, + 0.00682463264092803, + 0.009073985740542412, + 0.010768720880150795, + 0.009208525530993938, + 0.00537867471575737, + 0.01203136146068573, + 0.004381205886602402, + 0.009590458124876022, + 0.004703770857304335, + 0.006628394592553377, + 0.0073240348137915134, + 0.006633284501731396, + 0.0033848353195935488, + 0.015424102544784546, + 0.0070971595123410225, + 0.011520965956151485, + 0.006341255269944668, + 0.005360560957342386, + 0.005672472529113293, + 0.008492471650242805, + 0.005456829443573952, + 0.04154440760612488, + 0.021549830213189125, + 0.039063505828380585, + 0.027518194168806076, + 0.029965482652187347, + 0.042552754282951355, + 0.031923577189445496, + 0.0637039840221405, + 0.011989052407443523, + 0.00840668473392725, + 0.014919379726052284, + 0.009632591158151627, + 0.011958877556025982, + 0.010274723172187805, + 0.013562253676354885, + 0.010411273688077927, + 0.009691990911960602, + 0.002697952091693878, + 0.008964172564446926, + 0.005841068923473358, + 0.004799728747457266, + 0.005242670886218548, + 0.005079532042145729, + 0.005439207889139652, + 0.016360290348529816, + 0.007743273861706257, + 0.019846148788928986, + 0.011519160121679306, + 0.007428324781358242, + 0.00981920026242733, + 0.011208637617528439, + 0.01138211227953434, + 0.008347325026988983, + 0.004888364113867283, + 0.009659304283559322, + 0.008956650272011757, + 0.006526345387101173, + 0.006761061027646065, + 0.013011184521019459, + 0.005856509320437908, + 0.01030005980283022, + 0.006887066178023815, + 0.01359713077545166, + 0.00897624995559454, + 0.007109758909791708, + 0.008877066895365715, + 0.00947272963821888, + 0.008247677236795425 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 7 + ], + "selected_reward": -0.0015149556566029787 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a young purple dragon's claw pressing down on a jeweled egg?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the dragon's claw scales depicted as densely packed, fine, and shimmering with an iridescent amethyst sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a hairline crack appear on the eggshell and branch outward like a spiderweb?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a warm, luminous golden light pulsating from within the egg and casting sharp beams against a nearly black, out-of-focus background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera positioned at a slightly overhead perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03809351474046707 + ], + "opposite_logprob_tokens": [ + -3.2880935668945312 + ], + "logprob_reward": -0.03809351474046707, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03809351474046707, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23071812093257904, + 0.16437380015850067, + 0.11516884714365005, + 0.061374593526124954, + 0.09285882860422134, + 0.04814569279551506, + 0.09822069853544235, + 0.061017367988824844, + 0.0738871842622757, + 0.054234828799963 + ], + "credit_map": [ + 0.03156433254480362, + 0.019098693504929543, + 0.03460997715592384, + 0.020925069227814674, + 0.0357796847820282, + 0.02873898111283779, + 0.04114875569939613, + 0.01885264553129673, + 0.029976166784763336, + 0.013332545757293701, + 0.028075002133846283, + 0.02050609327852726, + 0.022620538249611855, + 0.01795872487127781, + 0.022773662582039833, + 0.00913107581436634, + 0.02095700614154339, + 0.010587030090391636, + 0.01847478561103344, + 0.012406200170516968, + 0.015763770788908005, + 0.011258522048592567, + 0.01900389790534973, + 0.00671763950958848, + 0.012701824307441711, + 0.004637609235942364, + 0.009697036817669868, + 0.005526564083993435, + 0.0078284852206707, + 0.006733998190611601, + 0.009027253836393356, + 0.005221829749643803, + 0.015945401042699814, + 0.009597971104085445, + 0.014123748056590557, + 0.009808901697397232, + 0.008010948076844215, + 0.012341238558292389, + 0.011061422526836395, + 0.011969204060733318, + 0.006942881736904383, + 0.003749037627130747, + 0.00816817581653595, + 0.005856018979102373, + 0.004479407798498869, + 0.004648579750210047, + 0.009162372909486294, + 0.00513922143727541, + 0.016572700813412666, + 0.007446590345352888, + 0.01510512176901102, + 0.01215599849820137, + 0.014704214408993721, + 0.0117584727704525, + 0.010797612369060516, + 0.009679997339844704, + 0.00969635508954525, + 0.004001756198704243, + 0.01152920164167881, + 0.007489002775400877, + 0.0056954724714159966, + 0.00626601604744792, + 0.007453854661434889, + 0.008885710500180721, + 0.013083448633551598, + 0.007044539786875248, + 0.014234324917197227, + 0.009649945423007011, + 0.006595460698008537, + 0.007394787855446339, + 0.007606887258589268, + 0.008277797140181065, + 0.008659428916871548, + 0.0058296420611441135, + 0.008211791515350342, + 0.00676504522562027, + 0.005935959983617067, + 0.006254766136407852, + 0.006330646108835936, + 0.006247553043067455 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.03809351474046707 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "stunning macro still shot, captured from a slightly overhead perspective, focuses the viewer's attention on the claw of a young purple dragon, carefully pressing down on the crown of a jeweled egg. The claw, though small, exudes power, covered in densely packed, fine scales that shimmer with the iridescent sheen of amethyst. Under steady pressure, a hairline crack appears on the surface of the eggshell, resembling raw, milky quartz. Then, with a slow yet steady rhythm, the crack begins to branch and spread outward, like a spiderweb. Truly breathtaking, a warm, luminous golden light pulsates from within the eggshell, filtering through each new crack and casting sharp beams of light against the nearly black, out-of-focus background. This holy light, originating from within, illuminates the dragon's claws from below in an almost magical way, outlining the edges of each purple scale with a dazzling golden edge. Its translucent texture and complex grain are displayed with unprecedented clarity, creating an atmosphere of anticipation, tranquility, and powerful vitality at the moment of magic's birth. The picture is extremely rich in details and full of epic fantasy.", + "seed": 27993, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000121/group_8/index_1633_seed_27993_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": -0.015056267008185386, + "vlm_logprob": -0.015056267008185386, + "vlm_yes_no_margin": 4.425, + "vlm_token_credit": -0.015056267008185386, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return 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 purple dragon's claw pressing down on a jeweled egg?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the dragon's claw scales depicted as densely packed, fine, and shimmering with an iridescent amethyst sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a hairline crack appear on the eggshell and branch outward like a spiderweb?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a warm, luminous golden light pulsating from within the egg and casting sharp beams against a nearly black, out-of-focus background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera positioned at a slightly overhead perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.017206907272338867 + ], + "opposite_logprob_tokens": [ + -4.392207145690918 + ], + "logprob_reward": -0.017206907272338867, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.017206907272338867, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3412567973136902, + 0.04354210942983627, + 0.08690860122442245, + 0.1237121969461441, + 0.0270011555403471, + 0.05659788101911545, + 0.03954858332872391, + 0.07753125578165054, + 0.0950678288936615, + 0.10883358120918274 + ], + "credit_map": [ + 0.04463198035955429, + 0.037435971200466156, + 0.050520628690719604, + 0.028864793479442596, + 0.038505133241415024, + 0.06905781477689743, + 0.03583597391843796, + 0.03640449047088623, + 0.0059372615069150925, + 0.005516628734767437, + 0.00722985016182065, + 0.004432587884366512, + 0.004817142616957426, + 0.0034747084137052298, + 0.005858260672539473, + 0.006275671999901533, + 0.010168534703552723, + 0.017243873327970505, + 0.012204157188534737, + 0.006376046221703291, + 0.019983790814876556, + 0.0074499803595244884, + 0.006965655833482742, + 0.006516569294035435, + 0.01926588825881481, + 0.017408030107617378, + 0.016330907121300697, + 0.00925096869468689, + 0.019169600680470467, + 0.015763597562909126, + 0.009469990618526936, + 0.01705319993197918, + 0.0038798910100013018, + 0.0022623878903687, + 0.004699050448834896, + 0.002881098072975874, + 0.002154409186914563, + 0.003952448256313801, + 0.0025946390815079212, + 0.004577231127768755, + 0.007892700843513012, + 0.007074588909745216, + 0.008631031960248947, + 0.008170424960553646, + 0.00535331666469574, + 0.0067840940319001675, + 0.006977453827857971, + 0.005714274011552334, + 0.006751479115337133, + 0.003837734693661332, + 0.004235742148011923, + 0.0048201545141637325, + 0.003965628799051046, + 0.004172885790467262, + 0.007078222464770079, + 0.004686734639108181, + 0.012910985387861729, + 0.008147435262799263, + 0.010311944410204887, + 0.010302433744072914, + 0.008055262267589569, + 0.008173735812306404, + 0.010468246415257454, + 0.009161215275526047, + 0.015478572808206081, + 0.01135291438549757, + 0.011613166891038418, + 0.00935771781951189, + 0.01600353978574276, + 0.011595646850764751, + 0.01093806978315115, + 0.00872819498181343, + 0.017766986042261124, + 0.010927018709480762, + 0.016760364174842834, + 0.011876176111400127, + 0.01190335676074028, + 0.012239769101142883, + 0.014899911358952522, + 0.01245999988168478 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 9 + ], + "selected_reward": -0.017206907272338867 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return 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 purple dragon's claw pressing down on a jeweled egg?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the dragon's claw scales depicted as densely packed, fine, and shimmering with an iridescent amethyst sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a hairline crack appear on the eggshell and branch outward like a spiderweb?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a warm, luminous golden light pulsating from within the egg and casting sharp beams against a nearly black, out-of-focus background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera positioned at a slightly overhead perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01610097847878933 + ], + "opposite_logprob_tokens": [ + -4.141100883483887 + ], + "logprob_reward": -0.01610097847878933, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.01610097847878933, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4601321220397949, + 0.048878107219934464, + 0.09782223403453827, + 0.07578366249799728, + 0.04075132682919502, + 0.06223955750465393, + 0.037626661360263824, + 0.039918918162584305, + 0.06248541548848152, + 0.07436198741197586 + ], + "credit_map": [ + 0.056806888431310654, + 0.05101492628455162, + 0.06623191386461258, + 0.04531574621796608, + 0.04689818248152733, + 0.0874267965555191, + 0.053422898054122925, + 0.05301478132605553, + 0.00747174397110939, + 0.009643783792853355, + 0.006556268781423569, + 0.004250417463481426, + 0.008758893236517906, + 0.004696503281593323, + 0.003308714833110571, + 0.004191788379102945, + 0.01424006000161171, + 0.015377664007246494, + 0.012201984412968159, + 0.006956126075237989, + 0.009558083489537239, + 0.02394990250468254, + 0.006088078022003174, + 0.009450336918234825, + 0.010856413282454014, + 0.007136213127523661, + 0.011128772981464863, + 0.00727528752759099, + 0.009397841989994049, + 0.014215080998837948, + 0.0054550389759242535, + 0.010319012217223644, + 0.00848002452403307, + 0.0046170256100595, + 0.010734278708696365, + 0.003495798446238041, + 0.0033710389398038387, + 0.0034640964586287737, + 0.0031553616281598806, + 0.0034337046090513468, + 0.009188584052026272, + 0.0072075678035616875, + 0.010324200615286827, + 0.008502645418047905, + 0.0048270197585225105, + 0.007420183625072241, + 0.006985738407820463, + 0.007783618755638599, + 0.006675673183053732, + 0.0030618098098784685, + 0.00541336927562952, + 0.003701471723616123, + 0.0033301750663667917, + 0.003654720028862357, + 0.007380278781056404, + 0.00440916558727622, + 0.006325502879917622, + 0.004723015241324902, + 0.0047436966560781, + 0.005813430529087782, + 0.004138363525271416, + 0.004266458563506603, + 0.005674426443874836, + 0.004234026651829481, + 0.00841542799025774, + 0.006058857310563326, + 0.006246370729058981, + 0.008576314896345139, + 0.0063974387012422085, + 0.0054438053630292416, + 0.009748248383402824, + 0.011598953045904636, + 0.014143680222332478, + 0.008424078114330769, + 0.014974069781601429, + 0.00697966618463397, + 0.007295047864317894, + 0.007427359465509653, + 0.00794035941362381, + 0.007177731487900019 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 3 + ], + "selected_reward": -0.01610097847878933 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return 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 purple dragon's claw pressing down on a jeweled egg?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the dragon's claw scales depicted as densely packed, fine, and shimmering with an iridescent amethyst sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a hairline crack appear on the eggshell and branch outward like a spiderweb?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a warm, luminous golden light pulsating from within the egg and casting sharp beams against a nearly black, out-of-focus background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera positioned at a slightly overhead perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018195146694779396 + ], + "opposite_logprob_tokens": [ + -4.018195152282715 + ], + "logprob_reward": -0.018195146694779396, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018195146694779396, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.26702457666397095, + 0.05293766409158707, + 0.04165917634963989, + 0.044032905250787735, + 0.06186356022953987, + 0.055911727249622345, + 0.04466262087225914, + 0.04581685736775398, + 0.22730600833892822, + 0.15878497064113617 + ], + "credit_map": [ + 0.033945128321647644, + 0.029124805703759193, + 0.03955307602882385, + 0.026653064414858818, + 0.02655601128935814, + 0.04846589267253876, + 0.02338598482310772, + 0.03934059292078018, + 0.007375303190201521, + 0.0082265455275774, + 0.0067530120722949505, + 0.004193766508251429, + 0.00898328423500061, + 0.0052398014813661575, + 0.00581843126565218, + 0.006347517482936382, + 0.005307191051542759, + 0.004112198483198881, + 0.004745362792164087, + 0.0036253423895686865, + 0.008503585122525692, + 0.00671420618891716, + 0.0033523859456181526, + 0.005298906471580267, + 0.007674485445022583, + 0.004082315135747194, + 0.009137592278420925, + 0.004277872387319803, + 0.003816803451627493, + 0.004422530997544527, + 0.004394028801470995, + 0.006227276753634214, + 0.01100107841193676, + 0.006763402372598648, + 0.011514214798808098, + 0.005028006620705128, + 0.005758588667958975, + 0.0058073862455785275, + 0.005696438252925873, + 0.010294445790350437, + 0.008489346131682396, + 0.005860277451574802, + 0.008122929371893406, + 0.009040957316756248, + 0.0038090141024440527, + 0.005528880283236504, + 0.007119227200746536, + 0.007941095158457756, + 0.007852060720324516, + 0.003808490699157119, + 0.007389213889837265, + 0.0053548612631857395, + 0.00345067260786891, + 0.004918438382446766, + 0.006293496582657099, + 0.005595382768660784, + 0.007268838584423065, + 0.004757265094667673, + 0.006280815228819847, + 0.007125816307961941, + 0.0047391424886882305, + 0.0033214481081813574, + 0.00674498500302434, + 0.005578543525189161, + 0.02955574356019497, + 0.02203887328505516, + 0.02140999585390091, + 0.03243601322174072, + 0.01951155625283718, + 0.0209723562002182, + 0.035303596407175064, + 0.04607786238193512, + 0.023533150553703308, + 0.018906334415078163, + 0.03584786877036095, + 0.01363999955356121, + 0.018596166744828224, + 0.01450745202600956, + 0.016005732119083405, + 0.017748268321156502 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 9 + ], + "selected_reward": -0.018195146694779396 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return 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 purple dragon's claw pressing down on a jeweled egg?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the dragon's claw scales depicted as densely packed, fine, and shimmering with an iridescent amethyst sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a hairline crack appear on the eggshell and branch outward like a spiderweb?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a warm, luminous golden light pulsating from within the egg and casting sharp beams against a nearly black, out-of-focus background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera positioned at a slightly overhead perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0031949449330568314 + ], + "opposite_logprob_tokens": [ + -5.753194808959961 + ], + "logprob_reward": -0.0031949449330568314, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.0031949449330568314, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.48567914962768555, + 0.049780141562223434, + 0.06512119621038437, + 0.09428303688764572, + 0.04060770571231842, + 0.04190312698483467, + 0.029911411926150322, + 0.03654970973730087, + 0.0724889487028122, + 0.0836755633354187 + ], + "credit_map": [ + 0.05932061746716499, + 0.061478614807128906, + 0.06072269007563591, + 0.04227278381586075, + 0.054169002920389175, + 0.09817957878112793, + 0.060808297246694565, + 0.04872756451368332, + 0.007348505314439535, + 0.00622511375695467, + 0.010073741897940636, + 0.004746284801512957, + 0.006765035912394524, + 0.004244628828018904, + 0.006082745268940926, + 0.004294090438634157, + 0.011288711801171303, + 0.006074215285480022, + 0.014335205778479576, + 0.0056932345032691956, + 0.0059787617065012455, + 0.006730349734425545, + 0.006824515759944916, + 0.00819620955735445, + 0.01637491211295128, + 0.009615818969905376, + 0.01669733226299286, + 0.009382001124322414, + 0.00834118202328682, + 0.01410000491887331, + 0.007433668244630098, + 0.012338117696344852, + 0.007593764923512936, + 0.0046743303537368774, + 0.008272437378764153, + 0.004372419789433479, + 0.0028571663424372673, + 0.004156512673944235, + 0.004014758393168449, + 0.004666318651288748, + 0.0065161315724253654, + 0.004837953485548496, + 0.006826208904385567, + 0.0067494940012693405, + 0.003449915675446391, + 0.0034690171014517546, + 0.004994262475520372, + 0.005060145631432533, + 0.003911016508936882, + 0.0027883369475603104, + 0.004037979058921337, + 0.0034499361645430326, + 0.004794502165168524, + 0.0033005482982844114, + 0.0040603987872600555, + 0.0035686958581209183, + 0.005796788260340691, + 0.004324326757341623, + 0.005633367225527763, + 0.0049385689198970795, + 0.00324057973921299, + 0.00516044395044446, + 0.003893553977832198, + 0.003562081838026643, + 0.01134059950709343, + 0.007922211661934853, + 0.01343615259975195, + 0.010421249084174633, + 0.007587439846247435, + 0.007350410334765911, + 0.007217098958790302, + 0.0072137946262955666, + 0.012313907966017723, + 0.006206005811691284, + 0.014254461973905563, + 0.010072587989270687, + 0.009583471342921257, + 0.008616427890956402, + 0.014489443972706795, + 0.008139264769852161 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 9 + ], + "selected_reward": -0.0031949449330568314 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return 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 purple dragon's claw pressing down on a jeweled egg?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the dragon's claw scales depicted as densely packed, fine, and shimmering with an iridescent amethyst sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a hairline crack appear on the eggshell and branch outward like a spiderweb?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a warm, luminous golden light pulsating from within the egg and casting sharp beams against a nearly black, out-of-focus background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera positioned at a slightly overhead perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02058335766196251 + ], + "opposite_logprob_tokens": [ + -3.895583391189575 + ], + "logprob_reward": -0.02058335766196251, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.02058335766196251, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2596336305141449, + 0.06055503338575363, + 0.07910461723804474, + 0.05161917582154274, + 0.0850677341222763, + 0.06309668719768524, + 0.044914864003658295, + 0.12291813641786575, + 0.10790208727121353, + 0.12518802285194397 + ], + "credit_map": [ + 0.03398768976330757, + 0.036386482417583466, + 0.04005981236696243, + 0.02560066618025303, + 0.037832994014024734, + 0.03123643808066845, + 0.021487342193722725, + 0.03304223343729973, + 0.01028265431523323, + 0.008208182640373707, + 0.009291443973779678, + 0.006320497486740351, + 0.004868660122156143, + 0.006543728988617659, + 0.00626997509971261, + 0.008769889362156391, + 0.01133129745721817, + 0.006109480746090412, + 0.011743522249162197, + 0.007651878986507654, + 0.015032053925096989, + 0.013390034437179565, + 0.005882049910724163, + 0.007964307442307472, + 0.00825112871825695, + 0.006572365760803223, + 0.008138066157698631, + 0.004230774473398924, + 0.0060904016718268394, + 0.008327223360538483, + 0.0038872677832841873, + 0.006121947895735502, + 0.014988387003540993, + 0.008677518926560879, + 0.018215375021100044, + 0.008830097503960133, + 0.0075486646965146065, + 0.011587608605623245, + 0.007440404500812292, + 0.007779676467180252, + 0.010565277189016342, + 0.004926464054733515, + 0.011187252588570118, + 0.006899350322782993, + 0.00527029437944293, + 0.010170447640120983, + 0.006402633152902126, + 0.007674965541809797, + 0.007153033744543791, + 0.0055918581783771515, + 0.006927171256393194, + 0.004311244469136, + 0.006188174709677696, + 0.004592384677380323, + 0.005347329191863537, + 0.004803668241947889, + 0.018463369458913803, + 0.014079203829169273, + 0.015726935118436813, + 0.02145872823894024, + 0.012606335803866386, + 0.01277944352477789, + 0.01545486506074667, + 0.012349258176982403, + 0.01732165552675724, + 0.011895429342985153, + 0.01247074268758297, + 0.0175689198076725, + 0.013185965828597546, + 0.008692597039043903, + 0.016362104564905167, + 0.010404668748378754, + 0.019393594935536385, + 0.015819266438484192, + 0.021466409787535667, + 0.01361172180622816, + 0.013547004200518131, + 0.013552861288189888, + 0.015153533779084682, + 0.01264362782239914 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 7 + ], + "selected_reward": -0.02058335766196251 + } + ], + "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/0000121/group_9/index_tensor([6916, 6916])_group_9_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000121/group_9/index_tensor([6916, 6916])_group_9_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..1cd453a26d5d8d462bff8a06338dbc0a1443a9df --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000121/group_9/index_tensor([6916, 6916])_group_9_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "dynamic, low-angle, handheld tracking shot closely follows a police officer sprinting at full speed. The scene is set in a narrow, grimy alley on a rainy night. The rain is pouring down, and the ground is filled with puddles. The officer wears a rain-soaked dark blue police uniform, covered with a tactical vest whose reflective strips flicker in the light. His short, wet black hair clings to his forehead, and his face is a mixture of rain and sweat. His expression is steely and hurried with intense concentration. Each step he takes slams against the asphalt, sending up circles of water mixed with neon reflections. The red, blue, and green neon lights from the shops at the alley's entrance intermittently cut through the darkness, painting the wet brick walls and scattered garbage with a cacophony of colors. Light and shadow shift rapidly as the camera shakes violently. His target is a blurry black figure that is barely visible at the end of the alley and almost merges into the shadows. The policeman leans forward, swings his arms vigorously, and the white breath exhaled from his mouth quickly dissipates in the cold air. The whole picture is decorated with psychedelic and chaotic neon lights against a cold background, full of the urgency of pursuit, uncertainty and suffocating tension.", + "seed": 7137, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000121/group_9/index_6916_seed_7137_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": -0.15228118216618897, + "vlm_logprob": -0.15228118216618897, + "vlm_yes_no_margin": 3.775, + "vlm_token_credit": -0.15228118216618897, + "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 police officer sprinting at full speed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a narrow, grimy alley on a rainy night with puddles on the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the officer wear a dark blue rain-soaked uniform with a tactical vest featuring flickering reflective strips?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera shot low-angle, handheld, and tracking the officer while shaking violently?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are red, blue, and green neon lights visible reflecting on wet surfaces and illuminating the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.6975152492523193 + ], + "opposite_logprob_tokens": [ + -0.6975152492523193 + ], + "logprob_reward": -0.6975152492523193, + "yes_no_margin_reward": 0.0, + "token_credit_reward": -0.6975152492523193, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.2222777009010315, + 0.05544637516140938, + 0.04130543768405914, + 0.1488390564918518, + 0.08273877948522568, + 0.15402571856975555, + 0.1038147583603859, + 0.11007827520370483, + 0.059975750744342804, + 0.021498192101716995 + ], + "credit_map": [ + 0.02463921159505844, + 0.012678844854235649, + 0.02552904561161995, + 0.017478564754128456, + 0.02046366222202778, + 0.04438628628849983, + 0.026118487119674683, + 0.05098358169198036, + 0.006363551132380962, + 0.0056389737874269485, + 0.010458292439579964, + 0.005538697354495525, + 0.010398097336292267, + 0.004221673123538494, + 0.009093500673770905, + 0.0037335881497710943, + 0.00358171877451241, + 0.004157590679824352, + 0.004447563551366329, + 0.004352064337581396, + 0.012588055804371834, + 0.004221606999635696, + 0.004131162539124489, + 0.0038256715051829815, + 0.014068603515625, + 0.017329147085547447, + 0.01599707640707493, + 0.021894799545407295, + 0.01706988923251629, + 0.020787682384252548, + 0.02364323101937771, + 0.018048621714115143, + 0.006535504944622517, + 0.010951654985547066, + 0.004169939085841179, + 0.016601871699094772, + 0.009380319155752659, + 0.011921797879040241, + 0.012951118871569633, + 0.010226568207144737, + 0.0061074914410710335, + 0.0041764420457184315, + 0.004955125041306019, + 0.011396553367376328, + 0.013971015810966492, + 0.05001421272754669, + 0.017282478511333466, + 0.046122413128614426, + 0.005744057707488537, + 0.007577846758067608, + 0.004622323904186487, + 0.009420914575457573, + 0.030108673498034477, + 0.01440618745982647, + 0.01929023303091526, + 0.012644519098103046, + 0.00770143186673522, + 0.011302182450890541, + 0.005608726758509874, + 0.01120531652122736, + 0.022360680624842644, + 0.015409644693136215, + 0.019515417516231537, + 0.01697487197816372, + 0.003994641359895468, + 0.008142003789544106, + 0.006022277288138866, + 0.008356270380318165, + 0.009280395694077015, + 0.007074485067278147, + 0.008530940860509872, + 0.008574738167226315, + 0.0018921296577900648, + 0.002240193309262395, + 0.002052755793556571, + 0.00177416589576751, + 0.0029237563721835613, + 0.0047795833088457584, + 0.0028545958921313286, + 0.00298101338557899 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 3 + ], + "selected_reward": -0.6975152492523193 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a police officer sprinting at full speed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a narrow, grimy alley on a rainy night with puddles on the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the officer wear a dark blue rain-soaked uniform with a tactical vest featuring flickering reflective strips?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera shot low-angle, handheld, and tracking the officer while shaking violently?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are red, blue, and green neon lights visible reflecting on wet surfaces and illuminating 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": [ + -0.005959243047982454 + ], + "opposite_logprob_tokens": [ + -5.1309590339660645 + ], + "logprob_reward": -0.005959243047982454, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005959243047982454, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16916429996490479, + 0.08075623214244843, + 0.062269922345876694, + 0.12641645967960358, + 0.09260198473930359, + 0.07352042943239212, + 0.14495469629764557, + 0.049688372761011124, + 0.1256941556930542, + 0.07493338733911514 + ], + "credit_map": [ + 0.01626167818903923, + 0.005846700165420771, + 0.02055954933166504, + 0.00700909411534667, + 0.009143195115029812, + 0.04711608961224556, + 0.011788354255259037, + 0.05143966153264046, + 0.010439489968121052, + 0.010116488672792912, + 0.012585816904902458, + 0.00860209483653307, + 0.012750519439578056, + 0.008037338964641094, + 0.010182528756558895, + 0.00804196484386921, + 0.00459320330992341, + 0.00501489732414484, + 0.004303370136767626, + 0.016432758420705795, + 0.019479528069496155, + 0.0033987765200436115, + 0.005117153748869896, + 0.0039302390068769455, + 0.018428310751914978, + 0.015336746349930763, + 0.017765074968338013, + 0.015147089026868343, + 0.011254023760557175, + 0.013991757296025753, + 0.017581023275852203, + 0.016912447288632393, + 0.014191344380378723, + 0.007789649534970522, + 0.0062653739005327225, + 0.014245889149606228, + 0.012426170520484447, + 0.014688276685774326, + 0.013598908670246601, + 0.009396378882229328, + 0.0036070419009774923, + 0.0034473019186407328, + 0.0026272167451679707, + 0.008454331196844578, + 0.006816534325480461, + 0.017675375565886497, + 0.014830278232693672, + 0.016062360256910324, + 0.011237242259085178, + 0.0221475251019001, + 0.008187603205442429, + 0.019737794995307922, + 0.02758091688156128, + 0.017232276499271393, + 0.020853465422987938, + 0.017977885901927948, + 0.003183871041983366, + 0.00953417457640171, + 0.0034078643657267094, + 0.005405896808952093, + 0.007067916449159384, + 0.0075511266477406025, + 0.006947834976017475, + 0.006589691154658794, + 0.009945377707481384, + 0.018812930211424828, + 0.018580563366413116, + 0.01631772704422474, + 0.014798400923609734, + 0.02676803059875965, + 0.009144360199570656, + 0.01132676750421524, + 0.009728113189339638, + 0.007439034525305033, + 0.009946614503860474, + 0.007654160726815462, + 0.016054127365350723, + 0.008038513362407684, + 0.009228792041540146, + 0.00684403907507658 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 3 + ], + "selected_reward": -0.005959243047982454 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a police officer sprinting at full speed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a narrow, grimy alley on a rainy night with puddles on the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the officer wear a dark blue rain-soaked uniform with a tactical vest featuring flickering reflective strips?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera shot low-angle, handheld, and tracking the officer while shaking violently?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are red, blue, and green neon lights visible reflecting on wet surfaces and illuminating 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": [ + -0.0046426597982645035 + ], + "opposite_logprob_tokens": [ + -5.379642486572266 + ], + "logprob_reward": -0.0046426597982645035, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.0046426597982645035, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2025945484638214, + 0.05934597924351692, + 0.13716381788253784, + 0.19310683012008667, + 0.06385321170091629, + 0.08633889257907867, + 0.12152139097452164, + 0.06173861399292946, + 0.031413253396749496, + 0.04292348772287369 + ], + "credit_map": [ + 0.008797929622232914, + 0.005506133660674095, + 0.00799258891493082, + 0.007385721430182457, + 0.008589663542807102, + 0.07414085417985916, + 0.011428041383624077, + 0.07875360548496246, + 0.010942311957478523, + 0.008220183663070202, + 0.006813333835452795, + 0.004262202885001898, + 0.007912629283964634, + 0.007399177644401789, + 0.009552033618092537, + 0.004244099836796522, + 0.00845203548669815, + 0.005992322228848934, + 0.00984155759215355, + 0.008124868385493755, + 0.051104817539453506, + 0.031078655272722244, + 0.012956872582435608, + 0.00961267575621605, + 0.027219008654356003, + 0.029330123215913773, + 0.02565004490315914, + 0.02250630035996437, + 0.015361649915575981, + 0.02240043878555298, + 0.027252867817878723, + 0.023386381566524506, + 0.007004374172538519, + 0.004688015673309565, + 0.0048515102826058865, + 0.011266790330410004, + 0.009410141967236996, + 0.009075744077563286, + 0.012247436679899693, + 0.005309192463755608, + 0.003447653027251363, + 0.0034874887205660343, + 0.005177371669560671, + 0.011292480863630772, + 0.008444187231361866, + 0.017986826598644257, + 0.021667448803782463, + 0.014835424721240997, + 0.0071070222184062, + 0.012257778085768223, + 0.008692784234881401, + 0.01248590275645256, + 0.03186443820595741, + 0.018393177539110184, + 0.017161546275019646, + 0.013558725826442242, + 0.003879532916471362, + 0.00693378783762455, + 0.004725942388176918, + 0.00539941992610693, + 0.01163166668266058, + 0.008760377764701843, + 0.012186788022518158, + 0.008221091702580452, + 0.003205783199518919, + 0.0034538335166871548, + 0.00489653367549181, + 0.003140705171972513, + 0.005116122774779797, + 0.00359717826358974, + 0.004398254211992025, + 0.003604837693274021, + 0.00500652426853776, + 0.005446015857160091, + 0.005609508138149977, + 0.004707092419266701, + 0.007703205104917288, + 0.00475471094250679, + 0.005524521227926016, + 0.004171904642134905 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 2 + ], + "selected_reward": -0.0046426597982645035 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a police officer sprinting at full speed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a narrow, grimy alley on a rainy night with puddles on the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the officer wear a dark blue rain-soaked uniform with a tactical vest featuring flickering reflective strips?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera shot low-angle, handheld, and tracking the officer while shaking violently?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are red, blue, and green neon lights visible reflecting on wet surfaces and illuminating 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": [ + -0.04864989593625069 + ], + "opposite_logprob_tokens": [ + -3.048649787902832 + ], + "logprob_reward": -0.04864989593625069, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04864989593625069, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12588369846343994, + 0.0652463436126709, + 0.2191266119480133, + 0.1413174569606781, + 0.051724981516599655, + 0.12088140100240707, + 0.09641019999980927, + 0.07143639028072357, + 0.057030998170375824, + 0.05094191059470177 + ], + "credit_map": [ + 0.00948005635291338, + 0.007108004298061132, + 0.0220059622079134, + 0.00933811441063881, + 0.012595056556165218, + 0.024523725733160973, + 0.013075246475636959, + 0.027757536619901657, + 0.008932658471167088, + 0.0053785708732903, + 0.010553923435509205, + 0.005300572607666254, + 0.012659003026783466, + 0.0055907247588038445, + 0.00989636592566967, + 0.006934518925845623, + 0.011041429825127125, + 0.011077913455665112, + 0.014661092311143875, + 0.04024991765618324, + 0.04771669954061508, + 0.05817742273211479, + 0.021090282127261162, + 0.015111846849322319, + 0.02475190907716751, + 0.014911788515746593, + 0.020179543644189835, + 0.015616443008184433, + 0.013651987537741661, + 0.01737150363624096, + 0.016241254284977913, + 0.018593037500977516, + 0.004638656508177519, + 0.005466753151267767, + 0.0037801191210746765, + 0.008376765064895153, + 0.007902457378804684, + 0.007075021974742413, + 0.008764167316257954, + 0.005721040535718203, + 0.004597990307956934, + 0.003462009597569704, + 0.004237108398228884, + 0.009715808555483818, + 0.009719510562717915, + 0.03742966428399086, + 0.017050204798579216, + 0.03466910868883133, + 0.004718898795545101, + 0.006149063818156719, + 0.002705618040636182, + 0.006322974804788828, + 0.032152533531188965, + 0.014392565935850143, + 0.017176320776343346, + 0.012792225927114487, + 0.00527176121249795, + 0.008175657130777836, + 0.005198013968765736, + 0.006866971030831337, + 0.01240753848105669, + 0.010749475099146366, + 0.014885283075273037, + 0.007881688885390759, + 0.005293537862598896, + 0.0064731077291071415, + 0.004926165100187063, + 0.006698851939290762, + 0.009839736856520176, + 0.01112192589789629, + 0.0058580427430570126, + 0.006819630041718483, + 0.005889215972274542, + 0.00483405776321888, + 0.00591706158593297, + 0.005447668023407459, + 0.007813128642737865, + 0.008852739818394184, + 0.006233831401914358, + 0.00595420366153121 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 3, + 0 + ], + "selected_reward": -0.04864989593625069 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a police officer sprinting at full speed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a narrow, grimy alley on a rainy night with puddles on the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the officer wear a dark blue rain-soaked uniform with a tactical vest featuring flickering reflective strips?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera shot low-angle, handheld, and tracking the officer while shaking violently?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are red, blue, and green neon lights visible reflecting on wet surfaces and illuminating 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": [ + -0.004638862796127796 + ], + "opposite_logprob_tokens": [ + -5.379638671875 + ], + "logprob_reward": -0.004638862796127796, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004638862796127796, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.26349273324012756, + 0.05927629396319389, + 0.06294641643762589, + 0.17464961111545563, + 0.05497850850224495, + 0.047551024705171585, + 0.16148881614208221, + 0.045755743980407715, + 0.08139125257730484, + 0.048469606786966324 + ], + "credit_map": [ + 0.015417765825986862, + 0.008963288739323616, + 0.016597576439380646, + 0.011166047304868698, + 0.01089076604694128, + 0.09076277166604996, + 0.014296662993729115, + 0.09539785236120224, + 0.0076393610797822475, + 0.004653507843613625, + 0.008578073233366013, + 0.005093811545521021, + 0.007378618232905865, + 0.007654556073248386, + 0.01381580252200365, + 0.0044625625014305115, + 0.0030502122826874256, + 0.0025994114112108946, + 0.004383633378893137, + 0.0041776616126298904, + 0.021680772304534912, + 0.01693163439631462, + 0.005658481270074844, + 0.004464615136384964, + 0.032809577882289886, + 0.017869343981146812, + 0.028404587879776955, + 0.017171891406178474, + 0.010964006185531616, + 0.023081433027982712, + 0.022050820291042328, + 0.022297950461506844, + 0.0038980578538030386, + 0.005093306768685579, + 0.004353041294962168, + 0.007945419289171696, + 0.008887450210750103, + 0.0104869045317173, + 0.008582470938563347, + 0.005731860641390085, + 0.001772950286976993, + 0.0017862023087218404, + 0.0023182358127087355, + 0.0052733211778104305, + 0.005107623990625143, + 0.012133371084928513, + 0.006291214842349291, + 0.01286811102181673, + 0.015298384241759777, + 0.020067378878593445, + 0.011916960589587688, + 0.017143264412879944, + 0.03146437183022499, + 0.018125684931874275, + 0.02547573484480381, + 0.02199704200029373, + 0.00338017288595438, + 0.004247695207595825, + 0.004402725026011467, + 0.005031371023505926, + 0.009055333212018013, + 0.006389915011823177, + 0.006800325121730566, + 0.006448211148381233, + 0.007412103470414877, + 0.01303866133093834, + 0.011242888867855072, + 0.010120104998350143, + 0.008622484281659126, + 0.011034871451556683, + 0.008283484727144241, + 0.01163665484637022, + 0.006850672420114279, + 0.0060573057271540165, + 0.007415750063955784, + 0.004456835333257914, + 0.008451327681541443, + 0.004736766219139099, + 0.005975867155939341, + 0.004525086376816034 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 6 + ], + "selected_reward": -0.004638862796127796 + } + ], + "vlm_decoded_outputs": [ + "No", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "dynamic, low-angle, handheld tracking shot closely follows a police officer sprinting at full speed. The scene is set in a narrow, grimy alley on a rainy night. The rain is pouring down, and the ground is filled with puddles. The officer wears a rain-soaked dark blue police uniform, covered with a tactical vest whose reflective strips flicker in the light. His short, wet black hair clings to his forehead, and his face is a mixture of rain and sweat. His expression is steely and hurried with intense concentration. Each step he takes slams against the asphalt, sending up circles of water mixed with neon reflections. The red, blue, and green neon lights from the shops at the alley's entrance intermittently cut through the darkness, painting the wet brick walls and scattered garbage with a cacophony of colors. Light and shadow shift rapidly as the camera shakes violently. His target is a blurry black figure that is barely visible at the end of the alley and almost merges into the shadows. The policeman leans forward, swings his arms vigorously, and the white breath exhaled from his mouth quickly dissipates in the cold air. The whole picture is decorated with psychedelic and chaotic neon lights against a cold background, full of the urgency of pursuit, uncertainty and suffocating tension.", + "seed": 7138, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000121/group_9/index_6916_seed_7138_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": -0.07673881398513913, + "vlm_logprob": -0.07673881398513913, + "vlm_yes_no_margin": 4.0, + "vlm_token_credit": -0.07673881398513913, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a police officer sprinting at full speed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a narrow, grimy alley on a rainy night with puddles on the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the officer wear a dark blue rain-soaked uniform with a tactical vest featuring flickering reflective strips?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera shot low-angle, handheld, and tracking the officer while shaking violently?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are red, blue, and green neon lights visible reflecting on wet surfaces and illuminating 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": [ + -0.3202138841152191 + ], + "opposite_logprob_tokens": [ + -1.3202139139175415 + ], + "logprob_reward": -0.3202138841152191, + "yes_no_margin_reward": 1.0, + "token_credit_reward": -0.3202138841152191, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07884316146373749, + 0.06663826107978821, + 0.03306012973189354, + 0.045345161110162735, + 0.37085190415382385, + 0.026053091511130333, + 0.048655878752470016, + 0.129904106259346, + 0.06281234323978424, + 0.13783590495586395 + ], + "credit_map": [ + 0.009672138839960098, + 0.010540845803916454, + 0.009599553421139717, + 0.015570552088320255, + 0.006219098344445229, + 0.007696870248764753, + 0.009454400278627872, + 0.01008970383554697, + 0.009384446777403355, + 0.003244374878704548, + 0.007338948082178831, + 0.0040537286549806595, + 0.004542254842817783, + 0.014628415927290916, + 0.007342210970818996, + 0.01610388793051243, + 0.003529773559421301, + 0.003767153713852167, + 0.005519388243556023, + 0.0037274817004799843, + 0.005000152625143528, + 0.003105396404862404, + 0.004470427520573139, + 0.003940360154956579, + 0.003298119641840458, + 0.003523891791701317, + 0.011125915683805943, + 0.003315028501674533, + 0.006010643672198057, + 0.007132347673177719, + 0.006911322008818388, + 0.00402789656072855, + 0.014141215942800045, + 0.009014887735247612, + 0.020582832396030426, + 0.022995078936219215, + 0.015639256685972214, + 0.10683641582727432, + 0.051356133073568344, + 0.1302860975265503, + 0.0035780591424554586, + 0.0017109282780438662, + 0.00318319583311677, + 0.0038869359996169806, + 0.003800653386861086, + 0.002087533939629793, + 0.0039628371596336365, + 0.0038429482374340296, + 0.004330754280090332, + 0.0031957728788256645, + 0.007151012774556875, + 0.005055633839219809, + 0.007999681867659092, + 0.007117660250514746, + 0.008105440996587276, + 0.005699926055967808, + 0.010107011534273624, + 0.009128040634095669, + 0.011202607303857803, + 0.012742148712277412, + 0.017651567235589027, + 0.019575055688619614, + 0.03037814050912857, + 0.019119536504149437, + 0.00444584246724844, + 0.004070327617228031, + 0.004619461949914694, + 0.004721830599009991, + 0.009787796065211296, + 0.01105685904622078, + 0.013051521964371204, + 0.011058712378144264, + 0.014448028989136219, + 0.013503006659448147, + 0.019195733591914177, + 0.01654992625117302, + 0.021223468706011772, + 0.0192624069750309, + 0.018899133428931236, + 0.014754215255379677 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 9, + 7 + ], + "selected_reward": -0.3202138841152191 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a police officer sprinting at full speed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a narrow, grimy alley on a rainy night with puddles on the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the officer wear a dark blue rain-soaked uniform with a tactical vest featuring flickering reflective strips?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera shot low-angle, handheld, and tracking the officer while shaking violently?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are red, blue, and green neon lights visible reflecting on wet surfaces and illuminating 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": [ + -0.005956398788839579 + ], + "opposite_logprob_tokens": [ + -5.130956172943115 + ], + "logprob_reward": -0.005956398788839579, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005956398788839579, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.32482650876045227, + 0.08085650950670242, + 0.11787983030080795, + 0.0710763931274414, + 0.07012633979320526, + 0.14842651784420013, + 0.050111304968595505, + 0.03214065730571747, + 0.04694265499711037, + 0.05761329457163811 + ], + "credit_map": [ + 0.05759814381599426, + 0.03376075252890587, + 0.04771384224295616, + 0.04951771721243858, + 0.03096543624997139, + 0.028078896924853325, + 0.03371607884764671, + 0.04347566142678261, + 0.008663038723170757, + 0.006746876053512096, + 0.009159809909760952, + 0.008063167333602905, + 0.010813918896019459, + 0.010074097663164139, + 0.014338123612105846, + 0.012997478246688843, + 0.012367009185254574, + 0.016736682504415512, + 0.019109278917312622, + 0.009923022240400314, + 0.013363041914999485, + 0.01306528877466917, + 0.013544043526053429, + 0.01977146603167057, + 0.005630169529467821, + 0.008020853623747826, + 0.013270181603729725, + 0.011732378974556923, + 0.012309419922530651, + 0.005098025780171156, + 0.010413858108222485, + 0.00460150046274066, + 0.0054429443553090096, + 0.005637072958052158, + 0.009984034113585949, + 0.00704111298546195, + 0.005368947982788086, + 0.01678423024713993, + 0.008891729637980461, + 0.010976264253258705, + 0.014187379740178585, + 0.02234750986099243, + 0.014436857774853706, + 0.022134512662887573, + 0.01108682993799448, + 0.02307499386370182, + 0.019692599773406982, + 0.021465841680765152, + 0.005224169231951237, + 0.005021125078201294, + 0.006532478146255016, + 0.006658325437456369, + 0.003879820229485631, + 0.009615939110517502, + 0.007526159752160311, + 0.005653285887092352, + 0.0029842949006706476, + 0.002529787365347147, + 0.0030076606199145317, + 0.0033237452153116465, + 0.00562194874510169, + 0.0049785044975578785, + 0.005424050148576498, + 0.0042706639505922794, + 0.002822912996634841, + 0.0034445642959326506, + 0.0034372983500361443, + 0.0024964632466435432, + 0.005295340903103352, + 0.006921017076820135, + 0.009711993858218193, + 0.012813067063689232, + 0.005054527428001165, + 0.005381458904594183, + 0.0051050614565610886, + 0.00706155551597476, + 0.009337303228676319, + 0.009816847741603851, + 0.01018532831221819, + 0.005671214312314987 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 2 + ], + "selected_reward": -0.005956398788839579 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a police officer sprinting at full speed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a narrow, grimy alley on a rainy night with puddles on the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the officer wear a dark blue rain-soaked uniform with a tactical vest featuring flickering reflective strips?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera shot low-angle, handheld, and tracking the officer while shaking violently?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are red, blue, and green neon lights visible reflecting on wet surfaces and illuminating 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": [ + -0.005258538294583559 + ], + "opposite_logprob_tokens": [ + -5.255258560180664 + ], + "logprob_reward": -0.005258538294583559, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005258538294583559, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11927429586648941, + 0.07474885135889053, + 0.055038344115018845, + 0.10340593010187149, + 0.33615800738334656, + 0.05999535322189331, + 0.03128756955265999, + 0.05138254165649414, + 0.05023711547255516, + 0.11847198009490967 + ], + "credit_map": [ + 0.014323570765554905, + 0.011134670116007328, + 0.021704575046896935, + 0.014671814627945423, + 0.010747546330094337, + 0.011211090721189976, + 0.010837490670382977, + 0.024643538519740105, + 0.011570905335247517, + 0.006272512953728437, + 0.010405143722891808, + 0.005566328298300505, + 0.004977373406291008, + 0.01336598489433527, + 0.008092631585896015, + 0.014497972093522549, + 0.005406079813838005, + 0.006858973763883114, + 0.00623510405421257, + 0.005016031209379435, + 0.012674471363425255, + 0.0057650841772556305, + 0.006914869882166386, + 0.006167732644826174, + 0.009679630398750305, + 0.01501382328569889, + 0.017269045114517212, + 0.012684695422649384, + 0.015537390485405922, + 0.011311184614896774, + 0.01394433993846178, + 0.007965821772813797, + 0.013774896040558815, + 0.010400409810245037, + 0.01749325543642044, + 0.02191697433590889, + 0.014091642573475838, + 0.09321814775466919, + 0.046353988349437714, + 0.11890866607427597, + 0.006395629141479731, + 0.006057017017155886, + 0.009797394275665283, + 0.00541093060746789, + 0.007745226379483938, + 0.00588307436555624, + 0.00972308311611414, + 0.008982999250292778, + 0.003001254750415683, + 0.0019563459791243076, + 0.0036883032880723476, + 0.0042022462002933025, + 0.003687396179884672, + 0.006150226108729839, + 0.005680492147803307, + 0.002921305363997817, + 0.0038492337334901094, + 0.003746096044778824, + 0.007680908776819706, + 0.006341231521219015, + 0.006616636645048857, + 0.007817188277840614, + 0.008528225123882294, + 0.006803018506616354, + 0.0038675242103636265, + 0.0042261467315256596, + 0.005314414855092764, + 0.004287463612854481, + 0.004585600458085537, + 0.008409325033426285, + 0.009344616904854774, + 0.010202020406723022, + 0.011557012796401978, + 0.012002875097095966, + 0.015040957368910313, + 0.010476733557879925, + 0.019389081746339798, + 0.020423799753189087, + 0.016985049471259117, + 0.012596472166478634 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 9 + ], + "selected_reward": -0.005258538294583559 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a police officer sprinting at full speed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a narrow, grimy alley on a rainy night with puddles on the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the officer wear a dark blue rain-soaked uniform with a tactical vest featuring flickering reflective strips?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera shot low-angle, handheld, and tracking the officer while shaking violently?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are red, blue, and green neon lights visible reflecting on wet surfaces and illuminating 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": [ + -0.04864819347858429 + ], + "opposite_logprob_tokens": [ + -3.0486481189727783 + ], + "logprob_reward": -0.04864819347858429, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04864819347858429, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10404397547245026, + 0.03638255223631859, + 0.06178143247961998, + 0.09318852424621582, + 0.3292919993400574, + 0.04579949378967285, + 0.028518030419945717, + 0.1485728770494461, + 0.06189645081758499, + 0.09052469581365585 + ], + "credit_map": [ + 0.014395647682249546, + 0.009786976501345634, + 0.019251132383942604, + 0.01654464192688465, + 0.009207417257130146, + 0.007334097288548946, + 0.009220479987561703, + 0.018303582444787025, + 0.0038562447298318148, + 0.0027877704706043005, + 0.005051833111792803, + 0.003813893301412463, + 0.003660049056634307, + 0.006487232632935047, + 0.005140434484928846, + 0.005585093982517719, + 0.00493762968108058, + 0.009534117765724659, + 0.00664684921503067, + 0.0055440193973481655, + 0.007011867128312588, + 0.00756384152919054, + 0.008132819086313248, + 0.012410292401909828, + 0.005801490042358637, + 0.006898685824126005, + 0.02253176085650921, + 0.009774910286068916, + 0.011294309981167316, + 0.017621232196688652, + 0.014384731650352478, + 0.004881402011960745, + 0.01044197753071785, + 0.0066025578416883945, + 0.014895853586494923, + 0.013203945010900497, + 0.010845299810171127, + 0.08920925855636597, + 0.051789842545986176, + 0.13230322301387787, + 0.0050777713768184185, + 0.004398571792989969, + 0.006698428187519312, + 0.0035842605866491795, + 0.004844351205974817, + 0.004732385743409395, + 0.007637307047843933, + 0.008826416917145252, + 0.0024181234184652567, + 0.002363236155360937, + 0.0035145545843988657, + 0.004140998236835003, + 0.003049531951546669, + 0.004561488050967455, + 0.004579061176627874, + 0.0038910366129130125, + 0.01022422220557928, + 0.010987012647092342, + 0.012329237535595894, + 0.015010258182883263, + 0.019182657822966576, + 0.023317066952586174, + 0.03520846366882324, + 0.022313954308629036, + 0.003454474499449134, + 0.004823895636945963, + 0.0033262763172388077, + 0.003539532423019409, + 0.010879801586270332, + 0.009391591884195805, + 0.012918674387037754, + 0.013562202453613281, + 0.008531593717634678, + 0.008426499553024769, + 0.009324132464826107, + 0.010223227553069592, + 0.013298599980771542, + 0.014810126274824142, + 0.016161993145942688, + 0.00974852591753006 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 7, + 0 + ], + "selected_reward": -0.04864819347858429 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a police officer sprinting at full speed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a narrow, grimy alley on a rainy night with puddles on the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the officer wear a dark blue rain-soaked uniform with a tactical vest featuring flickering reflective strips?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera shot low-angle, handheld, and tracking the officer while shaking violently?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are red, blue, and green neon lights visible reflecting on wet surfaces and illuminating 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": [ + -0.0036170552484691143 + ], + "opposite_logprob_tokens": [ + -5.628617286682129 + ], + "logprob_reward": -0.0036170552484691143, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.0036170552484691143, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0913999080657959, + 0.11734376847743988, + 0.031086532399058342, + 0.07499978691339493, + 0.4002658724784851, + 0.07225438207387924, + 0.03513330966234207, + 0.12857933342456818, + 0.02759503573179245, + 0.021342100575566292 + ], + "credit_map": [ + 0.013288344256579876, + 0.013354226015508175, + 0.014286658726632595, + 0.009747907519340515, + 0.008821469731628895, + 0.007272376213222742, + 0.009811529889702797, + 0.014817389659583569, + 0.012535087764263153, + 0.0128022450953722, + 0.01284685730934143, + 0.011962697841227055, + 0.017045550048351288, + 0.015179311856627464, + 0.01711435243487358, + 0.017857657745480537, + 0.004484105389565229, + 0.0034238314256072044, + 0.005232516210526228, + 0.0027807008009403944, + 0.003084859112277627, + 0.003245911793783307, + 0.0034464115742594004, + 0.005388194229453802, + 0.004846012685447931, + 0.0036929831840097904, + 0.025207003578543663, + 0.0051877754740417, + 0.005228129681199789, + 0.014322344213724136, + 0.013535450212657452, + 0.002980082528665662, + 0.013089640066027641, + 0.0075150663033127785, + 0.01583104394376278, + 0.014426556415855885, + 0.011822341941297054, + 0.11333955079317093, + 0.06301292777061462, + 0.1612287312746048, + 0.008407923392951488, + 0.011404124088585377, + 0.009152607060968876, + 0.008818984031677246, + 0.0061432975344359875, + 0.01020362414419651, + 0.008683527819812298, + 0.009440289810299873, + 0.003635056084021926, + 0.002469779225066304, + 0.0069037675857543945, + 0.0044379462487995625, + 0.004360630642622709, + 0.004203533288091421, + 0.005127474199980497, + 0.003995121456682682, + 0.008351687341928482, + 0.008245985954999924, + 0.010007273405790329, + 0.011966952122747898, + 0.014749303460121155, + 0.021016685292124748, + 0.03304032236337662, + 0.021201100200414658, + 0.002292115241289139, + 0.0020215159747749567, + 0.003014028538018465, + 0.0016293672379106283, + 0.0033303096424788237, + 0.0037258549127727747, + 0.00509891752153635, + 0.006482925731688738, + 0.0016093620797619224, + 0.0020963193383067846, + 0.0017897217767313123, + 0.0018160154577344656, + 0.0038747733924537897, + 0.0040600900538265705, + 0.0035446584224700928, + 0.002551159355789423 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 7, + 1 + ], + "selected_reward": -0.0036170552484691143 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "dynamic, low-angle, handheld tracking shot closely follows a police officer sprinting at full speed. The scene is set in a narrow, grimy alley on a rainy night. The rain is pouring down, and the ground is filled with puddles. The officer wears a rain-soaked dark blue police uniform, covered with a tactical vest whose reflective strips flicker in the light. His short, wet black hair clings to his forehead, and his face is a mixture of rain and sweat. His expression is steely and hurried with intense concentration. Each step he takes slams against the asphalt, sending up circles of water mixed with neon reflections. The red, blue, and green neon lights from the shops at the alley's entrance intermittently cut through the darkness, painting the wet brick walls and scattered garbage with a cacophony of colors. Light and shadow shift rapidly as the camera shakes violently. His target is a blurry black figure that is barely visible at the end of the alley and almost merges into the shadows. The policeman leans forward, swings his arms vigorously, and the white breath exhaled from his mouth quickly dissipates in the cold air. The whole picture is decorated with psychedelic and chaotic neon lights against a cold background, full of the urgency of pursuit, uncertainty and suffocating tension.", + "seed": 7139, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000121/group_9/index_6916_seed_7139_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": -0.039064737502485514, + "vlm_logprob": -0.039064737502485514, + "vlm_yes_no_margin": 4.650000023841858, + "vlm_token_credit": -0.039064737502485514, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a police officer sprinting at full speed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a narrow, grimy alley on a rainy night with puddles on the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the officer wear a dark blue rain-soaked uniform with a tactical vest featuring flickering reflective strips?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera shot low-angle, handheld, and tracking the officer while shaking violently?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are red, blue, and green neon lights visible reflecting on wet surfaces and illuminating 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": [ + -0.14943541586399078 + ], + "opposite_logprob_tokens": [ + -2.024435520172119 + ], + "logprob_reward": -0.14943541586399078, + "yes_no_margin_reward": 1.8750001192092896, + "token_credit_reward": -0.14943541586399078, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22151754796504974, + 0.13269175589084625, + 0.08915990591049194, + 0.11200431734323502, + 0.062335092574357986, + 0.08309462666511536, + 0.07100129127502441, + 0.03558618202805519, + 0.12985332310199738, + 0.06275596469640732 + ], + "credit_map": [ + 0.03930450975894928, + 0.01638161577284336, + 0.051746390759944916, + 0.02879149094223976, + 0.017555827274918556, + 0.020455878227949142, + 0.023936927318572998, + 0.02334490418434143, + 0.015917155891656876, + 0.014049878343939781, + 0.01594408228993416, + 0.015437250025570393, + 0.021176347509026527, + 0.019465146586298943, + 0.016562551259994507, + 0.014139344915747643, + 0.00877245981246233, + 0.008591612800955772, + 0.009722456336021423, + 0.008847932331264019, + 0.016002953052520752, + 0.005797476042062044, + 0.02179231122136116, + 0.009632700122892857, + 0.009774895384907722, + 0.010427156463265419, + 0.013999473303556442, + 0.014868099242448807, + 0.026139643043279648, + 0.010052105411887169, + 0.01663203537464142, + 0.010110907256603241, + 0.00864432193338871, + 0.008139005862176418, + 0.00878807995468378, + 0.009170041419565678, + 0.008932880125939846, + 0.0041031427681446075, + 0.009283510036766529, + 0.005274111870676279, + 0.0074448673985898495, + 0.003388972720131278, + 0.0070824106223881245, + 0.006563324946910143, + 0.004326045513153076, + 0.008405430242419243, + 0.019722528755664825, + 0.026161042973399162, + 0.004346164409071207, + 0.008464078418910503, + 0.005407792050391436, + 0.007541419938206673, + 0.012295646592974663, + 0.008203395642340183, + 0.013650214299559593, + 0.011092578992247581, + 0.005290667992085218, + 0.0029070645105093718, + 0.0034231289755553007, + 0.004409089218825102, + 0.00452172476798296, + 0.004021101631224155, + 0.005548050161451101, + 0.005465356633067131, + 0.015283922664821148, + 0.013067744672298431, + 0.009382974356412888, + 0.008991848677396774, + 0.010173069313168526, + 0.02573258988559246, + 0.019158102571964264, + 0.02806307002902031, + 0.009922774508595467, + 0.008424480445683002, + 0.008508985862135887, + 0.007810363080352545, + 0.005980213638395071, + 0.005403861869126558, + 0.010108685120940208, + 0.006596597842872143 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 8 + ], + "selected_reward": -0.14943541586399078 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a police officer sprinting at full speed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a narrow, grimy alley on a rainy night with puddles on the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the officer wear a dark blue rain-soaked uniform with a tactical vest featuring flickering reflective strips?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera shot low-angle, handheld, and tracking the officer while shaking violently?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are red, blue, and green neon lights visible reflecting on wet surfaces and illuminating 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": [ + -0.003625726094469428 + ], + "opposite_logprob_tokens": [ + -5.628625869750977 + ], + "logprob_reward": -0.003625726094469428, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.003625726094469428, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11415749043226242, + 0.2988981604576111, + 0.040099162608385086, + 0.0600130669772625, + 0.035624679177999496, + 0.060194142162799835, + 0.13138185441493988, + 0.15754002332687378, + 0.05852324888110161, + 0.043568205088377 + ], + "credit_map": [ + 0.01552809402346611, + 0.008398863486945629, + 0.021133750677108765, + 0.014015200547873974, + 0.01053666789084673, + 0.013964633457362652, + 0.016518965363502502, + 0.01406130287796259, + 0.04427672550082207, + 0.027835046872496605, + 0.033890239894390106, + 0.04026470333337784, + 0.05268935114145279, + 0.03668605536222458, + 0.036126285791397095, + 0.027129752561450005, + 0.005091333296149969, + 0.004271109122782946, + 0.004532619845122099, + 0.003943932708352804, + 0.0049412851221859455, + 0.0032057457137852907, + 0.007082516327500343, + 0.00703061930835247, + 0.007507180329412222, + 0.0068554044701159, + 0.007813851349055767, + 0.006103934720158577, + 0.009661559015512466, + 0.006897388957440853, + 0.009125438518822193, + 0.006048306357115507, + 0.004841894377022982, + 0.0059045166708528996, + 0.005461582913994789, + 0.0037485279608517885, + 0.003934262320399284, + 0.002300348598510027, + 0.003966021351516247, + 0.005467523820698261, + 0.00686772633343935, + 0.0035780256148427725, + 0.005469154100865126, + 0.0036732794251292944, + 0.0040247393772006035, + 0.007973135448992252, + 0.011895927600562572, + 0.01671215146780014, + 0.010892003774642944, + 0.012015034444630146, + 0.01859908364713192, + 0.009317503310739994, + 0.02203941158950329, + 0.02052966132760048, + 0.01862635836005211, + 0.019362784922122955, + 0.023284796625375748, + 0.012003814801573753, + 0.017124738544225693, + 0.016208287328481674, + 0.0179806686937809, + 0.022992605343461037, + 0.025123240426182747, + 0.02282184548676014, + 0.006553724408149719, + 0.003960590809583664, + 0.00567803205922246, + 0.004435096867382526, + 0.0062721907161176205, + 0.010261067189276218, + 0.008576418273150921, + 0.01278612669557333, + 0.006567086558789015, + 0.004734657239168882, + 0.006554913241416216, + 0.004374164622277021, + 0.003609479172155261, + 0.0043508014641702175, + 0.007275478448718786, + 0.006101623643189669 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 7, + 6 + ], + "selected_reward": -0.003625726094469428 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a police officer sprinting at full speed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a narrow, grimy alley on a rainy night with puddles on the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the officer wear a dark blue rain-soaked uniform with a tactical vest featuring flickering reflective strips?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera shot low-angle, handheld, and tracking the officer while shaking violently?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are red, blue, and green neon lights visible reflecting on wet surfaces and illuminating 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": [ + -0.0028279335238039494 + ], + "opposite_logprob_tokens": [ + -5.877828121185303 + ], + "logprob_reward": -0.0028279335238039494, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.0028279335238039494, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18239136040210724, + 0.3117775619029999, + 0.034658655524253845, + 0.0638388991355896, + 0.03734108805656433, + 0.058376915752887726, + 0.10806567221879959, + 0.08521921187639236, + 0.04896669089794159, + 0.06936399638652802 + ], + "credit_map": [ + 0.026133062317967415, + 0.013262095861136913, + 0.029571188613772392, + 0.02298082783818245, + 0.01712942123413086, + 0.01205326709896326, + 0.043058183044195175, + 0.018203310668468475, + 0.03978116810321808, + 0.03306432068347931, + 0.03699761629104614, + 0.0358605831861496, + 0.050345566123723984, + 0.04758952930569649, + 0.033103324472904205, + 0.03503544256091118, + 0.006422489415854216, + 0.004245743155479431, + 0.003922062460333109, + 0.0038781107868999243, + 0.003789733164012432, + 0.002787025412544608, + 0.005035742186009884, + 0.004577746149152517, + 0.005725707393139601, + 0.005446676164865494, + 0.008875171653926373, + 0.00845060870051384, + 0.008639899082481861, + 0.006400901358574629, + 0.011800459586083889, + 0.008499476127326488, + 0.005221856292337179, + 0.0038603427819907665, + 0.006145038641989231, + 0.005098193883895874, + 0.004294733516871929, + 0.0029992247000336647, + 0.005428317468613386, + 0.004293378442525864, + 0.007525509223341942, + 0.007886243052780628, + 0.007010071538388729, + 0.004736292641609907, + 0.005382617004215717, + 0.007732443977147341, + 0.007764098234474659, + 0.010339641012251377, + 0.013271501287817955, + 0.013101131655275822, + 0.01006408967077732, + 0.009610505774617195, + 0.01481502316892147, + 0.014505300670862198, + 0.014665822498500347, + 0.018032308667898178, + 0.003504903521388769, + 0.00615356070920825, + 0.005833123344928026, + 0.006437620148062706, + 0.013510280288755894, + 0.02317786030471325, + 0.012264356948435307, + 0.014337503351271152, + 0.004254499450325966, + 0.003644446609541774, + 0.00413035461679101, + 0.003958956338465214, + 0.005113479681313038, + 0.00918519776314497, + 0.007813257165253162, + 0.010866500437259674, + 0.008570089936256409, + 0.009859438054263592, + 0.0071258507668972015, + 0.005985652096569538, + 0.008189910091459751, + 0.007044622674584389, + 0.010606970638036728, + 0.011981463991105556 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 6 + ], + "selected_reward": -0.0028279335238039494 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a police officer sprinting at full speed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a narrow, grimy alley on a rainy night with puddles on the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the officer wear a dark blue rain-soaked uniform with a tactical vest featuring flickering reflective strips?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera shot low-angle, handheld, and tracking the officer while shaking violently?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are red, blue, and green neon lights visible reflecting on wet surfaces and illuminating 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": [ + -0.03809512034058571 + ], + "opposite_logprob_tokens": [ + -3.288095235824585 + ], + "logprob_reward": -0.03809512034058571, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03809512034058571, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11331149935722351, + 0.043943632394075394, + 0.04043733328580856, + 0.056542862206697464, + 0.08746294677257538, + 0.042106956243515015, + 0.07110428810119629, + 0.1631627380847931, + 0.34968075156211853, + 0.03224702924489975 + ], + "credit_map": [ + 0.013574576005339622, + 0.0059137181378901005, + 0.013781890273094177, + 0.014903093688189983, + 0.011375736445188522, + 0.012254259549081326, + 0.029819482937455177, + 0.011688734404742718, + 0.005894738249480724, + 0.005352227948606014, + 0.005007209721952677, + 0.008018326945602894, + 0.004813116975128651, + 0.0038418301846832037, + 0.006807241588830948, + 0.004208933562040329, + 0.003994097001850605, + 0.002960626734420657, + 0.005094848573207855, + 0.003665125695988536, + 0.00688143540173769, + 0.002820489462465048, + 0.010919850319623947, + 0.004100858699530363, + 0.006119807716459036, + 0.004937850404530764, + 0.007580416742712259, + 0.006891012657433748, + 0.009816998615860939, + 0.006531809456646442, + 0.008073280565440655, + 0.006591679062694311, + 0.01642150804400444, + 0.007907931692898273, + 0.012608693912625313, + 0.0101891178637743, + 0.011183250695466995, + 0.006281774491071701, + 0.010569901205599308, + 0.012300761416554451, + 0.00459128525108099, + 0.002465037163347006, + 0.0038556051440536976, + 0.0024977903813123703, + 0.0027199690230190754, + 0.007224537432193756, + 0.00788030307739973, + 0.010872420854866505, + 0.005247930996119976, + 0.006295205559581518, + 0.007616458926349878, + 0.007252383045852184, + 0.013048184104263783, + 0.010715619660913944, + 0.010369050316512585, + 0.010559441521763802, + 0.014663850888609886, + 0.014405816793441772, + 0.018978793174028397, + 0.02084968611598015, + 0.01923428848385811, + 0.026000438258051872, + 0.026441728696227074, + 0.022588111460208893, + 0.06553546339273453, + 0.026556197553873062, + 0.025732362642884254, + 0.020774375647306442, + 0.02821814827620983, + 0.067552849650383, + 0.04065483435988426, + 0.07465649396181107, + 0.0031123647931963205, + 0.0029381881467998028, + 0.0030462457798421383, + 0.002861279295757413, + 0.002609474351629615, + 0.003775151912122965, + 0.005391913000494242, + 0.00851240660995245 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 0 + ], + "selected_reward": -0.03809512034058571 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a police officer sprinting at full speed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a narrow, grimy alley on a rainy night with puddles on the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the officer wear a dark blue rain-soaked uniform with a tactical vest featuring flickering reflective strips?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera shot low-angle, handheld, and tracking the officer while shaking violently?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are red, blue, and green neon lights visible reflecting on wet surfaces and illuminating 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": [ + -0.0013394916895776987 + ], + "opposite_logprob_tokens": [ + -6.626339435577393 + ], + "logprob_reward": -0.0013394916895776987, + "yes_no_margin_reward": 6.625, + "token_credit_reward": -0.0013394916895776987, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12090139836072922, + 0.07765831798315048, + 0.10585762560367584, + 0.056332092732191086, + 0.06843768805265427, + 0.12795677781105042, + 0.1952400654554367, + 0.09881365299224854, + 0.07830436527729034, + 0.07049798965454102 + ], + "credit_map": [ + 0.02407432161271572, + 0.008412806317210197, + 0.023139357566833496, + 0.015472348779439926, + 0.008728825487196445, + 0.01126995962113142, + 0.014816902577877045, + 0.014986876398324966, + 0.01370148453861475, + 0.008481558412313461, + 0.0085629653185606, + 0.008900639601051807, + 0.012980212457478046, + 0.007425523828715086, + 0.009764747694134712, + 0.007841185666620731, + 0.02473677136003971, + 0.01687467284500599, + 0.013296808116137981, + 0.010282817296683788, + 0.007781817112118006, + 0.008147881366312504, + 0.012533551082015038, + 0.012203298509120941, + 0.007640411611646414, + 0.006576096173375845, + 0.008151322603225708, + 0.005987306125462055, + 0.005643072538077831, + 0.006651538424193859, + 0.010521525517106056, + 0.005160817410796881, + 0.00737666105851531, + 0.00729574728757143, + 0.011081215925514698, + 0.01014722604304552, + 0.012652493081986904, + 0.0051259552128612995, + 0.00855941791087389, + 0.006198969669640064, + 0.022534465417265892, + 0.015795692801475525, + 0.020482735708355904, + 0.008966648951172829, + 0.008819260634481907, + 0.026158051565289497, + 0.012108598835766315, + 0.013091325759887695, + 0.018031416460871696, + 0.01926564984023571, + 0.0281195268034935, + 0.016158897429704666, + 0.02851770631968975, + 0.030086366459727287, + 0.026033619418740273, + 0.029026875272393227, + 0.005092571023851633, + 0.008366216905415058, + 0.010665923357009888, + 0.00865727849304676, + 0.00942879356443882, + 0.02349955216050148, + 0.01739398017525673, + 0.015709344297647476, + 0.010818039067089558, + 0.007432996761053801, + 0.006230206694453955, + 0.005413329228758812, + 0.00819737184792757, + 0.014344013296067715, + 0.009236437268555164, + 0.016631970182061195, + 0.009789107367396355, + 0.008585678413510323, + 0.00856412760913372, + 0.006352643482387066, + 0.006812873296439648, + 0.00798041932284832, + 0.010422617197036743, + 0.011990523897111416 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 5, + 0 + ], + "selected_reward": -0.0013394916895776987 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "dynamic, low-angle, handheld tracking shot closely follows a police officer sprinting at full speed. The scene is set in a narrow, grimy alley on a rainy night. The rain is pouring down, and the ground is filled with puddles. The officer wears a rain-soaked dark blue police uniform, covered with a tactical vest whose reflective strips flicker in the light. His short, wet black hair clings to his forehead, and his face is a mixture of rain and sweat. His expression is steely and hurried with intense concentration. Each step he takes slams against the asphalt, sending up circles of water mixed with neon reflections. The red, blue, and green neon lights from the shops at the alley's entrance intermittently cut through the darkness, painting the wet brick walls and scattered garbage with a cacophony of colors. Light and shadow shift rapidly as the camera shakes violently. His target is a blurry black figure that is barely visible at the end of the alley and almost merges into the shadows. The policeman leans forward, swings his arms vigorously, and the white breath exhaled from his mouth quickly dissipates in the cold air. The whole picture is decorated with psychedelic and chaotic neon lights against a cold background, full of the urgency of pursuit, uncertainty and suffocating tension.", + "seed": 7140, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000121/group_9/index_6916_seed_7140_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": -0.16588298233691604, + "vlm_logprob": -0.16588298233691604, + "vlm_yes_no_margin": 3.975, + "vlm_token_credit": -0.16588298233691604, + "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 police officer sprinting at full speed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a narrow, grimy alley on a rainy night with puddles on the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the officer wear a dark blue rain-soaked uniform with a tactical vest featuring flickering reflective strips?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera shot low-angle, handheld, and tracking the officer while shaking violently?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are red, blue, and green neon lights visible reflecting on wet surfaces and illuminating the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.7676770091056824 + ], + "opposite_logprob_tokens": [ + -0.6426770091056824 + ], + "logprob_reward": -0.7676770091056824, + "yes_no_margin_reward": -0.125, + "token_credit_reward": -0.7676770091056824, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.2807607650756836, + 0.04438457638025284, + 0.03410809487104416, + 0.14102748036384583, + 0.09299574047327042, + 0.14374440908432007, + 0.07532324641942978, + 0.07469480484724045, + 0.06460075825452805, + 0.048360057175159454 + ], + "credit_map": [ + 0.027361232787370682, + 0.02163686417043209, + 0.02301803044974804, + 0.03274596855044365, + 0.042744919657707214, + 0.04240405559539795, + 0.05273249000310898, + 0.038117218762636185, + 0.005197515245527029, + 0.005160961300134659, + 0.005565469618886709, + 0.004975372459739447, + 0.0046621570363640785, + 0.006772727705538273, + 0.006260785274207592, + 0.005789587274193764, + 0.00471810856834054, + 0.0028410982340574265, + 0.005494336131960154, + 0.004363405518233776, + 0.004121503327041864, + 0.0043060630559921265, + 0.004665440879762173, + 0.0035981396213173866, + 0.019146211445331573, + 0.010315843857824802, + 0.027496542781591415, + 0.01696033403277397, + 0.019796306267380714, + 0.014691110700368881, + 0.011879588477313519, + 0.020741542801260948, + 0.005763423629105091, + 0.004349746275693178, + 0.008857340551912785, + 0.01513858512043953, + 0.01613573171198368, + 0.005022503901273012, + 0.02882792055606842, + 0.008900489658117294, + 0.010681846179068089, + 0.007418546825647354, + 0.009779609739780426, + 0.017250457778573036, + 0.0163927860558033, + 0.024714389815926552, + 0.04667048156261444, + 0.010836290195584297, + 0.00514235720038414, + 0.006941322237253189, + 0.010505350306630135, + 0.009980984032154083, + 0.01668916828930378, + 0.010100116953253746, + 0.008691636845469475, + 0.007272311020642519, + 0.005364210344851017, + 0.009762365370988846, + 0.008052625693380833, + 0.010040891356766224, + 0.019236991181969643, + 0.00517783360555768, + 0.01222867239266634, + 0.004831210244446993, + 0.00579017074778676, + 0.005301152355968952, + 0.0064513483084738255, + 0.00891940202564001, + 0.010458688251674175, + 0.008596166968345642, + 0.011357207782566547, + 0.007726623211055994, + 0.0023614661768078804, + 0.005118355620652437, + 0.003310315078124404, + 0.004095227923244238, + 0.006121082231402397, + 0.010863923467695713, + 0.0072536105290055275, + 0.009236079640686512 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 3 + ], + "selected_reward": -0.7676770091056824 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a police officer sprinting at full speed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a narrow, grimy alley on a rainy night with puddles on the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the officer wear a dark blue rain-soaked uniform with a tactical vest featuring flickering reflective strips?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera shot low-angle, handheld, and tracking the officer while shaking violently?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are red, blue, and green neon lights visible reflecting on wet surfaces and illuminating 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": [ + -0.008653759025037289 + ], + "opposite_logprob_tokens": [ + -4.75865364074707 + ], + "logprob_reward": -0.008653759025037289, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008653759025037289, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.35409271717071533, + 0.12437403202056885, + 0.05197696387767792, + 0.11271309107542038, + 0.0551522895693779, + 0.0812491923570633, + 0.08813773095607758, + 0.02246539294719696, + 0.06871727854013443, + 0.04112131893634796 + ], + "credit_map": [ + 0.034848302602767944, + 0.042054418474435806, + 0.039198003709316254, + 0.03292085975408554, + 0.048672791570425034, + 0.056719668209552765, + 0.0610300675034523, + 0.038648612797260284, + 0.01725560612976551, + 0.011820003390312195, + 0.01528793852776289, + 0.015116141177713871, + 0.014600876718759537, + 0.019497036933898926, + 0.01418081670999527, + 0.0166156105697155, + 0.006237468216568232, + 0.00564922858029604, + 0.006511786952614784, + 0.008049728348851204, + 0.007881192490458488, + 0.008024037815630436, + 0.006035218480974436, + 0.0035883060190826654, + 0.015810629352927208, + 0.011408322490751743, + 0.02310504950582981, + 0.011162929236888885, + 0.014582966454327106, + 0.01760830543935299, + 0.008157228119671345, + 0.010877677239477634, + 0.0039018408861011267, + 0.004127644933760166, + 0.007726936135441065, + 0.009305672720074654, + 0.005920250900089741, + 0.0038876794278621674, + 0.014429737813770771, + 0.0058525316417217255, + 0.012782823294401169, + 0.007809333968907595, + 0.01827097125351429, + 0.00781874917447567, + 0.006755067501217127, + 0.0117596834897995, + 0.010540323331952095, + 0.0055122533813118935, + 0.01042380090802908, + 0.00609921757131815, + 0.009479307569563389, + 0.009131252765655518, + 0.009718229062855244, + 0.01393011212348938, + 0.007159117143601179, + 0.022196704521775246, + 0.0035051917657256126, + 0.002927189925685525, + 0.0038669342175126076, + 0.0019759321585297585, + 0.0027697915211319923, + 0.0025800399016588926, + 0.0030467244796454906, + 0.0017935902578756213, + 0.006467570550739765, + 0.007441803812980652, + 0.008993926458060741, + 0.00858427956700325, + 0.006832668092101812, + 0.010016211308538914, + 0.012118963524699211, + 0.008261863142251968, + 0.0035375256557017565, + 0.0033732042647898197, + 0.0031908440869301558, + 0.002905789064243436, + 0.003936071414500475, + 0.01034395582973957, + 0.006462678778916597, + 0.007371252402663231 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 3 + ], + "selected_reward": -0.008653759025037289 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a police officer sprinting at full speed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a narrow, grimy alley on a rainy night with puddles on the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the officer wear a dark blue rain-soaked uniform with a tactical vest featuring flickering reflective strips?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera shot low-angle, handheld, and tracking the officer while shaking violently?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are red, blue, and green neon lights visible reflecting on wet surfaces and illuminating 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": [ + -0.0024953915271908045 + ], + "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": [ + 0.17989224195480347, + 0.08425389230251312, + 0.03849749267101288, + 0.32537463307380676, + 0.06240561231970787, + 0.06593038886785507, + 0.051757123321294785, + 0.05881256237626076, + 0.07678558677434921, + 0.056290436536073685 + ], + "credit_map": [ + 0.01512196846306324, + 0.011659163981676102, + 0.018830589950084686, + 0.019914895296096802, + 0.020180679857730865, + 0.036776844412088394, + 0.03340904042124748, + 0.02399907074868679, + 0.009387604892253876, + 0.004593671765178442, + 0.009432870894670486, + 0.009607195854187012, + 0.00476233521476388, + 0.013730368576943874, + 0.021063750609755516, + 0.011676104739308357, + 0.005493795964866877, + 0.0030427288729697466, + 0.005312354303896427, + 0.004233955405652523, + 0.005109523423016071, + 0.006675287149846554, + 0.004891166929155588, + 0.003738685045391321, + 0.037125226110219955, + 0.029991338029503822, + 0.04020468890666962, + 0.038592252880334854, + 0.05233131721615791, + 0.0337199792265892, + 0.027133364230394363, + 0.06627649813890457, + 0.00500207394361496, + 0.003944290801882744, + 0.005889344494789839, + 0.008464272134006023, + 0.018822019919753075, + 0.004522974602878094, + 0.007501250132918358, + 0.008259386755526066, + 0.00812885444611311, + 0.007308195345103741, + 0.0111285001039505, + 0.006472881883382797, + 0.006089063361287117, + 0.00908531341701746, + 0.009428570978343487, + 0.008289018645882607, + 0.002905532717704773, + 0.004282960668206215, + 0.005192146636545658, + 0.00656237592920661, + 0.010072089731693268, + 0.007747695315629244, + 0.006477317772805691, + 0.008517011068761349, + 0.004581431858241558, + 0.008199850097298622, + 0.008112186565995216, + 0.0069570173509418964, + 0.012775626964867115, + 0.0045507680624723434, + 0.009767608717083931, + 0.0038680739235132933, + 0.008022754453122616, + 0.00796943623572588, + 0.00917868036776781, + 0.010092229582369328, + 0.009294598363339901, + 0.011585267260670662, + 0.012292403727769852, + 0.008350225165486336, + 0.0068884240463376045, + 0.007944808341562748, + 0.006678068544715643, + 0.005850436631590128, + 0.004080584738403559, + 0.005469109863042831, + 0.010019185021519661, + 0.009359820745885372 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 1 + ], + "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 main subject a police officer sprinting at full speed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a narrow, grimy alley on a rainy night with puddles on the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the officer wear a dark blue rain-soaked uniform with a tactical vest featuring flickering reflective strips?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera shot low-angle, handheld, and tracking the officer while shaking violently?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are red, blue, and green neon lights visible reflecting on wet surfaces and illuminating 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": [ + -0.04864716902375221 + ], + "opposite_logprob_tokens": [ + -3.048647165298462 + ], + "logprob_reward": -0.04864716902375221, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04864716902375221, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19394995272159576, + 0.1258997917175293, + 0.06602050364017487, + 0.12029282748699188, + 0.06550174206495285, + 0.08369775861501694, + 0.06419359147548676, + 0.08787793666124344, + 0.10492563992738724, + 0.08764021843671799 + ], + "credit_map": [ + 0.030638132244348526, + 0.012753461487591267, + 0.02415883168578148, + 0.027541430667042732, + 0.019111258909106255, + 0.028815284371376038, + 0.02135927975177765, + 0.02957228384912014, + 0.01177230104804039, + 0.010696476325392723, + 0.013431419618427753, + 0.015208940021693707, + 0.018427420407533646, + 0.025607025250792503, + 0.01454237662255764, + 0.016213836148381233, + 0.00911675300449133, + 0.005422596819698811, + 0.009572296403348446, + 0.010120191611349583, + 0.010239859111607075, + 0.006900280714035034, + 0.007230908609926701, + 0.007417620625346899, + 0.019503027200698853, + 0.00864282064139843, + 0.030905526131391525, + 0.013532858341932297, + 0.017145482823252678, + 0.008725104853510857, + 0.008694377727806568, + 0.013143629767000675, + 0.005262298509478569, + 0.004205699544399977, + 0.007653975393623114, + 0.008908669464290142, + 0.014761009253561497, + 0.003775986609980464, + 0.013436300680041313, + 0.0074978056363761425, + 0.006485369987785816, + 0.004182861186563969, + 0.006579411681741476, + 0.012208052910864353, + 0.009423293173313141, + 0.014716004952788353, + 0.02436114102602005, + 0.005741622764617205, + 0.009193621575832367, + 0.004718515556305647, + 0.010397344827651978, + 0.00649985671043396, + 0.007165791932493448, + 0.008692102506756783, + 0.007834231480956078, + 0.009692124091088772, + 0.007831515744328499, + 0.01048372033983469, + 0.010058209300041199, + 0.012053043581545353, + 0.017044804990291595, + 0.011428466998040676, + 0.012348500080406666, + 0.006629672832787037, + 0.012212153524160385, + 0.008359754458069801, + 0.014593487605452538, + 0.00900338962674141, + 0.015327736735343933, + 0.016604505479335785, + 0.017389768734574318, + 0.011434837244451046, + 0.006192834582179785, + 0.012870985083281994, + 0.008709574118256569, + 0.0085576381534338, + 0.009309950284659863, + 0.008699024096131325, + 0.01662500761449337, + 0.016675200313329697 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 3 + ], + "selected_reward": -0.04864716902375221 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a police officer sprinting at full speed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a narrow, grimy alley on a rainy night with puddles on the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the officer wear a dark blue rain-soaked uniform with a tactical vest featuring flickering reflective strips?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera shot low-angle, handheld, and tracking the officer while shaking violently?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are red, blue, and green neon lights visible reflecting on wet surfaces and illuminating 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": [ + -0.0019415830029174685 + ], + "opposite_logprob_tokens": [ + -6.251941680908203 + ], + "logprob_reward": -0.0019415830029174685, + "yes_no_margin_reward": 6.25, + "token_credit_reward": -0.0019415830029174685, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.197605699300766, + 0.05870644003152847, + 0.05633221194148064, + 0.11320476979017258, + 0.08474939316511154, + 0.21188302338123322, + 0.050234850496053696, + 0.04981224611401558, + 0.09296926856040955, + 0.08450212329626083 + ], + "credit_map": [ + 0.019582495093345642, + 0.013923726975917816, + 0.020537687465548515, + 0.023616204038262367, + 0.020959867164492607, + 0.03159516677260399, + 0.025989241898059845, + 0.041401296854019165, + 0.005632662680000067, + 0.004108625929802656, + 0.006949265953153372, + 0.006010436452925205, + 0.005356860812753439, + 0.013480243273079395, + 0.009033427573740482, + 0.008134908974170685, + 0.008302845060825348, + 0.006197676528245211, + 0.009128131903707981, + 0.00810563750565052, + 0.006711967755109072, + 0.006934937555342913, + 0.006277319043874741, + 0.00467369519174099, + 0.011942297220230103, + 0.007325956132262945, + 0.023457616567611694, + 0.012416518293321133, + 0.01722916029393673, + 0.012271558865904808, + 0.010177897289395332, + 0.018383758142590523, + 0.006200495176017284, + 0.004530077334493399, + 0.00891728512942791, + 0.014618287794291973, + 0.023234089836478233, + 0.005719601176679134, + 0.011483757756650448, + 0.010045794770121574, + 0.03815050423145294, + 0.022101998329162598, + 0.0371602363884449, + 0.021153472363948822, + 0.025365184992551804, + 0.03040366992354393, + 0.019630732014775276, + 0.017917204648256302, + 0.00454063480719924, + 0.006413221824914217, + 0.005248952656984329, + 0.006218315102159977, + 0.009715138003230095, + 0.006568796932697296, + 0.004943135194480419, + 0.006586651783436537, + 0.005542837083339691, + 0.006702267099171877, + 0.00719909742474556, + 0.005858109798282385, + 0.007387471850961447, + 0.006212333682924509, + 0.006676558870822191, + 0.00423356844112277, + 0.009465980343520641, + 0.009056808426976204, + 0.011105185374617577, + 0.012661546468734741, + 0.01380160916596651, + 0.011086063459515572, + 0.016332048922777176, + 0.009460018947720528, + 0.0068673607893288136, + 0.005817013327032328, + 0.00644491333514452, + 0.004602497443556786, + 0.010690653696656227, + 0.020811866968870163, + 0.015101855620741844, + 0.014165951870381832 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 3 + ], + "selected_reward": -0.0019415830029174685 + } + ], + "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/0000151/group_8/index_tensor([4038, 4038])_group_8_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000151/group_8/index_tensor([4038, 4038])_group_8_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..cdeef43e44ea99348efbbd634a34fe6e6fbae44f --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000151/group_8/index_tensor([4038, 4038])_group_8_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "smooth, uninterrupted medium-long shot begins, focusing on a handsome, scaly-faced snake prince. With graceful, silent grace, he meanders along an ancient, rain-soaked stone path. He possesses a sharp jawline and a prominent nose. His long, jet-black hair sways gently with his movements, a few damp strands plastering his forehead. From his cheekbones to his neck, a dense, shimmering emerald-green scales shimmer in the hazy light, reflecting an ever-shifting shimmer. He clads a dark green silk robe, slightly slit at the collar. The soft fabric clings to his muscular torso, rising and falling rhythmically as his powerful serpentine tail glides along the ground in a fluid S-curve. The camera steadily and deliberately approaches him. The background is a narrow alleyway lined with quaint wooden buildings. Red paper lanterns hanging from the eaves provide the only light source. Their warm red glow penetrates the misty air, casting soft, shifting reflections on the damp, uneven stone slabs. As the camera advances, the image gradually tightens into a close-up. His eyes, like golden glass, are filled with profound emotion. The entire scene is enveloped in a tranquil, mysterious atmosphere imbued with the whimsical charm of the East.", + "seed": 11139, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000151/group_8/index_4038_seed_11139_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": -0.05482395393773913, + "vlm_logprob": -0.05482395393773913, + "vlm_yes_no_margin": 3.65, + "vlm_token_credit": -0.05482395393773913, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a handsome man with a scaly face, sharp jawline, and prominent nose?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the character have long jet-black hair and emerald-green scales covering his face and neck?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character wearing a dark green silk robe that is slightly slit at the collar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the character possess a powerful serpentine tail that glides along the ground in a fluid S-curve?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start as a medium-long shot and steadily approach the subject until it becomes a 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": [ + -0.20417122542858124 + ], + "opposite_logprob_tokens": [ + -1.7041711807250977 + ], + "logprob_reward": -0.20417122542858124, + "yes_no_margin_reward": 1.5, + "token_credit_reward": -0.20417122542858124, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.02793143317103386, + 0.3715240955352783, + 0.013562493026256561, + 0.02587353065609932, + 0.10327903926372528, + 0.05596974119544029, + 0.21773412823677063, + 0.01767887733876705, + 0.02875402942299843, + 0.1376926451921463 + ], + "credit_map": [ + 0.003828595159575343, + 0.002491000108420849, + 0.003566280473023653, + 0.004391816444694996, + 0.0027591113466769457, + 0.004753234330564737, + 0.004134170711040497, + 0.0020072252955287695, + 0.04703957587480545, + 0.011313225142657757, + 0.11679837852716446, + 0.01875981129705906, + 0.03736016899347305, + 0.1164889708161354, + 0.010147636756300926, + 0.013616347685456276, + 0.0014917366206645966, + 0.0021224720403552055, + 0.0014879850205034018, + 0.0026878979988396168, + 0.0014162063598632812, + 0.001371765392832458, + 0.0017788191325962543, + 0.0012056094128638506, + 0.00313414609991014, + 0.0023530826438218355, + 0.003269101958721876, + 0.00537321250885725, + 0.0027879630215466022, + 0.0025470643304288387, + 0.003976122941821814, + 0.002432834589853883, + 0.028479265049099922, + 0.004889954812824726, + 0.03706444427371025, + 0.007680629845708609, + 0.004672281444072723, + 0.007611776236444712, + 0.005208904389292002, + 0.007671785075217485, + 0.006790868006646633, + 0.011142699979245663, + 0.00927306991070509, + 0.007010110188275576, + 0.005708056502044201, + 0.003523464547470212, + 0.005772419273853302, + 0.006749048829078674, + 0.06020171940326691, + 0.008688694797456264, + 0.02172692120075226, + 0.013811826705932617, + 0.005757142789661884, + 0.01170635037124157, + 0.006620022933930159, + 0.08922145515680313, + 0.0022327222395688295, + 0.002000720240175724, + 0.001969733741134405, + 0.002574163954705, + 0.0016244525322690606, + 0.0022410801611840725, + 0.001295233378186822, + 0.00374077120795846, + 0.008980954065918922, + 0.0017915525240823627, + 0.006496336776763201, + 0.00196472043171525, + 0.0015914991963654757, + 0.00259381253272295, + 0.0019961402285844088, + 0.0033390142489224672, + 0.01184229925274849, + 0.004998624324798584, + 0.008240415714681149, + 0.02140199765563011, + 0.008297128602862358, + 0.010118609294295311, + 0.055050838738679886, + 0.01774272881448269 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 6, + 9 + ], + "selected_reward": -0.20417122542858124 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a handsome man with a scaly face, sharp jawline, and prominent nose?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the character have long jet-black hair and emerald-green scales covering his face and neck?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character wearing a dark green silk robe that is slightly slit at the collar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the character possess a powerful serpentine tail that glides along the ground in a fluid S-curve?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start as a medium-long shot and steadily approach the subject until it becomes a 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": [ + -0.012556896544992924 + ], + "opposite_logprob_tokens": [ + -4.387557029724121 + ], + "logprob_reward": -0.012556896544992924, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012556896544992924, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.029109695926308632, + 0.31233078241348267, + 0.01380747091025114, + 0.019268929958343506, + 0.030675498768687248, + 0.08262322843074799, + 0.13619187474250793, + 0.027431199327111244, + 0.0963393822312355, + 0.2522219121456146 + ], + "credit_map": [ + 0.0038696536794304848, + 0.003478338709101081, + 0.003493329044431448, + 0.0050182100385427475, + 0.002996759954839945, + 0.0049054427072405815, + 0.0032303775660693645, + 0.0021175825968384743, + 0.038534682244062424, + 0.00997204054147005, + 0.09800144284963608, + 0.01666329987347126, + 0.03162291646003723, + 0.09726670384407043, + 0.008677501231431961, + 0.011592208407819271, + 0.0016167440917342901, + 0.0015263253590092063, + 0.0015757718356326222, + 0.002711883280426264, + 0.002121416851878166, + 0.0016105205286294222, + 0.0015766550786793232, + 0.0010681534186005592, + 0.0028237332589924335, + 0.0020375398453325033, + 0.0038154355715960264, + 0.003557153046131134, + 0.0016291951760649681, + 0.0014991523930802941, + 0.0020818791817873716, + 0.0018248418346047401, + 0.0061157019808888435, + 0.0031339803244918585, + 0.007651735562831163, + 0.0040682037360966206, + 0.0022806290071457624, + 0.0022621084935963154, + 0.002369495341554284, + 0.00279364176094532, + 0.014120164327323437, + 0.009105210192501545, + 0.028576279059052467, + 0.0064023821614682674, + 0.005334597080945969, + 0.0037680414970964193, + 0.0040107062086462975, + 0.011305853724479675, + 0.038127072155475616, + 0.010108561255037785, + 0.011472871527075768, + 0.008356696926057339, + 0.004923662170767784, + 0.007751661352813244, + 0.00486518070101738, + 0.05058617517352104, + 0.002975467825308442, + 0.0029000993818044662, + 0.003497972385957837, + 0.004351709503680468, + 0.0021074379328638315, + 0.0023363123182207346, + 0.003307765582576394, + 0.005954434163868427, + 0.029900701716542244, + 0.004322424530982971, + 0.025830652564764023, + 0.006890166085213423, + 0.004094327799975872, + 0.013359793461859226, + 0.005771030206233263, + 0.006170285865664482, + 0.022297119721770287, + 0.0077954246662557125, + 0.014450334012508392, + 0.03967184200882912, + 0.015211126767098904, + 0.019570549950003624, + 0.10264156758785248, + 0.030583936721086502 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 9, + 6 + ], + "selected_reward": -0.012556896544992924 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a handsome man with a scaly face, sharp jawline, and prominent nose?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the character have long jet-black hair and emerald-green scales covering his face and neck?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character wearing a dark green silk robe that is slightly slit at the collar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the character possess a powerful serpentine tail that glides along the ground in a fluid S-curve?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start as a medium-long shot and steadily approach the subject until it becomes a 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": [ + -0.03369786962866783 + ], + "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": [ + 0.09773354232311249, + 0.11021721363067627, + 0.05726844072341919, + 0.03901388496160507, + 0.13239583373069763, + 0.07236095517873764, + 0.1312243640422821, + 0.0856703445315361, + 0.11574812233448029, + 0.1583673357963562 + ], + "credit_map": [ + 0.014212754555046558, + 0.009213835932314396, + 0.011636624112725258, + 0.009420104324817657, + 0.015234805643558502, + 0.013193856924772263, + 0.0152368089184165, + 0.009584750048816204, + 0.012342212721705437, + 0.00652712257578969, + 0.028323719277977943, + 0.008388889953494072, + 0.011548050679266453, + 0.029563115909695625, + 0.008183883503079414, + 0.005340222269296646, + 0.0068933553993701935, + 0.0062723178416490555, + 0.006686248350888491, + 0.01160037238150835, + 0.007689444348216057, + 0.006597389001399279, + 0.006360618397593498, + 0.005168699659407139, + 0.00480011897161603, + 0.004783865064382553, + 0.009381952695548534, + 0.005982525181025267, + 0.0033545815385878086, + 0.003789933631196618, + 0.0029783204663544893, + 0.003942587878555059, + 0.031634557992219925, + 0.00943030696362257, + 0.04327581077814102, + 0.012422237545251846, + 0.0064211320132017136, + 0.0089079849421978, + 0.00950569100677967, + 0.010798105970025063, + 0.015499476343393326, + 0.007185226306319237, + 0.023860730230808258, + 0.005033961031585932, + 0.0045891511254012585, + 0.005627436097711325, + 0.0036214864812791348, + 0.006943488027900457, + 0.035941023379564285, + 0.00750813540071249, + 0.012173688039183617, + 0.007910508662462234, + 0.0039901467971503735, + 0.008047747425734997, + 0.004419749602675438, + 0.05123336240649223, + 0.008289630524814129, + 0.011153489351272583, + 0.01048986241221428, + 0.01395900547504425, + 0.007338745519518852, + 0.0048876176588237286, + 0.006523870397359133, + 0.02302812971174717, + 0.01620953530073166, + 0.010238304734230042, + 0.0226103775203228, + 0.014371346682310104, + 0.010000603273510933, + 0.019940098747611046, + 0.011198705062270164, + 0.011179150082170963, + 0.015482759103178978, + 0.004845605231821537, + 0.01003357209265232, + 0.023984944447875023, + 0.010108978487551212, + 0.0093022296205163, + 0.0640396997332573, + 0.020569538697600365 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 4, + 6 + ], + "selected_reward": -0.03369786962866783 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a handsome man with a scaly face, sharp jawline, and prominent nose?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the character have long jet-black hair and emerald-green scales covering his face and neck?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character wearing a dark green silk robe that is slightly slit at the collar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the character possess a powerful serpentine tail that glides along the ground in a fluid S-curve?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start as a medium-long shot and steadily approach the subject until it becomes a 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": [ + -0.016067076474428177 + ], + "opposite_logprob_tokens": [ + -4.141067028045654 + ], + "logprob_reward": -0.016067076474428177, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016067076474428177, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15056638419628143, + 0.3398386836051941, + 0.022542905062437057, + 0.02392209880053997, + 0.023642523214221, + 0.01845093071460724, + 0.2328871190547943, + 0.028371263295412064, + 0.11354927718639374, + 0.0462288036942482 + ], + "credit_map": [ + 0.016095707193017006, + 0.014856799505650997, + 0.015796896070241928, + 0.014143458567559719, + 0.0174079742282629, + 0.022692976519465446, + 0.030961332842707634, + 0.0186112429946661, + 0.0417240709066391, + 0.010461239144206047, + 0.10666953772306442, + 0.01761578768491745, + 0.03470781072974205, + 0.10663799196481705, + 0.009251347742974758, + 0.012770875357091427, + 0.0030043257866054773, + 0.0016962006920948625, + 0.003271566703915596, + 0.0022198217920958996, + 0.0028008692897856236, + 0.0028108255937695503, + 0.003301094751805067, + 0.0034381996374577284, + 0.0030616940930485725, + 0.0024246498942375183, + 0.004823084454983473, + 0.0026033956091850996, + 0.0024351058527827263, + 0.0019930987618863583, + 0.0037644808180630207, + 0.0028165883850306273, + 0.005027661100029945, + 0.0013866305816918612, + 0.0076411752961575985, + 0.0017755022272467613, + 0.0017565222224220634, + 0.0022018649615347385, + 0.001647570519708097, + 0.002205595374107361, + 0.0025056833401322365, + 0.0026850823778659105, + 0.0040333871729671955, + 0.001544076600112021, + 0.001993359997868538, + 0.0016583969118073583, + 0.0015977027360349894, + 0.002433242043480277, + 0.06557071208953857, + 0.010879012756049633, + 0.02221659943461418, + 0.013499850407242775, + 0.005840301979333162, + 0.012671394273638725, + 0.006755761802196503, + 0.09545349329710007, + 0.003930637147277594, + 0.004271958023309708, + 0.0036636539734899998, + 0.004170327913016081, + 0.002773806219920516, + 0.0029388873372226954, + 0.002621545922011137, + 0.004000445827841759, + 0.019533559679985046, + 0.004985353443771601, + 0.04344598203897476, + 0.015618031844496727, + 0.004114329349249601, + 0.007470474112778902, + 0.007807686924934387, + 0.010573859326541424, + 0.004276685416698456, + 0.001986283343285322, + 0.0030595192220062017, + 0.0068143378011882305, + 0.0029359792824834585, + 0.002221573842689395, + 0.017551371827721596, + 0.00738305039703846 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 6, + 0 + ], + "selected_reward": -0.016067076474428177 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a handsome man with a scaly face, sharp jawline, and prominent nose?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the character have long jet-black hair and emerald-green scales covering his face and neck?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character wearing a dark green silk robe that is slightly slit at the collar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the character possess a powerful serpentine tail that glides along the ground in a fluid S-curve?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start as a medium-long shot and steadily approach the subject until it becomes a 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": [ + -0.007626701612025499 + ], + "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": [ + 0.04313673451542854, + 0.038334496319293976, + 0.011130401864647865, + 0.02294343151152134, + 0.03285505995154381, + 0.06418320536613464, + 0.4016213119029999, + 0.07112748920917511, + 0.28879085183143616, + 0.025877034291625023 + ], + "credit_map": [ + 0.005557200405746698, + 0.003630982944741845, + 0.00610382528975606, + 0.005730805452913046, + 0.004363776184618473, + 0.005675890017300844, + 0.007081079296767712, + 0.004993177950382233, + 0.005195822101086378, + 0.0031727785244584084, + 0.006586435716599226, + 0.005383656360208988, + 0.0032911603339016438, + 0.0076937805861234665, + 0.00433787377551198, + 0.002672990085557103, + 0.0016351307276636362, + 0.0014127710601314902, + 0.0014032134786248207, + 0.0017725343350321054, + 0.0011237263679504395, + 0.0012392856879159808, + 0.0012910662917420268, + 0.0012526731006801128, + 0.0029866539407521486, + 0.0020862978417426348, + 0.0037772951181977987, + 0.004875903949141502, + 0.0020527478773146868, + 0.0024955160915851593, + 0.0026759784668684006, + 0.001993038458749652, + 0.007430534344166517, + 0.002854001009836793, + 0.009630855172872543, + 0.003714781953021884, + 0.0017703203484416008, + 0.002888049930334091, + 0.001920147449709475, + 0.0026463696267455816, + 0.010596070438623428, + 0.007856618613004684, + 0.019496310502290726, + 0.004941979888826609, + 0.005179942585527897, + 0.0038488400168716908, + 0.00344830472022295, + 0.008815133944153786, + 0.11431503295898438, + 0.017694072797894478, + 0.0374942272901535, + 0.023127354681491852, + 0.010548674501478672, + 0.022525101900100708, + 0.012314733117818832, + 0.16360212862491608, + 0.007088696584105492, + 0.00850373413413763, + 0.007789932191371918, + 0.011533326469361782, + 0.005283987615257502, + 0.004281264264136553, + 0.005999489687383175, + 0.02064705640077591, + 0.047146737575531006, + 0.022743023931980133, + 0.0678759217262268, + 0.0310586616396904, + 0.02110343798995018, + 0.04549086466431618, + 0.023928366601467133, + 0.02944381907582283, + 0.003719884669408202, + 0.001398595399223268, + 0.003041809191927314, + 0.0029693362303078175, + 0.0017652647802606225, + 0.002206535544246435, + 0.007899594493210316, + 0.0028760135173797607 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 8, + 7 + ], + "selected_reward": -0.007626701612025499 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "smooth, uninterrupted medium-long shot begins, focusing on a handsome, scaly-faced snake prince. With graceful, silent grace, he meanders along an ancient, rain-soaked stone path. He possesses a sharp jawline and a prominent nose. His long, jet-black hair sways gently with his movements, a few damp strands plastering his forehead. From his cheekbones to his neck, a dense, shimmering emerald-green scales shimmer in the hazy light, reflecting an ever-shifting shimmer. He clads a dark green silk robe, slightly slit at the collar. The soft fabric clings to his muscular torso, rising and falling rhythmically as his powerful serpentine tail glides along the ground in a fluid S-curve. The camera steadily and deliberately approaches him. The background is a narrow alleyway lined with quaint wooden buildings. Red paper lanterns hanging from the eaves provide the only light source. Their warm red glow penetrates the misty air, casting soft, shifting reflections on the damp, uneven stone slabs. As the camera advances, the image gradually tightens into a close-up. His eyes, like golden glass, are filled with profound emotion. The entire scene is enveloped in a tranquil, mysterious atmosphere imbued with the whimsical charm of the East.", + "seed": 11140, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000151/group_8/index_4038_seed_11140_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": -0.021416112780570984, + "vlm_logprob": -0.021416112780570984, + "vlm_yes_no_margin": 4.4, + "vlm_token_credit": -0.021416112780570984, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a handsome man with a scaly face, sharp jawline, and prominent nose?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the character have long jet-black hair and emerald-green scales covering his face and neck?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character wearing a dark green silk robe that is slightly slit at the collar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the character possess a powerful serpentine tail that glides along the ground in a fluid S-curve?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start as a medium-long shot and steadily approach the subject until it becomes a 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": [ + -0.0715341717004776 + ], + "opposite_logprob_tokens": [ + -2.6965341567993164 + ], + "logprob_reward": -0.0715341717004776, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.0715341717004776, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.058080654591321945, + 0.013255473226308823, + 0.06319129467010498, + 0.02819928526878357, + 0.027078675106167793, + 0.04772089421749115, + 0.03478558734059334, + 0.024593740701675415, + 0.1704399436712265, + 0.5326544642448425 + ], + "credit_map": [ + 0.007869808003306389, + 0.0065967729315161705, + 0.006126080173999071, + 0.0065652793273329735, + 0.009188669733703136, + 0.007469122298061848, + 0.006684241816401482, + 0.007580675650388002, + 0.0018793842755258083, + 0.0018271857406944036, + 0.0019339252030476928, + 0.0015414035879075527, + 0.0018267675768584013, + 0.0015607122331857681, + 0.0014179295394569635, + 0.0012681657681241632, + 0.010083772242069244, + 0.0033234828151762486, + 0.007011798210442066, + 0.00181890360545367, + 0.0024063107557594776, + 0.02244166098535061, + 0.0023399873171001673, + 0.013765373267233372, + 0.003998011816293001, + 0.0034552705474197865, + 0.004760850220918655, + 0.0028688833117485046, + 0.003498550271615386, + 0.0035852761939167976, + 0.002851655473932624, + 0.003180787432938814, + 0.004822478163987398, + 0.002918949583545327, + 0.004021807108074427, + 0.0019213085761293769, + 0.003953647334128618, + 0.003963567316532135, + 0.0029694384429603815, + 0.0025074779987335205, + 0.009009124711155891, + 0.003246922278776765, + 0.014792349189519882, + 0.005517429206520319, + 0.0029834189917892218, + 0.003357294946908951, + 0.00626782001927495, + 0.0025465332437306643, + 0.005972990766167641, + 0.0036572415847331285, + 0.003389062127098441, + 0.005691767204552889, + 0.004633577074855566, + 0.003183304797858, + 0.005541638936847448, + 0.002716006711125374, + 0.0029219184070825577, + 0.00293676252476871, + 0.0030415046494454145, + 0.0034709919709712267, + 0.0035801841877400875, + 0.003397664986550808, + 0.002671126974746585, + 0.0025735872332006693, + 0.045254580676555634, + 0.007740848697721958, + 0.007327334024012089, + 0.009837477467954159, + 0.009229286573827267, + 0.030230985954403877, + 0.009200917556881905, + 0.051618512719869614, + 0.14805521070957184, + 0.02574123442173004, + 0.030904529616236687, + 0.019277676939964294, + 0.04793527349829674, + 0.043002963066101074, + 0.04193854704499245, + 0.17579902708530426 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 2 + ], + "selected_reward": -0.0715341717004776 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a handsome man with a scaly face, sharp jawline, and prominent nose?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the character have long jet-black hair and emerald-green scales covering his face and neck?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character wearing a dark green silk robe that is slightly slit at the collar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the character possess a powerful serpentine tail that glides along the ground in a fluid S-curve?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start as a medium-long shot and steadily approach the subject until it becomes a 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": [ + -0.006751582026481628 + ], + "opposite_logprob_tokens": [ + -5.006751537322998 + ], + "logprob_reward": -0.006751582026481628, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006751582026481628, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07066810876131058, + 0.016558947041630745, + 0.06513670086860657, + 0.055916715413331985, + 0.027334149926900864, + 0.0271452646702528, + 0.04826299473643303, + 0.09553170949220657, + 0.07618110626935959, + 0.5172642469406128 + ], + "credit_map": [ + 0.01108723133802414, + 0.009729137644171715, + 0.00740515161305666, + 0.008385325782001019, + 0.008940794505178928, + 0.009243016131222248, + 0.006248821038752794, + 0.009628637693822384, + 0.0032707788050174713, + 0.0017594924429431558, + 0.0024658851325511932, + 0.0015940293669700623, + 0.0016547746490687132, + 0.0024737075436860323, + 0.001692113932222128, + 0.001648166449740529, + 0.00721134664490819, + 0.0030464306473731995, + 0.007419988978654146, + 0.002167933387681842, + 0.0021750496234744787, + 0.02182834967970848, + 0.0023624086752533913, + 0.018925197422504425, + 0.008840754628181458, + 0.005671922117471695, + 0.008932389318943024, + 0.004617101978510618, + 0.007912127301096916, + 0.006776059046387672, + 0.005387049168348312, + 0.007779314182698727, + 0.003813965944573283, + 0.0037664088886231184, + 0.0046160100027918816, + 0.0034496765583753586, + 0.003885438432916999, + 0.003457466373220086, + 0.0024104523472487926, + 0.0019347338238731027, + 0.004851524718105793, + 0.0025651538744568825, + 0.0061197164468467236, + 0.0031738830730319023, + 0.0016922680661082268, + 0.00391816534101963, + 0.003145365510135889, + 0.001679191249422729, + 0.010283436626195908, + 0.004973210394382477, + 0.005202254746109247, + 0.009422611445188522, + 0.00419915746897459, + 0.005608707666397095, + 0.005086125340312719, + 0.003487497102469206, + 0.01085633784532547, + 0.011458354070782661, + 0.013952157460153103, + 0.013196391053497791, + 0.010043918155133724, + 0.0145955178886652, + 0.007995037361979485, + 0.013434004038572311, + 0.01987437531352043, + 0.004358124453574419, + 0.005297979339957237, + 0.006245257332921028, + 0.004014671314507723, + 0.006210251711308956, + 0.005944317206740379, + 0.02423613891005516, + 0.14924053847789764, + 0.021696673706173897, + 0.024806106463074684, + 0.02043493650853634, + 0.03788446635007858, + 0.050356268882751465, + 0.035747263580560684, + 0.17709802091121674 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 7, + 8 + ], + "selected_reward": -0.006751582026481628 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a handsome man with a scaly face, sharp jawline, and prominent nose?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the character have long jet-black hair and emerald-green scales covering his face and neck?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character wearing a dark green silk robe that is slightly slit at the collar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the character possess a powerful serpentine tail that glides along the ground in a fluid S-curve?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start as a medium-long shot and steadily approach the subject until it becomes a 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": [ + -0.016067545861005783 + ], + "opposite_logprob_tokens": [ + -4.1410675048828125 + ], + "logprob_reward": -0.016067545861005783, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016067545861005783, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11053977906703949, + 0.049152109771966934, + 0.05462000146508217, + 0.05478796362876892, + 0.11596928536891937, + 0.050980210304260254, + 0.07200078666210175, + 0.1373392790555954, + 0.24149931967258453, + 0.11311132460832596 + ], + "credit_map": [ + 0.019839879125356674, + 0.00994170643389225, + 0.01335878949612379, + 0.010221266187727451, + 0.013917005620896816, + 0.01845410466194153, + 0.010277282446622849, + 0.014529740437865257, + 0.008012608624994755, + 0.004991852678358555, + 0.004992258735001087, + 0.0045494320802390575, + 0.00937095656991005, + 0.007188479416072369, + 0.004731188993901014, + 0.005315329879522324, + 0.012944838963449001, + 0.004162497352808714, + 0.008079849183559418, + 0.0033357723150402308, + 0.004126784857362509, + 0.006697443779557943, + 0.003923253156244755, + 0.011349565349519253, + 0.007833915762603283, + 0.005514597985893488, + 0.00912308506667614, + 0.005007964093238115, + 0.008532554842531681, + 0.006763940677046776, + 0.007188200019299984, + 0.004823706578463316, + 0.014061440713703632, + 0.01427632849663496, + 0.01917477883398533, + 0.01776166819036007, + 0.01303625013679266, + 0.011897847056388855, + 0.0178113654255867, + 0.007949598133563995, + 0.008629740215837955, + 0.004396096337586641, + 0.009991491213440895, + 0.00591597193852067, + 0.0050678797997534275, + 0.005141066387295723, + 0.008018242195248604, + 0.00381972617469728, + 0.014079352840781212, + 0.0059807877987623215, + 0.00763231236487627, + 0.0138044822961092, + 0.006856431253254414, + 0.007546493783593178, + 0.010081643238663673, + 0.006019274704158306, + 0.015868432819843292, + 0.01323937252163887, + 0.01692378520965576, + 0.01966887153685093, + 0.01481329184025526, + 0.01836816780269146, + 0.01567716896533966, + 0.022780179977416992, + 0.06714147329330444, + 0.01290785800665617, + 0.020113084465265274, + 0.016969842836260796, + 0.014365841634571552, + 0.01629410684108734, + 0.0174576323479414, + 0.07624947279691696, + 0.017822211608290672, + 0.010516045615077019, + 0.013199304230511189, + 0.00916806049644947, + 0.015560384839773178, + 0.02037973329424858, + 0.008681787177920341, + 0.017783796414732933 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 4 + ], + "selected_reward": -0.016067545861005783 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a handsome man with a scaly face, sharp jawline, and prominent nose?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the character have long jet-black hair and emerald-green scales covering his face and neck?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character wearing a dark green silk robe that is slightly slit at the collar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the character possess a powerful serpentine tail that glides along the ground in a fluid S-curve?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start as a medium-long shot and steadily approach the subject until it becomes a 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": [ + -0.008634612895548344 + ], + "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": [ + 0.0899464562535286, + 0.07691620290279388, + 0.13841086626052856, + 0.051620520651340485, + 0.024281449615955353, + 0.04313921183347702, + 0.08882477134466171, + 0.21166303753852844, + 0.1257825493812561, + 0.14941492676734924 + ], + "credit_map": [ + 0.014971671625971794, + 0.008661281317472458, + 0.010899960994720459, + 0.012723980471491814, + 0.01155849453061819, + 0.011569835245609283, + 0.008437113836407661, + 0.011124110780656338, + 0.017983367666602135, + 0.008235825225710869, + 0.01029029581695795, + 0.0067405542358756065, + 0.009008378721773624, + 0.009036501869559288, + 0.00756522361189127, + 0.00805605761706829, + 0.019128073006868362, + 0.005633774679154158, + 0.015417751856148243, + 0.003221607767045498, + 0.004520044662058353, + 0.05115721374750137, + 0.004718781914561987, + 0.03461362421512604, + 0.006750653497874737, + 0.005614158697426319, + 0.008298731409013271, + 0.006547108758240938, + 0.00879499688744545, + 0.004925527609884739, + 0.006566385272890329, + 0.004122954327613115, + 0.004343564156442881, + 0.0030907453037798405, + 0.004167844075709581, + 0.0030754688195884228, + 0.0024334662593901157, + 0.0025052058044821024, + 0.002676323289051652, + 0.0019888330716639757, + 0.006984420120716095, + 0.003712079022079706, + 0.005068105645477772, + 0.005520000122487545, + 0.004907690454274416, + 0.005176342558115721, + 0.008029346354305744, + 0.0037412261590361595, + 0.009126839227974415, + 0.008492287248373032, + 0.006535263266414404, + 0.014194886200129986, + 0.015060102567076683, + 0.00909457728266716, + 0.018773863092064857, + 0.007546946872025728, + 0.0255037359893322, + 0.020933011546730995, + 0.026025816798210144, + 0.037846870720386505, + 0.025241412222385406, + 0.021639464423060417, + 0.021663038060069084, + 0.032809678465127945, + 0.032253462821245193, + 0.0058395168744027615, + 0.008914255537092686, + 0.010697842575609684, + 0.007054390385746956, + 0.012787810526788235, + 0.008053560741245747, + 0.040181715041399, + 0.040049098432064056, + 0.00796241220086813, + 0.00926154013723135, + 0.006128485314548016, + 0.016035746783018112, + 0.0109021021053195, + 0.01153990812599659, + 0.047535642981529236 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 9, + 2 + ], + "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 a handsome man with a scaly face, sharp jawline, and prominent nose?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the character have long jet-black hair and emerald-green scales covering his face and neck?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character wearing a dark green silk robe that is slightly slit at the collar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the character possess a powerful serpentine tail that glides along the ground in a fluid S-curve?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start as a medium-long shot and steadily approach the subject until it becomes a 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": [ + -0.004092651419341564 + ], + "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": [ + 0.05823014676570892, + 0.028520580381155014, + 0.13630448281764984, + 0.0358896479010582, + 0.0688195526599884, + 0.10731738060712814, + 0.08632111549377441, + 0.2175845503807068, + 0.05959807336330414, + 0.20141448080539703 + ], + "credit_map": [ + 0.011072324588894844, + 0.005105501040816307, + 0.007838289253413677, + 0.0057228305377066135, + 0.0055356440134346485, + 0.010326437652111053, + 0.00608857162296772, + 0.00654054107144475, + 0.00412742281332612, + 0.0035113664343953133, + 0.0036265167873352766, + 0.003084841649979353, + 0.0029840152710676193, + 0.0036145939957350492, + 0.004352659452706575, + 0.003219161182641983, + 0.01892944611608982, + 0.006329690106213093, + 0.015605336986482143, + 0.004096160177141428, + 0.005536681041121483, + 0.04815550521016121, + 0.0051932670176029205, + 0.03245839104056358, + 0.005303245969116688, + 0.0036599235609173775, + 0.0061327083967626095, + 0.0029012339655309916, + 0.006608378142118454, + 0.003182994434610009, + 0.005038142669945955, + 0.0030630205292254686, + 0.00987935159355402, + 0.008821536786854267, + 0.012943092733621597, + 0.013123204931616783, + 0.005265107378363609, + 0.006958706304430962, + 0.0066522834822535515, + 0.005176261067390442, + 0.014464929699897766, + 0.009517588652670383, + 0.02549687959253788, + 0.011681033298373222, + 0.008892586454749107, + 0.012926220893859863, + 0.017338648438453674, + 0.006999482866376638, + 0.013597956858575344, + 0.007726484443992376, + 0.010440109297633171, + 0.014859016053378582, + 0.0102997412905097, + 0.009388481266796589, + 0.013811753131449223, + 0.006197564769536257, + 0.028023002669215202, + 0.025671042501926422, + 0.030061114579439163, + 0.03622802719473839, + 0.03074190765619278, + 0.0258423313498497, + 0.015793226659297943, + 0.025223901495337486, + 0.008240816183388233, + 0.004182318225502968, + 0.006377502344548702, + 0.006106250919401646, + 0.0064238980412483215, + 0.016962271183729172, + 0.004619543440639973, + 0.006685469765216112, + 0.06116534769535065, + 0.00914405845105648, + 0.010935022495687008, + 0.008043942973017693, + 0.013497523963451385, + 0.014315783977508545, + 0.011888478882610798, + 0.07242430001497269 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 9, + 2 + ], + "selected_reward": -0.004092651419341564 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "smooth, uninterrupted medium-long shot begins, focusing on a handsome, scaly-faced snake prince. With graceful, silent grace, he meanders along an ancient, rain-soaked stone path. He possesses a sharp jawline and a prominent nose. His long, jet-black hair sways gently with his movements, a few damp strands plastering his forehead. From his cheekbones to his neck, a dense, shimmering emerald-green scales shimmer in the hazy light, reflecting an ever-shifting shimmer. He clads a dark green silk robe, slightly slit at the collar. The soft fabric clings to his muscular torso, rising and falling rhythmically as his powerful serpentine tail glides along the ground in a fluid S-curve. The camera steadily and deliberately approaches him. The background is a narrow alleyway lined with quaint wooden buildings. Red paper lanterns hanging from the eaves provide the only light source. Their warm red glow penetrates the misty air, casting soft, shifting reflections on the damp, uneven stone slabs. As the camera advances, the image gradually tightens into a close-up. His eyes, like golden glass, are filled with profound emotion. The entire scene is enveloped in a tranquil, mysterious atmosphere imbued with the whimsical charm of the East.", + "seed": 11141, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000151/group_8/index_4038_seed_11141_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": -0.07802777867764235, + "vlm_logprob": -0.07802777867764235, + "vlm_yes_no_margin": 3.949999952316284, + "vlm_token_credit": -0.07802777867764235, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a handsome man with a scaly face, sharp jawline, and prominent nose?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the character have long jet-black hair and emerald-green scales covering his face and neck?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character wearing a dark green silk robe that is slightly slit at the collar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the character possess a powerful serpentine tail that glides along the ground in a fluid S-curve?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start as a medium-long shot and steadily approach the subject until it becomes a 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": [ + -0.35075244307518005 + ], + "opposite_logprob_tokens": [ + -1.2257524728775024 + ], + "logprob_reward": -0.35075244307518005, + "yes_no_margin_reward": 0.875, + "token_credit_reward": -0.35075244307518005, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1208786889910698, + 0.05326717719435692, + 0.15675124526023865, + 0.11762399971485138, + 0.022741373628377914, + 0.016130536794662476, + 0.029599247500300407, + 0.04066869616508484, + 0.15735049545764923, + 0.28498855233192444 + ], + "credit_map": [ + 0.015615114942193031, + 0.01476216595619917, + 0.016085343435406685, + 0.010501738637685776, + 0.015285841189324856, + 0.01015909481793642, + 0.014934047125279903, + 0.02353535033762455, + 0.015664754435420036, + 0.0034096240997314453, + 0.004827091470360756, + 0.00202167103998363, + 0.004293309058994055, + 0.006327732931822538, + 0.0046118805184960365, + 0.01211111806333065, + 0.023089732974767685, + 0.0053519015200436115, + 0.024405047297477722, + 0.005342837423086166, + 0.005642585922032595, + 0.04026744142174721, + 0.009669175371527672, + 0.04298253357410431, + 0.009753790684044361, + 0.004227687604725361, + 0.03151920437812805, + 0.0047806198708713055, + 0.0035964155104011297, + 0.006872405298054218, + 0.015747081488370895, + 0.041126787662506104, + 0.002669884590432048, + 0.002242519287392497, + 0.0034387074410915375, + 0.0025163053069263697, + 0.0021512864623218775, + 0.0029218795243650675, + 0.004029047209769487, + 0.002771743107587099, + 0.0021531591191887856, + 0.0015003667213022709, + 0.0021957303397357464, + 0.0015821269480511546, + 0.0023520556278526783, + 0.0019634105265140533, + 0.0026244951877743006, + 0.001759192906320095, + 0.003828126471489668, + 0.003306786762550473, + 0.006277392618358135, + 0.003120215842500329, + 0.003106958931311965, + 0.0029993208590894938, + 0.003935690503567457, + 0.003024754347279668, + 0.006796037312597036, + 0.0037723176646977663, + 0.00684675108641386, + 0.004608504474163055, + 0.0027810607571154833, + 0.003131438512355089, + 0.008398439735174179, + 0.004334148019552231, + 0.028135307133197784, + 0.01221293956041336, + 0.014585091732442379, + 0.016354817897081375, + 0.016417648643255234, + 0.015569701790809631, + 0.03340250626206398, + 0.02067248709499836, + 0.04017348960042, + 0.02182711288332939, + 0.05468593165278435, + 0.025959348306059837, + 0.019817683845758438, + 0.03683506324887276, + 0.06112084910273552, + 0.024569066241383553 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 2 + ], + "selected_reward": -0.35075244307518005 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a handsome man with a scaly face, sharp jawline, and prominent nose?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the character have long jet-black hair and emerald-green scales covering his face and neck?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character wearing a dark green silk robe that is slightly slit at the collar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the character possess a powerful serpentine tail that glides along the ground in a fluid S-curve?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start as a medium-long shot and steadily approach the subject until it becomes a 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": [ + -0.005963390693068504 + ], + "opposite_logprob_tokens": [ + -5.130963325500488 + ], + "logprob_reward": -0.005963390693068504, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005963390693068504, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1195531114935875, + 0.1216360479593277, + 0.2594655752182007, + 0.03253869339823723, + 0.02278888039290905, + 0.05992212891578674, + 0.032822124660015106, + 0.07091420143842697, + 0.1457471251487732, + 0.1346120983362198 + ], + "credit_map": [ + 0.013076888397336006, + 0.01129397377371788, + 0.01583090052008629, + 0.013394388370215893, + 0.01725670136511326, + 0.01322305016219616, + 0.015783751383423805, + 0.019693449139595032, + 0.0390855148434639, + 0.007459554355591536, + 0.010834050364792347, + 0.004381557460874319, + 0.007486889138817787, + 0.01434145774692297, + 0.008894993923604488, + 0.029152024537324905, + 0.03705036640167236, + 0.009569711983203888, + 0.04329841956496239, + 0.00862274318933487, + 0.009609242901206017, + 0.05948728322982788, + 0.016196291893720627, + 0.0756315141916275, + 0.0031252100598067045, + 0.0021814110223203897, + 0.007189982570707798, + 0.0017701759934425354, + 0.002396075986325741, + 0.0023429321590811014, + 0.004134611692279577, + 0.009398291818797588, + 0.0024870496708899736, + 0.0028380644507706165, + 0.0037174029275774956, + 0.0026905089616775513, + 0.0021452561486512423, + 0.003140837186947465, + 0.0034767198376357555, + 0.002293040743097663, + 0.007626350037753582, + 0.0045568617060780525, + 0.014196404255926609, + 0.004414882976561785, + 0.006200683768838644, + 0.01198550034314394, + 0.006234160158783197, + 0.0047072856687009335, + 0.004538032691925764, + 0.0024982579052448273, + 0.006410552654415369, + 0.00305393454618752, + 0.0029482839163392782, + 0.003346731187775731, + 0.0068203723058104515, + 0.0032059592194855213, + 0.011500688269734383, + 0.005766787566244602, + 0.007513319607824087, + 0.00860205851495266, + 0.0071486374363303185, + 0.0064185988157987595, + 0.017606427893042564, + 0.006357679143548012, + 0.023262090981006622, + 0.012621446512639523, + 0.011149842292070389, + 0.018244454637169838, + 0.018995095044374466, + 0.015336387790739536, + 0.02930668368935585, + 0.016831135377287865, + 0.020862676203250885, + 0.01374114491045475, + 0.015127068385481834, + 0.013405793346464634, + 0.01010268647223711, + 0.01685665175318718, + 0.03260420635342598, + 0.011911879293620586 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 8, + 9 + ], + "selected_reward": -0.005963390693068504 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a handsome man with a scaly face, sharp jawline, and prominent nose?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the character have long jet-black hair and emerald-green scales covering his face and neck?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character wearing a dark green silk robe that is slightly slit at the collar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the character possess a powerful serpentine tail that glides along the ground in a fluid S-curve?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start as a medium-long shot and steadily approach the subject until it becomes a 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": [ + -0.018178174272179604 + ], + "opposite_logprob_tokens": [ + -4.0181779861450195 + ], + "logprob_reward": -0.018178174272179604, + "yes_no_margin_reward": 3.999999761581421, + "token_credit_reward": -0.018178174272179604, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12273754924535751, + 0.04651990532875061, + 0.018506024032831192, + 0.16537228226661682, + 0.05691654980182648, + 0.0992933139204979, + 0.027297768741846085, + 0.1394115537405014, + 0.08680321276187897, + 0.23714181780815125 + ], + "credit_map": [ + 0.01462365873157978, + 0.013289598748087883, + 0.014665954746305943, + 0.016106421127915382, + 0.013133515603840351, + 0.012553966604173183, + 0.016843438148498535, + 0.02152099274098873, + 0.009405609220266342, + 0.005978369154036045, + 0.0054091173224151134, + 0.0032020488288253546, + 0.0047831544652581215, + 0.006189335137605667, + 0.003916656132787466, + 0.00763561949133873, + 0.00249165715649724, + 0.0019825834315270185, + 0.0023868056014180183, + 0.0017293185228481889, + 0.0015833823708817363, + 0.0034849424846470356, + 0.001859995536506176, + 0.0029873393941670656, + 0.013441143557429314, + 0.006578686181455851, + 0.032203856855630875, + 0.009126849472522736, + 0.0067431083880364895, + 0.010055668652057648, + 0.0358029268682003, + 0.05142002925276756, + 0.006690938491374254, + 0.007310445420444012, + 0.00986759178340435, + 0.006752713117748499, + 0.006764027290046215, + 0.006013565696775913, + 0.008619701489806175, + 0.004897569306194782, + 0.01416703313589096, + 0.004181527532637119, + 0.0328158438205719, + 0.004688132554292679, + 0.00440332293510437, + 0.027433017268776894, + 0.003970648627728224, + 0.007633785251528025, + 0.0036717383190989494, + 0.0022585312835872173, + 0.004101612139493227, + 0.0026987011078745127, + 0.003426362294703722, + 0.0030352515168488026, + 0.005318408366292715, + 0.002787163248285651, + 0.02346727065742016, + 0.011957383714616299, + 0.027847502380609512, + 0.0176639836281538, + 0.011712959967553616, + 0.014866292476654053, + 0.01881253719329834, + 0.01308363676071167, + 0.013880748301744461, + 0.005188967101275921, + 0.012450853362679482, + 0.0071367137134075165, + 0.008926887065172195, + 0.009225166402757168, + 0.019051341339945793, + 0.010942538268864155, + 0.03254719823598862, + 0.03282264620065689, + 0.05927838757634163, + 0.02190462127327919, + 0.013445035554468632, + 0.025513941422104836, + 0.02975759655237198, + 0.021872399374842644 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 3, + 7 + ], + "selected_reward": -0.018178174272179604 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a handsome man with a scaly face, sharp jawline, and prominent nose?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the character have long jet-black hair and emerald-green scales covering his face and neck?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character wearing a dark green silk robe that is slightly slit at the collar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the character possess a powerful serpentine tail that glides along the ground in a fluid S-curve?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start as a medium-long shot and steadily approach the subject until it becomes a 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": [ + -0.007622560951858759 + ], + "opposite_logprob_tokens": [ + -4.882622718811035 + ], + "logprob_reward": -0.007622560951858759, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007622560951858759, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06373562663793564, + 0.06248737871646881, + 0.015905797481536865, + 0.0600646547973156, + 0.019691944122314453, + 0.07090307027101517, + 0.0882304310798645, + 0.4374234974384308, + 0.042420148849487305, + 0.13913749158382416 + ], + "credit_map": [ + 0.008214914239943027, + 0.006991889793425798, + 0.007918785326182842, + 0.005394881125539541, + 0.009195256046950817, + 0.006904053967446089, + 0.006814418360590935, + 0.01230142917484045, + 0.020835543051362038, + 0.0035451268777251244, + 0.005467374809086323, + 0.002015163190662861, + 0.0029585852753371, + 0.007870635017752647, + 0.004014154896140099, + 0.015780793502926826, + 0.0021186836529523134, + 0.0010298234410583973, + 0.002419349504634738, + 0.0010687229223549366, + 0.0010869542602449656, + 0.0035039677750319242, + 0.0011038987431675196, + 0.003574396949261427, + 0.004882734268903732, + 0.001954083563759923, + 0.017022717744112015, + 0.002093184506520629, + 0.00197968608699739, + 0.0034175708424299955, + 0.006781155709177256, + 0.021933522075414658, + 0.0020463585387915373, + 0.0022360787261277437, + 0.004026554990559816, + 0.00270497752353549, + 0.0021550552919507027, + 0.0020056029316037893, + 0.002257860731333494, + 0.0022594539914280176, + 0.009636280126869678, + 0.0029785577207803726, + 0.02122480422258377, + 0.0038695826660841703, + 0.0038331057876348495, + 0.021119654178619385, + 0.002728548599407077, + 0.005512535106390715, + 0.01153470203280449, + 0.007949165999889374, + 0.01628441922366619, + 0.007720767054706812, + 0.009478176943957806, + 0.006878243759274483, + 0.020419003441929817, + 0.007965954020619392, + 0.06761182099580765, + 0.0499662347137928, + 0.07182478159666061, + 0.05987212061882019, + 0.03898129612207413, + 0.026691069826483727, + 0.0819656178355217, + 0.04051050916314125, + 0.007262208964675665, + 0.0042679188773036, + 0.004163308069109917, + 0.005843176506459713, + 0.003341034520417452, + 0.00532691553235054, + 0.008137061260640621, + 0.004078525584191084, + 0.019543791189789772, + 0.015226456336677074, + 0.028140205889940262, + 0.013458924368023872, + 0.009108728729188442, + 0.01903117448091507, + 0.023965753614902496, + 0.010662458837032318 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 9, + 6 + ], + "selected_reward": -0.007622560951858759 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a handsome man with a scaly face, sharp jawline, and prominent nose?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the character have long jet-black hair and emerald-green scales covering his face and neck?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character wearing a dark green silk robe that is slightly slit at the collar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the character possess a powerful serpentine tail that glides along the ground in a fluid S-curve?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start as a medium-long shot and steadily approach the subject until it becomes a 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": [ + -0.007622324395924807 + ], + "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": [ + 0.0364510677754879, + 0.07288990169763565, + 0.12974326312541962, + 0.07323088496923447, + 0.03755030035972595, + 0.058103296905756, + 0.03089170902967453, + 0.10044151544570923, + 0.06698688864707947, + 0.3937112092971802 + ], + "credit_map": [ + 0.004636470228433609, + 0.0037048147059977055, + 0.0054857260547578335, + 0.002967354841530323, + 0.004784794989973307, + 0.0042769163846969604, + 0.004023652523756027, + 0.0065713366493582726, + 0.023732686415314674, + 0.004386331420391798, + 0.0060844020918011665, + 0.003018307965248823, + 0.004160270560532808, + 0.0090641425922513, + 0.005002228543162346, + 0.017441526055336, + 0.019336959347128868, + 0.00453003216534853, + 0.019085630774497986, + 0.0042962320148944855, + 0.004874631762504578, + 0.035221606492996216, + 0.008101324550807476, + 0.03429684042930603, + 0.006563153583556414, + 0.003347936784848571, + 0.0199128407984972, + 0.0035135955549776554, + 0.0034677148796617985, + 0.004254468716681004, + 0.007912577129900455, + 0.02425859123468399, + 0.004002705216407776, + 0.005474892910569906, + 0.008398164063692093, + 0.004006399307399988, + 0.002829004544764757, + 0.002419254509732127, + 0.007160781417042017, + 0.003259095596149564, + 0.007771384436637163, + 0.0027885509189218283, + 0.013544883579015732, + 0.0035808891989290714, + 0.003856196068227291, + 0.017225787043571472, + 0.0045328461565077305, + 0.004802756477147341, + 0.004392987582832575, + 0.00443124957382679, + 0.004832036793231964, + 0.0039343698881566525, + 0.0038030412979424, + 0.002729782834649086, + 0.0037353639490902424, + 0.0030328764114528894, + 0.02256636694073677, + 0.0038175364024937153, + 0.03447968140244484, + 0.005542600993067026, + 0.004852159880101681, + 0.005162716843187809, + 0.013862685300409794, + 0.010157769545912743, + 0.009611454792320728, + 0.004488965962082148, + 0.012203110381960869, + 0.006986498832702637, + 0.007013109978288412, + 0.006215974222868681, + 0.011407846584916115, + 0.00905992928892374, + 0.05489735305309296, + 0.0593678317964077, + 0.10744744539260864, + 0.03676110506057739, + 0.018987536430358887, + 0.034674521535634995, + 0.04863235726952553, + 0.03294307366013527 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 2, + 7 + ], + "selected_reward": -0.007622324395924807 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "smooth, uninterrupted medium-long shot begins, focusing on a handsome, scaly-faced snake prince. With graceful, silent grace, he meanders along an ancient, rain-soaked stone path. He possesses a sharp jawline and a prominent nose. His long, jet-black hair sways gently with his movements, a few damp strands plastering his forehead. From his cheekbones to his neck, a dense, shimmering emerald-green scales shimmer in the hazy light, reflecting an ever-shifting shimmer. He clads a dark green silk robe, slightly slit at the collar. The soft fabric clings to his muscular torso, rising and falling rhythmically as his powerful serpentine tail glides along the ground in a fluid S-curve. The camera steadily and deliberately approaches him. The background is a narrow alleyway lined with quaint wooden buildings. Red paper lanterns hanging from the eaves provide the only light source. Their warm red glow penetrates the misty air, casting soft, shifting reflections on the damp, uneven stone slabs. As the camera advances, the image gradually tightens into a close-up. His eyes, like golden glass, are filled with profound emotion. The entire scene is enveloped in a tranquil, mysterious atmosphere imbued with the whimsical charm of the East.", + "seed": 11142, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000151/group_8/index_4038_seed_11142_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": -0.0277984575368464, + "vlm_logprob": -0.0277984575368464, + "vlm_yes_no_margin": 4.15, + "vlm_token_credit": -0.0277984575368464, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a handsome man with a scaly face, sharp jawline, and prominent nose?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the character have long jet-black hair and emerald-green scales covering his face and neck?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character wearing a dark green silk robe that is slightly slit at the collar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the character possess a powerful serpentine tail that glides along the ground in a fluid S-curve?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start as a medium-long shot and steadily approach the subject until it becomes a 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": [ + -0.09117648750543594 + ], + "opposite_logprob_tokens": [ + -2.4661765098571777 + ], + "logprob_reward": -0.09117648750543594, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.09117648750543594, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15037156641483307, + 0.04402034357190132, + 0.031080590561032295, + 0.02797931805253029, + 0.04660143703222275, + 0.03746357187628746, + 0.06973806023597717, + 0.0328405424952507, + 0.22152133285999298, + 0.3383832275867462 + ], + "credit_map": [ + 0.016259387135505676, + 0.020914748311042786, + 0.01963367871940136, + 0.015107226558029652, + 0.020335664972662926, + 0.016184933483600616, + 0.02004515938460827, + 0.021890779957175255, + 0.003784645115956664, + 0.005004031118005514, + 0.007550496142357588, + 0.004099362995475531, + 0.005026193801313639, + 0.004063664935529232, + 0.005838056560605764, + 0.008653892204165459, + 0.0028193362522870302, + 0.00454559875652194, + 0.004693526308983564, + 0.0042340634390711784, + 0.003640458919107914, + 0.0028595132753252983, + 0.0035762619227170944, + 0.00471183517947793, + 0.0034829142969101667, + 0.004291512072086334, + 0.003484650980681181, + 0.0029407679103314877, + 0.004973342176526785, + 0.0025343759916722775, + 0.0037228798028081656, + 0.0025488778483122587, + 0.005712465848773718, + 0.005234388634562492, + 0.0068884724751114845, + 0.004911701660603285, + 0.005081513896584511, + 0.003870123764500022, + 0.007462457288056612, + 0.007440314628183842, + 0.004067775793373585, + 0.004895064514130354, + 0.004409012384712696, + 0.003675352083519101, + 0.006409227382391691, + 0.003442821791395545, + 0.005879220087081194, + 0.004685098771005869, + 0.008018985390663147, + 0.006999662611633539, + 0.0077516064047813416, + 0.012032639235258102, + 0.008836885914206505, + 0.008412268944084644, + 0.009704448282718658, + 0.007981574162840843, + 0.0043677375651896, + 0.0028802002780139446, + 0.004341002553701401, + 0.00472598522901535, + 0.0033527466002851725, + 0.003297411371022463, + 0.0053754220716655254, + 0.004500041715800762, + 0.018771357834339142, + 0.02377479523420334, + 0.017897848039865494, + 0.02450079657137394, + 0.035900987684726715, + 0.030137456953525543, + 0.03712552413344383, + 0.03341258689761162, + 0.023984434083104134, + 0.042615167796611786, + 0.05142519250512123, + 0.05589659512042999, + 0.027326395735144615, + 0.03255455568432808, + 0.06392428278923035, + 0.04065662622451782 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 0 + ], + "selected_reward": -0.09117648750543594 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a handsome man with a scaly face, sharp jawline, and prominent nose?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the character have long jet-black hair and emerald-green scales covering his face and neck?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character wearing a dark green silk robe that is slightly slit at the collar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the character possess a powerful serpentine tail that glides along the ground in a fluid S-curve?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start as a medium-long shot and steadily approach the subject until it becomes a 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": [ + -0.005265178624540567 + ], + "opposite_logprob_tokens": [ + -5.255265235900879 + ], + "logprob_reward": -0.005265178624540567, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005265178624540567, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16508366167545319, + 0.034786663949489594, + 0.027275796979665756, + 0.02213977836072445, + 0.03112516552209854, + 0.019889377057552338, + 0.035056136548519135, + 0.05876750499010086, + 0.22053395211696625, + 0.38534191250801086 + ], + "credit_map": [ + 0.01577579230070114, + 0.016476353630423546, + 0.040265731513500214, + 0.01420737523585558, + 0.013382812030613422, + 0.013510877266526222, + 0.014416093938052654, + 0.03704863786697388, + 0.00365421618334949, + 0.006177981849759817, + 0.005331729538738728, + 0.0037862251047044992, + 0.005089092534035444, + 0.0032262008171528578, + 0.004187568090856075, + 0.0033336554188281298, + 0.0034312447533011436, + 0.00585240637883544, + 0.003299971343949437, + 0.0024452023208141327, + 0.0037304179277271032, + 0.0024262878578156233, + 0.0028471339028328657, + 0.003243135754019022, + 0.0022575485054403543, + 0.0020156807731837034, + 0.004302013199776411, + 0.0023929644376039505, + 0.002613767981529236, + 0.0019039707258343697, + 0.002627797657623887, + 0.0040260376408696175, + 0.0037417621351778507, + 0.0031005197670310736, + 0.005311331246048212, + 0.0029811945278197527, + 0.004419153556227684, + 0.0023658808786422014, + 0.0040620421059429646, + 0.00514328433200717, + 0.0026413444429636, + 0.0024776398204267025, + 0.002538334811106324, + 0.0022613529581576586, + 0.0029467009007930756, + 0.0020124095026403666, + 0.002348527777940035, + 0.00266307033598423, + 0.003190188203006983, + 0.0037491575349122286, + 0.0044272346422076225, + 0.007557955104857683, + 0.0047240424901247025, + 0.0033567980863153934, + 0.0038993824273347855, + 0.004151380155235529, + 0.003370343241840601, + 0.0024349549785256386, + 0.017986902967095375, + 0.004047516733407974, + 0.003201043698936701, + 0.004342313390225172, + 0.005040003452450037, + 0.01834443397819996, + 0.013987279497087002, + 0.025696417316794395, + 0.01656237430870533, + 0.021229911595582962, + 0.03684684634208679, + 0.0385558120906353, + 0.03702022135257721, + 0.030635103583335876, + 0.020511692389845848, + 0.024573341012001038, + 0.022476976737380028, + 0.12080380320549011, + 0.024508822709321976, + 0.024360185489058495, + 0.04481745883822441, + 0.10328967869281769 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 0 + ], + "selected_reward": -0.005265178624540567 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a handsome man with a scaly face, sharp jawline, and prominent nose?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the character have long jet-black hair and emerald-green scales covering his face and neck?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character wearing a dark green silk robe that is slightly slit at the collar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the character possess a powerful serpentine tail that glides along the ground in a fluid S-curve?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start as a medium-long shot and steadily approach the subject until it becomes a 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": [ + -0.012536530382931232 + ], + "opposite_logprob_tokens": [ + -4.387536525726318 + ], + "logprob_reward": -0.012536530382931232, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012536530382931232, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.26430660486221313, + 0.055308062583208084, + 0.05127716064453125, + 0.029958246275782585, + 0.0367414690554142, + 0.04290437698364258, + 0.08939999341964722, + 0.10634789615869522, + 0.15905438363552094, + 0.16470177471637726 + ], + "credit_map": [ + 0.026626653969287872, + 0.018892722204327583, + 0.07125097513198853, + 0.01826036535203457, + 0.029185492545366287, + 0.018149789422750473, + 0.01796906441450119, + 0.06397156417369843, + 0.004482717718929052, + 0.006634890101850033, + 0.011921091936528683, + 0.004470341373234987, + 0.006136851850897074, + 0.0041648019105196, + 0.007906734012067318, + 0.009590638801455498, + 0.006386754103004932, + 0.009522815234959126, + 0.0073967608623206615, + 0.004149278625845909, + 0.007361105177551508, + 0.0045027416199445724, + 0.005532519426196814, + 0.006425191182643175, + 0.003342807525768876, + 0.0024933386594057083, + 0.006068090442568064, + 0.0029559284448623657, + 0.0028021736070513725, + 0.002151049906387925, + 0.0030218265019357204, + 0.00712303351610899, + 0.003951167222112417, + 0.003119457047432661, + 0.004891838412731886, + 0.0037832637317478657, + 0.007227228954434395, + 0.00267542595975101, + 0.007839301601052284, + 0.003253794042393565, + 0.007110636681318283, + 0.0064254030585289, + 0.0055378335528075695, + 0.0042913854122161865, + 0.005859116557985544, + 0.0038201988209038973, + 0.004535865969955921, + 0.005323943216353655, + 0.009746928699314594, + 0.011050467379391193, + 0.013554721139371395, + 0.008669569157063961, + 0.009650491178035736, + 0.017323443666100502, + 0.010233331471681595, + 0.009171048179268837, + 0.006342995911836624, + 0.004134466405957937, + 0.03432378172874451, + 0.005763579625636339, + 0.004699603654444218, + 0.00822541769593954, + 0.0066700587049126625, + 0.03618800640106201, + 0.017104284837841988, + 0.014510308392345905, + 0.01617324724793434, + 0.021374870091676712, + 0.01546100527048111, + 0.023028340190649033, + 0.028420237824320793, + 0.022982122376561165, + 0.011954736895859241, + 0.0156722255051136, + 0.014030537568032742, + 0.03681690990924835, + 0.012696769088506699, + 0.012910284101963043, + 0.030451862141489983, + 0.03016846254467964 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 8 + ], + "selected_reward": -0.012536530382931232 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a handsome man with a scaly face, sharp jawline, and prominent nose?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the character have long jet-black hair and emerald-green scales covering his face and neck?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character wearing a dark green silk robe that is slightly slit at the collar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the character possess a powerful serpentine tail that glides along the ground in a fluid S-curve?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start as a medium-long shot and steadily approach the subject until it becomes a 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": [ + -0.023280270397663116 + ], + "opposite_logprob_tokens": [ + -3.773280382156372 + ], + "logprob_reward": -0.023280270397663116, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023280270397663116, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1594768911600113, + 0.07653507590293884, + 0.08785531669855118, + 0.1676747351884842, + 0.058183249086141586, + 0.046132396906614304, + 0.09957767277956009, + 0.04684547334909439, + 0.11352463066577911, + 0.14419448375701904 + ], + "credit_map": [ + 0.016857361420989037, + 0.017447909340262413, + 0.0263332761824131, + 0.013095127418637276, + 0.01797374151647091, + 0.011950626038014889, + 0.028779996559023857, + 0.02703886851668358, + 0.006897288840264082, + 0.008327177725732327, + 0.018284928053617477, + 0.005810871720314026, + 0.006211642175912857, + 0.005668032914400101, + 0.009924668818712234, + 0.01541046891361475, + 0.014459644444286823, + 0.016798511147499084, + 0.012074488215148449, + 0.00907090399414301, + 0.0064617604948580265, + 0.009241493418812752, + 0.006978795398026705, + 0.012769733555614948, + 0.01791522651910782, + 0.012078845873475075, + 0.038049839437007904, + 0.019202444702386856, + 0.01343055535107851, + 0.012615692801773548, + 0.013707233592867851, + 0.04067491367459297, + 0.004730618558824062, + 0.0069653685204684734, + 0.007795269135385752, + 0.009112386032938957, + 0.009871212765574455, + 0.003970279823988676, + 0.010042180307209492, + 0.005695942789316177, + 0.0063123502768576145, + 0.006507578771561384, + 0.006600912194699049, + 0.004625280853360891, + 0.005746117327362299, + 0.004807635210454464, + 0.00558751355856657, + 0.005945011042058468, + 0.012239261530339718, + 0.011094154790043831, + 0.01279949676245451, + 0.014985421672463417, + 0.009029126726090908, + 0.015714561566710472, + 0.010574255138635635, + 0.013141400180757046, + 0.003259810619056225, + 0.0049740965478122234, + 0.007893688045442104, + 0.004263352137058973, + 0.005625269375741482, + 0.00434717396274209, + 0.005575316492468119, + 0.010906772688031197, + 0.013826609589159489, + 0.010583369992673397, + 0.011816095560789108, + 0.018024124205112457, + 0.01009832601994276, + 0.012913502752780914, + 0.018280399963259697, + 0.01798221841454506, + 0.009424710646271706, + 0.013541575521230698, + 0.015568448230624199, + 0.037517473101615906, + 0.008244708180427551, + 0.009201022796332836, + 0.014168924652040005, + 0.036527637392282486 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 9 + ], + "selected_reward": -0.023280270397663116 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a handsome man with a scaly face, sharp jawline, and prominent nose?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the character have long jet-black hair and emerald-green scales covering his face and neck?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character wearing a dark green silk robe that is slightly slit at the collar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the character possess a powerful serpentine tail that glides along the ground in a fluid S-curve?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start as a medium-long shot and steadily approach the subject until it becomes a 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": [ + -0.006733820773661137 + ], + "opposite_logprob_tokens": [ + -5.0067338943481445 + ], + "logprob_reward": -0.006733820773661137, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006733820773661137, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11163730174303055, + 0.049582589417696, + 0.03235233575105667, + 0.13871759176254272, + 0.049570269882678986, + 0.06799933314323425, + 0.11456745117902756, + 0.23122529685497284, + 0.1202431321144104, + 0.0841047614812851 + ], + "credit_map": [ + 0.01297523733228445, + 0.010613782331347466, + 0.022191302850842476, + 0.008773593232035637, + 0.018904950469732285, + 0.00916743092238903, + 0.012724722735583782, + 0.01628626324236393, + 0.005116816144436598, + 0.006008869502693415, + 0.010383013635873795, + 0.0051381089724600315, + 0.005905864294618368, + 0.0037194222677499056, + 0.004907544236630201, + 0.00840294361114502, + 0.004609312862157822, + 0.004205789882689714, + 0.004649566486477852, + 0.002821024041622877, + 0.004998207092285156, + 0.0030070673674345016, + 0.004603104200214148, + 0.0034582577645778656, + 0.014798814430832863, + 0.007099485956132412, + 0.043227292597293854, + 0.006974853575229645, + 0.004487005993723869, + 0.007985861971974373, + 0.004826925694942474, + 0.04931732639670372, + 0.008313809521496296, + 0.005734203848987818, + 0.008220158517360687, + 0.004834301769733429, + 0.0041196225211024284, + 0.004244325682520866, + 0.005899765063077211, + 0.008204078301787376, + 0.012160453014075756, + 0.009051216766238213, + 0.00959179550409317, + 0.006349837873131037, + 0.007713488303124905, + 0.006571799982339144, + 0.006756485905498266, + 0.00980424601584673, + 0.016926737502217293, + 0.011388096027076244, + 0.023563992232084274, + 0.00812993012368679, + 0.011398021131753922, + 0.017928406596183777, + 0.010288710705935955, + 0.014943544752895832, + 0.016913404688239098, + 0.010629240423440933, + 0.07064854353666306, + 0.01390808168798685, + 0.011253353208303452, + 0.01860073022544384, + 0.011267644353210926, + 0.07800427079200745, + 0.013656260445713997, + 0.012371026910841465, + 0.013288210146129131, + 0.013724539428949356, + 0.015950748696923256, + 0.02202674001455307, + 0.013315926305949688, + 0.015909673646092415, + 0.009107275865972042, + 0.005632334388792515, + 0.010354269295930862, + 0.018955739215016365, + 0.005258659366518259, + 0.00949785578995943, + 0.006887845229357481, + 0.018410783261060715 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 3, + 8 + ], + "selected_reward": -0.006733820773661137 + } + ], + "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/0000156/group_0/index_tensor([211, 211])_group_0_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000156/group_0/index_tensor([211, 211])_group_0_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..db4443c711177b451ff5dc8d7ccf67dae1fdb0aa --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000156/group_0/index_tensor([211, 211])_group_0_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "An immersive, slowly moving tracking shot plunges the viewer into the bottomless depths of the inky blue ocean, home to a kingdom of countless translucent, luminescent jellyfish. These creatures' jelly-like bells emanate a steady, soft bioluminescence, predominantly a pale blue hue, occasionally tinged with a soft pink-purple. This light penetrates their translucent bodies, illuminating the delicate contours of their internal structures and casting a dreamy halo against the inky water. They pulsate gracefully with a leisurely rhythm, each contraction and expansion like a gentle heartbeat, steadily propelling them forward through the nearly still water. As their bodies rise and fall, their long, slender, equally luminescent tentacles dance behind them like silken ribbons. The camera smoothly moves among this huge group, sometimes approaching one so that we can clearly see the smooth texture of its bell-shaped surface, and sometimes retreating to show the spectacular sight of hundreds of light points breathing synchronously in the vast darkness. The whole picture is shrouded in an extreme tranquility and mystery, like a silent and magnificent underwater ballet, creating a fascinating, almost hypnotic visual experience.", + "seed": 8980, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000156/group_0/index_211_seed_8980_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": -0.06774864553008228, + "vlm_logprob": -0.06774864553008228, + "vlm_yes_no_margin": 4.6, + "vlm_token_credit": -0.06774864553008228, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 inky blue ocean?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are translucent jellyfish with pale blue and pink-purple bioluminescence visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the jellyfish exhibit a pulsating motion with long, slender tentacles trailing behind them?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera move among the group, alternating between close-ups of individual bells and wide shots of many jellyfish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the water nearly still with no visible currents or turbulence?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00417444808408618 + ], + "opposite_logprob_tokens": [ + -7.50417423248291 + ], + "logprob_reward": -0.00417444808408618, + "yes_no_margin_reward": 7.5, + "token_credit_reward": -0.00417444808408618, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11175716668367386, + 0.08789346367120743, + 0.06495245546102524, + 0.12819522619247437, + 0.11633483320474625, + 0.039027851074934006, + 0.045326367020606995, + 0.11805959790945053, + 0.1752643883228302, + 0.11318867653608322 + ], + "credit_map": [ + 0.012972512282431126, + 0.009262974373996258, + 0.023661935701966286, + 0.021588459610939026, + 0.009101389907300472, + 0.012175027281045914, + 0.010674512945115566, + 0.012320361100137234, + 0.006532737519592047, + 0.006239349488168955, + 0.0042378841899335384, + 0.012438739649951458, + 0.006031289231032133, + 0.018894411623477936, + 0.01042767334729433, + 0.023091377690434456, + 0.006553618237376213, + 0.006082701031118631, + 0.004008694551885128, + 0.006013930309563875, + 0.013835631310939789, + 0.01077713817358017, + 0.007851951755583286, + 0.009828791953623295, + 0.009709011763334274, + 0.02192923054099083, + 0.01322892028838396, + 0.011011750437319279, + 0.021922552958130836, + 0.01585620455443859, + 0.019821681082248688, + 0.014715870842337608, + 0.009760405868291855, + 0.00900792796164751, + 0.018478546291589737, + 0.018859872594475746, + 0.01420772448182106, + 0.028235722333192825, + 0.007334992755204439, + 0.010449642315506935, + 0.004452559165656567, + 0.004793059080839157, + 0.00749510433524847, + 0.004966103006154299, + 0.006040281616151333, + 0.004902172833681107, + 0.0025647066067904234, + 0.0038138634990900755, + 0.008426728658378124, + 0.005592834670096636, + 0.00558488629758358, + 0.008926420472562313, + 0.004358104430139065, + 0.003756400430575013, + 0.004646114073693752, + 0.004034882411360741, + 0.015345425345003605, + 0.015344376675784588, + 0.022736066952347755, + 0.016920963302254677, + 0.010660097934305668, + 0.016288841143250465, + 0.00902851577848196, + 0.011735322885215282, + 0.008368278853595257, + 0.009454515762627125, + 0.009696288965642452, + 0.029779313132166862, + 0.009876864962279797, + 0.006466052494943142, + 0.063548743724823, + 0.0380743183195591, + 0.01109708845615387, + 0.01994366943836212, + 0.018320471048355103, + 0.017520977184176445, + 0.012517853640019894, + 0.012757121585309505, + 0.011838901787996292, + 0.009192590601742268 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 3, + 7 + ], + "selected_reward": -0.00417444808408618 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 inky blue ocean?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are translucent jellyfish with pale blue and pink-purple bioluminescence visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the jellyfish exhibit a pulsating motion with long, slender tentacles trailing behind them?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera move among the group, alternating between close-ups of individual bells and wide shots of many jellyfish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the water nearly still with no visible currents or turbulence?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.001522692502476275 + ], + "opposite_logprob_tokens": [ + -6.501522541046143 + ], + "logprob_reward": -0.001522692502476275, + "yes_no_margin_reward": 6.5, + "token_credit_reward": -0.001522692502476275, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10796666890382767, + 0.07915787398815155, + 0.06281919777393341, + 0.22055640816688538, + 0.061330098658800125, + 0.049362510442733765, + 0.11850310862064362, + 0.058181606233119965, + 0.16736747324466705, + 0.07475502043962479 + ], + "credit_map": [ + 0.012953348457813263, + 0.00851376261562109, + 0.01777108944952488, + 0.017039960250258446, + 0.00872140284627676, + 0.010984056629240513, + 0.008095251396298409, + 0.023887792602181435, + 0.006960235070437193, + 0.008873037062585354, + 0.010359682142734528, + 0.014044051058590412, + 0.008438060991466045, + 0.00873810239136219, + 0.011056733317673206, + 0.010687977075576782, + 0.006427220068871975, + 0.006166811566799879, + 0.004031454212963581, + 0.009700368158519268, + 0.012190726585686207, + 0.008390640839934349, + 0.007205126341432333, + 0.00870684813708067, + 0.01614067330956459, + 0.04044158011674881, + 0.01974552683532238, + 0.019934650510549545, + 0.0391022264957428, + 0.026496674865484238, + 0.035755373537540436, + 0.02293970249593258, + 0.0044708154164254665, + 0.004357139114290476, + 0.005225600674748421, + 0.011738177388906479, + 0.0072407484985888, + 0.017923695966601372, + 0.00425131618976593, + 0.006122606340795755, + 0.007605277933180332, + 0.008162328042089939, + 0.008731739595532417, + 0.007193801924586296, + 0.004869355354458094, + 0.0051268320530653, + 0.0032535092905163765, + 0.0044196657836437225, + 0.014965886250138283, + 0.010122447274625301, + 0.012737972661852837, + 0.03626476973295212, + 0.009881442412734032, + 0.005422502756118774, + 0.01831820420920849, + 0.01078987866640091, + 0.006270560435950756, + 0.007143767084926367, + 0.011100229807198048, + 0.007678428664803505, + 0.005529571790248156, + 0.009447888471186161, + 0.0055876728147268295, + 0.005423489958047867, + 0.009552299976348877, + 0.00880552176386118, + 0.005635987035930157, + 0.015696244314312935, + 0.006776315625756979, + 0.007334108930081129, + 0.06303518265485764, + 0.05053180456161499, + 0.0059001208283007145, + 0.012300974689424038, + 0.010392031632363796, + 0.017149409279227257, + 0.007723240647464991, + 0.007694258354604244, + 0.00861993245780468, + 0.004975060001015663 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 8, + 6 + ], + "selected_reward": -0.001522692502476275 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 inky blue ocean?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are translucent jellyfish with pale blue and pink-purple bioluminescence visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the jellyfish exhibit a pulsating motion with long, slender tentacles trailing behind them?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera move among the group, alternating between close-ups of individual bells and wide shots of many jellyfish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the water nearly still with no visible currents or turbulence?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008643240667879581 + ], + "opposite_logprob_tokens": [ + -4.75864315032959 + ], + "logprob_reward": -0.008643240667879581, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008643240667879581, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1408662050962448, + 0.09602267295122147, + 0.13046298921108246, + 0.06171233206987381, + 0.09829164296388626, + 0.13327547907829285, + 0.04974593594670296, + 0.04711885005235672, + 0.17185227572917938, + 0.07065164297819138 + ], + "credit_map": [ + 0.02335314080119133, + 0.010222508572041988, + 0.026860784739255905, + 0.02886289544403553, + 0.00830751471221447, + 0.012944119051098824, + 0.011582025326788425, + 0.018733207136392593, + 0.008674557320773602, + 0.008943919092416763, + 0.011364420875906944, + 0.012427540495991707, + 0.00952974148094654, + 0.01239696703851223, + 0.015137589536607265, + 0.017547931522130966, + 0.00784775149077177, + 0.013849707320332527, + 0.007071599829941988, + 0.011840143240988255, + 0.037789423018693924, + 0.017158987000584602, + 0.01333507988601923, + 0.021570298820734024, + 0.007052894681692123, + 0.008835822343826294, + 0.010315421968698502, + 0.005979839246720076, + 0.009700736962258816, + 0.007533467840403318, + 0.006136246956884861, + 0.0061578997410833836, + 0.0066002546809613705, + 0.007102424278855324, + 0.009705871343612671, + 0.018703123554587364, + 0.010403874330222607, + 0.0285809226334095, + 0.006390093360096216, + 0.01080507505685091, + 0.017740992829203606, + 0.011819462291896343, + 0.027889326214790344, + 0.01974368840456009, + 0.01715194061398506, + 0.013174459338188171, + 0.009222371503710747, + 0.016533223912119865, + 0.004963853862136602, + 0.0035313067492097616, + 0.007102619856595993, + 0.010159778408706188, + 0.004461970645934343, + 0.0029265224002301693, + 0.0061825416050851345, + 0.010417338460683823, + 0.005651745479553938, + 0.0056112962774932384, + 0.010189300402998924, + 0.006317277438938618, + 0.004392051603645086, + 0.0062811654061079025, + 0.004722402431070805, + 0.003953611943870783, + 0.00583350146189332, + 0.006143474485725164, + 0.00488723861053586, + 0.04005184397101402, + 0.00555290374904871, + 0.0059694950468838215, + 0.07125255465507507, + 0.03216124325990677, + 0.007330434396862984, + 0.010015455074608326, + 0.011905338615179062, + 0.011002381332218647, + 0.008062899112701416, + 0.007484819274395704, + 0.008008473552763462, + 0.006841838359832764 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 5 + ], + "selected_reward": -0.008643240667879581 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 inky blue ocean?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are translucent jellyfish with pale blue and pink-purple bioluminescence visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the jellyfish exhibit a pulsating motion with long, slender tentacles trailing behind them?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera move among the group, alternating between close-ups of individual bells and wide shots of many jellyfish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the water nearly still with no visible currents or turbulence?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04306042939424515 + ], + "opposite_logprob_tokens": [ + -3.168060541152954 + ], + "logprob_reward": -0.04306042939424515, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04306042939424515, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10634306818246841, + 0.05437146872282028, + 0.08030189573764801, + 0.16760148108005524, + 0.07794756442308426, + 0.035685665905475616, + 0.06242326647043228, + 0.08972594887018204, + 0.1891830712556839, + 0.13641655445098877 + ], + "credit_map": [ + 0.020837193354964256, + 0.009415174834430218, + 0.02913576364517212, + 0.014947542920708656, + 0.005989146884530783, + 0.00723454263061285, + 0.008233373053371906, + 0.010550329461693764, + 0.004154540598392487, + 0.005801524966955185, + 0.003132664365693927, + 0.0065899319015443325, + 0.00507635110989213, + 0.008729944005608559, + 0.00579654797911644, + 0.015089965425431728, + 0.008694907650351524, + 0.006507184822112322, + 0.0046452987007796764, + 0.011797986924648285, + 0.01661796309053898, + 0.01028288621455431, + 0.008647183887660503, + 0.01310847606509924, + 0.012783084064722061, + 0.032334402203559875, + 0.014562422409653664, + 0.015234588645398617, + 0.0269115399569273, + 0.020470552146434784, + 0.028113223612308502, + 0.017191674560308456, + 0.006935809273272753, + 0.006192858796566725, + 0.013579662889242172, + 0.010838737711310387, + 0.009162642061710358, + 0.01171971671283245, + 0.006995280273258686, + 0.012522859498858452, + 0.0045278011821210384, + 0.002737491624429822, + 0.006080778315663338, + 0.00432294886559248, + 0.004942784551531076, + 0.0034499962348490953, + 0.0025719953700900078, + 0.007051871623843908, + 0.008618823252618313, + 0.005870303139090538, + 0.007070114836096764, + 0.009699095040559769, + 0.006823924835771322, + 0.006166346836835146, + 0.009162353351712227, + 0.009012306109070778, + 0.006022463086992502, + 0.006776845082640648, + 0.020970942452549934, + 0.012553675100207329, + 0.011013108305633068, + 0.014035585336387157, + 0.010704821906983852, + 0.007648509927093983, + 0.009114647284150124, + 0.01593398116528988, + 0.01199014950543642, + 0.02539377100765705, + 0.01145616453140974, + 0.012286252342164516, + 0.06309761852025986, + 0.03991048410534859, + 0.0115280132740736, + 0.0234551802277565, + 0.016641288995742798, + 0.03330802917480469, + 0.011625885963439941, + 0.011603865772485733, + 0.019065944477915764, + 0.009188340976834297 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 3, + 9 + ], + "selected_reward": -0.04306042939424515 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 inky blue ocean?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are translucent jellyfish with pale blue and pink-purple bioluminescence visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the jellyfish exhibit a pulsating motion with long, slender tentacles trailing behind them?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera move among the group, alternating between close-ups of individual bells and wide shots of many jellyfish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the water nearly still with no visible currents or turbulence?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.28134241700172424 + ], + "opposite_logprob_tokens": [ + -1.4063423871994019 + ], + "logprob_reward": -0.28134241700172424, + "yes_no_margin_reward": 1.125, + "token_credit_reward": -0.28134241700172424, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17689552903175354, + 0.11335913836956024, + 0.16678908467292786, + 0.10945291072130203, + 0.07632529735565186, + 0.09841800481081009, + 0.05993069335818291, + 0.07429726421833038, + 0.07573547959327698, + 0.04879656434059143 + ], + "credit_map": [ + 0.022776294499635696, + 0.010600419715046883, + 0.039663806557655334, + 0.03894735500216484, + 0.0062535610049963, + 0.008082754909992218, + 0.011915233917534351, + 0.03865610808134079, + 0.00906684622168541, + 0.01298051979392767, + 0.010262856259942055, + 0.021836385130882263, + 0.01092575117945671, + 0.01911384053528309, + 0.009046671912074089, + 0.02012627199292183, + 0.015640532597899437, + 0.01950715109705925, + 0.008245043456554413, + 0.028018493205308914, + 0.02827351726591587, + 0.023106232285499573, + 0.020694727078080177, + 0.023303387686610222, + 0.00752750039100647, + 0.01968400739133358, + 0.010259361937642097, + 0.011188121512532234, + 0.014448203146457672, + 0.013466906733810902, + 0.01975671947002411, + 0.013122095726430416, + 0.004262612201273441, + 0.005563623737543821, + 0.006108919158577919, + 0.013156143017113209, + 0.011663881130516529, + 0.018886778503656387, + 0.006037542130798101, + 0.010645795613527298, + 0.011530116200447083, + 0.012586081400513649, + 0.015551872551441193, + 0.015172049403190613, + 0.01322487834841013, + 0.0138719966635108, + 0.006622759159654379, + 0.009858246892690659, + 0.006789174396544695, + 0.005083276890218258, + 0.0068669100292027, + 0.009912446141242981, + 0.00578183913603425, + 0.006154157686978579, + 0.008640206418931484, + 0.010702679865062237, + 0.008040274493396282, + 0.008526758290827274, + 0.01055915467441082, + 0.012354299426078796, + 0.009433230385184288, + 0.01440229918807745, + 0.006298546679317951, + 0.0046827043406665325, + 0.004164157435297966, + 0.0034717193339020014, + 0.003711791941896081, + 0.013732215389609337, + 0.004125586245208979, + 0.003830079222097993, + 0.026303814724087715, + 0.016396114602684975, + 0.004771212115883827, + 0.009338462725281715, + 0.006455516908317804, + 0.007648266851902008, + 0.005112844984978437, + 0.005084368400275707, + 0.00620828615501523, + 0.00417760806158185 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 1 + ], + "selected_reward": -0.28134241700172424 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An immersive, slowly moving tracking shot plunges the viewer into the bottomless depths of the inky blue ocean, home to a kingdom of countless translucent, luminescent jellyfish. These creatures' jelly-like bells emanate a steady, soft bioluminescence, predominantly a pale blue hue, occasionally tinged with a soft pink-purple. This light penetrates their translucent bodies, illuminating the delicate contours of their internal structures and casting a dreamy halo against the inky water. They pulsate gracefully with a leisurely rhythm, each contraction and expansion like a gentle heartbeat, steadily propelling them forward through the nearly still water. As their bodies rise and fall, their long, slender, equally luminescent tentacles dance behind them like silken ribbons. The camera smoothly moves among this huge group, sometimes approaching one so that we can clearly see the smooth texture of its bell-shaped surface, and sometimes retreating to show the spectacular sight of hundreds of light points breathing synchronously in the vast darkness. The whole picture is shrouded in an extreme tranquility and mystery, like a silent and magnificent underwater ballet, creating a fascinating, almost hypnotic visual experience.", + "seed": 8981, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000156/group_0/index_211_seed_8981_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": -0.0521246389602311, + "vlm_logprob": -0.0521246389602311, + "vlm_yes_no_margin": 4.425, + "vlm_token_credit": -0.0521246389602311, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 inky blue ocean?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are translucent jellyfish with pale blue and pink-purple bioluminescence visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the jellyfish exhibit a pulsating motion with long, slender tentacles trailing behind them?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera move among the group, alternating between close-ups of individual bells and wide shots of many jellyfish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the water nearly still with no visible currents or turbulence?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005816557444632053 + ], + "opposite_logprob_tokens": [ + -6.130816459655762 + ], + "logprob_reward": -0.005816557444632053, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.005816557444632053, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10577484965324402, + 0.04920045658946037, + 0.04397350177168846, + 0.05120384693145752, + 0.14508946239948273, + 0.06507638096809387, + 0.10908404737710953, + 0.055326223373413086, + 0.05792535841464996, + 0.31734591722488403 + ], + "credit_map": [ + 0.014568001963198185, + 0.011724462732672691, + 0.014043096452951431, + 0.014727793633937836, + 0.014280030503869057, + 0.009241008199751377, + 0.01258314959704876, + 0.014607304707169533, + 0.004993046168237925, + 0.005053122993558645, + 0.004892697092145681, + 0.006130589172244072, + 0.00493217445909977, + 0.004963607992976904, + 0.012172351591289043, + 0.006062863860279322, + 0.0030372601468116045, + 0.00905236229300499, + 0.004220036324113607, + 0.006075182929635048, + 0.0029713346157222986, + 0.0035409971605986357, + 0.010921705514192581, + 0.004154616966843605, + 0.00733585562556982, + 0.005254603922367096, + 0.007942315191030502, + 0.013388419523835182, + 0.002789756515994668, + 0.00304851564578712, + 0.0072816647589206696, + 0.004162712953984737, + 0.010370193049311638, + 0.012782005593180656, + 0.026713911443948746, + 0.01844891905784607, + 0.01905851997435093, + 0.019869985058903694, + 0.011245940811932087, + 0.026599979028105736, + 0.0054319752380251884, + 0.006376130040735006, + 0.017267294228076935, + 0.005045751109719276, + 0.0041376203298568726, + 0.004720140714198351, + 0.0028558121994137764, + 0.01924165152013302, + 0.011931437999010086, + 0.013864428736269474, + 0.01754920184612274, + 0.014304406009614468, + 0.015761680901050568, + 0.012504508718848228, + 0.013651111163198948, + 0.009517263621091843, + 0.006383388768881559, + 0.008270053192973137, + 0.008904462680220604, + 0.006766776088625193, + 0.005753207951784134, + 0.005622401367872953, + 0.006355184130370617, + 0.00727074034512043, + 0.005379443522542715, + 0.0065435427241027355, + 0.007135099731385708, + 0.009013490751385689, + 0.004735090769827366, + 0.007593432441353798, + 0.006809194106608629, + 0.010716057382524014, + 0.009928109124302864, + 0.009024228900671005, + 0.01213851012289524, + 0.00812874361872673, + 0.010664508678019047, + 0.11763810366392136, + 0.13206593692302704, + 0.017757749184966087 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 4, + 6 + ], + "selected_reward": -0.005816557444632053 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 inky blue ocean?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are translucent jellyfish with pale blue and pink-purple bioluminescence visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the jellyfish exhibit a pulsating motion with long, slender tentacles trailing behind them?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera move among the group, alternating between close-ups of individual bells and wide shots of many jellyfish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the water nearly still with no visible currents or turbulence?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.000935950840357691 + ], + "opposite_logprob_tokens": [ + -7.000936031341553 + ], + "logprob_reward": -0.000935950840357691, + "yes_no_margin_reward": 7.0, + "token_credit_reward": -0.000935950840357691, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0948973149061203, + 0.10650854557752609, + 0.07806284725666046, + 0.05402752757072449, + 0.039454057812690735, + 0.10862097889184952, + 0.0435604453086853, + 0.03829827904701233, + 0.12806318700313568, + 0.3085067868232727 + ], + "credit_map": [ + 0.01593562215566635, + 0.012001359835267067, + 0.013071347959339619, + 0.011263738386332989, + 0.009705841541290283, + 0.007025857921689749, + 0.012347865849733353, + 0.013545680791139603, + 0.010944406501948833, + 0.009239647537469864, + 0.01338998507708311, + 0.013130436651408672, + 0.005074255168437958, + 0.006224243436008692, + 0.03524250537157059, + 0.013263064436614513, + 0.004401958081871271, + 0.028184132650494576, + 0.003292471868917346, + 0.004880796652287245, + 0.0041846283711493015, + 0.0049019125290215015, + 0.023955723270773888, + 0.004261218477040529, + 0.006969761103391647, + 0.008803990669548512, + 0.009794449433684349, + 0.008133651688694954, + 0.0036637294106185436, + 0.004464485216885805, + 0.004815683234483004, + 0.0073817758820950985, + 0.004612863063812256, + 0.006472745444625616, + 0.00536963390186429, + 0.005822124425321817, + 0.0037829477805644274, + 0.004650401882827282, + 0.00470758555456996, + 0.00403575599193573, + 0.012245521880686283, + 0.01620156690478325, + 0.017561107873916626, + 0.012557004578411579, + 0.01118812058120966, + 0.011556234210729599, + 0.006912296172231436, + 0.020399130880832672, + 0.005828011780977249, + 0.005909624043852091, + 0.004962204024195671, + 0.004432880785316229, + 0.009178628213703632, + 0.004071309696882963, + 0.0047157215885818005, + 0.004462066106498241, + 0.003253954229876399, + 0.006995814852416515, + 0.00676753232255578, + 0.005780529696494341, + 0.004741572309285402, + 0.0034594861790537834, + 0.003726147348061204, + 0.0035732395481318235, + 0.011877939105033875, + 0.01679460145533085, + 0.01391944196075201, + 0.017105212435126305, + 0.011262062005698681, + 0.013755193911492825, + 0.019684603437781334, + 0.02366412803530693, + 0.01042250543832779, + 0.010724921710789204, + 0.0099039850756526, + 0.00911011267453432, + 0.014184434898197651, + 0.1178397685289383, + 0.12267500907182693, + 0.013646041974425316 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 5 + ], + "selected_reward": -0.000935950840357691 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 inky blue ocean?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are translucent jellyfish with pale blue and pink-purple bioluminescence visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the jellyfish exhibit a pulsating motion with long, slender tentacles trailing behind them?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera move among the group, alternating between close-ups of individual bells and wide shots of many jellyfish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the water nearly still with no visible currents or turbulence?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014194360002875328 + ], + "opposite_logprob_tokens": [ + -4.264194488525391 + ], + "logprob_reward": -0.014194360002875328, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014194360002875328, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10032261162996292, + 0.20424045622348785, + 0.0372038260102272, + 0.029714012518525124, + 0.036383867263793945, + 0.0680554136633873, + 0.06306634843349457, + 0.04420217126607895, + 0.1275051236152649, + 0.2893061935901642 + ], + "credit_map": [ + 0.017725706100463867, + 0.009176747873425484, + 0.015092405490577221, + 0.014520470052957535, + 0.00814991258084774, + 0.008127536624670029, + 0.015351497568190098, + 0.012178332544863224, + 0.02042914740741253, + 0.016293196007609367, + 0.01952330581843853, + 0.019748888909816742, + 0.010314786806702614, + 0.009443629533052444, + 0.07741545885801315, + 0.031072037294507027, + 0.005473138298839331, + 0.008460737764835358, + 0.002987097715958953, + 0.004078766796737909, + 0.005159182008355856, + 0.0034968345426023006, + 0.0039965989999473095, + 0.003551467787474394, + 0.0034491082187741995, + 0.003945428878068924, + 0.00474386615678668, + 0.00726255914196372, + 0.0017688520019873977, + 0.002277631778270006, + 0.0029182960279285908, + 0.003348269732668996, + 0.0029695702251046896, + 0.0036870185285806656, + 0.005380110815167427, + 0.0053146518766880035, + 0.004994596820324659, + 0.0047817532904446125, + 0.00512712262570858, + 0.004129044711589813, + 0.006993764545768499, + 0.008984824642539024, + 0.009494313970208168, + 0.007985546253621578, + 0.010935891419649124, + 0.008835089392960072, + 0.0055258609354496, + 0.009300129488110542, + 0.008786235004663467, + 0.008799356408417225, + 0.006953652016818523, + 0.00583239272236824, + 0.013387237675487995, + 0.007139135617762804, + 0.005677311215549707, + 0.00649102870374918, + 0.0034967863466590643, + 0.009050332941114902, + 0.007164178881794214, + 0.007225572597235441, + 0.005115579813718796, + 0.002845786279067397, + 0.004869278986006975, + 0.0044346521608531475, + 0.01177685521543026, + 0.0175099465996027, + 0.014662680216133595, + 0.017682572826743126, + 0.011171434074640274, + 0.01375812478363514, + 0.019738800823688507, + 0.021204711869359016, + 0.00799680408090353, + 0.007161607034504414, + 0.008833627216517925, + 0.006805714685469866, + 0.010277213528752327, + 0.11283230036497116, + 0.11934686452150345, + 0.016052039340138435 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 1, + 8 + ], + "selected_reward": -0.014194360002875328 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 inky blue ocean?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are translucent jellyfish with pale blue and pink-purple bioluminescence visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the jellyfish exhibit a pulsating motion with long, slender tentacles trailing behind them?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera move among the group, alternating between close-ups of individual bells and wide shots of many jellyfish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the water nearly still with no visible currents or turbulence?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"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": [ + 0.052222494035959244, + 0.08728863298892975, + 0.23346494138240814, + 0.05724311247467995, + 0.05343705788254738, + 0.10052712261676788, + 0.05358801409602165, + 0.0504770502448082, + 0.2220296859741211, + 0.08972186595201492 + ], + "credit_map": [ + 0.009417325258255005, + 0.005997648928314447, + 0.006664206273853779, + 0.007187035400420427, + 0.003269644919782877, + 0.0034765962045639753, + 0.010784663259983063, + 0.005425369367003441, + 0.010561653412878513, + 0.008868305943906307, + 0.00939153041690588, + 0.013597683981060982, + 0.006204748060554266, + 0.005738078150898218, + 0.024349259212613106, + 0.008577374741435051, + 0.016941385343670845, + 0.0898195430636406, + 0.010910128243267536, + 0.017719637602567673, + 0.009090382605791092, + 0.012397090904414654, + 0.0635586753487587, + 0.013028089888393879, + 0.00897353794425726, + 0.006922407075762749, + 0.00974730122834444, + 0.009265962988138199, + 0.004119098652154207, + 0.004888923838734627, + 0.008556424640119076, + 0.0047694542445242405, + 0.0048096273094415665, + 0.00636129779741168, + 0.007613415829837322, + 0.007812127470970154, + 0.005529047921299934, + 0.005978364497423172, + 0.005749222822487354, + 0.009583953768014908, + 0.005858056712895632, + 0.005387288052588701, + 0.033654239028692245, + 0.0055291298776865005, + 0.00460836524143815, + 0.005143890157341957, + 0.004063121974468231, + 0.03628303110599518, + 0.006561168935149908, + 0.007871819660067558, + 0.005946049932390451, + 0.006025753449648619, + 0.008327116258442402, + 0.00652721244841814, + 0.006551555823534727, + 0.0057773408479988575, + 0.0035998583771288395, + 0.008027992211282253, + 0.008214333094656467, + 0.006350185256451368, + 0.005809597671031952, + 0.004597228951752186, + 0.004815796390175819, + 0.00906206015497446, + 0.019815439358353615, + 0.02545440010726452, + 0.024111144244670868, + 0.028405101969838142, + 0.018264848738908768, + 0.024622034281492233, + 0.035023968666791916, + 0.04633275792002678, + 0.006962296087294817, + 0.007435146253556013, + 0.005917938891798258, + 0.004342145752161741, + 0.005171196535229683, + 0.0284983292222023, + 0.024696549400687218, + 0.006698267068713903 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 8, + 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\": \"Is the scene set in an inky blue ocean?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are translucent jellyfish with pale blue and pink-purple bioluminescence visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the jellyfish exhibit a pulsating motion with long, slender tentacles trailing behind them?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera move among the group, alternating between close-ups of individual bells and wide shots of many jellyfish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the water nearly still with no visible currents or turbulence?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.20157764852046967 + ], + "opposite_logprob_tokens": [ + -1.7015776634216309 + ], + "logprob_reward": -0.20157764852046967, + "yes_no_margin_reward": 1.5, + "token_credit_reward": -0.20157764852046967, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1887391209602356, + 0.18158555030822754, + 0.08124189078807831, + 0.04089073836803436, + 0.12464691698551178, + 0.03918052464723587, + 0.06875927746295929, + 0.050037283450365067, + 0.061726611107587814, + 0.16319210827350616 + ], + "credit_map": [ + 0.020705141127109528, + 0.01702403835952282, + 0.019684771075844765, + 0.02517491765320301, + 0.011522973887622356, + 0.012408704496920109, + 0.04426061362028122, + 0.03795795515179634, + 0.01586901769042015, + 0.013336438685655594, + 0.01615431345999241, + 0.01567075587809086, + 0.006926048547029495, + 0.006906424183398485, + 0.0705384612083435, + 0.03618409112095833, + 0.005887317005544901, + 0.029350943863391876, + 0.004049739334732294, + 0.006841158494353294, + 0.0038374813739210367, + 0.0032988903112709522, + 0.0240433719009161, + 0.003932987339794636, + 0.004671208560466766, + 0.006795964669436216, + 0.006832895800471306, + 0.0063104466535151005, + 0.002753258915618062, + 0.0023048981092870235, + 0.007278598845005035, + 0.003943464253097773, + 0.01079624518752098, + 0.012099321000277996, + 0.023532480001449585, + 0.01975010521709919, + 0.012356976978480816, + 0.014441169798374176, + 0.009827188216149807, + 0.02184343710541725, + 0.0051229652017354965, + 0.0051449197344481945, + 0.004703837912529707, + 0.005925594363361597, + 0.004414206370711327, + 0.0027794879861176014, + 0.00362962344661355, + 0.007459887303411961, + 0.00595458596944809, + 0.00978582538664341, + 0.010879452340304852, + 0.009265796281397343, + 0.008237958885729313, + 0.008371873758733273, + 0.00906447321176529, + 0.0071993074379861355, + 0.00373582704924047, + 0.006793368607759476, + 0.01074095070362091, + 0.008579345420002937, + 0.005149570293724537, + 0.005340162664651871, + 0.006296329665929079, + 0.003401731373742223, + 0.006300162989646196, + 0.006828530225902796, + 0.005898178089410067, + 0.007891523651778698, + 0.005288025364279747, + 0.0074315122328698635, + 0.010176288895308971, + 0.0119123924523592, + 0.004517874680459499, + 0.003870363114401698, + 0.00545844342559576, + 0.00368850352242589, + 0.005596351809799671, + 0.06348788738250732, + 0.0678027868270874, + 0.008769883774220943 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 9 + ], + "selected_reward": -0.20157764852046967 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An immersive, slowly moving tracking shot plunges the viewer into the bottomless depths of the inky blue ocean, home to a kingdom of countless translucent, luminescent jellyfish. These creatures' jelly-like bells emanate a steady, soft bioluminescence, predominantly a pale blue hue, occasionally tinged with a soft pink-purple. This light penetrates their translucent bodies, illuminating the delicate contours of their internal structures and casting a dreamy halo against the inky water. They pulsate gracefully with a leisurely rhythm, each contraction and expansion like a gentle heartbeat, steadily propelling them forward through the nearly still water. As their bodies rise and fall, their long, slender, equally luminescent tentacles dance behind them like silken ribbons. The camera smoothly moves among this huge group, sometimes approaching one so that we can clearly see the smooth texture of its bell-shaped surface, and sometimes retreating to show the spectacular sight of hundreds of light points breathing synchronously in the vast darkness. The whole picture is shrouded in an extreme tranquility and mystery, like a silent and magnificent underwater ballet, creating a fascinating, almost hypnotic visual experience.", + "seed": 8982, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000156/group_0/index_211_seed_8982_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": -0.05799343234393746, + "vlm_logprob": -0.05799343234393746, + "vlm_yes_no_margin": 4.575, + "vlm_token_credit": -0.05799343234393746, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 inky blue ocean?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are translucent jellyfish with pale blue and pink-purple bioluminescence visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the jellyfish exhibit a pulsating motion with long, slender tentacles trailing behind them?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera move among the group, alternating between close-ups of individual bells and wide shots of many jellyfish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the water nearly still with no visible currents or turbulence?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004331611096858978 + ], + "opposite_logprob_tokens": [ + -7.254331588745117 + ], + "logprob_reward": -0.004331611096858978, + "yes_no_margin_reward": 7.25, + "token_credit_reward": -0.004331611096858978, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.020079288631677628, + 0.805325984954834, + 0.02349504642188549, + 0.014368784613907337, + 0.01995505392551422, + 0.022436155006289482, + 0.05298076197504997, + 0.00805389042943716, + 0.017392082139849663, + 0.015913013368844986 + ], + "credit_map": [ + 0.001629573991522193, + 0.0015445190947502851, + 0.0030554598197340965, + 0.003482573200017214, + 0.002447261242195964, + 0.0018972450634464622, + 0.003467832924798131, + 0.0025548238772898912, + 0.014681111089885235, + 0.02111802063882351, + 0.04044202342629433, + 0.31645211577415466, + 0.026929782703518867, + 0.0286849532276392, + 0.3278588652610779, + 0.029159052297472954, + 0.0014964165166020393, + 0.0029349052347242832, + 0.0035727715585380793, + 0.002396403346210718, + 0.002552533056586981, + 0.0021967224311083555, + 0.004482158925384283, + 0.003863136749714613, + 0.0012905801413580775, + 0.0016026324592530727, + 0.0031790623906999826, + 0.001946469652466476, + 0.0016097555635496974, + 0.001507620676420629, + 0.0017443796386942267, + 0.001488285488449037, + 0.0015631510177627206, + 0.00308457319624722, + 0.002563886111602187, + 0.002729312051087618, + 0.003014654153957963, + 0.0021271270234137774, + 0.0025781996082514524, + 0.0022941508796066046, + 0.0017533154459670186, + 0.0028310632333159447, + 0.003360728733241558, + 0.0015398758696392179, + 0.0062335277907550335, + 0.0024493569508194923, + 0.0025864485651254654, + 0.0016818371368572116, + 0.0026340922340750694, + 0.001279365853406489, + 0.01929553411900997, + 0.0024734956678003073, + 0.018374357372522354, + 0.002502809977158904, + 0.0030235315207391977, + 0.003397575579583645, + 0.0006077757570892572, + 0.0004614011850208044, + 0.0014121190179139376, + 0.000874900259077549, + 0.0012253724271431565, + 0.000756327121052891, + 0.0017467224970459938, + 0.0009692718740552664, + 0.0015225645620375872, + 0.0016841687029227614, + 0.001891318941488862, + 0.002156652510166168, + 0.0019022433552891016, + 0.0020331856794655323, + 0.00240454007871449, + 0.0037974086590111256, + 0.001177811878733337, + 0.0014647794887423515, + 0.00441680196672678, + 0.0022026351653039455, + 0.001641555456444621, + 0.0011702480260282755, + 0.0020947849843651056, + 0.0017443968681618571 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 6, + 2 + ], + "selected_reward": -0.004331611096858978 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 inky blue ocean?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are translucent jellyfish with pale blue and pink-purple bioluminescence visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the jellyfish exhibit a pulsating motion with long, slender tentacles trailing behind them?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera move among the group, alternating between close-ups of individual bells and wide shots of many jellyfish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the water nearly still with no visible currents or turbulence?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0009323778795078397 + ], + "opposite_logprob_tokens": [ + -7.000932216644287 + ], + "logprob_reward": -0.0009323778795078397, + "yes_no_margin_reward": 7.0, + "token_credit_reward": -0.0009323778795078397, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11567426472902298, + 0.20136021077632904, + 0.032937273383140564, + 0.02221606858074665, + 0.020250368863344193, + 0.03373211249709129, + 0.5480329394340515, + 0.006573086604475975, + 0.006007423158735037, + 0.013216249644756317 + ], + "credit_map": [ + 0.008227459155023098, + 0.010475605726242065, + 0.01372659020125866, + 0.01572420634329319, + 0.025576705113053322, + 0.013633894734084606, + 0.01665998436510563, + 0.01164981909096241, + 0.003970331512391567, + 0.005640972405672073, + 0.010545308701694012, + 0.07926501333713531, + 0.006309655029326677, + 0.006761210039258003, + 0.08162044733762741, + 0.007247260306030512, + 0.002138121286407113, + 0.00438857963308692, + 0.004162230994552374, + 0.003932211548089981, + 0.0032828599214553833, + 0.00335554126650095, + 0.005754244048148394, + 0.0059234825894236565, + 0.0019357787678018212, + 0.0024289365392178297, + 0.005601275246590376, + 0.0023095321375876665, + 0.0024289891589432955, + 0.0023106560111045837, + 0.0029474180191755295, + 0.0022534828167408705, + 0.0017508991295471787, + 0.00413860660046339, + 0.002637740457430482, + 0.0026598649565130472, + 0.0031001137103885412, + 0.001854793750680983, + 0.0025480843614786863, + 0.0015602660132572055, + 0.0025905733928084373, + 0.004530701320618391, + 0.00520219886675477, + 0.0021009508054703474, + 0.009544767439365387, + 0.0031121314968913794, + 0.004266954492777586, + 0.0023838349152356386, + 0.021072834730148315, + 0.009192174300551414, + 0.23050707578659058, + 0.01631235145032406, + 0.23051416873931885, + 0.015972383320331573, + 0.015262911096215248, + 0.00919901393353939, + 0.000488920253701508, + 0.0004904099623672664, + 0.0008651982061564922, + 0.00069762475322932, + 0.0008934294455684721, + 0.0007707641343586147, + 0.0015267370035871863, + 0.0008400027290917933, + 0.0005817107739858329, + 0.0005786851979792118, + 0.0006433449452742934, + 0.0007672910578548908, + 0.0008913533529266715, + 0.0007902509532868862, + 0.0007618076633661985, + 0.0009929791558533907, + 0.0012241742806509137, + 0.0011470497120171785, + 0.0016597220674157143, + 0.0021290502045303583, + 0.0019209322053939104, + 0.0013551260344684124, + 0.0017523065907880664, + 0.0020278887823224068 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 1, + 0 + ], + "selected_reward": -0.0009323778795078397 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 inky blue ocean?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are translucent jellyfish with pale blue and pink-purple bioluminescence visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the jellyfish exhibit a pulsating motion with long, slender tentacles trailing behind them?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera move among the group, alternating between close-ups of individual bells and wide shots of many jellyfish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the water nearly still with no visible currents or turbulence?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"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": [ + 0.019585425034165382, + 0.03111831657588482, + 0.037819523364305496, + 0.008289171382784843, + 0.006450108718127012, + 0.010015293955802917, + 0.8695869445800781, + 0.0037987332325428724, + 0.0032595486845821142, + 0.010076849721372128 + ], + "credit_map": [ + 0.0014468533918261528, + 0.002057489473372698, + 0.0023773214779794216, + 0.0027257506735622883, + 0.004000532440841198, + 0.001679045264609158, + 0.0026444129180163145, + 0.002654018346220255, + 0.002584770554676652, + 0.003310908330604434, + 0.00495539978146553, + 0.003691689111292362, + 0.0026157295797020197, + 0.005559603217989206, + 0.00511224614456296, + 0.0032879714854061604, + 0.0021947058849036694, + 0.0038037083577364683, + 0.0036987189669162035, + 0.004871001001447439, + 0.005353759042918682, + 0.005281481891870499, + 0.006714629475027323, + 0.0059015206061303616, + 0.0007340253796428442, + 0.0018130582757294178, + 0.0009630396380089223, + 0.001190772047266364, + 0.0010339065920561552, + 0.0006504355114884675, + 0.001009245403110981, + 0.0008946884772740304, + 0.000619466241914779, + 0.0015252804150804877, + 0.0007513901218771935, + 0.0008471183246001601, + 0.0008775230962783098, + 0.0005575469695031643, + 0.0007205186993815005, + 0.0005512646748684347, + 0.000853832985740155, + 0.0010724000167101622, + 0.0015566486399620771, + 0.0009362910059280694, + 0.001968007069081068, + 0.0009663112577982247, + 0.001567914616316557, + 0.0010938886553049088, + 0.03326910361647606, + 0.01456052903085947, + 0.3656551241874695, + 0.025883043184876442, + 0.36651647090911865, + 0.02507912553846836, + 0.02427039109170437, + 0.014353237114846706, + 0.00035038997884839773, + 0.00029749772511422634, + 0.00037180446088314056, + 0.000367691129213199, + 0.0006006084731779993, + 0.0005424044793471694, + 0.0008518979884684086, + 0.0004164390848018229, + 0.00029194491798989475, + 0.00035161079722456634, + 0.00035766122164204717, + 0.0004669821064453572, + 0.00039818836376070976, + 0.00038145019789226353, + 0.0004382650076877326, + 0.0005734459846280515, + 0.0008336749160662293, + 0.0009267581626772881, + 0.001434003934264183, + 0.0011654500849545002, + 0.0013864055508747697, + 0.001316112233325839, + 0.0013155213091522455, + 0.0016989238793030381 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 2, + 1 + ], + "selected_reward": -0.016065552830696106 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 inky blue ocean?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are translucent jellyfish with pale blue and pink-purple bioluminescence visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the jellyfish exhibit a pulsating motion with long, slender tentacles trailing behind them?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera move among the group, alternating between close-ups of individual bells and wide shots of many jellyfish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the water nearly still with no visible currents or turbulence?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04305335134267807 + ], + "opposite_logprob_tokens": [ + -3.168053388595581 + ], + "logprob_reward": -0.04305335134267807, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04305335134267807, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.005278683267533779, + 0.9082127809524536, + 0.012124714441597462, + 0.006639775354415178, + 0.008943288587033749, + 0.010112341493368149, + 0.040543098002672195, + 0.0022765665780752897, + 0.0038370671682059765, + 0.00203154981136322 + ], + "credit_map": [ + 0.0004960732767358422, + 0.0004734405083581805, + 0.0005433348705992103, + 0.0008012749603949487, + 0.0008645285852253437, + 0.0006602851208299398, + 0.0006779793184250593, + 0.0007617676164954901, + 0.01569470576941967, + 0.023205433040857315, + 0.046649642288684845, + 0.36223316192626953, + 0.02850530296564102, + 0.029424870386719704, + 0.3712770342826843, + 0.03122282586991787, + 0.0007332931854762137, + 0.0011943033896386623, + 0.001290828105993569, + 0.0014906410360708833, + 0.001810684334486723, + 0.0020408895798027515, + 0.002357366494834423, + 0.0012067100033164024, + 0.0005804104148410261, + 0.0008561710710637271, + 0.0014582815347239375, + 0.0009488970972597599, + 0.0006750925676897168, + 0.0006588551914319396, + 0.0007755373371765018, + 0.0006865316536277533, + 0.0005949970218352973, + 0.0022251498885452747, + 0.0008263236959464848, + 0.0015227154362946749, + 0.0010972553864121437, + 0.0007667348254472017, + 0.0012101569445803761, + 0.000699956261087209, + 0.0007008052198216319, + 0.0014355230377987027, + 0.0015153184067457914, + 0.0006532601546496153, + 0.0028594755567610264, + 0.0008287288364954293, + 0.0014114256482571363, + 0.0007078065536916256, + 0.0015781568363308907, + 0.0008003634866327047, + 0.016012823209166527, + 0.0015520858578383923, + 0.01594802923500538, + 0.0015546251088380814, + 0.0016652063932269812, + 0.0014318120665848255, + 0.00016674984362907708, + 0.00016711432544980198, + 0.00028136715991422534, + 0.00025508401449769735, + 0.00031059025786817074, + 0.00022590544540435076, + 0.0005340962670743465, + 0.000335659715346992, + 0.00029861749499104917, + 0.0005311162676662207, + 0.00040973458089865744, + 0.000371741276467219, + 0.0005148060154169798, + 0.00047077328781597316, + 0.0005437047802843153, + 0.000696574104949832, + 0.00013285971363075078, + 0.00019072923169005662, + 0.00040708950837142766, + 0.00040357824764214456, + 0.0001661590940784663, + 0.00025263181305490434, + 0.0002410145098110661, + 0.00023748815874569118 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 6, + 2 + ], + "selected_reward": -0.04305335134267807 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 inky blue ocean?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are translucent jellyfish with pale blue and pink-purple bioluminescence visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the jellyfish exhibit a pulsating motion with long, slender tentacles trailing behind them?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera move among the group, alternating between close-ups of individual bells and wide shots of many jellyfish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the water nearly still with no visible currents or turbulence?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2255842685699463 + ], + "opposite_logprob_tokens": [ + -1.6005842685699463 + ], + "logprob_reward": -0.2255842685699463, + "yes_no_margin_reward": 1.375, + "token_credit_reward": -0.2255842685699463, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.012880468741059303, + 0.727719247341156, + 0.06762111932039261, + 0.009398956783115864, + 0.006834668107330799, + 0.02089753746986389, + 0.12204349040985107, + 0.007988432422280312, + 0.006280019413679838, + 0.01833612471818924 + ], + "credit_map": [ + 0.0009236148907802999, + 0.001528874272480607, + 0.0015081632882356644, + 0.0016202222323045135, + 0.0019744127057492733, + 0.0008863526745699346, + 0.002787449164316058, + 0.0016513790469616652, + 0.012695146724581718, + 0.019100863486528397, + 0.03825610503554344, + 0.2911388874053955, + 0.022401856258511543, + 0.02333952486515045, + 0.29585668444633484, + 0.02493012696504593, + 0.0033340584486722946, + 0.00495630269870162, + 0.0072241974994540215, + 0.008065581321716309, + 0.014607066288590431, + 0.01268860325217247, + 0.011284077540040016, + 0.005461229011416435, + 0.0008232753607444465, + 0.0016350112855434418, + 0.00123179005458951, + 0.0016631223261356354, + 0.0011270992690697312, + 0.0007810289971530437, + 0.0011925370199605823, + 0.0009450912475585938, + 0.0006022042944096029, + 0.0015579990576952696, + 0.0009182180510833859, + 0.0009872539667412639, + 0.0007119571673683822, + 0.0006092389812693, + 0.0007705952739343047, + 0.0006772011402063072, + 0.0015742956893518567, + 0.002782694762572646, + 0.0034951919224113226, + 0.0012827377067878842, + 0.005835963878780603, + 0.0017267366638407111, + 0.0026897804345935583, + 0.0015101362951099873, + 0.005108660086989403, + 0.0023326468653976917, + 0.04945676401257515, + 0.004801344126462936, + 0.04887218028306961, + 0.0041729360818862915, + 0.003879878204315901, + 0.0034190688747912645, + 0.0006950476672500372, + 0.0007268329500220716, + 0.0011106146266683936, + 0.0008982711588032544, + 0.0012144454522058368, + 0.0009560570470057428, + 0.0012014827225357294, + 0.0011856802739202976, + 0.0005852376925759017, + 0.000881408283021301, + 0.0007636564550921321, + 0.0006641455693170428, + 0.0009149103425443172, + 0.0006694160401821136, + 0.0010578123619779944, + 0.0007434321451000869, + 0.0012468426721170545, + 0.001993246842175722, + 0.004192017484456301, + 0.0027603106573224068, + 0.0012727316934615374, + 0.001557295210659504, + 0.0027808123268187046, + 0.0025328651536256075 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 6, + 2 + ], + "selected_reward": -0.2255842685699463 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An immersive, slowly moving tracking shot plunges the viewer into the bottomless depths of the inky blue ocean, home to a kingdom of countless translucent, luminescent jellyfish. These creatures' jelly-like bells emanate a steady, soft bioluminescence, predominantly a pale blue hue, occasionally tinged with a soft pink-purple. This light penetrates their translucent bodies, illuminating the delicate contours of their internal structures and casting a dreamy halo against the inky water. They pulsate gracefully with a leisurely rhythm, each contraction and expansion like a gentle heartbeat, steadily propelling them forward through the nearly still water. As their bodies rise and fall, their long, slender, equally luminescent tentacles dance behind them like silken ribbons. The camera smoothly moves among this huge group, sometimes approaching one so that we can clearly see the smooth texture of its bell-shaped surface, and sometimes retreating to show the spectacular sight of hundreds of light points breathing synchronously in the vast darkness. The whole picture is shrouded in an extreme tranquility and mystery, like a silent and magnificent underwater ballet, creating a fascinating, almost hypnotic visual experience.", + "seed": 8983, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000156/group_0/index_211_seed_8983_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": -0.061521971505135296, + "vlm_logprob": -0.061521971505135296, + "vlm_yes_no_margin": 4.75, + "vlm_token_credit": -0.061521971505135296, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 inky blue ocean?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are translucent jellyfish with pale blue and pink-purple bioluminescence visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the jellyfish exhibit a pulsating motion with long, slender tentacles trailing behind them?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera move among the group, alternating between close-ups of individual bells and wide shots of many jellyfish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the water nearly still with no visible currents or turbulence?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0036272702272981405 + ], + "opposite_logprob_tokens": [ + -7.753627300262451 + ], + "logprob_reward": -0.0036272702272981405, + "yes_no_margin_reward": 7.75, + "token_credit_reward": -0.0036272702272981405, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0751463770866394, + 0.06226634606719017, + 0.10047081857919693, + 0.16603969037532806, + 0.09820359945297241, + 0.04671299457550049, + 0.24710969626903534, + 0.119707852602005, + 0.03791895881295204, + 0.04642370343208313 + ], + "credit_map": [ + 0.006289495620876551, + 0.00510597787797451, + 0.005426597315818071, + 0.007731543388217688, + 0.013151885010302067, + 0.013095717877149582, + 0.012037325650453568, + 0.012307839468121529, + 0.005257942248135805, + 0.005791591480374336, + 0.0073339669033885, + 0.008466671220958233, + 0.01381581462919712, + 0.004753565415740013, + 0.007984069176018238, + 0.008862721733748913, + 0.011133776977658272, + 0.01185356080532074, + 0.014945064671337605, + 0.008817320689558983, + 0.016986919566988945, + 0.017420360818505287, + 0.011117711663246155, + 0.00819610059261322, + 0.008276759646832943, + 0.012456809170544147, + 0.04884037747979164, + 0.009009955450892448, + 0.04543149098753929, + 0.021924704313278198, + 0.007634849287569523, + 0.012464728206396103, + 0.008877372369170189, + 0.016966907307505608, + 0.01161911804229021, + 0.010623485781252384, + 0.010408207774162292, + 0.01778438501060009, + 0.009174495004117489, + 0.012749626301229, + 0.0035081072710454464, + 0.00511750765144825, + 0.005122463684529066, + 0.005860746372491121, + 0.005503093358129263, + 0.00489423843100667, + 0.007856032811105251, + 0.008850803598761559, + 0.01339047309011221, + 0.01197308860719204, + 0.07692211121320724, + 0.06403083354234695, + 0.00851241871714592, + 0.033135075122117996, + 0.009596595540642738, + 0.02954910881817341, + 0.007661708164960146, + 0.008800720795989037, + 0.03365626186132431, + 0.03121764399111271, + 0.009162668138742447, + 0.010333826765418053, + 0.01037806086242199, + 0.008496960625052452, + 0.0029632162768393755, + 0.005163676105439663, + 0.0036380328238010406, + 0.00445930939167738, + 0.005081988405436277, + 0.00672355480492115, + 0.0050100237131118774, + 0.004879159387201071, + 0.0038243120070546865, + 0.004687791224569082, + 0.004713280126452446, + 0.005911372136324644, + 0.004252694547176361, + 0.00975365936756134, + 0.004868968855589628, + 0.008411627262830734 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 3, + 7 + ], + "selected_reward": -0.0036272702272981405 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 inky blue ocean?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are translucent jellyfish with pale blue and pink-purple bioluminescence visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the jellyfish exhibit a pulsating motion with long, slender tentacles trailing behind them?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera move among the group, alternating between close-ups of individual bells and wide shots of many jellyfish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the water nearly still with no visible currents or turbulence?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0011881680693477392 + ], + "opposite_logprob_tokens": [ + -6.751188278198242 + ], + "logprob_reward": -0.0011881680693477392, + "yes_no_margin_reward": 6.75, + "token_credit_reward": -0.0011881680693477392, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06319219619035721, + 0.10788595676422119, + 0.09168702363967896, + 0.20461387932300568, + 0.02416500635445118, + 0.06549308449029922, + 0.30741921067237854, + 0.051271554082632065, + 0.027145065367221832, + 0.057127051055431366 + ], + "credit_map": [ + 0.008744652383029461, + 0.007655834313482046, + 0.005205689929425716, + 0.0054372763261199, + 0.008465507067739964, + 0.007344333920627832, + 0.009733211249113083, + 0.010605700314044952, + 0.006041883956640959, + 0.00915837287902832, + 0.011198475956916809, + 0.012162904255092144, + 0.025275040417909622, + 0.0069357468746602535, + 0.014958486892282963, + 0.022155048325657845, + 0.007635584566742182, + 0.009423446841537952, + 0.015640532597899437, + 0.011648518033325672, + 0.012427465058863163, + 0.012279235757887363, + 0.010465904138982296, + 0.012166338972747326, + 0.008337522856891155, + 0.016942596063017845, + 0.07263488322496414, + 0.010480659082531929, + 0.06718970835208893, + 0.014126600697636604, + 0.006473786197602749, + 0.008428128436207771, + 0.0022969550918787718, + 0.0031806465703994036, + 0.003176348749548197, + 0.003054671920835972, + 0.00271904026158154, + 0.003554169088602066, + 0.00339782633818686, + 0.002785347867757082, + 0.0048721302300691605, + 0.005694300401955843, + 0.006821704097092152, + 0.0060068098828196526, + 0.006146694533526897, + 0.008787763305008411, + 0.013516043312847614, + 0.01364764291793108, + 0.018207242712378502, + 0.01581444963812828, + 0.08309125155210495, + 0.06534457951784134, + 0.015138071961700916, + 0.04834591597318649, + 0.015498832799494267, + 0.04597886651754379, + 0.004277346655726433, + 0.0049490295350551605, + 0.01330018974840641, + 0.01286280620843172, + 0.003737898776307702, + 0.0038591620977967978, + 0.0038570892065763474, + 0.004428028594702482, + 0.0015891521470621228, + 0.0029320784378796816, + 0.006469033658504486, + 0.005305929109454155, + 0.002624855376780033, + 0.002827006159350276, + 0.0027835501823574305, + 0.0026134606450796127, + 0.0044601368717849255, + 0.0059763588942587376, + 0.005439642816781998, + 0.006675421260297298, + 0.007361089810729027, + 0.007225502748042345, + 0.008506136015057564, + 0.011482763104140759 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 3, + 1 + ], + "selected_reward": -0.0011881680693477392 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 inky blue ocean?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are translucent jellyfish with pale blue and pink-purple bioluminescence visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the jellyfish exhibit a pulsating motion with long, slender tentacles trailing behind them?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera move among the group, alternating between close-ups of individual bells and wide shots of many jellyfish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the water nearly still with no visible currents or turbulence?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007628239691257477 + ], + "opposite_logprob_tokens": [ + -4.882628440856934 + ], + "logprob_reward": -0.007628239691257477, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007628239691257477, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09757998585700989, + 0.1003529280424118, + 0.06927600502967834, + 0.06473740935325623, + 0.06756291538476944, + 0.045538175851106644, + 0.35975754261016846, + 0.0895928293466568, + 0.033902961760759354, + 0.07169931381940842 + ], + "credit_map": [ + 0.007745479233562946, + 0.007455532439053059, + 0.008057539351284504, + 0.011339935474097729, + 0.019436189904808998, + 0.012054129503667355, + 0.013100353069603443, + 0.01839081197977066, + 0.004306958056986332, + 0.0055359588004648685, + 0.007031772285699844, + 0.008693239651620388, + 0.030061569064855576, + 0.005869599059224129, + 0.010405545122921467, + 0.028448283672332764, + 0.00642324099317193, + 0.008553694933652878, + 0.009019982069730759, + 0.009387695230543613, + 0.010379398241639137, + 0.009889970533549786, + 0.005530045833438635, + 0.010091976262629032, + 0.004359110724180937, + 0.008389798924326897, + 0.010724464431405067, + 0.006183315999805927, + 0.012511714361608028, + 0.01112186349928379, + 0.004897895269095898, + 0.006549241486936808, + 0.003608583938330412, + 0.013344673439860344, + 0.011787351220846176, + 0.00916037056595087, + 0.007085935678333044, + 0.008474738337099552, + 0.007995429448783398, + 0.0061058285646140575, + 0.004321630112826824, + 0.0034106727689504623, + 0.006060994695872068, + 0.005954860709607601, + 0.003752634162083268, + 0.005935184191912413, + 0.005589665379375219, + 0.01051252894103527, + 0.01756066456437111, + 0.016713542863726616, + 0.11021079868078232, + 0.09379138052463531, + 0.012494148686528206, + 0.049009907990694046, + 0.014235323294997215, + 0.04574175551533699, + 0.004409478977322578, + 0.007555357180535793, + 0.027053549885749817, + 0.025035537779331207, + 0.005654899403452873, + 0.007241838611662388, + 0.0060861678794026375, + 0.006555984728038311, + 0.0018261189106851816, + 0.003128977958112955, + 0.0077166203409433365, + 0.007035245653241873, + 0.003217529272660613, + 0.003651463659480214, + 0.0026917997747659683, + 0.004635202698409557, + 0.005322441458702087, + 0.0075047132559120655, + 0.006969755981117487, + 0.008778762072324753, + 0.010359699837863445, + 0.009607751853764057, + 0.011527126654982567, + 0.011629057116806507 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 1, + 0 + ], + "selected_reward": -0.007628239691257477 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 inky blue ocean?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are translucent jellyfish with pale blue and pink-purple bioluminescence visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the jellyfish exhibit a pulsating motion with long, slender tentacles trailing behind them?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera move among the group, alternating between close-ups of individual bells and wide shots of many jellyfish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the water nearly still with no visible currents or turbulence?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04305015504360199 + ], + "opposite_logprob_tokens": [ + -3.1680500507354736 + ], + "logprob_reward": -0.04305015504360199, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04305015504360199, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.050256967544555664, + 0.11445926129817963, + 0.09047967940568924, + 0.08081114292144775, + 0.15011867880821228, + 0.15758438408374786, + 0.19602034986019135, + 0.038341671228408813, + 0.0652034729719162, + 0.056724339723587036 + ], + "credit_map": [ + 0.005794941913336515, + 0.004873286932706833, + 0.0038550442550331354, + 0.005101657472550869, + 0.00817909650504589, + 0.00654092151671648, + 0.008486554957926273, + 0.007425462361425161, + 0.0058700917288661, + 0.004848781041800976, + 0.006695004645735025, + 0.00721384771168232, + 0.04030122235417366, + 0.006804926320910454, + 0.007879371754825115, + 0.03484600782394409, + 0.008164657279849052, + 0.008096168749034405, + 0.012396657839417458, + 0.010386420413851738, + 0.0153824957087636, + 0.013673979789018631, + 0.00829430390149355, + 0.014084992930293083, + 0.005166011396795511, + 0.018654173240065575, + 0.011903425678610802, + 0.007355548907071352, + 0.010635645128786564, + 0.010433346964418888, + 0.008404337801039219, + 0.008258653804659843, + 0.01186690665781498, + 0.031457334756851196, + 0.024419575929641724, + 0.021991796791553497, + 0.014879656955599785, + 0.014300258830189705, + 0.014308074489235878, + 0.016895079985260963, + 0.008293637074530125, + 0.013091878965497017, + 0.015066380612552166, + 0.015488680452108383, + 0.013792869634926319, + 0.01616797223687172, + 0.038007039576768875, + 0.03767592832446098, + 0.01326936949044466, + 0.014511890709400177, + 0.05286356806755066, + 0.04757744073867798, + 0.010342003777623177, + 0.024858400225639343, + 0.009678435511887074, + 0.02291923761367798, + 0.002789705991744995, + 0.0026150771882385015, + 0.01074284128844738, + 0.009491945616900921, + 0.0026907811406999826, + 0.002557339845225215, + 0.004323532339185476, + 0.003130449214950204, + 0.005668875761330128, + 0.009330076165497303, + 0.009682000614702702, + 0.010364256799221039, + 0.005766439251601696, + 0.008109087124466896, + 0.006208675913512707, + 0.010074063204228878, + 0.00428827153518796, + 0.006013742182403803, + 0.0053286245092749596, + 0.008072521537542343, + 0.007024877704679966, + 0.00816107727587223, + 0.011249497532844543, + 0.006585726514458656 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 5, + 4 + ], + "selected_reward": -0.04305015504360199 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in an inky blue ocean?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are translucent jellyfish with pale blue and pink-purple bioluminescence visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the jellyfish exhibit a pulsating motion with long, slender tentacles trailing behind them?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera move among the group, alternating between close-ups of individual bells and wide shots of many jellyfish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the water nearly still with no visible currents or turbulence?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.25211602449417114 + ], + "opposite_logprob_tokens": [ + -1.5021159648895264 + ], + "logprob_reward": -0.25211602449417114, + "yes_no_margin_reward": 1.25, + "token_credit_reward": -0.25211602449417114, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1283552348613739, + 0.16972270607948303, + 0.1373158097267151, + 0.11400347203016281, + 0.0758618488907814, + 0.04453641176223755, + 0.10497766733169556, + 0.06870339810848236, + 0.027934180572628975, + 0.12858928740024567 + ], + "credit_map": [ + 0.013264751061797142, + 0.01555909588932991, + 0.01333034597337246, + 0.014726217836141586, + 0.017473872750997543, + 0.012301615439355373, + 0.022171679884195328, + 0.019527649506926537, + 0.008812475018203259, + 0.007548687979578972, + 0.013468971475958824, + 0.013341288082301617, + 0.052910059690475464, + 0.009809422306716442, + 0.017560699954628944, + 0.04627109691500664, + 0.011631140485405922, + 0.012212203815579414, + 0.018999094143509865, + 0.016229620203375816, + 0.02147163450717926, + 0.023501338437199593, + 0.01811871863901615, + 0.01515205204486847, + 0.004669550806283951, + 0.0071671269834041595, + 0.03362927585840225, + 0.005887217354029417, + 0.03391347452998161, + 0.01636691577732563, + 0.005287920590490103, + 0.007081988733261824, + 0.004761837888509035, + 0.009354681707918644, + 0.010501766577363014, + 0.010429099202156067, + 0.007868638262152672, + 0.014598280191421509, + 0.009727359749376774, + 0.008620190434157848, + 0.0031962899956852198, + 0.004694818984717131, + 0.004870153497904539, + 0.005355670116841793, + 0.004533467348664999, + 0.007905127480626106, + 0.007420205045491457, + 0.006560679525136948, + 0.007827257737517357, + 0.008328620344400406, + 0.031584326177835464, + 0.027056129649281502, + 0.005758803803473711, + 0.0077774724923074245, + 0.00539396982640028, + 0.011251090094447136, + 0.008080924861133099, + 0.008191627450287342, + 0.009384281933307648, + 0.01006721705198288, + 0.006406690459698439, + 0.012120086699724197, + 0.007350474130362272, + 0.007102098315954208, + 0.0021117115393280983, + 0.0034953546710312366, + 0.0031644762493669987, + 0.004039018880575895, + 0.0038961295504122972, + 0.004120017867535353, + 0.0038547494914382696, + 0.0032527234870940447, + 0.009249888360500336, + 0.012781086377799511, + 0.014350136741995811, + 0.01681017130613327, + 0.0209409911185503, + 0.013804055750370026, + 0.019851235672831535, + 0.020801715552806854 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 2, + 9 + ], + "selected_reward": -0.25211602449417114 + } + ], + "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/0000156/group_1/index_tensor([7733, 7733])_group_1_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000156/group_1/index_tensor([7733, 7733])_group_1_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..7d3566a94194c2b1e935ea1be1963e847dde368b --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000156/group_1/index_tensor([7733, 7733])_group_1_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "high-speed, dynamic swooping shot, composed with the rigor of Arnold Newman, plummets downward from the sky, capturing a moment of primal power. Beneath the frame, an industrial ruin of broken concrete slabs, rusted rebar, and massive metal trusses casts sharp, interlocking geometric shadows in the low sunlight. Above this decaying landscape of gray and ochre, a nimble boy is frozen in midair, performing a cartwheel with breathtaking force. He wears a garment meticulously woven from vibrant, deep green vines and leaves, tightly hugging his powerful muscular lines, some of the leaves trembling slightly with the violent twirling. His body stretches in the air into a perfect straight line, parallel to the ground, a display of supreme control. Most striking are his eyes: not human, but two burning flames, gleaming with a fiery orange-red glow, even sparks emanating from them, casting an ominous yet warm glow on his resolute features. The entire scene is painted in a cold and desaturated tone, highlighting the desolation of the ruins. Only the vibrant green of the boy’s clothes and the fire in his eyes become the visual center of the picture, creating a strong dramatic conflict and tense atmosphere of wild vitality bursting out from the dead structure.", + "seed": 29645, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000156/group_1/index_7733_seed_29645_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": -0.49538393029943106, + "vlm_logprob": -0.49538393029943106, + "vlm_yes_no_margin": 2.625, + "vlm_token_credit": -0.49538393029943106, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a high-speed, dynamic swooping movement plunging downward from the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background an industrial ruin featuring broken concrete slabs, rusted rebar, and massive metal trusses?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a boy frozen in midair performing a cartwheel with his body stretched into a straight line parallel to the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the boy wearing a garment woven from vibrant, deep green vines and leaves, and do his eyes glow with fiery orange-red flames?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the overall scene tone cold and desaturated, with the green clothing and fire in the eyes serving as the only vibrant visual centers?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"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": [ + -2.240396022796631 + ], + "opposite_logprob_tokens": [ + -0.11539611220359802 + ], + "logprob_reward": -2.240396022796631, + "yes_no_margin_reward": -2.125, + "token_credit_reward": -2.240396022796631, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.20965610444545746, + 0.09454822540283203, + 0.04320157319307327, + 0.05495539307594299, + 0.03510846197605133, + 0.0726727843284607, + 0.046581413596868515, + 0.04837961867451668, + 0.21031686663627625, + 0.1845795214176178 + ], + "credit_map": [ + 0.02764110453426838, + 0.02021455578505993, + 0.023451639339327812, + 0.024347985163331032, + 0.02515951171517372, + 0.028375552967190742, + 0.03884841501712799, + 0.021617351099848747, + 0.01252178568392992, + 0.008820103481411934, + 0.013723626732826233, + 0.01192798838019371, + 0.011147339828312397, + 0.010357791557908058, + 0.01329065952450037, + 0.012758932076394558, + 0.004845461342483759, + 0.004231472499668598, + 0.003470224794000387, + 0.010517619550228119, + 0.005863646976649761, + 0.0025266550946980715, + 0.008028213866055012, + 0.0037182795349508524, + 0.00769388722255826, + 0.009279330261051655, + 0.006906819064170122, + 0.00737015949562192, + 0.00548688555136323, + 0.005998173728585243, + 0.006399780046194792, + 0.0058203572407364845, + 0.0029077138751745224, + 0.003567081643268466, + 0.005986104719340801, + 0.005445989314466715, + 0.0034967849496752024, + 0.004342494532465935, + 0.004492443520575762, + 0.004869853612035513, + 0.010949545539915562, + 0.007941819727420807, + 0.011823348701000214, + 0.0053558992221951485, + 0.007362174801528454, + 0.013620015233755112, + 0.006924543064087629, + 0.008695434778928757, + 0.004753482528030872, + 0.006066900212317705, + 0.0060236952267587185, + 0.0057074809446930885, + 0.00410403311252594, + 0.008819286711513996, + 0.0052583408541977406, + 0.005848192144185305, + 0.006617546081542969, + 0.004201245494186878, + 0.005479485262185335, + 0.005713080987334251, + 0.007207532413303852, + 0.0061696190387010574, + 0.009100849740207195, + 0.003890260122716427, + 0.049999285489320755, + 0.027246061712503433, + 0.05399448052048683, + 0.02225055918097496, + 0.010243250988423824, + 0.01580568216741085, + 0.015380432829260826, + 0.015397109091281891, + 0.02332169935107231, + 0.017705777660012245, + 0.021063292399048805, + 0.029126450419425964, + 0.025744836777448654, + 0.01904095709323883, + 0.024669330567121506, + 0.023907171562314034 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 9 + ], + "selected_reward": -2.240396022796631 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 high-speed, dynamic swooping movement plunging downward from the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background an industrial ruin featuring broken concrete slabs, rusted rebar, and massive metal trusses?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a boy frozen in midair performing a cartwheel with his body stretched into a straight line parallel to the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the boy wearing a garment woven from vibrant, deep green vines and leaves, and do his eyes glow with fiery orange-red flames?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the overall scene tone cold and desaturated, with the green clothing and fire in the eyes serving as the only vibrant visual centers?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"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": [ + 0.19504421949386597, + 0.16312703490257263, + 0.04864085465669632, + 0.06352751702070236, + 0.04672612249851227, + 0.04753454402089119, + 0.05567162483930588, + 0.08898583054542542, + 0.23032619059085846, + 0.060416094958782196 + ], + "credit_map": [ + 0.02211344987154007, + 0.019957436248660088, + 0.02357148565351963, + 0.03077283687889576, + 0.02196543477475643, + 0.023111645132303238, + 0.03212488070130348, + 0.021427031606435776, + 0.025884706526994705, + 0.014419828541576862, + 0.028645098209381104, + 0.020747467875480652, + 0.01562326867133379, + 0.018852688372135162, + 0.017597025260329247, + 0.021356942132115364, + 0.006146120373159647, + 0.004923271480947733, + 0.005269079934805632, + 0.009047886356711388, + 0.005718061234802008, + 0.0037118324544280767, + 0.008335019461810589, + 0.0054895817302167416, + 0.010154442861676216, + 0.010484570637345314, + 0.010388411581516266, + 0.006169210188090801, + 0.004122429061681032, + 0.009803145192563534, + 0.006852091755717993, + 0.00555321155115962, + 0.004997056908905506, + 0.006826004013419151, + 0.005415735300630331, + 0.00499621732160449, + 0.004786260891705751, + 0.006287619471549988, + 0.004240204114466906, + 0.00917702168226242, + 0.005755553487688303, + 0.0065668439492583275, + 0.007594636641442776, + 0.0036301088985055685, + 0.004825861193239689, + 0.007878451608121395, + 0.00500764325261116, + 0.0062754456885159016, + 0.010246006771922112, + 0.0068579260259866714, + 0.008729355409741402, + 0.004566643387079239, + 0.006605120375752449, + 0.006936523597687483, + 0.0070227207615971565, + 0.0047073266468942165, + 0.010160665959119797, + 0.008440850302577019, + 0.009986786171793938, + 0.015293100848793983, + 0.013851037248969078, + 0.007757934276014566, + 0.017089826986193657, + 0.00640562828630209, + 0.06478787958621979, + 0.02383541874587536, + 0.06719490885734558, + 0.023350490257143974, + 0.00945989415049553, + 0.01359996385872364, + 0.014064349234104156, + 0.014033280313014984, + 0.006938894744962454, + 0.005867743864655495, + 0.006780613213777542, + 0.008349952287971973, + 0.009481889195740223, + 0.0074180425144732, + 0.007347654551267624, + 0.008231299929320812 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 1 + ], + "selected_reward": -0.016087958589196205 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 high-speed, dynamic swooping movement plunging downward from the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background an industrial ruin featuring broken concrete slabs, rusted rebar, and massive metal trusses?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a boy frozen in midair performing a cartwheel with his body stretched into a straight line parallel to the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the boy wearing a garment woven from vibrant, deep green vines and leaves, and do his eyes glow with fiery orange-red flames?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the overall scene tone cold and desaturated, with the green clothing and fire in the eyes serving as the only vibrant visual centers?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2016032636165619 + ], + "opposite_logprob_tokens": [ + -1.7016032934188843 + ], + "logprob_reward": -0.2016032636165619, + "yes_no_margin_reward": 1.5, + "token_credit_reward": -0.2016032636165619, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1268819272518158, + 0.062082439661026, + 0.08316361159086227, + 0.0589095875620842, + 0.06020789220929146, + 0.09136107563972473, + 0.12369444221258163, + 0.18493999540805817, + 0.10255655646324158, + 0.10620248317718506 + ], + "credit_map": [ + 0.016774587333202362, + 0.0143613675609231, + 0.014928936026990414, + 0.013107583858072758, + 0.019806692376732826, + 0.016471020877361298, + 0.021002139896154404, + 0.010429594665765762, + 0.006735464092344046, + 0.007273608818650246, + 0.006295512896031141, + 0.007080402225255966, + 0.012654379941523075, + 0.006906005088239908, + 0.008863027207553387, + 0.00627403799444437, + 0.008490665815770626, + 0.007789960131049156, + 0.010778343304991722, + 0.017909197136759758, + 0.008595033548772335, + 0.0070400177501142025, + 0.014989604242146015, + 0.007570790592581034, + 0.007241240236908197, + 0.008140449412167072, + 0.01014876738190651, + 0.011196285486221313, + 0.005085745360702276, + 0.005395413842052221, + 0.007244090549647808, + 0.004457593895494938, + 0.0036391238681972027, + 0.008215061388909817, + 0.009959230199456215, + 0.00476435711607337, + 0.009276721626520157, + 0.010360760614275932, + 0.004846381023526192, + 0.009146254509687424, + 0.007195198908448219, + 0.012586291879415512, + 0.011745885014533997, + 0.011204035952687263, + 0.0145058399066329, + 0.010124576278030872, + 0.012800540775060654, + 0.011198706924915314, + 0.019555898383259773, + 0.01756443828344345, + 0.018689289689064026, + 0.012042596004903316, + 0.01551610603928566, + 0.01575569063425064, + 0.015169165097177029, + 0.00940125621855259, + 0.02212398499250412, + 0.01605949178338051, + 0.02021973952651024, + 0.029579661786556244, + 0.031613532453775406, + 0.018461667001247406, + 0.03445463255047798, + 0.01242728903889656, + 0.023003891110420227, + 0.013585873879492283, + 0.023354532197117805, + 0.009991358034312725, + 0.007733410224318504, + 0.008793031796813011, + 0.007300644647330046, + 0.008793818764388561, + 0.013396325521171093, + 0.009961680509150028, + 0.011782177723944187, + 0.012925636023283005, + 0.020297866314649582, + 0.011929208412766457, + 0.012660699896514416, + 0.013248881325125694 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 6 + ], + "selected_reward": -0.2016032636165619 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 high-speed, dynamic swooping movement plunging downward from the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background an industrial ruin featuring broken concrete slabs, rusted rebar, and massive metal trusses?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a boy frozen in midair performing a cartwheel with his body stretched into a straight line parallel to the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the boy wearing a garment woven from vibrant, deep green vines and leaves, and do his eyes glow with fiery orange-red flames?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the overall scene tone cold and desaturated, with the green clothing and fire in the eyes serving as the only vibrant visual centers?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00463957479223609 + ], + "opposite_logprob_tokens": [ + -5.379639625549316 + ], + "logprob_reward": -0.00463957479223609, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.00463957479223609, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1469629406929016, + 0.11726952344179153, + 0.04039323702454567, + 0.08474960178136826, + 0.05012211203575134, + 0.08752813190221786, + 0.08911512047052383, + 0.12247616797685623, + 0.13845586776733398, + 0.12292737513780594 + ], + "credit_map": [ + 0.020705359056591988, + 0.017029142007231712, + 0.015866458415985107, + 0.016812819987535477, + 0.013840962201356888, + 0.01861371286213398, + 0.02744186855852604, + 0.016652623191475868, + 0.010729181580245495, + 0.016769668087363243, + 0.010454951785504818, + 0.01158696599304676, + 0.020839938893914223, + 0.016896380111575127, + 0.016592834144830704, + 0.013399594463407993, + 0.00494789844378829, + 0.0040626730769872665, + 0.006477033719420433, + 0.00566685339435935, + 0.0048175775445997715, + 0.003947427496314049, + 0.004875633865594864, + 0.00559814041480422, + 0.010449696332216263, + 0.011272036470472813, + 0.010961976833641529, + 0.014199623838067055, + 0.009949203580617905, + 0.008049411699175835, + 0.011085721664130688, + 0.008781936019659042, + 0.002862001769244671, + 0.007167409639805555, + 0.009866973385214806, + 0.006682717707008123, + 0.006830956321209669, + 0.0071220421232283115, + 0.00417634891346097, + 0.005413659382611513, + 0.013502499088644981, + 0.010916121304035187, + 0.01343205850571394, + 0.006181133911013603, + 0.008653935976326466, + 0.01564570516347885, + 0.009308981709182262, + 0.00988769344985485, + 0.012186674401164055, + 0.00872599147260189, + 0.013224928639829159, + 0.009741530753672123, + 0.009583104401826859, + 0.011596394702792168, + 0.014585614204406738, + 0.009470878168940544, + 0.016495445743203163, + 0.009972375817596912, + 0.013611996546387672, + 0.015630260109901428, + 0.019116945564746857, + 0.013394644483923912, + 0.024558255448937416, + 0.009696249850094318, + 0.03747720643877983, + 0.013079569675028324, + 0.03764846920967102, + 0.013544362969696522, + 0.007616288494318724, + 0.010527278296649456, + 0.009812903590500355, + 0.008749788627028465, + 0.015046489425003529, + 0.009148736484348774, + 0.016316644847393036, + 0.009891843423247337, + 0.025114087387919426, + 0.01670263148844242, + 0.01463507954031229, + 0.016071857884526253 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 9 + ], + "selected_reward": -0.00463957479223609 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 high-speed, dynamic swooping movement plunging downward from the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background an industrial ruin featuring broken concrete slabs, rusted rebar, and massive metal trusses?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a boy frozen in midair performing a cartwheel with his body stretched into a straight line parallel to the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the boy wearing a garment woven from vibrant, deep green vines and leaves, and do his eyes glow with fiery orange-red flames?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the overall scene tone cold and desaturated, with the green clothing and fire in the eyes serving as the only vibrant visual centers?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"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": [ + 0.17038966715335846, + 0.09340263903141022, + 0.04434893652796745, + 0.09470890462398529, + 0.06396176666021347, + 0.0773906335234642, + 0.09397906064987183, + 0.07396683096885681, + 0.1561446189880371, + 0.13170692324638367 + ], + "credit_map": [ + 0.02118426002562046, + 0.019605441018939018, + 0.017480986192822456, + 0.025342833250761032, + 0.013756821863353252, + 0.021592887118458748, + 0.030021201819181442, + 0.02140524610877037, + 0.00811343640089035, + 0.012231928296387196, + 0.011071587912738323, + 0.011272195726633072, + 0.012931670993566513, + 0.012736564502120018, + 0.014638201333582401, + 0.010407054796814919, + 0.005023727659136057, + 0.003865133272483945, + 0.005982962436974049, + 0.00925111211836338, + 0.006181297358125448, + 0.003059291746467352, + 0.007720342371612787, + 0.0032650723587721586, + 0.01617920771241188, + 0.013432992622256279, + 0.01131605263799429, + 0.008185683749616146, + 0.006838872097432613, + 0.019358685240149498, + 0.007124499883502722, + 0.01227291114628315, + 0.00487145408987999, + 0.007847879081964493, + 0.008305536583065987, + 0.011458712629973888, + 0.007747361436486244, + 0.006901630666106939, + 0.00873454287648201, + 0.008094651624560356, + 0.011116241104900837, + 0.008809376507997513, + 0.012796107679605484, + 0.005881506949663162, + 0.005723900161683559, + 0.014520997181534767, + 0.008564389310777187, + 0.009978111833333969, + 0.00883506704121828, + 0.013242592103779316, + 0.012359365820884705, + 0.01061516348272562, + 0.008173417299985886, + 0.016938237473368645, + 0.011174398474395275, + 0.0126408152282238, + 0.009455441497266293, + 0.006130983587354422, + 0.00837914738804102, + 0.009486080147325993, + 0.012044797651469707, + 0.008862980641424656, + 0.014158220961689949, + 0.005449180956929922, + 0.03346695005893707, + 0.022252090275287628, + 0.036986518651247025, + 0.017130134627223015, + 0.008971834555268288, + 0.013761724345386028, + 0.011455371044576168, + 0.012119999155402184, + 0.016805090010166168, + 0.011334539391100407, + 0.014205779880285263, + 0.013538472354412079, + 0.02809790149331093, + 0.01522307563573122, + 0.01578274369239807, + 0.016719315201044083 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 9 + ], + "selected_reward": -0.014192831702530384 + } + ], + "vlm_decoded_outputs": [ + "No", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "high-speed, dynamic swooping shot, composed with the rigor of Arnold Newman, plummets downward from the sky, capturing a moment of primal power. Beneath the frame, an industrial ruin of broken concrete slabs, rusted rebar, and massive metal trusses casts sharp, interlocking geometric shadows in the low sunlight. Above this decaying landscape of gray and ochre, a nimble boy is frozen in midair, performing a cartwheel with breathtaking force. He wears a garment meticulously woven from vibrant, deep green vines and leaves, tightly hugging his powerful muscular lines, some of the leaves trembling slightly with the violent twirling. His body stretches in the air into a perfect straight line, parallel to the ground, a display of supreme control. Most striking are his eyes: not human, but two burning flames, gleaming with a fiery orange-red glow, even sparks emanating from them, casting an ominous yet warm glow on his resolute features. The entire scene is painted in a cold and desaturated tone, highlighting the desolation of the ruins. Only the vibrant green of the boy’s clothes and the fire in his eyes become the visual center of the picture, creating a strong dramatic conflict and tense atmosphere of wild vitality bursting out from the dead structure.", + "seed": 29646, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000156/group_1/index_7733_seed_29646_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": -0.09263206184841692, + "vlm_logprob": -0.09263206184841692, + "vlm_yes_no_margin": 3.55, + "vlm_token_credit": -0.09263206184841692, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 high-speed, dynamic swooping movement plunging downward from the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background an industrial ruin featuring broken concrete slabs, rusted rebar, and massive metal trusses?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a boy frozen in midair performing a cartwheel with his body stretched into a straight line parallel to the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the boy wearing a garment woven from vibrant, deep green vines and leaves, and do his eyes glow with fiery orange-red flames?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the overall scene tone cold and desaturated, with the green clothing and fire in the eyes serving as the only vibrant visual centers?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.23045025765895844 + ], + "opposite_logprob_tokens": [ + -1.6054502725601196 + ], + "logprob_reward": -0.23045025765895844, + "yes_no_margin_reward": 1.375, + "token_credit_reward": -0.23045025765895844, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07802431285381317, + 0.05875266343355179, + 0.0676891952753067, + 0.15892404317855835, + 0.17989549040794373, + 0.19984713196754456, + 0.038407184183597565, + 0.04742419719696045, + 0.12084785103797913, + 0.050187963992357254 + ], + "credit_map": [ + 0.008606105111539364, + 0.01140725426375866, + 0.008594589307904243, + 0.009118362329900265, + 0.009578067809343338, + 0.008854944258928299, + 0.012178479693830013, + 0.009686507284641266, + 0.006210813298821449, + 0.007684820797294378, + 0.007795047480612993, + 0.008165030740201473, + 0.0074038258753716946, + 0.008222706615924835, + 0.007530672010034323, + 0.005739746615290642, + 0.008173913694918156, + 0.007356521673500538, + 0.010966477915644646, + 0.005551373586058617, + 0.008824137039482594, + 0.010711346752941608, + 0.007495786529034376, + 0.008609641343355179, + 0.018288981169462204, + 0.013482431881129742, + 0.023871680721640587, + 0.023767786100506783, + 0.017359042540192604, + 0.02063797228038311, + 0.02245277911424637, + 0.019063372164964676, + 0.01328209601342678, + 0.0164487361907959, + 0.02248215489089489, + 0.02664719521999359, + 0.02429024502635002, + 0.014312580227851868, + 0.034724149852991104, + 0.027708329260349274, + 0.012218896299600601, + 0.04690839722752571, + 0.059839971363544464, + 0.01785208284854889, + 0.010032322257757187, + 0.011425845324993134, + 0.01471454743295908, + 0.02685506083071232, + 0.003694612067192793, + 0.005194040481001139, + 0.006111066322773695, + 0.00700344517827034, + 0.003775110701099038, + 0.005082843825221062, + 0.003737851046025753, + 0.003808214794844389, + 0.005260152742266655, + 0.006313366815447807, + 0.004771443549543619, + 0.006016906350851059, + 0.0062341829761862755, + 0.008374557830393314, + 0.004092801362276077, + 0.006360785569995642, + 0.013242457993328571, + 0.018178189173340797, + 0.010799532756209373, + 0.010437129996716976, + 0.013379507698118687, + 0.02304067648947239, + 0.016117025166749954, + 0.015653328970074654, + 0.004773659631609917, + 0.00745950685814023, + 0.004387733992189169, + 0.006894088350236416, + 0.009040621109306812, + 0.006025979295372963, + 0.0066479905508458614, + 0.004958378616720438 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 4, + 3 + ], + "selected_reward": -0.23045025765895844 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 high-speed, dynamic swooping movement plunging downward from the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background an industrial ruin featuring broken concrete slabs, rusted rebar, and massive metal trusses?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a boy frozen in midair performing a cartwheel with his body stretched into a straight line parallel to the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the boy wearing a garment woven from vibrant, deep green vines and leaves, and do his eyes glow with fiery orange-red flames?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the overall scene tone cold and desaturated, with the green clothing and fire in the eyes serving as the only vibrant visual centers?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020588964223861694 + ], + "opposite_logprob_tokens": [ + -3.8955888748168945 + ], + "logprob_reward": -0.020588964223861694, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020588964223861694, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.33475103974342346, + 0.13084007799625397, + 0.17130166292190552, + 0.07915865629911423, + 0.06149682030081749, + 0.07790954411029816, + 0.08194789290428162, + 0.02251378819346428, + 0.023511603474617004, + 0.01656891219317913 + ], + "credit_map": [ + 0.037154894322156906, + 0.03812349587678909, + 0.0385223813354969, + 0.040457144379615784, + 0.05416176840662956, + 0.04188300296664238, + 0.04221191629767418, + 0.04223642870783806, + 0.013015338219702244, + 0.012858724221587181, + 0.02855830080807209, + 0.015709932893514633, + 0.01651344634592533, + 0.013465636409819126, + 0.015728143975138664, + 0.014990558847784996, + 0.018728192895650864, + 0.01828833855688572, + 0.032143183052539825, + 0.013020681217312813, + 0.017468120902776718, + 0.02542397938668728, + 0.01951480656862259, + 0.026714365929365158, + 0.012563985772430897, + 0.006477365270256996, + 0.011835531331598759, + 0.007194850593805313, + 0.0072897570207715034, + 0.009536710567772388, + 0.012317271903157234, + 0.011943180114030838, + 0.004166030324995518, + 0.008115732111036777, + 0.007416002918034792, + 0.013430600054562092, + 0.008605509996414185, + 0.004595500882714987, + 0.007519479840993881, + 0.007647969294339418, + 0.005430242512375116, + 0.019775129854679108, + 0.01021108403801918, + 0.007584320846945047, + 0.0046417913399636745, + 0.004053861368447542, + 0.010994832962751389, + 0.015218283049762249, + 0.006561956834048033, + 0.007561870384961367, + 0.010536467656493187, + 0.011221407912671566, + 0.009185553528368473, + 0.01802681013941765, + 0.006365453824400902, + 0.01248837634921074, + 0.0032866750843822956, + 0.003407084848731756, + 0.002444493817165494, + 0.003042948665097356, + 0.0026808951515704393, + 0.0026366179808974266, + 0.002421728102490306, + 0.0025933426804840565, + 0.0036449593026190996, + 0.002485346980392933, + 0.002878220984712243, + 0.002377857454121113, + 0.003222546773031354, + 0.0031724688597023487, + 0.003493205877020955, + 0.00223699607886374, + 0.0012741830432787538, + 0.001594570348970592, + 0.00144813337828964, + 0.0018065649783238769, + 0.002336905337870121, + 0.003437955630943179, + 0.0023823913652449846, + 0.0022882085759192705 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 1 + ], + "selected_reward": -0.020588964223861694 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 high-speed, dynamic swooping movement plunging downward from the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background an industrial ruin featuring broken concrete slabs, rusted rebar, and massive metal trusses?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a boy frozen in midair performing a cartwheel with his body stretched into a straight line parallel to the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the boy wearing a garment woven from vibrant, deep green vines and leaves, and do his eyes glow with fiery orange-red flames?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the overall scene tone cold and desaturated, with the green clothing and fire in the eyes serving as the only vibrant visual centers?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.20154987275600433 + ], + "opposite_logprob_tokens": [ + -1.7015498876571655 + ], + "logprob_reward": -0.20154987275600433, + "yes_no_margin_reward": 1.5, + "token_credit_reward": -0.20154987275600433, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11837203800678253, + 0.07604353874921799, + 0.06045908108353615, + 0.07409758120775223, + 0.11862889677286148, + 0.0913461372256279, + 0.1208137795329094, + 0.0717475488781929, + 0.1334315687417984, + 0.13505986332893372 + ], + "credit_map": [ + 0.012833102606236935, + 0.014017082750797272, + 0.010182073339819908, + 0.013124111108481884, + 0.019671913236379623, + 0.01721608266234398, + 0.015779484063386917, + 0.015548184514045715, + 0.009916292503476143, + 0.009869122877717018, + 0.013131658546626568, + 0.009844114072620869, + 0.008879849687218666, + 0.009953240863978863, + 0.006634932942688465, + 0.007814335636794567, + 0.006878843531012535, + 0.004766005091369152, + 0.009918678551912308, + 0.004452198278158903, + 0.005308075807988644, + 0.011531728319823742, + 0.006935075391083956, + 0.010668473318219185, + 0.010326096788048744, + 0.005604381673038006, + 0.010813442058861256, + 0.009791233576834202, + 0.009814879857003689, + 0.009832673706114292, + 0.00918610394001007, + 0.008728774264454842, + 0.007555413991212845, + 0.009677592664957047, + 0.014352577738463879, + 0.014045873656868935, + 0.009345544502139091, + 0.014058919623494148, + 0.019962066784501076, + 0.029630905017256737, + 0.0063666668720543385, + 0.01992332562804222, + 0.01996525563299656, + 0.008840437047183514, + 0.009301194921135902, + 0.005318581126630306, + 0.012442337349057198, + 0.009188331663608551, + 0.009402343071997166, + 0.020524758845567703, + 0.017343871295452118, + 0.02573280781507492, + 0.01395262498408556, + 0.011724734678864479, + 0.012423144653439522, + 0.009709503501653671, + 0.0073693785816431046, + 0.010796829126775265, + 0.007170391734689474, + 0.008654283359646797, + 0.009983394294977188, + 0.009510962292551994, + 0.007912247441709042, + 0.010350058786571026, + 0.01984274387359619, + 0.02564089372754097, + 0.011879914440214634, + 0.014234447851777077, + 0.015811996534466743, + 0.013622630387544632, + 0.021402541548013687, + 0.010996411554515362, + 0.017387909814715385, + 0.02584090642631054, + 0.010898744687438011, + 0.015267709270119667, + 0.01838771440088749, + 0.01699851267039776, + 0.017662519589066505, + 0.01261584460735321 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 6 + ], + "selected_reward": -0.20154987275600433 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 high-speed, dynamic swooping movement plunging downward from the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background an industrial ruin featuring broken concrete slabs, rusted rebar, and massive metal trusses?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a boy frozen in midair performing a cartwheel with his body stretched into a straight line parallel to the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the boy wearing a garment woven from vibrant, deep green vines and leaves, and do his eyes glow with fiery orange-red flames?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the overall scene tone cold and desaturated, with the green clothing and fire in the eyes serving as the only vibrant visual centers?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00193896540440619 + ], + "opposite_logprob_tokens": [ + -6.251938819885254 + ], + "logprob_reward": -0.00193896540440619, + "yes_no_margin_reward": 6.25, + "token_credit_reward": -0.00193896540440619, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0850939005613327, + 0.05957677587866783, + 0.060484495013952255, + 0.05780879780650139, + 0.15906144678592682, + 0.0867680013179779, + 0.09167665988206863, + 0.11185272783041, + 0.18655157089233398, + 0.10112562775611877 + ], + "credit_map": [ + 0.010275083594024181, + 0.010141931474208832, + 0.008686644025146961, + 0.00965211819857359, + 0.011771820485591888, + 0.010862895287573338, + 0.01449744775891304, + 0.009205964393913746, + 0.005526288412511349, + 0.0064797671511769295, + 0.006926488131284714, + 0.006109394133090973, + 0.007683658041059971, + 0.009323627687990665, + 0.009836038574576378, + 0.007691509556025267, + 0.00616067461669445, + 0.004944666754454374, + 0.010976315476000309, + 0.004739529453217983, + 0.006502105388790369, + 0.009379708208143711, + 0.008406373672187328, + 0.009375125169754028, + 0.007198086474090815, + 0.004714812617748976, + 0.00688150292262435, + 0.008714292198419571, + 0.009377353824675083, + 0.006476213689893484, + 0.008088722825050354, + 0.006357813719660044, + 0.011565512046217918, + 0.011812258511781693, + 0.02415907382965088, + 0.023871243000030518, + 0.027269147336483, + 0.012912816368043423, + 0.01597403921186924, + 0.03149735555052757, + 0.007406649645417929, + 0.016349438577890396, + 0.008278136141598225, + 0.009656498208642006, + 0.0072638047859072685, + 0.005493297707289457, + 0.011328292079269886, + 0.02099187858402729, + 0.006663529668003321, + 0.012859252281486988, + 0.011267508380115032, + 0.014056841842830181, + 0.012023841962218285, + 0.015994122251868248, + 0.007407141383737326, + 0.01140441745519638, + 0.008680194616317749, + 0.009192616678774357, + 0.014629543758928776, + 0.014666513539850712, + 0.014899940229952335, + 0.012946061789989471, + 0.019088856875896454, + 0.01774900034070015, + 0.041839033365249634, + 0.023846836760640144, + 0.02363518811762333, + 0.02026381716132164, + 0.021775316447019577, + 0.021511703729629517, + 0.022142525762319565, + 0.011537134647369385, + 0.012319065630435944, + 0.021510621532797813, + 0.008661770261824131, + 0.011561588384211063, + 0.011029485613107681, + 0.013287592679262161, + 0.013999424874782562, + 0.008756084367632866 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 4, + 7 + ], + "selected_reward": -0.00193896540440619 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 high-speed, dynamic swooping movement plunging downward from the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background an industrial ruin featuring broken concrete slabs, rusted rebar, and massive metal trusses?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a boy frozen in midair performing a cartwheel with his body stretched into a straight line parallel to the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the boy wearing a garment woven from vibrant, deep green vines and leaves, and do his eyes glow with fiery orange-red flames?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the overall scene tone cold and desaturated, with the green clothing and fire in the eyes serving as the only vibrant visual centers?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00863224919885397 + ], + "opposite_logprob_tokens": [ + -4.758632183074951 + ], + "logprob_reward": -0.00863224919885397, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.00863224919885397, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10565484315156937, + 0.06578967720270157, + 0.1302890032529831, + 0.10281854122877121, + 0.2094201296567917, + 0.14618931710720062, + 0.07184837758541107, + 0.041985511779785156, + 0.07441011071205139, + 0.05159448832273483 + ], + "credit_map": [ + 0.01101062074303627, + 0.011213666759431362, + 0.009228235110640526, + 0.012990803457796574, + 0.018783213570713997, + 0.015152966603636742, + 0.011749502271413803, + 0.015525829046964645, + 0.005210044328123331, + 0.006795835215598345, + 0.00784809235483408, + 0.004876749124377966, + 0.009570670314133167, + 0.007453472353518009, + 0.014246201142668724, + 0.00978861004114151, + 0.014938905835151672, + 0.013393078930675983, + 0.024951200932264328, + 0.009099015034735203, + 0.01213360670953989, + 0.017276812344789505, + 0.017722301185131073, + 0.02077406272292137, + 0.013081325218081474, + 0.009075166657567024, + 0.014267032966017723, + 0.010470105335116386, + 0.01383492536842823, + 0.01071104221045971, + 0.01793833076953888, + 0.013440608978271484, + 0.009356780909001827, + 0.014595153741538525, + 0.019489210098981857, + 0.01840505748987198, + 0.031368400901556015, + 0.0363689549267292, + 0.021960748359560966, + 0.05787583068013191, + 0.013352468609809875, + 0.036645062267780304, + 0.02617703564465046, + 0.016842134296894073, + 0.006525589618831873, + 0.0073921591974794865, + 0.01853124052286148, + 0.020723620429635048, + 0.006767048500478268, + 0.010175365023314953, + 0.008572946302592754, + 0.008821402676403522, + 0.00883969571441412, + 0.012631281279027462, + 0.00514940032735467, + 0.010891237296164036, + 0.004747677594423294, + 0.005173911340534687, + 0.004737925250083208, + 0.004968592897057533, + 0.0065920064225792885, + 0.005466004367917776, + 0.0043809381313622, + 0.005918450187891722, + 0.009106813929975033, + 0.016342326998710632, + 0.006250875070691109, + 0.007883835583925247, + 0.007870187051594257, + 0.00747731514275074, + 0.011386044323444366, + 0.008092708885669708, + 0.006139386910945177, + 0.007389434613287449, + 0.004139163997024298, + 0.006516765337437391, + 0.0077736349776387215, + 0.0073862457647919655, + 0.007176219020038843, + 0.005073634907603264 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 5, + 2 + ], + "selected_reward": -0.00863224919885397 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "high-speed, dynamic swooping shot, composed with the rigor of Arnold Newman, plummets downward from the sky, capturing a moment of primal power. Beneath the frame, an industrial ruin of broken concrete slabs, rusted rebar, and massive metal trusses casts sharp, interlocking geometric shadows in the low sunlight. Above this decaying landscape of gray and ochre, a nimble boy is frozen in midair, performing a cartwheel with breathtaking force. He wears a garment meticulously woven from vibrant, deep green vines and leaves, tightly hugging his powerful muscular lines, some of the leaves trembling slightly with the violent twirling. His body stretches in the air into a perfect straight line, parallel to the ground, a display of supreme control. Most striking are his eyes: not human, but two burning flames, gleaming with a fiery orange-red glow, even sparks emanating from them, casting an ominous yet warm glow on his resolute features. The entire scene is painted in a cold and desaturated tone, highlighting the desolation of the ruins. Only the vibrant green of the boy’s clothes and the fire in his eyes become the visual center of the picture, creating a strong dramatic conflict and tense atmosphere of wild vitality bursting out from the dead structure.", + "seed": 29647, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000156/group_1/index_7733_seed_29647_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": -0.165350538212806, + "vlm_logprob": -0.165350538212806, + "vlm_yes_no_margin": 3.025, + "vlm_token_credit": -0.165350538212806, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 high-speed, dynamic swooping movement plunging downward from the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background an industrial ruin featuring broken concrete slabs, rusted rebar, and massive metal trusses?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a boy frozen in midair performing a cartwheel with his body stretched into a straight line parallel to the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the boy wearing a garment woven from vibrant, deep green vines and leaves, and do his eyes glow with fiery orange-red flames?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the overall scene tone cold and desaturated, with the green clothing and fire in the eyes serving as the only vibrant visual centers?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.5273041725158691 + ], + "opposite_logprob_tokens": [ + -0.9023041725158691 + ], + "logprob_reward": -0.5273041725158691, + "yes_no_margin_reward": 0.375, + "token_credit_reward": -0.5273041725158691, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18417170643806458, + 0.17546577751636505, + 0.07152825593948364, + 0.05964672937989235, + 0.04789082705974579, + 0.24395833909511566, + 0.02349383570253849, + 0.035575613379478455, + 0.08519375324249268, + 0.07307516783475876 + ], + "credit_map": [ + 0.022938886657357216, + 0.020528433844447136, + 0.025320090353488922, + 0.015886768698692322, + 0.037767451256513596, + 0.02300407737493515, + 0.016556266695261, + 0.022169725969433784, + 0.01640077866613865, + 0.03634515032172203, + 0.023996727541089058, + 0.017439324408769608, + 0.022701619192957878, + 0.0212422925978899, + 0.014191105961799622, + 0.02314877323806286, + 0.007127363700419664, + 0.010179024189710617, + 0.009158112108707428, + 0.007573430892080069, + 0.006521958392113447, + 0.009308726526796818, + 0.009704655967652798, + 0.011954983696341515, + 0.005804577376693487, + 0.011440832167863846, + 0.005107907112687826, + 0.0072934250347316265, + 0.00825630221515894, + 0.004786358680576086, + 0.006101605482399464, + 0.010855716653168201, + 0.00469635333865881, + 0.007268352434039116, + 0.005255427211523056, + 0.0052258106879889965, + 0.009982999414205551, + 0.006308097392320633, + 0.004834119696170092, + 0.0043196724727749825, + 0.02226834185421467, + 0.03251011669635773, + 0.034495171159505844, + 0.027506645768880844, + 0.025031860917806625, + 0.050601813942193985, + 0.0229885745793581, + 0.028555816039443016, + 0.0023423258680850267, + 0.0021917347330600023, + 0.003380734706297517, + 0.0030592463444918394, + 0.0032072989270091057, + 0.0033962305169552565, + 0.0029867570847272873, + 0.0029295103158801794, + 0.0037256155628710985, + 0.004648849833756685, + 0.0057814959436655045, + 0.004572590813040733, + 0.00643949257209897, + 0.002792136976495385, + 0.003981409594416618, + 0.003634022083133459, + 0.008474336937069893, + 0.0070501575246453285, + 0.009954879060387611, + 0.010341555811464787, + 0.009041287936270237, + 0.017344145104289055, + 0.012203361839056015, + 0.010784025304019451, + 0.00846016500145197, + 0.0051574381068348885, + 0.010319286026060581, + 0.007701056078076363, + 0.010128384456038475, + 0.013931453227996826, + 0.00945583637803793, + 0.007921547628939152 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 1 + ], + "selected_reward": -0.5273041725158691 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 high-speed, dynamic swooping movement plunging downward from the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background an industrial ruin featuring broken concrete slabs, rusted rebar, and massive metal trusses?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a boy frozen in midair performing a cartwheel with his body stretched into a straight line parallel to the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the boy wearing a garment woven from vibrant, deep green vines and leaves, and do his eyes glow with fiery orange-red flames?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the overall scene tone cold and desaturated, with the green clothing and fire in the eyes serving as the only vibrant visual centers?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02635362558066845 + ], + "opposite_logprob_tokens": [ + -3.651353597640991 + ], + "logprob_reward": -0.02635362558066845, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.02635362558066845, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.35419154167175293, + 0.19784599542617798, + 0.09459153562784195, + 0.02963908202946186, + 0.1523667573928833, + 0.047934792935848236, + 0.0350002758204937, + 0.012982526794075966, + 0.044803619384765625, + 0.030643943697214127 + ], + "credit_map": [ + 0.030043113976716995, + 0.05727003514766693, + 0.052296895533800125, + 0.05399177968502045, + 0.034820664674043655, + 0.02898787520825863, + 0.04341926425695419, + 0.0533619187772274, + 0.024442804977297783, + 0.012569817714393139, + 0.04714880511164665, + 0.01198816578835249, + 0.018134959042072296, + 0.01847662962973118, + 0.012655695900321007, + 0.05242910981178284, + 0.007682566065341234, + 0.010822753421962261, + 0.011672139167785645, + 0.012658091261982918, + 0.010803023353219032, + 0.010672767646610737, + 0.013030325062572956, + 0.01724986359477043, + 0.0024039866402745247, + 0.00502101331949234, + 0.003340855473652482, + 0.0033021322451531887, + 0.0032277635764330626, + 0.003700196510180831, + 0.003913834225386381, + 0.004729297943413258, + 0.014955051243305206, + 0.018389180302619934, + 0.01931791938841343, + 0.015604475513100624, + 0.036157816648483276, + 0.02074478007853031, + 0.013377174735069275, + 0.0138203464448452, + 0.0034507582895457745, + 0.007011084351688623, + 0.005572427995502949, + 0.005480150692164898, + 0.005827919114381075, + 0.010492334142327309, + 0.00441950187087059, + 0.005680613219738007, + 0.004275502637028694, + 0.00394790805876255, + 0.003726030234247446, + 0.004280053079128265, + 0.008132071234285831, + 0.0024943032767623663, + 0.004475951194763184, + 0.0036684577353298664, + 0.001495352596975863, + 0.0017416714690625668, + 0.0023430688306689262, + 0.0017178670968860388, + 0.0015308965230360627, + 0.0010844931239262223, + 0.0013465097872540355, + 0.001722667133435607, + 0.004828441422432661, + 0.003693017875775695, + 0.005652576219290495, + 0.005572783760726452, + 0.005225140135735273, + 0.008818421512842178, + 0.0060381535440683365, + 0.004975087009370327, + 0.003957482520490885, + 0.0024900941643863916, + 0.003914430737495422, + 0.0026162811554968357, + 0.006108906120061874, + 0.004699922166764736, + 0.004418530501425266, + 0.0024382956326007843 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 4 + ], + "selected_reward": -0.02635362558066845 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 high-speed, dynamic swooping movement plunging downward from the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background an industrial ruin featuring broken concrete slabs, rusted rebar, and massive metal trusses?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a boy frozen in midair performing a cartwheel with his body stretched into a straight line parallel to the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the boy wearing a garment woven from vibrant, deep green vines and leaves, and do his eyes glow with fiery orange-red flames?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the overall scene tone cold and desaturated, with the green clothing and fire in the eyes serving as the only vibrant visual centers?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.25209712982177734 + ], + "opposite_logprob_tokens": [ + -1.5020971298217773 + ], + "logprob_reward": -0.25209712982177734, + "yes_no_margin_reward": 1.25, + "token_credit_reward": -0.25209712982177734, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16377851366996765, + 0.08176473528146744, + 0.042961232364177704, + 0.032370343804359436, + 0.13989461958408356, + 0.16833925247192383, + 0.07159547507762909, + 0.0395275242626667, + 0.08325541019439697, + 0.17651288211345673 + ], + "credit_map": [ + 0.012943530455231667, + 0.030975213274359703, + 0.018958894535899162, + 0.017440015450119972, + 0.021595953032374382, + 0.011235825717449188, + 0.02880789339542389, + 0.021821195259690285, + 0.009272661991417408, + 0.01474793255329132, + 0.012152574956417084, + 0.006724874023348093, + 0.008794955909252167, + 0.009692012332379818, + 0.007207129616290331, + 0.013172590173780918, + 0.003703071502968669, + 0.004559193272143602, + 0.00545920105651021, + 0.004748205188661814, + 0.003734402358531952, + 0.0076183439232409, + 0.0066347201354801655, + 0.006504093296825886, + 0.0026940025854855776, + 0.005315789021551609, + 0.004280385561287403, + 0.003981791902333498, + 0.00425704475492239, + 0.004612554796040058, + 0.0038746329955756664, + 0.0033541410230100155, + 0.013433012180030346, + 0.018111160025000572, + 0.01684776321053505, + 0.014785883016884327, + 0.03558213636279106, + 0.018430039286613464, + 0.010846096090972424, + 0.01185853686183691, + 0.01691415347158909, + 0.02442239411175251, + 0.0246474239975214, + 0.020266825333237648, + 0.017778759822249413, + 0.030954604968428612, + 0.015244880691170692, + 0.01811019517481327, + 0.008081204257905483, + 0.006145560648292303, + 0.007863919250667095, + 0.009667588397860527, + 0.01167858112603426, + 0.008108081296086311, + 0.009562884457409382, + 0.01048765704035759, + 0.003992339596152306, + 0.0049383193254470825, + 0.0059760757721960545, + 0.005229293368756771, + 0.007666804362088442, + 0.003262374782934785, + 0.004396412521600723, + 0.004065906163305044, + 0.008276727050542831, + 0.007306376472115517, + 0.00891707744449377, + 0.0093041667714715, + 0.011654908768832684, + 0.01719694398343563, + 0.010857569985091686, + 0.009741643443703651, + 0.020564019680023193, + 0.01631169766187668, + 0.025766165927052498, + 0.015805630013346672, + 0.03088608756661415, + 0.023917030543088913, + 0.02732851728796959, + 0.01593373902142048 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 5, + 0 + ], + "selected_reward": -0.25209712982177734 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 high-speed, dynamic swooping movement plunging downward from the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background an industrial ruin featuring broken concrete slabs, rusted rebar, and massive metal trusses?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a boy frozen in midair performing a cartwheel with his body stretched into a straight line parallel to the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the boy wearing a garment woven from vibrant, deep green vines and leaves, and do his eyes glow with fiery orange-red flames?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the overall scene tone cold and desaturated, with the green clothing and fire in the eyes serving as the only vibrant visual centers?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002818067092448473 + ], + "opposite_logprob_tokens": [ + -5.8778181076049805 + ], + "logprob_reward": -0.002818067092448473, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.002818067092448473, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07250652462244034, + 0.059643615037202835, + 0.06059763580560684, + 0.07787348330020905, + 0.0319824181497097, + 0.31747883558273315, + 0.1401396244764328, + 0.05107848346233368, + 0.06648965179920197, + 0.12220965325832367 + ], + "credit_map": [ + 0.00927563477307558, + 0.00787307694554329, + 0.013201622292399406, + 0.0077979848720133305, + 0.009629203006625175, + 0.008165978826582432, + 0.00694660609588027, + 0.009616421535611153, + 0.006518047768622637, + 0.009241574443876743, + 0.012300802394747734, + 0.005292718298733234, + 0.006075703073292971, + 0.003950667101889849, + 0.003505525179207325, + 0.012758585624396801, + 0.0065709189511835575, + 0.009446117095649242, + 0.007935370318591595, + 0.008517787791788578, + 0.007804695051163435, + 0.009631806053221226, + 0.005022491328418255, + 0.00566845340654254, + 0.007880651392042637, + 0.015122651122510433, + 0.0054203360341489315, + 0.009921388700604439, + 0.012921502813696861, + 0.007295949850231409, + 0.008796941488981247, + 0.01051406841725111, + 0.003428239608183503, + 0.005303830374032259, + 0.0041094496846199036, + 0.0037954444997012615, + 0.003981558606028557, + 0.005082401446998119, + 0.003267771564424038, + 0.0030137235298752785, + 0.03209560737013817, + 0.06217718496918678, + 0.04887867346405983, + 0.044468894600868225, + 0.028226692229509354, + 0.04314765706658363, + 0.03129465878009796, + 0.0271894708275795, + 0.016841817647218704, + 0.016471078619360924, + 0.01711259037256241, + 0.020329028367996216, + 0.022102829068899155, + 0.014162230305373669, + 0.014217855408787727, + 0.018902208656072617, + 0.006047936622053385, + 0.004575803410261869, + 0.008685005828738213, + 0.005916247610002756, + 0.007498559076339006, + 0.0038384725339710712, + 0.005578470882028341, + 0.008937989361584187, + 0.007278873585164547, + 0.005543732549995184, + 0.010698309168219566, + 0.007472718600183725, + 0.007545229513198137, + 0.0117656160145998, + 0.009655628353357315, + 0.006529550068080425, + 0.012545937672257423, + 0.01420862227678299, + 0.014767161570489407, + 0.009960636496543884, + 0.027739187702536583, + 0.014781546778976917, + 0.017916595563292503, + 0.010289973579347134 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 6, + 9 + ], + "selected_reward": -0.002818067092448473 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 high-speed, dynamic swooping movement plunging downward from the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background an industrial ruin featuring broken concrete slabs, rusted rebar, and massive metal trusses?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a boy frozen in midair performing a cartwheel with his body stretched into a straight line parallel to the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the boy wearing a garment woven from vibrant, deep green vines and leaves, and do his eyes glow with fiery orange-red flames?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the overall scene tone cold and desaturated, with the green clothing and fire in the eyes serving as the only vibrant visual centers?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018179696053266525 + ], + "opposite_logprob_tokens": [ + -4.018179893493652 + ], + "logprob_reward": -0.018179696053266525, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018179696053266525, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2549690008163452, + 0.24011366069316864, + 0.1054060235619545, + 0.06794261187314987, + 0.02847759611904621, + 0.129696324467659, + 0.06810365617275238, + 0.02742510661482811, + 0.03855344280600548, + 0.03931259736418724 + ], + "credit_map": [ + 0.023132091388106346, + 0.04294103384017944, + 0.03576705604791641, + 0.034285783767700195, + 0.028739765286445618, + 0.022345595061779022, + 0.032315194606781006, + 0.03544246032834053, + 0.025324825197458267, + 0.032551176846027374, + 0.05321476608514786, + 0.01568511128425598, + 0.024716895073652267, + 0.01889556646347046, + 0.011839724145829678, + 0.057885583490133286, + 0.007427924312651157, + 0.014180611819028854, + 0.010493484325706959, + 0.014448930509388447, + 0.012545280158519745, + 0.01522900816053152, + 0.012443987652659416, + 0.01863679476082325, + 0.006469557993113995, + 0.01381093543022871, + 0.005722387228161097, + 0.008580818772315979, + 0.008837149478495121, + 0.0068684713914990425, + 0.006791999563574791, + 0.01086129155009985, + 0.0027962152380496264, + 0.003910483792424202, + 0.003582938574254513, + 0.0035787089727818966, + 0.003815798554569483, + 0.005271357484161854, + 0.002535999519750476, + 0.002986091421917081, + 0.01205445546656847, + 0.020011359825730324, + 0.018073953688144684, + 0.01808198355138302, + 0.009368257597088814, + 0.02373293973505497, + 0.011205322109162807, + 0.01716805435717106, + 0.0067465160973370075, + 0.007148523349314928, + 0.009308374486863613, + 0.009028475731611252, + 0.01328481175005436, + 0.006578573491424322, + 0.00761085981503129, + 0.008397514931857586, + 0.004165512975305319, + 0.002790396101772785, + 0.004282600712031126, + 0.0036264355294406414, + 0.004649064969271421, + 0.0023126662708818913, + 0.003068485762923956, + 0.0025299431290477514, + 0.003923684358596802, + 0.00349682685919106, + 0.006524388678371906, + 0.005109877325594425, + 0.005576347000896931, + 0.00454822788015008, + 0.004450754262506962, + 0.004923333879560232, + 0.0042585586197674274, + 0.003143404610455036, + 0.006090975366532803, + 0.002645662287250161, + 0.005317366681993008, + 0.008155643939971924, + 0.0061995042487978935, + 0.00350148044526577 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 5 + ], + "selected_reward": -0.018179696053266525 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "high-speed, dynamic swooping shot, composed with the rigor of Arnold Newman, plummets downward from the sky, capturing a moment of primal power. Beneath the frame, an industrial ruin of broken concrete slabs, rusted rebar, and massive metal trusses casts sharp, interlocking geometric shadows in the low sunlight. Above this decaying landscape of gray and ochre, a nimble boy is frozen in midair, performing a cartwheel with breathtaking force. He wears a garment meticulously woven from vibrant, deep green vines and leaves, tightly hugging his powerful muscular lines, some of the leaves trembling slightly with the violent twirling. His body stretches in the air into a perfect straight line, parallel to the ground, a display of supreme control. Most striking are his eyes: not human, but two burning flames, gleaming with a fiery orange-red glow, even sparks emanating from them, casting an ominous yet warm glow on his resolute features. The entire scene is painted in a cold and desaturated tone, highlighting the desolation of the ruins. Only the vibrant green of the boy’s clothes and the fire in his eyes become the visual center of the picture, creating a strong dramatic conflict and tense atmosphere of wild vitality bursting out from the dead structure.", + "seed": 29648, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000156/group_1/index_7733_seed_29648_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": -0.06112081515602767, + "vlm_logprob": -0.06112081515602767, + "vlm_yes_no_margin": 3.624999976158142, + "vlm_token_credit": -0.06112081515602767, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 high-speed, dynamic swooping movement plunging downward from the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background an industrial ruin featuring broken concrete slabs, rusted rebar, and massive metal trusses?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a boy frozen in midair performing a cartwheel with his body stretched into a straight line parallel to the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the boy wearing a garment woven from vibrant, deep green vines and leaves, and do his eyes glow with fiery orange-red flames?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the overall scene tone cold and desaturated, with the green clothing and fire in the eyes serving as the only vibrant visual centers?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.14681346714496613 + ], + "opposite_logprob_tokens": [ + -2.02181339263916 + ], + "logprob_reward": -0.14681346714496613, + "yes_no_margin_reward": 1.8749998807907104, + "token_credit_reward": -0.14681346714496613, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13799594342708588, + 0.24784520268440247, + 0.04151136055588722, + 0.05518808588385582, + 0.06332620978355408, + 0.04635923355817795, + 0.09977064281702042, + 0.22978699207305908, + 0.04299415647983551, + 0.035222169011831284 + ], + "credit_map": [ + 0.01855866238474846, + 0.023096246644854546, + 0.021415075287222862, + 0.01536561083048582, + 0.017663760110735893, + 0.01339889969676733, + 0.011442864313721657, + 0.017054805532097816, + 0.02701410837471485, + 0.01881757751107216, + 0.04753902927041054, + 0.04196982458233833, + 0.019613776355981827, + 0.0385337695479393, + 0.021787606179714203, + 0.03256950527429581, + 0.0030794430058449507, + 0.003116196021437645, + 0.008155819959938526, + 0.006435359828174114, + 0.005091093014925718, + 0.007555672433227301, + 0.004425697028636932, + 0.0036520743742585182, + 0.006753061432391405, + 0.006200364325195551, + 0.007646099664270878, + 0.006534331478178501, + 0.006773616652935743, + 0.007647472899407148, + 0.007783944718539715, + 0.005849186796694994, + 0.00651909876614809, + 0.009302300401031971, + 0.006529851816594601, + 0.008396524004638195, + 0.009939919225871563, + 0.007858924567699432, + 0.005783886183053255, + 0.00899569597095251, + 0.0064112236723303795, + 0.006684069521725178, + 0.0058056204579770565, + 0.006785341538488865, + 0.00447967741638422, + 0.005759004503488541, + 0.004605100490152836, + 0.005829192232340574, + 0.007883681915700436, + 0.00899938028305769, + 0.009138895198702812, + 0.012487762607634068, + 0.02185947820544243, + 0.013841125182807446, + 0.01303509809076786, + 0.012525211088359356, + 0.013401142321527004, + 0.009885343722999096, + 0.07788977026939392, + 0.0133393919095397, + 0.016000904142856598, + 0.07848823070526123, + 0.009114982560276985, + 0.011667223647236824, + 0.011534195393323898, + 0.004243849311023951, + 0.012161720544099808, + 0.002170517574995756, + 0.00399528443813324, + 0.003143971087411046, + 0.002631473820656538, + 0.0031131417490541935, + 0.006205792538821697, + 0.005597118753939867, + 0.00705721415579319, + 0.0032047743443399668, + 0.004236577544361353, + 0.0032520454842597246, + 0.0023224863689392805, + 0.0033461563289165497 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 7, + 0 + ], + "selected_reward": -0.14681346714496613 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 high-speed, dynamic swooping movement plunging downward from the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background an industrial ruin featuring broken concrete slabs, rusted rebar, and massive metal trusses?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a boy frozen in midair performing a cartwheel with his body stretched into a straight line parallel to the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the boy wearing a garment woven from vibrant, deep green vines and leaves, and do his eyes glow with fiery orange-red flames?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the overall scene tone cold and desaturated, with the green clothing and fire in the eyes serving as the only vibrant visual centers?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04866454377770424 + ], + "opposite_logprob_tokens": [ + -3.0486645698547363 + ], + "logprob_reward": -0.04866454377770424, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04866454377770424, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10772483795881271, + 0.09759163856506348, + 0.27424439787864685, + 0.12803669273853302, + 0.06308489292860031, + 0.05220329761505127, + 0.03867219015955925, + 0.13174262642860413, + 0.0799511969089508, + 0.026748163625597954 + ], + "credit_map": [ + 0.022622907534241676, + 0.013784050941467285, + 0.019446684047579765, + 0.011371297761797905, + 0.011493184603750706, + 0.009335000067949295, + 0.008265374228358269, + 0.01140634436160326, + 0.009891327470541, + 0.009011654183268547, + 0.0243903286755085, + 0.011378385126590729, + 0.008140264078974724, + 0.010495648719370365, + 0.011876001954078674, + 0.012408026494085789, + 0.02271222695708275, + 0.008221358992159367, + 0.10562562942504883, + 0.013155112974345684, + 0.010632420890033245, + 0.09320151805877686, + 0.009213263168931007, + 0.011482877656817436, + 0.01083945482969284, + 0.024257652461528778, + 0.015522236935794353, + 0.0160121601074934, + 0.015148581936955452, + 0.015572043135762215, + 0.012763588689267635, + 0.01792098954319954, + 0.0058170510455966, + 0.011089848354458809, + 0.007463262882083654, + 0.006738258525729179, + 0.00912459846585989, + 0.007792433723807335, + 0.008215217851102352, + 0.006844226736575365, + 0.00828214269131422, + 0.0073567950166761875, + 0.004716978874057531, + 0.0058255670592188835, + 0.0038109917659312487, + 0.008511967957019806, + 0.004716827999800444, + 0.008982032537460327, + 0.0037467770744115114, + 0.0036845915019512177, + 0.004109119065105915, + 0.007553240284323692, + 0.006149886641651392, + 0.0055261822417378426, + 0.004916229750961065, + 0.002986165462061763, + 0.007939827628433704, + 0.00838114507496357, + 0.04016359522938728, + 0.009265845641493797, + 0.01256400253623724, + 0.038147032260894775, + 0.0068715717643499374, + 0.008409610949456692, + 0.016416318714618683, + 0.007202413864433765, + 0.018419088795781136, + 0.006077038589864969, + 0.0057517727836966515, + 0.008414443582296371, + 0.010008118115365505, + 0.007662008050829172, + 0.003684282535687089, + 0.0036831116303801537, + 0.003278970019891858, + 0.003612517612054944, + 0.002191077684983611, + 0.0027565110940486193, + 0.0032188668847084045, + 0.004322827327996492 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 7, + 3 + ], + "selected_reward": -0.04866454377770424 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 high-speed, dynamic swooping movement plunging downward from the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background an industrial ruin featuring broken concrete slabs, rusted rebar, and massive metal trusses?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a boy frozen in midair performing a cartwheel with his body stretched into a straight line parallel to the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the boy wearing a garment woven from vibrant, deep green vines and leaves, and do his eyes glow with fiery orange-red flames?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the overall scene tone cold and desaturated, with the green clothing and fire in the eyes serving as the only vibrant visual centers?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10030568391084671 + ], + "opposite_logprob_tokens": [ + -2.3503057956695557 + ], + "logprob_reward": -0.10030568391084671, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10030568391084671, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18246445059776306, + 0.2589993476867676, + 0.07569237798452377, + 0.08691416680812836, + 0.08487673848867416, + 0.06758023053407669, + 0.07571633160114288, + 0.04178183525800705, + 0.056502509862184525, + 0.06947208940982819 + ], + "credit_map": [ + 0.02889453060925007, + 0.018720095977187157, + 0.025881845504045486, + 0.019953694194555283, + 0.022350210696458817, + 0.020745644345879555, + 0.01973472349345684, + 0.026183679699897766, + 0.0359463095664978, + 0.02309132181107998, + 0.05276431515812874, + 0.037069715559482574, + 0.01753951981663704, + 0.03153838589787483, + 0.028437281027436256, + 0.03261246904730797, + 0.007050783839076757, + 0.010828394442796707, + 0.013741891831159592, + 0.009109804406762123, + 0.0068763564340770245, + 0.01602274365723133, + 0.007725250441581011, + 0.004337139893323183, + 0.013567018322646618, + 0.013648414984345436, + 0.015718864277005196, + 0.011808743700385094, + 0.008316528052091599, + 0.0071613178588449955, + 0.007126417942345142, + 0.00956684909760952, + 0.00910765677690506, + 0.014312807470560074, + 0.011986597441136837, + 0.008306360803544521, + 0.011861433275043964, + 0.009956875815987587, + 0.009101875126361847, + 0.010243122465908527, + 0.011040442623198032, + 0.011542066000401974, + 0.00621009012684226, + 0.008158378303050995, + 0.006552509963512421, + 0.0071642110124230385, + 0.0061002750881016254, + 0.010812259279191494, + 0.006010760553181171, + 0.00784916803240776, + 0.008228552527725697, + 0.01112182904034853, + 0.016807086765766144, + 0.008809697814285755, + 0.00812226627022028, + 0.0087669612839818, + 0.003054016502574086, + 0.003696870757266879, + 0.009049208834767342, + 0.004197014030069113, + 0.005324168596416712, + 0.009153197519481182, + 0.003133319551125169, + 0.004174037370830774, + 0.007793492171913385, + 0.006723848171532154, + 0.008500616066157818, + 0.0050753457471728325, + 0.00698848208412528, + 0.00825649406760931, + 0.007307776715606451, + 0.005856449715793133, + 0.012670253403484821, + 0.011730118654668331, + 0.013706211932003498, + 0.006597875151783228, + 0.007335122674703598, + 0.007068511564284563, + 0.005042162258177996, + 0.005321831908077002 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 3 + ], + "selected_reward": -0.10030568391084671 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 high-speed, dynamic swooping movement plunging downward from the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background an industrial ruin featuring broken concrete slabs, rusted rebar, and massive metal trusses?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a boy frozen in midair performing a cartwheel with his body stretched into a straight line parallel to the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the boy wearing a garment woven from vibrant, deep green vines and leaves, and do his eyes glow with fiery orange-red flames?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the overall scene tone cold and desaturated, with the green clothing and fire in the eyes serving as the only vibrant visual centers?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0021961636375635862 + ], + "opposite_logprob_tokens": [ + -6.127196311950684 + ], + "logprob_reward": -0.0021961636375635862, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.0021961636375635862, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08097179234027863, + 0.04486287757754326, + 0.0349547415971756, + 0.0709896832704544, + 0.06707999855279922, + 0.04278272017836571, + 0.07628011703491211, + 0.5130233764648438, + 0.03739377483725548, + 0.03166091442108154 + ], + "credit_map": [ + 0.011531825177371502, + 0.010271209292113781, + 0.010286225005984306, + 0.01054761465638876, + 0.011186777614057064, + 0.011150008998811245, + 0.007564822677522898, + 0.00843330379575491, + 0.0068659232929348946, + 0.00604198919609189, + 0.005982667673379183, + 0.00786370225250721, + 0.003975841216742992, + 0.0037740764673799276, + 0.004153416957706213, + 0.006205259822309017, + 0.004577455576509237, + 0.004168341401964426, + 0.006489449180662632, + 0.003393525257706642, + 0.003992146346718073, + 0.006788537837564945, + 0.0029609601479023695, + 0.0025843267794698477, + 0.006904647219926119, + 0.008462585508823395, + 0.006693223956972361, + 0.008601288311183453, + 0.008374710567295551, + 0.010953172110021114, + 0.010072404518723488, + 0.010927656665444374, + 0.0071305627934634686, + 0.010820484720170498, + 0.005802927538752556, + 0.008560526184737682, + 0.008554667234420776, + 0.009288373403251171, + 0.006198910064995289, + 0.010723548009991646, + 0.005898736417293549, + 0.007294868119060993, + 0.00426489906385541, + 0.006135710049420595, + 0.003062264062464237, + 0.005622268654406071, + 0.0035240482538938522, + 0.006979925557971001, + 0.006721031852066517, + 0.00882659200578928, + 0.007954411208629608, + 0.01398402452468872, + 0.017005931586027145, + 0.007123357616364956, + 0.008678718470036983, + 0.00598605303093791, + 0.029631884768605232, + 0.020140767097473145, + 0.1794346123933792, + 0.028011025860905647, + 0.024060651659965515, + 0.18805289268493652, + 0.019850848242640495, + 0.023840680718421936, + 0.005985413212329149, + 0.005396872293204069, + 0.006481284275650978, + 0.002663444960489869, + 0.004101793747395277, + 0.00461575947701931, + 0.005203824490308762, + 0.0029453830793499947, + 0.006424711551517248, + 0.005688657518476248, + 0.004807711113244295, + 0.0022455838043242693, + 0.003914300352334976, + 0.0031783529557287693, + 0.002170416759327054, + 0.0032311805989593267 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 6 + ], + "selected_reward": -0.0021961636375635862 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 high-speed, dynamic swooping movement plunging downward from the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background an industrial ruin featuring broken concrete slabs, rusted rebar, and massive metal trusses?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a boy frozen in midair performing a cartwheel with his body stretched into a straight line parallel to the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the boy wearing a garment woven from vibrant, deep green vines and leaves, and do his eyes glow with fiery orange-red flames?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the overall scene tone cold and desaturated, with the green clothing and fire in the eyes serving as the only vibrant visual centers?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0076242173090577126 + ], + "opposite_logprob_tokens": [ + -4.88262414932251 + ], + "logprob_reward": -0.0076242173090577126, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.0076242173090577126, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05892379209399223, + 0.08696677535772324, + 0.16918188333511353, + 0.08156203478574753, + 0.07699326425790787, + 0.0650927796959877, + 0.04902315139770508, + 0.3594208061695099, + 0.033072371035814285, + 0.019763119518756866 + ], + "credit_map": [ + 0.008085213601589203, + 0.004787950310856104, + 0.007909972220659256, + 0.008398868143558502, + 0.008163000456988811, + 0.007150924764573574, + 0.009433952160179615, + 0.004993912298232317, + 0.009221464395523071, + 0.008500480093061924, + 0.02040213532745838, + 0.010370816104114056, + 0.009312290698289871, + 0.008481521159410477, + 0.009048697538673878, + 0.011629370972514153, + 0.016654126346111298, + 0.013959256000816822, + 0.052117157727479935, + 0.00970680546015501, + 0.013366776518523693, + 0.0480973906815052, + 0.007981343194842339, + 0.007299022749066353, + 0.013187216594815254, + 0.00986908283084631, + 0.01810499094426632, + 0.007955889217555523, + 0.0066797188483178616, + 0.009370333515107632, + 0.008051449432969093, + 0.008343354798853397, + 0.00620981166139245, + 0.014859138987958431, + 0.009947272948920727, + 0.01059331838041544, + 0.008876320905983448, + 0.008954761549830437, + 0.008679097518324852, + 0.008873539976775646, + 0.008810139261186123, + 0.008608122356235981, + 0.006105202250182629, + 0.007838021032512188, + 0.005471233278512955, + 0.010700500570237637, + 0.006766581442207098, + 0.010792981833219528, + 0.0045635090209543705, + 0.004009265452623367, + 0.006757246796041727, + 0.008173255249857903, + 0.007453246507793665, + 0.00767942052334547, + 0.005138750653713942, + 0.005248455796390772, + 0.021064188331365585, + 0.013717878609895706, + 0.12762773036956787, + 0.020516417920589447, + 0.013640089891850948, + 0.13251791894435883, + 0.013937834650278091, + 0.016398757696151733, + 0.0070016696117818356, + 0.003304357174783945, + 0.007698460016399622, + 0.002389539498835802, + 0.0032801677007228136, + 0.003336368827149272, + 0.003296343144029379, + 0.002765463897958398, + 0.002634296892210841, + 0.0021424018777906895, + 0.0029770119581371546, + 0.0021228052210062742, + 0.0020581604912877083, + 0.003214888973161578, + 0.002233531791716814, + 0.0023800223134458065 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 2, + 1 + ], + "selected_reward": -0.0076242173090577126 + } + ], + "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/0000156/group_10/index_tensor([6862, 6862])_group_10_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000156/group_10/index_tensor([6862, 6862])_group_10_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..3c9082e5f20708215e499e3b02005a2d561fa2ab --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000156/group_10/index_tensor([6862, 6862])_group_10_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "As the camera smoothly glides from a high-angle aerial perspective over gently rolling hills bathed in the warm golden glow of the afternoon sun, a vast, cinematic panorama unfolds. The landscape is a tapestry of meticulously arranged, sun-drenched grapevines, their vibrant green leaves laden with clusters of deep purple grapes, arranged in perfectly parallel rows following the rolling contours of the land. A narrow, light-brown dirt path, its surface dry and dusted with fine dust and loose pebbles, gracefully winds between the rows, its sinuous shape drawing the eye deeply into the vast landscape. The low-angle sunlight casts long, soft shadows from the vines lining the path, creating rhythmic patterns of light and dark that accentuate the terrain. The entire scene is imbued with a profound tranquility and idyllic atmosphere, capturing the serene, breathtaking beauty of nature beneath a brilliant, cloudless sky.", + "seed": 21514, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000156/group_10/index_6862_seed_21514_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": -0.03766844240017235, + "vlm_logprob": -0.03766844240017235, + "vlm_yes_no_margin": 3.9, + "vlm_token_credit": -0.03766844240017235, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 high-angle aerial view gliding over the landscape?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the hills covered with rows of grapevines bearing deep purple grapes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a narrow, light-brown dirt path winding between the grapevine rows visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by low-angle afternoon sunlight casting long, soft shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the sky in the scene brilliant and cloudless?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.032614655792713165 + ], + "opposite_logprob_tokens": [ + -3.5326147079467773 + ], + "logprob_reward": -0.032614655792713165, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.032614655792713165, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16621068120002747, + 0.08438658714294434, + 0.05901237949728966, + 0.15559816360473633, + 0.07914945483207703, + 0.0575421005487442, + 0.13557958602905273, + 0.1397988200187683, + 0.06472872942686081, + 0.057993464171886444 + ], + "credit_map": [ + 0.020531663671135902, + 0.014702149666845798, + 0.016733061522245407, + 0.018186120316386223, + 0.01426626741886139, + 0.020008228719234467, + 0.03389505669474602, + 0.027888145297765732, + 0.0071303690783679485, + 0.011015682481229305, + 0.012501264922320843, + 0.015601620078086853, + 0.007552051451057196, + 0.006826261058449745, + 0.013083348982036114, + 0.010675991885364056, + 0.009355565533041954, + 0.008866199292242527, + 0.005450278054922819, + 0.005476762540638447, + 0.008126327767968178, + 0.0073366365395486355, + 0.008164494298398495, + 0.006236113142222166, + 0.01776236481964588, + 0.01666072942316532, + 0.019366789609193802, + 0.01792919635772705, + 0.011908185668289661, + 0.028343990445137024, + 0.029579443857073784, + 0.0140474708750844, + 0.009916158393025398, + 0.01145996991544962, + 0.013406305573880672, + 0.011077672243118286, + 0.007248015142977238, + 0.00783819891512394, + 0.00958267506211996, + 0.008620457723736763, + 0.007249041926115751, + 0.008110538125038147, + 0.00491194985806942, + 0.00426208833232522, + 0.004400925245136023, + 0.005914404056966305, + 0.010500303469598293, + 0.012192850932478905, + 0.012074539437890053, + 0.010564994066953659, + 0.0388248972594738, + 0.015932254493236542, + 0.006072666496038437, + 0.010481518693268299, + 0.008872662670910358, + 0.03275606036186218, + 0.01712394505739212, + 0.01231475081294775, + 0.01742825098335743, + 0.026253394782543182, + 0.02058771811425686, + 0.011081933975219727, + 0.02313942089676857, + 0.011869389563798904, + 0.005903997458517551, + 0.0028472193516790867, + 0.01803085207939148, + 0.0052789608016610146, + 0.01704738289117813, + 0.0042893728241324425, + 0.004765855614095926, + 0.006565087009221315, + 0.0074519626796245575, + 0.007132346276193857, + 0.005518211051821709, + 0.005179123487323523, + 0.008900550194084644, + 0.006440399680286646, + 0.009715577587485313, + 0.0076552932150661945 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 7 + ], + "selected_reward": -0.032614655792713165 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 high-angle aerial view gliding over the landscape?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the hills covered with rows of grapevines bearing deep purple grapes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a narrow, light-brown dirt path winding between the grapevine rows visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by low-angle afternoon sunlight casting long, soft shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the sky in the scene brilliant and cloudless?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05497133359313011 + ], + "opposite_logprob_tokens": [ + -2.929971218109131 + ], + "logprob_reward": -0.05497133359313011, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05497133359313011, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13029921054840088, + 0.20885343849658966, + 0.03828601539134979, + 0.04019948095083237, + 0.09784200042486191, + 0.08920899033546448, + 0.21974562108516693, + 0.07937174290418625, + 0.015913458541035652, + 0.08028008788824081 + ], + "credit_map": [ + 0.01375515665858984, + 0.016281450167298317, + 0.01994885876774788, + 0.015694385394454002, + 0.018667545169591904, + 0.009809565730392933, + 0.01946367882192135, + 0.01667855866253376, + 0.02535800263285637, + 0.023799145594239235, + 0.034144140779972076, + 0.03669431433081627, + 0.016953395679593086, + 0.013682329095900059, + 0.02974063716828823, + 0.02848144806921482, + 0.006292625796049833, + 0.005130969453603029, + 0.0041535417549312115, + 0.0033534241374582052, + 0.004959136247634888, + 0.006318544037640095, + 0.0036091043148189783, + 0.004468667320907116, + 0.006702296435832977, + 0.005079754628241062, + 0.004361313302069902, + 0.003656717948615551, + 0.0032753096893429756, + 0.0069711399264633656, + 0.006683129817247391, + 0.0034698171075433493, + 0.010723263025283813, + 0.017176806926727295, + 0.01679247058928013, + 0.011964298784732819, + 0.007990503683686256, + 0.007147990167140961, + 0.01457454264163971, + 0.011472112499177456, + 0.012219595722854137, + 0.010933195240795612, + 0.00741847138851881, + 0.008326783776283264, + 0.009326151572167873, + 0.013966578990221024, + 0.017534947022795677, + 0.00948325451463461, + 0.0170145221054554, + 0.01331835612654686, + 0.06248292326927185, + 0.022377047687768936, + 0.009605752304196358, + 0.011795851401984692, + 0.014138171449303627, + 0.06901296228170395, + 0.008615422993898392, + 0.0059599969536066055, + 0.007808804977685213, + 0.008728375658392906, + 0.01361654419451952, + 0.00710503663867712, + 0.01845700852572918, + 0.009080548770725727, + 0.0020843790844082832, + 0.001374334329739213, + 0.0014212202513590455, + 0.002499375957995653, + 0.0030160353053361177, + 0.0015569939278066158, + 0.0019276903476566076, + 0.0020334285218268633, + 0.010572614148259163, + 0.006356755271553993, + 0.009033158421516418, + 0.007916856557130814, + 0.010360478423535824, + 0.008892725221812725, + 0.014093070290982723, + 0.013054421171545982 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 1, + 0 + ], + "selected_reward": -0.05497133359313011 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 high-angle aerial view gliding over the landscape?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the hills covered with rows of grapevines bearing deep purple grapes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a narrow, light-brown dirt path winding between the grapevine rows visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by low-angle afternoon sunlight casting long, soft shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the sky in the scene brilliant and cloudless?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0011799524072557688 + ], + "opposite_logprob_tokens": [ + -6.751180171966553 + ], + "logprob_reward": -0.0011799524072557688, + "yes_no_margin_reward": 6.75, + "token_credit_reward": -0.0011799524072557688, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15397252142429352, + 0.1336108148097992, + 0.06506957858800888, + 0.10418536514043808, + 0.06312002241611481, + 0.03749105706810951, + 0.045193493366241455, + 0.16080984473228455, + 0.17766423523426056, + 0.05888303741812706 + ], + "credit_map": [ + 0.023550163954496384, + 0.018034376204013824, + 0.01479396317154169, + 0.01660710945725441, + 0.012711020186543465, + 0.011948912404477596, + 0.036799270659685135, + 0.01952771656215191, + 0.01350948866456747, + 0.014995968900620937, + 0.01938381791114807, + 0.020733490586280823, + 0.012757852673530579, + 0.00973084568977356, + 0.026996303349733353, + 0.015503060072660446, + 0.00740836700424552, + 0.006629751995205879, + 0.007831905968487263, + 0.007551920600235462, + 0.006418547593057156, + 0.015498469583690166, + 0.006399947684258223, + 0.007330672349780798, + 0.015159056521952152, + 0.01459529995918274, + 0.013988873921334743, + 0.012966576963663101, + 0.009381375275552273, + 0.008906440809369087, + 0.01527952030301094, + 0.013908235356211662, + 0.008307636715471745, + 0.009393317624926567, + 0.006264669820666313, + 0.007957047782838345, + 0.00596030056476593, + 0.009398968890309334, + 0.008427233435213566, + 0.00741084897890687, + 0.0036183111369609833, + 0.0033937853295356035, + 0.003968918230384588, + 0.003356993431225419, + 0.004772641230374575, + 0.006775830872356892, + 0.007014211732894182, + 0.004590365570038557, + 0.008278337307274342, + 0.0041946955025196075, + 0.00973918754607439, + 0.004180502146482468, + 0.004025763366371393, + 0.0048865084536373615, + 0.005627072881907225, + 0.00426142942160368, + 0.016365768387913704, + 0.014100833795964718, + 0.022953854873776436, + 0.03125534951686859, + 0.020665787160396576, + 0.015236944891512394, + 0.020873291417956352, + 0.019358042627573013, + 0.02146781235933304, + 0.007674600929021835, + 0.03947293013334274, + 0.010543170385062695, + 0.05622624233365059, + 0.013665266335010529, + 0.010794437490403652, + 0.017819784581661224, + 0.005962524097412825, + 0.0046809096820652485, + 0.005562777630984783, + 0.006485656835138798, + 0.010976338759064674, + 0.005495011806488037, + 0.009030215442180634, + 0.010689605958759785 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 0 + ], + "selected_reward": -0.0011799524072557688 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 high-angle aerial view gliding over the landscape?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the hills covered with rows of grapevines bearing deep purple grapes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a narrow, light-brown dirt path winding between the grapevine rows visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by low-angle afternoon sunlight casting long, soft shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the sky in the scene brilliant and cloudless?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020580438897013664 + ], + "opposite_logprob_tokens": [ + -3.895580530166626 + ], + "logprob_reward": -0.020580438897013664, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020580438897013664, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06370276212692261, + 0.1866077333688736, + 0.2423730343580246, + 0.040510766208171844, + 0.038290124386548996, + 0.10376684367656708, + 0.19754834473133087, + 0.05964932590723038, + 0.04030797630548477, + 0.027243033051490784 + ], + "credit_map": [ + 0.009460797533392906, + 0.009339500218629837, + 0.008177385665476322, + 0.007297610864043236, + 0.005914685782045126, + 0.005072306375950575, + 0.00791863352060318, + 0.010521848686039448, + 0.017738599330186844, + 0.020448895171284676, + 0.026952678337693214, + 0.03275115787982941, + 0.01603611372411251, + 0.012593237683176994, + 0.03909977898001671, + 0.020987290889024734, + 0.0456562414765358, + 0.026355478912591934, + 0.027115346863865852, + 0.02661757543683052, + 0.029627447947859764, + 0.01932637393474579, + 0.03483306244015694, + 0.0328415147960186, + 0.006623005960136652, + 0.004046643618494272, + 0.004804464057087898, + 0.004235421773046255, + 0.003768899943679571, + 0.006096397526562214, + 0.007143544033169746, + 0.003792389063164592, + 0.004603233188390732, + 0.006520049646496773, + 0.006760984659194946, + 0.004321056883782148, + 0.002991537796333432, + 0.003963073715567589, + 0.004754033405333757, + 0.004376154858618975, + 0.013732017017900944, + 0.012944535352289677, + 0.010867861099541187, + 0.013448070734739304, + 0.010388493537902832, + 0.01240178756415844, + 0.020288122817873955, + 0.00969595555216074, + 0.023590940982103348, + 0.01401966717094183, + 0.03680147975683212, + 0.018149444833397865, + 0.009473133832216263, + 0.013778732158243656, + 0.012993511743843555, + 0.06874144077301025, + 0.008056019432842731, + 0.005637215916067362, + 0.007659498136490583, + 0.009234224446117878, + 0.0082711735740304, + 0.005957925692200661, + 0.00942736491560936, + 0.005405904725193977, + 0.005073012318462133, + 0.002535365056246519, + 0.009525002911686897, + 0.003195864614099264, + 0.008798315189778805, + 0.004252434708178043, + 0.0034464243799448013, + 0.00348155596293509, + 0.0031669859308749437, + 0.002597382990643382, + 0.003106725635007024, + 0.002641415223479271, + 0.004837317857891321, + 0.002733038505539298, + 0.004374957177788019, + 0.0037852113600820303 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 6, + 1 + ], + "selected_reward": -0.020580438897013664 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 high-angle aerial view gliding over the landscape?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the hills covered with rows of grapevines bearing deep purple grapes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a narrow, light-brown dirt path winding between the grapevine rows visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by low-angle afternoon sunlight casting long, soft shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the sky in the scene brilliant and cloudless?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07899583131074905 + ], + "opposite_logprob_tokens": [ + -2.578995943069458 + ], + "logprob_reward": -0.07899583131074905, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07899583131074905, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09502038359642029, + 0.3145928978919983, + 0.25053954124450684, + 0.08146601915359497, + 0.04178998991847038, + 0.04483994096517563, + 0.0452684611082077, + 0.07630623877048492, + 0.0311244148761034, + 0.019052080810070038 + ], + "credit_map": [ + 0.01191727165132761, + 0.01112412754446268, + 0.014553030021488667, + 0.014100924134254456, + 0.00968484953045845, + 0.006438673473894596, + 0.016040103510022163, + 0.01116140279918909, + 0.03282065689563751, + 0.036965884268283844, + 0.040868937969207764, + 0.056403618305921555, + 0.0285467728972435, + 0.026823662221431732, + 0.054823484271764755, + 0.03733990341424942, + 0.0470198318362236, + 0.03481246158480644, + 0.03610500320792198, + 0.03203364834189415, + 0.022864066064357758, + 0.02237616293132305, + 0.02944757416844368, + 0.025880789384245872, + 0.013686076737940311, + 0.009929649531841278, + 0.011920423246920109, + 0.013983295299112797, + 0.006279203575104475, + 0.005843163467943668, + 0.010029102675616741, + 0.009795106016099453, + 0.00590751226991415, + 0.006659632548689842, + 0.006416765507310629, + 0.005463968496769667, + 0.003411773592233658, + 0.003973045852035284, + 0.005513899028301239, + 0.004443393088877201, + 0.005510097835212946, + 0.004770077299326658, + 0.0049441163428127766, + 0.006227097939699888, + 0.005162988789379597, + 0.004604534711688757, + 0.0075541953556239605, + 0.006066831760108471, + 0.005387333687394857, + 0.005680250935256481, + 0.007481961511075497, + 0.005686413962393999, + 0.0038248326163738966, + 0.005749173928052187, + 0.004464342724531889, + 0.006994151044636965, + 0.006147427950054407, + 0.006790617015212774, + 0.009905405342578888, + 0.014463928528130054, + 0.013019951060414314, + 0.006111982278525829, + 0.011952955275774002, + 0.007913975045084953, + 0.005201530177146196, + 0.0024516074918210506, + 0.0031092276331037283, + 0.0034562675282359123, + 0.005896410439163446, + 0.003799647791311145, + 0.0038546076975762844, + 0.0033551184460520744, + 0.002216423861682415, + 0.001574191963300109, + 0.001918705878779292, + 0.0015819971449673176, + 0.002191604347899556, + 0.0025178983341902494, + 0.003583574900403619, + 0.003467684844508767 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 2, + 0 + ], + "selected_reward": -0.07899583131074905 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "As the camera smoothly glides from a high-angle aerial perspective over gently rolling hills bathed in the warm golden glow of the afternoon sun, a vast, cinematic panorama unfolds. The landscape is a tapestry of meticulously arranged, sun-drenched grapevines, their vibrant green leaves laden with clusters of deep purple grapes, arranged in perfectly parallel rows following the rolling contours of the land. A narrow, light-brown dirt path, its surface dry and dusted with fine dust and loose pebbles, gracefully winds between the rows, its sinuous shape drawing the eye deeply into the vast landscape. The low-angle sunlight casts long, soft shadows from the vines lining the path, creating rhythmic patterns of light and dark that accentuate the terrain. The entire scene is imbued with a profound tranquility and idyllic atmosphere, capturing the serene, breathtaking beauty of nature beneath a brilliant, cloudless sky.", + "seed": 21515, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000156/group_10/index_6862_seed_21515_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": -0.05332963766995817, + "vlm_logprob": -0.05332963766995817, + "vlm_yes_no_margin": 3.65, + "vlm_token_credit": -0.05332963766995817, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 high-angle aerial view gliding over the landscape?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the hills covered with rows of grapevines bearing deep purple grapes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a narrow, light-brown dirt path winding between the grapevine rows visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by low-angle afternoon sunlight casting long, soft shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the sky in the scene brilliant and cloudless?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04585405811667442 + ], + "opposite_logprob_tokens": [ + -3.170854091644287 + ], + "logprob_reward": -0.04585405811667442, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04585405811667442, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2011851817369461, + 0.09229489415884018, + 0.1305796205997467, + 0.11045069992542267, + 0.0916358008980751, + 0.03332377225160599, + 0.13075928390026093, + 0.06390921026468277, + 0.036146536469459534, + 0.10971499979496002 + ], + "credit_map": [ + 0.025833649560809135, + 0.019595211371779442, + 0.022260654717683792, + 0.02114882692694664, + 0.01300760731101036, + 0.021547820419073105, + 0.03385113552212715, + 0.04394029825925827, + 0.009858112782239914, + 0.008613143116235733, + 0.012389381416141987, + 0.02172335796058178, + 0.00868277344852686, + 0.010854206047952175, + 0.011192897334694862, + 0.00898103415966034, + 0.013783366419374943, + 0.016583995893597603, + 0.01947062276303768, + 0.02436339482665062, + 0.011900875717401505, + 0.02399854175746441, + 0.011176500469446182, + 0.00930233858525753, + 0.01286289468407631, + 0.016960494220256805, + 0.019552169367671013, + 0.02383572608232498, + 0.0069918860681355, + 0.013559062033891678, + 0.008753941394388676, + 0.007934538647532463, + 0.01070873998105526, + 0.008061055094003677, + 0.02217443473637104, + 0.012262215837836266, + 0.005727957468479872, + 0.017259642481803894, + 0.0075691016390919685, + 0.00787266343832016, + 0.0030245499219745398, + 0.003196818521246314, + 0.004713499452918768, + 0.005119845271110535, + 0.0022952118888497353, + 0.006514568347483873, + 0.004851500503718853, + 0.0036077804397791624, + 0.03466387838125229, + 0.018709594383835793, + 0.026652848348021507, + 0.014512203633785248, + 0.007253591902554035, + 0.008926422335207462, + 0.010317671112716198, + 0.009723089635372162, + 0.010075604543089867, + 0.0048233033157885075, + 0.0226416178047657, + 0.0096996258944273, + 0.001974476035684347, + 0.003938749432563782, + 0.003606297541409731, + 0.007149538490921259, + 0.0041429209522902966, + 0.0029644514434039593, + 0.011714955791831017, + 0.006713134236633778, + 0.002053905511274934, + 0.002307737013325095, + 0.003464912995696068, + 0.0027845234144479036, + 0.014500608667731285, + 0.011577608995139599, + 0.029061613604426384, + 0.018161574378609657, + 0.008302509784698486, + 0.009886367246508598, + 0.010404326021671295, + 0.007820403203368187 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 2 + ], + "selected_reward": -0.04585405811667442 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 high-angle aerial view gliding over the landscape?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the hills covered with rows of grapevines bearing deep purple grapes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a narrow, light-brown dirt path winding between the grapevine rows visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by low-angle afternoon sunlight casting long, soft shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the sky in the scene brilliant and cloudless?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.043083954602479935 + ], + "opposite_logprob_tokens": [ + -3.168083906173706 + ], + "logprob_reward": -0.043083954602479935, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.043083954602479935, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3149518072605133, + 0.03525327518582344, + 0.20871703326702118, + 0.036215245723724365, + 0.029070260003209114, + 0.033264148980379105, + 0.15273913741111755, + 0.14337678253650665, + 0.025089168921113014, + 0.021323196589946747 + ], + "credit_map": [ + 0.036972541362047195, + 0.027463657781481743, + 0.04236249998211861, + 0.0306024719029665, + 0.01577005535364151, + 0.029890183359384537, + 0.07725364714860916, + 0.05463672801852226, + 0.004241711925715208, + 0.004998385906219482, + 0.005889012478291988, + 0.007453636731952429, + 0.0022556616459041834, + 0.004667904693633318, + 0.0029777390882372856, + 0.002769219921901822, + 0.018372531980276108, + 0.024164976552128792, + 0.030885400250554085, + 0.04376927390694618, + 0.017912304028868675, + 0.03912702202796936, + 0.017350032925605774, + 0.01713547855615616, + 0.002919204067438841, + 0.00393266836181283, + 0.006277037784457207, + 0.007670385763049126, + 0.002982006873935461, + 0.004720055032521486, + 0.0035555828362703323, + 0.004158302675932646, + 0.0029345655348151922, + 0.0043542212806642056, + 0.0034629441797733307, + 0.004776451271027327, + 0.00258233561180532, + 0.005247459281235933, + 0.0031413657125085592, + 0.002570914104580879, + 0.004304930567741394, + 0.004398784134536982, + 0.006204033270478249, + 0.004821429960429668, + 0.002347580622881651, + 0.004506731405854225, + 0.0033718894701451063, + 0.0033087662886828184, + 0.0422043651342392, + 0.023932112380862236, + 0.013800251297652721, + 0.007941336371004581, + 0.003955631051212549, + 0.0068279532715678215, + 0.013981679454445839, + 0.040095794945955276, + 0.030230848118662834, + 0.008961201645433903, + 0.05064897984266281, + 0.021362632513046265, + 0.003854360431432724, + 0.010406161658465862, + 0.005825057625770569, + 0.012087521143257618, + 0.003463768633082509, + 0.002578195882961154, + 0.007029184140264988, + 0.004401992075145245, + 0.0013440420152619481, + 0.0016168461879715323, + 0.0022191768512129784, + 0.0024359605740755796, + 0.0030272649601101875, + 0.0018968055956065655, + 0.005794224329292774, + 0.003756890306249261, + 0.001611497369594872, + 0.0018875455716624856, + 0.0019490644335746765, + 0.001399902394041419 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 6 + ], + "selected_reward": -0.043083954602479935 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 high-angle aerial view gliding over the landscape?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the hills covered with rows of grapevines bearing deep purple grapes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a narrow, light-brown dirt path winding between the grapevine rows visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by low-angle afternoon sunlight casting long, soft shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the sky in the scene brilliant and cloudless?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0011800715001299977 + ], + "opposite_logprob_tokens": [ + -6.751180171966553 + ], + "logprob_reward": -0.0011800715001299977, + "yes_no_margin_reward": 6.75, + "token_credit_reward": -0.0011800715001299977, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1105642095208168, + 0.12453001737594604, + 0.0729597955942154, + 0.08275818079710007, + 0.04817068204283714, + 0.03475640341639519, + 0.35534363985061646, + 0.03319360688328743, + 0.04800081625580788, + 0.0897226631641388 + ], + "credit_map": [ + 0.01515581738203764, + 0.007041325327008963, + 0.02863006666302681, + 0.012385178357362747, + 0.006316978018730879, + 0.010960009880363941, + 0.010093624703586102, + 0.019981205463409424, + 0.016790959984064102, + 0.01718014106154442, + 0.016511596739292145, + 0.025714438408613205, + 0.010371790267527103, + 0.018792487680912018, + 0.00909147784113884, + 0.01007713284343481, + 0.0076347701251506805, + 0.00975022092461586, + 0.01233170460909605, + 0.012466132640838623, + 0.0070414189249277115, + 0.011221587657928467, + 0.006244654767215252, + 0.006269306410104036, + 0.008227965794503689, + 0.009596004150807858, + 0.015077457763254642, + 0.020045258104801178, + 0.004771159030497074, + 0.011376923881471157, + 0.006950044073164463, + 0.0067133731208741665, + 0.004869905300438404, + 0.004767448175698519, + 0.006908892188221216, + 0.005864973645657301, + 0.005146486684679985, + 0.005642727483063936, + 0.008136478252708912, + 0.006833769381046295, + 0.0029725548811256886, + 0.0024123932234942913, + 0.006049772258847952, + 0.0059227715246379375, + 0.0030815331265330315, + 0.004361326340585947, + 0.0050794691778719425, + 0.004876583348959684, + 0.10458517074584961, + 0.0547090545296669, + 0.02620624005794525, + 0.017162932083010674, + 0.007592770736664534, + 0.015447806566953659, + 0.03352687880396843, + 0.09611280262470245, + 0.0031392143573611975, + 0.002307512564584613, + 0.005180739797651768, + 0.002963730599731207, + 0.0031163578387349844, + 0.005716256331652403, + 0.0038753459230065346, + 0.0068944478407502174, + 0.0033432666677981615, + 0.00342498067766428, + 0.010508785024285316, + 0.007085122633725405, + 0.003377809887751937, + 0.0046560741029679775, + 0.007126748561859131, + 0.008478029631078243, + 0.011799388565123081, + 0.008485171012580395, + 0.023138169199228287, + 0.014684179797768593, + 0.006367568392306566, + 0.009001871570944786, + 0.008946601301431656, + 0.007299712859094143 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 1, + 0 + ], + "selected_reward": -0.0011800715001299977 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 high-angle aerial view gliding over the landscape?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the hills covered with rows of grapevines bearing deep purple grapes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a narrow, light-brown dirt path winding between the grapevine rows visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by low-angle afternoon sunlight casting long, soft shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the sky in the scene brilliant and cloudless?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"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": [ + 0.17304278910160065, + 0.025761105120182037, + 0.019843317568302155, + 0.05294550955295563, + 0.02048865333199501, + 0.0192596223205328, + 0.24882107973098755, + 0.3669523000717163, + 0.023288074880838394, + 0.049597498029470444 + ], + "credit_map": [ + 0.03386928141117096, + 0.0100348936393857, + 0.031678516417741776, + 0.010774394497275352, + 0.007206242997199297, + 0.008728764951229095, + 0.04264647886157036, + 0.028104234486818314, + 0.002571270801126957, + 0.0032237046398222446, + 0.0035395738668739796, + 0.00411849282681942, + 0.0023639092687517405, + 0.003959829919040203, + 0.002946361666545272, + 0.003037963528186083, + 0.002004929119721055, + 0.0016232890775427222, + 0.003271460300311446, + 0.004979611374437809, + 0.0012540165334939957, + 0.0031820221338421106, + 0.0021047405898571014, + 0.0014232505345717072, + 0.005771611351519823, + 0.007230897434055805, + 0.009426127187907696, + 0.012145589105784893, + 0.003048319136723876, + 0.007146283518522978, + 0.003836180316284299, + 0.004340504761785269, + 0.002040269784629345, + 0.0016772752860561013, + 0.0044485656544566154, + 0.0028778475243598223, + 0.0014989457558840513, + 0.00329490820877254, + 0.0023613811936229467, + 0.0022894626017659903, + 0.0017096728552132845, + 0.0024134807754307985, + 0.0025423814076930285, + 0.003391700331121683, + 0.0012975388672202826, + 0.002782425843179226, + 0.0027349432930350304, + 0.0023874803446233273, + 0.06451475620269775, + 0.03319767862558365, + 0.02493935078382492, + 0.01187406200915575, + 0.00563029432669282, + 0.012642893940210342, + 0.02421925961971283, + 0.07180279493331909, + 0.08278382569551468, + 0.02422763966023922, + 0.1322331428527832, + 0.05533154308795929, + 0.010044380091130733, + 0.02416732721030712, + 0.013136940076947212, + 0.02502749115228653, + 0.003758261911571026, + 0.00222689937800169, + 0.00353596406057477, + 0.002892381977289915, + 0.001962875947356224, + 0.0021008558105677366, + 0.0035592252388596535, + 0.0032516142819076777, + 0.006421372294425964, + 0.0050023384392261505, + 0.008585799485445023, + 0.007499368395656347, + 0.003966132178902626, + 0.006629389710724354, + 0.005737049039453268, + 0.005756055004894733 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 6, + 0 + ], + "selected_reward": -0.03369833156466484 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 high-angle aerial view gliding over the landscape?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the hills covered with rows of grapevines bearing deep purple grapes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a narrow, light-brown dirt path winding between the grapevine rows visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by low-angle afternoon sunlight casting long, soft shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the sky in the scene brilliant and cloudless?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.14283177256584167 + ], + "opposite_logprob_tokens": [ + -2.017831802368164 + ], + "logprob_reward": -0.14283177256584167, + "yes_no_margin_reward": 1.875, + "token_credit_reward": -0.14283177256584167, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19645439088344574, + 0.10777263343334198, + 0.21349187195301056, + 0.2810703217983246, + 0.014866129495203495, + 0.019083773717284203, + 0.06358754634857178, + 0.03938718140125275, + 0.013230294920504093, + 0.05105585232377052 + ], + "credit_map": [ + 0.037694696336984634, + 0.016148073598742485, + 0.03399500623345375, + 0.020941779017448425, + 0.009290198795497417, + 0.015443175099790096, + 0.03583146631717682, + 0.027109971269965172, + 0.013082011602818966, + 0.01730458252131939, + 0.01353787537664175, + 0.02400781586766243, + 0.007585897110402584, + 0.014831077307462692, + 0.009334824047982693, + 0.008088544942438602, + 0.022986721247434616, + 0.030378662049770355, + 0.031173504889011383, + 0.03732544928789139, + 0.019144384190440178, + 0.03888498619198799, + 0.017082056030631065, + 0.016516083851456642, + 0.030878901481628418, + 0.038707539439201355, + 0.05114338919520378, + 0.06476885825395584, + 0.01492750272154808, + 0.038726914674043655, + 0.019962497055530548, + 0.02195470593869686, + 0.0015794815262779593, + 0.0017495405627414584, + 0.0016326523618772626, + 0.001928039942868054, + 0.0014963984722271562, + 0.0023505454882979393, + 0.0019119445933029056, + 0.0022175260819494724, + 0.0017158585833385587, + 0.0020124102011322975, + 0.0025784866884350777, + 0.0030504551250487566, + 0.0012614228762686253, + 0.0024996749125421047, + 0.002956981537863612, + 0.0030084813479334116, + 0.015960458666086197, + 0.007634934037923813, + 0.006793289911001921, + 0.004508672747761011, + 0.0021027487237006426, + 0.0038197168614715338, + 0.005991015583276749, + 0.016776708886027336, + 0.004785727243870497, + 0.004500291310250759, + 0.008992291986942291, + 0.005094023887068033, + 0.0032594981603324413, + 0.005100709851831198, + 0.0032123741693794727, + 0.004442261531949043, + 0.001397550106048584, + 0.0010700542479753494, + 0.002370338188484311, + 0.0015349920140579343, + 0.0008231009705923498, + 0.00121116463560611, + 0.002131775254383683, + 0.0026913187466561794, + 0.006320971995592117, + 0.008769193664193153, + 0.008594335988163948, + 0.007537970319390297, + 0.005955756641924381, + 0.005769699811935425, + 0.004516052082180977, + 0.003591868793591857 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 2, + 0 + ], + "selected_reward": -0.14283177256584167 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "As the camera smoothly glides from a high-angle aerial perspective over gently rolling hills bathed in the warm golden glow of the afternoon sun, a vast, cinematic panorama unfolds. The landscape is a tapestry of meticulously arranged, sun-drenched grapevines, their vibrant green leaves laden with clusters of deep purple grapes, arranged in perfectly parallel rows following the rolling contours of the land. A narrow, light-brown dirt path, its surface dry and dusted with fine dust and loose pebbles, gracefully winds between the rows, its sinuous shape drawing the eye deeply into the vast landscape. The low-angle sunlight casts long, soft shadows from the vines lining the path, creating rhythmic patterns of light and dark that accentuate the terrain. The entire scene is imbued with a profound tranquility and idyllic atmosphere, capturing the serene, breathtaking beauty of nature beneath a brilliant, cloudless sky.", + "seed": 21516, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000156/group_10/index_6862_seed_21516_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": -0.09898719547782094, + "vlm_logprob": -0.09898719547782094, + "vlm_yes_no_margin": 2.975, + "vlm_token_credit": -0.09898719547782094, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 high-angle aerial view gliding over the landscape?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the hills covered with rows of grapevines bearing deep purple grapes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a narrow, light-brown dirt path winding between the grapevine rows visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by low-angle afternoon sunlight casting long, soft shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the sky in the scene brilliant and cloudless?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11564982682466507 + ], + "opposite_logprob_tokens": [ + -2.240649938583374 + ], + "logprob_reward": -0.11564982682466507, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11564982682466507, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4789314866065979, + 0.03542199730873108, + 0.07277128100395203, + 0.058894675225019455, + 0.09576907008886337, + 0.030438488349318504, + 0.07837702333927155, + 0.015483386814594269, + 0.11229036748409271, + 0.021622242406010628 + ], + "credit_map": [ + 0.03587908670306206, + 0.020543092861771584, + 0.03156678378582001, + 0.02691231667995453, + 0.06218107417225838, + 0.04368499293923378, + 0.16066958010196686, + 0.09749456495046616, + 0.002871485659852624, + 0.0037170907016843557, + 0.003298211609944701, + 0.004248569719493389, + 0.00402015820145607, + 0.0034329707268625498, + 0.009165278635919094, + 0.004668228793889284, + 0.005910531617701054, + 0.005761224776506424, + 0.013351102359592915, + 0.010876428335905075, + 0.006647247355431318, + 0.011643817648291588, + 0.009433778002858162, + 0.009147154167294502, + 0.0035003884695470333, + 0.005123175214976072, + 0.00867037009447813, + 0.009427233599126339, + 0.009522663429379463, + 0.00560007942840457, + 0.007680680137127638, + 0.009370080195367336, + 0.004934377036988735, + 0.004499343689531088, + 0.02165922522544861, + 0.01328277587890625, + 0.006928459741175175, + 0.007764387875795364, + 0.011920895427465439, + 0.024779608473181725, + 0.0030130359809845686, + 0.001683067181147635, + 0.00281692435964942, + 0.0030049490742385387, + 0.005513643845915794, + 0.004331659991294146, + 0.006389245390892029, + 0.003685964969918132, + 0.00910754781216383, + 0.005332883447408676, + 0.006869717501103878, + 0.013873542658984661, + 0.01577729918062687, + 0.009294970892369747, + 0.00844558048993349, + 0.009675484150648117, + 0.001660128473304212, + 0.0012092614779248834, + 0.001712613389827311, + 0.0018331517931073904, + 0.0021845675073564053, + 0.0022687057498842478, + 0.0018662589136511087, + 0.0027486991602927446, + 0.004722230602055788, + 0.003791367169469595, + 0.007064325734972954, + 0.00854606181383133, + 0.015621147118508816, + 0.007904566824436188, + 0.0389888621866703, + 0.025651808828115463, + 0.0016742907464504242, + 0.001168190035969019, + 0.0045298244804143906, + 0.0019144732505083084, + 0.0016001093899831176, + 0.0016850678948685527, + 0.0023154006339609623, + 0.006734886672347784 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 4 + ], + "selected_reward": -0.11564982682466507 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 high-angle aerial view gliding over the landscape?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the hills covered with rows of grapevines bearing deep purple grapes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a narrow, light-brown dirt path winding between the grapevine rows visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by low-angle afternoon sunlight casting long, soft shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the sky in the scene brilliant and cloudless?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.12706127762794495 + ], + "opposite_logprob_tokens": [ + -2.127061367034912 + ], + "logprob_reward": -0.12706127762794495, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.12706127762794495, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24494294822216034, + 0.07024257630109787, + 0.13000424206256866, + 0.12219991534948349, + 0.2072421759366989, + 0.11610434949398041, + 0.02451685257256031, + 0.03530241549015045, + 0.03537381812930107, + 0.014070688746869564 + ], + "credit_map": [ + 0.021542375907301903, + 0.02278740704059601, + 0.0251678004860878, + 0.028190530836582184, + 0.0229084063321352, + 0.040642060339450836, + 0.024193523451685905, + 0.05951085314154625, + 0.00874966662377119, + 0.0072230082005262375, + 0.010861489921808243, + 0.009220409207046032, + 0.006361236795783043, + 0.009728627279400826, + 0.01036718487739563, + 0.007730951067060232, + 0.012188533321022987, + 0.012951219454407692, + 0.012383478693664074, + 0.014565235935151577, + 0.0195175651460886, + 0.021786782890558243, + 0.02094372734427452, + 0.01566769927740097, + 0.009966730140149593, + 0.010081973858177662, + 0.01700417697429657, + 0.02321787364780903, + 0.01337938942015171, + 0.017638614401221275, + 0.01527696754783392, + 0.01563418097794056, + 0.018161360174417496, + 0.015499981120228767, + 0.053732749074697495, + 0.02771691046655178, + 0.02011982910335064, + 0.015021362341940403, + 0.026327082887291908, + 0.030662881210446358, + 0.011068295687437057, + 0.011083833873271942, + 0.011798080056905746, + 0.015435818582773209, + 0.013397350907325745, + 0.01706487312912941, + 0.01699712872505188, + 0.01925896666944027, + 0.0025461181066930294, + 0.0023256985004991293, + 0.0038352233823388815, + 0.004123709164559841, + 0.0031173424795269966, + 0.0027217818424105644, + 0.002845228649675846, + 0.0030017525423318148, + 0.003472402924671769, + 0.002204151824116707, + 0.0042806765995919704, + 0.0029755043797194958, + 0.005732417106628418, + 0.003185773966833949, + 0.004549771081656218, + 0.008901716209948063, + 0.0018305755220353603, + 0.0020372108556330204, + 0.003003254532814026, + 0.009063484147191048, + 0.0045303418301045895, + 0.003264698665589094, + 0.007989645004272461, + 0.003654608502984047, + 0.0009427701006643474, + 0.001490800641477108, + 0.0013949681306257844, + 0.0025346530601382256, + 0.0017361536156386137, + 0.0018320176750421524, + 0.0024692295119166374, + 0.0016700946725904942 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 2 + ], + "selected_reward": -0.12706127762794495 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 high-angle aerial view gliding over the landscape?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the hills covered with rows of grapevines bearing deep purple grapes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a narrow, light-brown dirt path winding between the grapevine rows visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by low-angle afternoon sunlight casting long, soft shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the sky in the scene brilliant and cloudless?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0019414640264585614 + ], + "opposite_logprob_tokens": [ + -6.251941680908203 + ], + "logprob_reward": -0.0019414640264585614, + "yes_no_margin_reward": 6.25, + "token_credit_reward": -0.0019414640264585614, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11633968353271484, + 0.16534806787967682, + 0.08238085359334946, + 0.08596687018871307, + 0.1795027107000351, + 0.049759332090616226, + 0.10752765834331512, + 0.06201926991343498, + 0.11779719591140747, + 0.033358342945575714 + ], + "credit_map": [ + 0.008543313480913639, + 0.006486943457275629, + 0.01555150467902422, + 0.015697479248046875, + 0.01266931276768446, + 0.017806261777877808, + 0.016311386600136757, + 0.023273492231965065, + 0.015025920234620571, + 0.015770666301250458, + 0.01677551679313183, + 0.02193276770412922, + 0.019774645566940308, + 0.024228984490036964, + 0.036041583865880966, + 0.01579800620675087, + 0.0076290941797196865, + 0.005701146554201841, + 0.011068563908338547, + 0.016558226197957993, + 0.011087535880506039, + 0.012097752653062344, + 0.011706551536917686, + 0.006531986873596907, + 0.004767861682921648, + 0.006020830478519201, + 0.009434870444238186, + 0.010933841578662395, + 0.015909867361187935, + 0.00812035147100687, + 0.014817931689321995, + 0.01596132665872574, + 0.010424100793898106, + 0.007439590059220791, + 0.03922915831208229, + 0.01992582343518734, + 0.02241053804755211, + 0.016665779054164886, + 0.025878116488456726, + 0.03752962499856949, + 0.004631801508367062, + 0.003985001239925623, + 0.005093086045235395, + 0.007000720594078302, + 0.008426266722381115, + 0.0068274326622486115, + 0.006718083750456572, + 0.00707694236189127, + 0.00817847903817892, + 0.004918977152556181, + 0.00798342376947403, + 0.01444070041179657, + 0.02641346864402294, + 0.01641777530312538, + 0.013560580089688301, + 0.015614275820553303, + 0.004010504577308893, + 0.0025029736571013927, + 0.00645882124081254, + 0.0037333262152969837, + 0.008622310124337673, + 0.0036203551571816206, + 0.017710519954562187, + 0.015360469929873943, + 0.00670976098626852, + 0.003117287065833807, + 0.00629570335149765, + 0.01689044013619423, + 0.011999478563666344, + 0.006635541561990976, + 0.04105045273900032, + 0.02509855106472969, + 0.001954580657184124, + 0.0018901112489402294, + 0.0058904108591377735, + 0.004911650903522968, + 0.005435430444777012, + 0.004425219725817442, + 0.00497626094147563, + 0.0038746818900108337 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 1, + 8 + ], + "selected_reward": -0.0019414640264585614 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 high-angle aerial view gliding over the landscape?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the hills covered with rows of grapevines bearing deep purple grapes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a narrow, light-brown dirt path winding between the grapevine rows visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by low-angle afternoon sunlight casting long, soft shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the sky in the scene brilliant and cloudless?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04864807799458504 + ], + "opposite_logprob_tokens": [ + -3.0486481189727783 + ], + "logprob_reward": -0.04864807799458504, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04864807799458504, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.107863649725914, + 0.26044344902038574, + 0.04965011402964592, + 0.048016030341386795, + 0.3171818256378174, + 0.02779623679816723, + 0.06619217991828918, + 0.021722940728068352, + 0.07381749153137207, + 0.027316059917211533 + ], + "credit_map": [ + 0.007056872826069593, + 0.011165746487677097, + 0.010036909021437168, + 0.010595272295176983, + 0.011102179065346718, + 0.011903150007128716, + 0.019140031188726425, + 0.02686348930001259, + 0.015837987884879112, + 0.013723315671086311, + 0.027438092976808548, + 0.0338834747672081, + 0.022297672927379608, + 0.021205775439739227, + 0.10318806767463684, + 0.022869063541293144, + 0.003212594659999013, + 0.003179504768922925, + 0.011918089352548122, + 0.004999107681214809, + 0.004941366612911224, + 0.004644411616027355, + 0.005822084844112396, + 0.010932957753539085, + 0.003213556483387947, + 0.0034155934117734432, + 0.005459984298795462, + 0.006105310749262571, + 0.007784784305840731, + 0.004508848302066326, + 0.011595509015023708, + 0.005932443309575319, + 0.016526440158486366, + 0.0123904999345541, + 0.09437841176986694, + 0.03595995530486107, + 0.01998559758067131, + 0.01815154403448105, + 0.033633697777986526, + 0.08615570515394211, + 0.0028718241956084967, + 0.002316006924957037, + 0.0033583000767976046, + 0.0033692640718072653, + 0.005493270233273506, + 0.0038507930003106594, + 0.004142244812101126, + 0.0023945330176502466, + 0.007735959719866514, + 0.0048791454173624516, + 0.009045545011758804, + 0.010212792083621025, + 0.009230210445821285, + 0.007826258428394794, + 0.006956057157367468, + 0.010306205600500107, + 0.0015094188274815679, + 0.0014707179507240653, + 0.00295359012670815, + 0.002752244006842375, + 0.003972863778471947, + 0.002227212069556117, + 0.003457298967987299, + 0.0033795943018049, + 0.0027189687825739384, + 0.0026917101349681616, + 0.004689668770879507, + 0.012574550695717335, + 0.015987904742360115, + 0.005422470159828663, + 0.016733340919017792, + 0.012998872436583042, + 0.0012713412288576365, + 0.0012682478409260511, + 0.006988053675740957, + 0.004131955094635487, + 0.0019256985979154706, + 0.0022636447101831436, + 0.002703289268538356, + 0.006763828452676535 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 1, + 0 + ], + "selected_reward": -0.04864807799458504 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 high-angle aerial view gliding over the landscape?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the hills covered with rows of grapevines bearing deep purple grapes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a narrow, light-brown dirt path winding between the grapevine rows visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by low-angle afternoon sunlight casting long, soft shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the sky in the scene brilliant and cloudless?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.20163533091545105 + ], + "opposite_logprob_tokens": [ + -1.7016353607177734 + ], + "logprob_reward": -0.20163533091545105, + "yes_no_margin_reward": 1.5, + "token_credit_reward": -0.20163533091545105, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.28522056341171265, + 0.2012939751148224, + 0.08595327287912369, + 0.05717353895306587, + 0.07282239198684692, + 0.02469922974705696, + 0.0485062375664711, + 0.02419600635766983, + 0.18141479790210724, + 0.018719950690865517 + ], + "credit_map": [ + 0.019386380910873413, + 0.013416470028460026, + 0.023516617715358734, + 0.018096569925546646, + 0.03534303233027458, + 0.023332791402935982, + 0.08981969952583313, + 0.062309008091688156, + 0.01493454072624445, + 0.010437515564262867, + 0.022316936403512955, + 0.028111755847930908, + 0.017637010663747787, + 0.0169697105884552, + 0.0745653510093689, + 0.016321150586009026, + 0.007212778087705374, + 0.0049727484583854675, + 0.02117411233484745, + 0.013520737178623676, + 0.007494846824556589, + 0.009398720227181911, + 0.010433406569063663, + 0.011745919473469257, + 0.00386095535941422, + 0.00335874711163342, + 0.006933174096047878, + 0.006735015194863081, + 0.009443485178053379, + 0.007501985877752304, + 0.012841986492276192, + 0.006498188711702824, + 0.0043534268625080585, + 0.005064970813691616, + 0.009401975199580193, + 0.011900801211595535, + 0.009863477200269699, + 0.007168530020862818, + 0.016225194558501244, + 0.00884401984512806, + 0.002286709612235427, + 0.0015957524301484227, + 0.002715836511924863, + 0.002198390429839492, + 0.0035870808642357588, + 0.0031753734219819307, + 0.006050922907888889, + 0.003089164150878787, + 0.005138562060892582, + 0.003114813705906272, + 0.004608740098774433, + 0.007051218766719103, + 0.008599421009421349, + 0.006194124463945627, + 0.005908722523599863, + 0.007890637032687664, + 0.0017459159716963768, + 0.0017485112184658647, + 0.0031964313238859177, + 0.002843015594407916, + 0.004061525221914053, + 0.0031173410825431347, + 0.002654797164723277, + 0.004828468896448612, + 0.010625237599015236, + 0.004504361655563116, + 0.00952632911503315, + 0.01542742270976305, + 0.015522395260632038, + 0.010910269804298878, + 0.06858114898204803, + 0.04631762579083443, + 0.0014034940395504236, + 0.0009025423787534237, + 0.003712084610015154, + 0.002226915443316102, + 0.00154491001740098, + 0.002047342248260975, + 0.0018467818154022098, + 0.00503587955608964 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 8 + ], + "selected_reward": -0.20163533091545105 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "As the camera smoothly glides from a high-angle aerial perspective over gently rolling hills bathed in the warm golden glow of the afternoon sun, a vast, cinematic panorama unfolds. The landscape is a tapestry of meticulously arranged, sun-drenched grapevines, their vibrant green leaves laden with clusters of deep purple grapes, arranged in perfectly parallel rows following the rolling contours of the land. A narrow, light-brown dirt path, its surface dry and dusted with fine dust and loose pebbles, gracefully winds between the rows, its sinuous shape drawing the eye deeply into the vast landscape. The low-angle sunlight casts long, soft shadows from the vines lining the path, creating rhythmic patterns of light and dark that accentuate the terrain. The entire scene is imbued with a profound tranquility and idyllic atmosphere, capturing the serene, breathtaking beauty of nature beneath a brilliant, cloudless sky.", + "seed": 21517, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000156/group_10/index_6862_seed_21517_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": -0.08707608665572479, + "vlm_logprob": -0.08707608665572479, + "vlm_yes_no_margin": 3.55, + "vlm_token_credit": -0.08707608665572479, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 high-angle aerial view gliding over the landscape?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the hills covered with rows of grapevines bearing deep purple grapes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a narrow, light-brown dirt path winding between the grapevine rows visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by low-angle afternoon sunlight casting long, soft shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the sky in the scene brilliant and cloudless?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.040895652025938034 + ], + "opposite_logprob_tokens": [ + -3.290895700454712 + ], + "logprob_reward": -0.040895652025938034, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.040895652025938034, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.7415699362754822, + 0.027348585426807404, + 0.025025323033332825, + 0.04053092375397682, + 0.04253233224153519, + 0.05303383991122246, + 0.02169135957956314, + 0.013537370599806309, + 0.025054320693016052, + 0.009676019661128521 + ], + "credit_map": [ + 0.05623568966984749, + 0.031243015080690384, + 0.027273474261164665, + 0.015583645552396774, + 0.025243639945983887, + 0.28361526131629944, + 0.046727996319532394, + 0.2556471824645996, + 0.0026151277124881744, + 0.002012585988268256, + 0.0026052435860037804, + 0.0013471601996570826, + 0.006932602263987064, + 0.004950993228703737, + 0.0034322189167141914, + 0.003452654229477048, + 0.0033787607681006193, + 0.002452487125992775, + 0.002900500316172838, + 0.0030647311359643936, + 0.003075319342315197, + 0.0038581043481826782, + 0.00326096941716969, + 0.0030344505794346333, + 0.002761989599093795, + 0.0014793704031035304, + 0.003555909264832735, + 0.001691727084107697, + 0.0021518401335924864, + 0.015604401007294655, + 0.0029808548279106617, + 0.010304833762347698, + 0.003642266383394599, + 0.002366406377404928, + 0.0033859878312796354, + 0.0023188001941889524, + 0.004915640223771334, + 0.01147349365055561, + 0.005806573200970888, + 0.00862316507846117, + 0.0025615179911255836, + 0.0011135019594803452, + 0.004372682422399521, + 0.00230844272300601, + 0.0020702199544757605, + 0.02379552088677883, + 0.001523649669252336, + 0.015288304537534714, + 0.003310861298814416, + 0.0016263247234746814, + 0.001762443920597434, + 0.0018617940368130803, + 0.0022832255344837904, + 0.004576481878757477, + 0.0027809273451566696, + 0.0034893015399575233, + 0.001081935130059719, + 0.0008529530605301261, + 0.0012577176094055176, + 0.001348174992017448, + 0.0020404341630637646, + 0.002061411738395691, + 0.0031115764286369085, + 0.0017831671284511685, + 0.0026784746441990137, + 0.0018580462783575058, + 0.0018902224255725741, + 0.001973762409761548, + 0.0033359334338456392, + 0.004068086855113506, + 0.004327785689383745, + 0.004922007210552692, + 0.0011150079080834985, + 0.0004431505221873522, + 0.0010317486012354493, + 0.0005765461828559637, + 0.0014444409171119332, + 0.0029362128116190434, + 0.001182270934805274, + 0.000946641550399363 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 4 + ], + "selected_reward": -0.040895652025938034 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 high-angle aerial view gliding over the landscape?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the hills covered with rows of grapevines bearing deep purple grapes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a narrow, light-brown dirt path winding between the grapevine rows visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by low-angle afternoon sunlight casting long, soft shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the sky in the scene brilliant and cloudless?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08906836062669754 + ], + "opposite_logprob_tokens": [ + -2.4640684127807617 + ], + "logprob_reward": -0.08906836062669754, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08906836062669754, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.41288653016090393, + 0.0960659608244896, + 0.14360679686069489, + 0.04927197843790054, + 0.02434706501662731, + 0.2130904197692871, + 0.016240479424595833, + 0.016321733593940735, + 0.015289290808141232, + 0.012879724614322186 + ], + "credit_map": [ + 0.03149870038032532, + 0.024794332683086395, + 0.021441586315631866, + 0.023933691903948784, + 0.02602780982851982, + 0.1392083466053009, + 0.03207557275891304, + 0.11390652507543564, + 0.010599484667181969, + 0.011459802277386189, + 0.012420081533491611, + 0.009515135549008846, + 0.012992730364203453, + 0.014215326867997646, + 0.008377761580049992, + 0.016485633328557014, + 0.01823773793876171, + 0.015222311951220036, + 0.020529910922050476, + 0.02027990110218525, + 0.017349105328321457, + 0.02287309430539608, + 0.01325216330587864, + 0.015862561762332916, + 0.0024945323821157217, + 0.0016959031345322728, + 0.0028737594839185476, + 0.0016042512143030763, + 0.002819154877215624, + 0.018055345863103867, + 0.003099247580394149, + 0.016629787161946297, + 0.005186052992939949, + 0.0027189727406948805, + 0.0023557685781270266, + 0.002800555666908622, + 0.0026941848918795586, + 0.00413140095770359, + 0.0024488396011292934, + 0.002011289354413748, + 0.009176773019134998, + 0.004604879766702652, + 0.01507261861115694, + 0.008310923352837563, + 0.008106819353997707, + 0.09203508496284485, + 0.009125477634370327, + 0.06665784120559692, + 0.0019611292518675327, + 0.001285175676457584, + 0.00160291139036417, + 0.0017373997252434492, + 0.0021531011443585157, + 0.0029991602059453726, + 0.002369649475440383, + 0.00213195220567286, + 0.0013801904860883951, + 0.0013650917680934072, + 0.0011786073446273804, + 0.001470191404223442, + 0.0023285800125449896, + 0.0022157994098961353, + 0.0028920418117195368, + 0.0034912305418401957, + 0.0016830357490107417, + 0.001044633099809289, + 0.0014227695064619184, + 0.0012270844308659434, + 0.0018519931472837925, + 0.0025494194123893976, + 0.0025999979116022587, + 0.0029103574343025684, + 0.0015274188481271267, + 0.000964890408795327, + 0.0017720434116199613, + 0.0012083970941603184, + 0.0019197061192244291, + 0.002141001168638468, + 0.0016628915909677744, + 0.001683375216089189 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 2 + ], + "selected_reward": -0.08906836062669754 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 high-angle aerial view gliding over the landscape?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the hills covered with rows of grapevines bearing deep purple grapes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a narrow, light-brown dirt path winding between the grapevine rows visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by low-angle afternoon sunlight casting long, soft shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the sky in the scene brilliant and cloudless?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0007188120507635176 + ], + "opposite_logprob_tokens": [ + -7.250718593597412 + ], + "logprob_reward": -0.0007188120507635176, + "yes_no_margin_reward": 7.25, + "token_credit_reward": -0.0007188120507635176, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5022062659263611, + 0.04796094074845314, + 0.014977692626416683, + 0.09435510635375977, + 0.03439304605126381, + 0.10057447105646133, + 0.0982528105378151, + 0.042881086468696594, + 0.039897117763757706, + 0.024501478299498558 + ], + "credit_map": [ + 0.036069128662347794, + 0.022427041083574295, + 0.017075881361961365, + 0.008795320987701416, + 0.015124324709177017, + 0.19640260934829712, + 0.025857793167233467, + 0.18045414984226227, + 0.004678759723901749, + 0.00381457363255322, + 0.005085770972073078, + 0.003505551954731345, + 0.0067559597082436085, + 0.009251057170331478, + 0.006201561540365219, + 0.008667705580592155, + 0.0018876001704484224, + 0.0014002209063619375, + 0.0018037850968539715, + 0.0022763365413993597, + 0.0018050589133054018, + 0.0018774758791550994, + 0.0019138094503432512, + 0.0020134039223194122, + 0.006333514116704464, + 0.004625928588211536, + 0.00723164202645421, + 0.005968686658889055, + 0.004998715128749609, + 0.030792001634836197, + 0.006490013096481562, + 0.02791459858417511, + 0.004645387642085552, + 0.001963833812624216, + 0.003236588789150119, + 0.002805153839290142, + 0.003220646409317851, + 0.008818432688713074, + 0.003824810963124037, + 0.005878190975636244, + 0.0038943630643188953, + 0.0021880764979869127, + 0.0067452252842485905, + 0.0038945518899708986, + 0.004313292447477579, + 0.041204459965229034, + 0.006778541021049023, + 0.03155595436692238, + 0.009754829108715057, + 0.006172498222440481, + 0.00962143950164318, + 0.008573118597269058, + 0.013775167055428028, + 0.02431146427989006, + 0.011856060475111008, + 0.01418822817504406, + 0.004244941286742687, + 0.0030008929315954447, + 0.004481676500290632, + 0.0044250451028347015, + 0.005979583133012056, + 0.006903066299855709, + 0.008361064828932285, + 0.0054848152212798595, + 0.0037394894752651453, + 0.002755211666226387, + 0.0031592780724167824, + 0.0037097015883773565, + 0.0055389562621712685, + 0.007394724525511265, + 0.007198868319392204, + 0.00640088552609086, + 0.0028813467361032963, + 0.0012944983318448067, + 0.0021655033342540264, + 0.0015323290135711432, + 0.0036600008606910706, + 0.006944871973246336, + 0.003088352968916297, + 0.002934571821242571 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 6 + ], + "selected_reward": -0.0007188120507635176 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 high-angle aerial view gliding over the landscape?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the hills covered with rows of grapevines bearing deep purple grapes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a narrow, light-brown dirt path winding between the grapevine rows visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by low-angle afternoon sunlight casting long, soft shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the sky in the scene brilliant and cloudless?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023290520533919334 + ], + "opposite_logprob_tokens": [ + -3.7732906341552734 + ], + "logprob_reward": -0.023290520533919334, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023290520533919334, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11720466613769531, + 0.04199206456542015, + 0.05611754208803177, + 0.08867400139570236, + 0.04062362015247345, + 0.45554274320602417, + 0.02624526247382164, + 0.06754893064498901, + 0.07634630054235458, + 0.02970488928258419 + ], + "credit_map": [ + 0.014004837721586227, + 0.009378541260957718, + 0.008974854834377766, + 0.011175921186804771, + 0.015409570187330246, + 0.018212612718343735, + 0.01967390812933445, + 0.020374413579702377, + 0.004375840537250042, + 0.005967294331640005, + 0.006554668769240379, + 0.004041227512061596, + 0.005505030509084463, + 0.006233735475689173, + 0.004493989050388336, + 0.004820273257791996, + 0.007603500969707966, + 0.00437608826905489, + 0.007451994810253382, + 0.0058699618093669415, + 0.00947174895554781, + 0.009581206366419792, + 0.006418139208108187, + 0.005344896111637354, + 0.007349926978349686, + 0.003508737776428461, + 0.009863143786787987, + 0.005289990920573473, + 0.005493489094078541, + 0.033896610140800476, + 0.006005265284329653, + 0.01726682484149933, + 0.005407020915299654, + 0.0025998344644904137, + 0.0039887684397399426, + 0.004937769845128059, + 0.0049227504059672356, + 0.007418923079967499, + 0.0061498479917645454, + 0.005198700353503227, + 0.0176093652844429, + 0.013370157219469547, + 0.028501057997345924, + 0.026703739538788795, + 0.018205126747488976, + 0.16909991204738617, + 0.026627635583281517, + 0.15542572736740112, + 0.0032340558245778084, + 0.0018576105358079076, + 0.0027678594924509525, + 0.0022280823905020952, + 0.003797531593590975, + 0.005538887344300747, + 0.0035376157611608505, + 0.0032836166210472584, + 0.010519085451960564, + 0.00494717201218009, + 0.0061494107358157635, + 0.005526247899979353, + 0.007121517788618803, + 0.010113651864230633, + 0.011458280496299267, + 0.011713553220033646, + 0.009253996424376965, + 0.0052964529022574425, + 0.006967298686504364, + 0.007086705416440964, + 0.00982704572379589, + 0.012983156368136406, + 0.013746523298323154, + 0.0111851179972291, + 0.0044778455048799515, + 0.0014617895940318704, + 0.0026842758525162935, + 0.0022301231510937214, + 0.004183873534202576, + 0.008940997533500195, + 0.0025057545863091946, + 0.003220224054530263 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 3 + ], + "selected_reward": -0.023290520533919334 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 high-angle aerial view gliding over the landscape?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the hills covered with rows of grapevines bearing deep purple grapes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a narrow, light-brown dirt path winding between the grapevine rows visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by low-angle afternoon sunlight casting long, soft shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the sky in the scene brilliant and cloudless?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.28140708804130554 + ], + "opposite_logprob_tokens": [ + -1.406407117843628 + ], + "logprob_reward": -0.28140708804130554, + "yes_no_margin_reward": 1.125, + "token_credit_reward": -0.28140708804130554, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5555331110954285, + 0.029581455513834953, + 0.021886898204684258, + 0.03745774179697037, + 0.05117807909846306, + 0.069830022752285, + 0.11362546682357788, + 0.07859760522842407, + 0.024051303043961525, + 0.01825830154120922 + ], + "credit_map": [ + 0.043052420020103455, + 0.024460550397634506, + 0.02110821008682251, + 0.011200673878192902, + 0.02024995908141136, + 0.210889995098114, + 0.03300360217690468, + 0.19156768918037415, + 0.004000499378889799, + 0.002629661001265049, + 0.003161277621984482, + 0.0020900366362184286, + 0.005269233137369156, + 0.005159982480108738, + 0.003978920169174671, + 0.0032918453216552734, + 0.0028362080920487642, + 0.002221854403614998, + 0.00234437664039433, + 0.0024150279350578785, + 0.0028711389750242233, + 0.003566921455785632, + 0.0030132445972412825, + 0.0026181284338235855, + 0.0037677220534533262, + 0.0030118150170892477, + 0.004041980020701885, + 0.0035317917354404926, + 0.004195279907435179, + 0.006137879565358162, + 0.0037430417723953724, + 0.009028235450387001, + 0.008081937208771706, + 0.003522074082866311, + 0.0046309819445014, + 0.003824780462309718, + 0.005484649445861578, + 0.010880552232265472, + 0.006840633228421211, + 0.007912473753094673, + 0.0046913777478039265, + 0.003627551021054387, + 0.009104007855057716, + 0.00538184680044651, + 0.007271257694810629, + 0.020624713972210884, + 0.01175536960363388, + 0.0073739029467105865, + 0.013384705409407616, + 0.007458210457116365, + 0.010099338367581367, + 0.009589262306690216, + 0.014566837809979916, + 0.026368577033281326, + 0.01266083400696516, + 0.019497711211442947, + 0.013150631450116634, + 0.00633760541677475, + 0.006121392827481031, + 0.006644630338996649, + 0.00890263356268406, + 0.011256932280957699, + 0.013463563285768032, + 0.01272022444754839, + 0.002294673817232251, + 0.00149360834620893, + 0.002189252059906721, + 0.00243234820663929, + 0.0033057709224522114, + 0.0041917054913938046, + 0.004902880173176527, + 0.003241065191105008, + 0.002441854914650321, + 0.001041607465595007, + 0.0014221309684216976, + 0.0011790466960519552, + 0.0028208859730511904, + 0.006120471283793449, + 0.0013485878007486463, + 0.001883718534372747 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 7 + ], + "selected_reward": -0.28140708804130554 + } + ], + "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/0000156/group_11/index_tensor([5590, 5590])_group_11_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000156/group_11/index_tensor([5590, 5590])_group_11_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..f3364214157af99f0b1fb0f40b5f56804edf6d0d --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000156/group_11/index_tensor([5590, 5590])_group_11_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "An initially static, low-angle, wide-angle shot begins to tremble violently, capturing the interior of a vast, cavernous hall within a futuristic base on an alien world, designed in a stark, dark, and minimalist style. The floor, a seamless expanse of polished, high-gloss black material that reflects the entire structure like a black mirror, suddenly begins to crack at its center. A glowing spiderweb of cracks emanating from pure white light erupts across the surface, spreading at a terrifying pace and branching out in jagged, unpredictable patterns. As the violence throughout the structure intensifies, the polished floor begins to bulge and tilt. The only illumination comes from emergency lights, which flicker erratically, casting the scene in a disorienting, stroboscopic glare, creating long, dancing shadows and moments of absolute darkness, establishing an atmosphere of chaotic and utterly urgent catastrophic structural failure.", + "seed": 4450, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000156/group_11/index_5590_seed_4450_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": -0.027438854053616523, + "vlm_logprob": -0.027438854053616523, + "vlm_yes_no_margin": 4.425, + "vlm_token_credit": -0.027438854053616523, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 vast, cavernous hall within a futuristic base on an alien world?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the interior designed in a stark, dark, and minimalist style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the floor consist of a seamless, high-gloss black material that reflects the structure like a black mirror?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do glowing white cracks spread rapidly across the floor in jagged, unpredictable patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot initially static, low-angle, and wide-angle before the scene begins to tremble violently?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.09651143103837967 + ], + "opposite_logprob_tokens": [ + -2.4715113639831543 + ], + "logprob_reward": -0.09651143103837967, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.09651143103837967, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07535497099161148, + 0.06842420995235443, + 0.08148199319839478, + 0.0477600060403347, + 0.4334845244884491, + 0.05681874230504036, + 0.01835760287940502, + 0.10222470760345459, + 0.09260733425617218, + 0.02348591946065426 + ], + "credit_map": [ + 0.009574211202561855, + 0.009211186319589615, + 0.009852022863924503, + 0.008737319149076939, + 0.006493992637842894, + 0.010167947970330715, + 0.010444365441799164, + 0.010873937048017979, + 0.007939842529594898, + 0.006276624742895365, + 0.01586948335170746, + 0.011704538017511368, + 0.008207921870052814, + 0.0045458548702299595, + 0.007928476668894291, + 0.005951469298452139, + 0.009976774454116821, + 0.006305905058979988, + 0.008330184035003185, + 0.01161197666078806, + 0.008636988699436188, + 0.008793281391263008, + 0.014672483317553997, + 0.013154403306543827, + 0.00572587875649333, + 0.003091187449172139, + 0.0035209727939218283, + 0.004478727001696825, + 0.00580903235822916, + 0.007176746614277363, + 0.0057961586862802505, + 0.012161307968199253, + 0.015667052939534187, + 0.020567072555422783, + 0.04773604869842529, + 0.013008667156100273, + 0.024113886058330536, + 0.018876897171139717, + 0.1511799544095993, + 0.1423349529504776, + 0.0061782426200807095, + 0.005541348364204168, + 0.009883790276944637, + 0.004629223141819239, + 0.00409815926104784, + 0.010583953000605106, + 0.0080034751445055, + 0.007900552824139595, + 0.0017537825042381883, + 0.002626374363899231, + 0.002927683759480715, + 0.001332728425040841, + 0.0025286544114351273, + 0.0025554057210683823, + 0.002205598633736372, + 0.0024273765739053488, + 0.008998410776257515, + 0.011820632964372635, + 0.015111205168068409, + 0.017901131883263588, + 0.010377684608101845, + 0.00786998588591814, + 0.01890316978096962, + 0.01124249305576086, + 0.011650399304926395, + 0.0255841463804245, + 0.011444334872066975, + 0.007574351504445076, + 0.007015875540673733, + 0.01454427745193243, + 0.008176526986062527, + 0.006617419887334108, + 0.003750381525605917, + 0.0036459406837821007, + 0.0036261638160794973, + 0.0032947203144431114, + 0.0021438971161842346, + 0.0027427845634520054, + 0.002043438609689474, + 0.0022385946940630674 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 7, + 8 + ], + "selected_reward": -0.09651143103837967 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 vast, cavernous hall within a futuristic base on an alien world?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the interior designed in a stark, dark, and minimalist style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the floor consist of a seamless, high-gloss black material that reflects the structure like a black mirror?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do glowing white cracks spread rapidly across the floor in jagged, unpredictable patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot initially static, low-angle, and wide-angle before the scene begins to tremble violently?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014219745993614197 + ], + "opposite_logprob_tokens": [ + -4.264219760894775 + ], + "logprob_reward": -0.014219745993614197, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014219745993614197, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10689353197813034, + 0.24934273958206177, + 0.08730880171060562, + 0.03789690136909485, + 0.09085632860660553, + 0.08690787851810455, + 0.0452609658241272, + 0.06069188565015793, + 0.16567684710025787, + 0.06916414946317673 + ], + "credit_map": [ + 0.010856226086616516, + 0.013972886838018894, + 0.00957375019788742, + 0.01112366747111082, + 0.008250685408711433, + 0.022007929161190987, + 0.01641102135181427, + 0.014697355218231678, + 0.013282730244100094, + 0.012191878631711006, + 0.08638089895248413, + 0.07681147009134293, + 0.01884833164513111, + 0.011133207008242607, + 0.016199782490730286, + 0.014494438655674458, + 0.011985227465629578, + 0.007140928879380226, + 0.010135126300156116, + 0.00883932039141655, + 0.008887205272912979, + 0.008128751069307327, + 0.017248013988137245, + 0.014944227412343025, + 0.004184809513390064, + 0.003658450208604336, + 0.00577580276876688, + 0.005028316751122475, + 0.004233842715620995, + 0.005003083497285843, + 0.005058335606008768, + 0.004954261239618063, + 0.005874456372112036, + 0.002789054997265339, + 0.014928776770830154, + 0.0029763965867459774, + 0.004072874318808317, + 0.015192834660410881, + 0.016908295452594757, + 0.02811364270746708, + 0.00802505761384964, + 0.00746344868093729, + 0.013046853244304657, + 0.008366122841835022, + 0.0057900878600776196, + 0.02023465186357498, + 0.011106402613222599, + 0.012875255197286606, + 0.006469296291470528, + 0.0035914964973926544, + 0.008176338858902454, + 0.0026937841903418303, + 0.008558385074138641, + 0.006404077168554068, + 0.003700838889926672, + 0.005666748620569706, + 0.0071992515586316586, + 0.011578180827200413, + 0.007984546013176441, + 0.007635178975760937, + 0.004864371381700039, + 0.006556744687259197, + 0.008460946381092072, + 0.006412666290998459, + 0.016274407505989075, + 0.05122387781739235, + 0.022141877561807632, + 0.009448547847568989, + 0.013718634843826294, + 0.02520725689828396, + 0.012887546792626381, + 0.01477469876408577, + 0.009504684247076511, + 0.009231691248714924, + 0.009453420527279377, + 0.009035114198923111, + 0.008969363756477833, + 0.007661682087928057, + 0.008962828665971756, + 0.006345360539853573 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 8, + 0 + ], + "selected_reward": -0.014219745993614197 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 vast, cavernous hall within a futuristic base on an alien world?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the interior designed in a stark, dark, and minimalist style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the floor consist of a seamless, high-gloss black material that reflects the structure like a black mirror?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do glowing white cracks spread rapidly across the floor in jagged, unpredictable patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot initially static, low-angle, and wide-angle before the scene begins to tremble violently?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006755726411938667 + ], + "opposite_logprob_tokens": [ + -5.006755828857422 + ], + "logprob_reward": -0.006755726411938667, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006755726411938667, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07248983532190323, + 0.05579037964344025, + 0.5341026782989502, + 0.0416080579161644, + 0.05983502417802811, + 0.04739842191338539, + 0.026598399505019188, + 0.023086145520210266, + 0.09961584955453873, + 0.03947526961565018 + ], + "credit_map": [ + 0.011086750775575638, + 0.009089539758861065, + 0.006383505649864674, + 0.009312677197158337, + 0.007292604073882103, + 0.009320646524429321, + 0.01148690190166235, + 0.00851720292121172, + 0.006712109781801701, + 0.008440567180514336, + 0.007450551260262728, + 0.00853180605918169, + 0.00846856739372015, + 0.005039819981902838, + 0.006255193613469601, + 0.004891762975603342, + 0.055837273597717285, + 0.03907112777233124, + 0.05262631177902222, + 0.04593747481703758, + 0.06937738507986069, + 0.07390379160642624, + 0.09421423077583313, + 0.10313505679368973, + 0.003924168646335602, + 0.002668574685230851, + 0.004069829359650612, + 0.004154452588409185, + 0.003509912174195051, + 0.006860613357275724, + 0.006350683979690075, + 0.010069819167256355, + 0.0037339329719543457, + 0.004631944000720978, + 0.009295792318880558, + 0.003598943818360567, + 0.004268025979399681, + 0.005227165296673775, + 0.0109872967004776, + 0.018091918900609016, + 0.004480512347072363, + 0.003934330306947231, + 0.007820513099431992, + 0.003777315840125084, + 0.0035980853717774153, + 0.01155853271484375, + 0.005676195491105318, + 0.006552936974912882, + 0.003973802551627159, + 0.003294525668025017, + 0.0031416367273777723, + 0.0017367932014167309, + 0.00325248995795846, + 0.005540831014513969, + 0.0020227304194122553, + 0.0036355897318571806, + 0.002754616318270564, + 0.003912971355021, + 0.0033457670360803604, + 0.0024248631671071053, + 0.0023407922126352787, + 0.0029853901360183954, + 0.0030085209291428328, + 0.0023132229689508677, + 0.012177401222288609, + 0.030031368136405945, + 0.013299860991537571, + 0.005735160317271948, + 0.007978233508765697, + 0.014949577860534191, + 0.00754661625251174, + 0.007897622883319855, + 0.004305284935981035, + 0.0049569508992135525, + 0.005712952930480242, + 0.005838275421410799, + 0.004265923053026199, + 0.0038322368636727333, + 0.005296074319630861, + 0.005267568398267031 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 8, + 0 + ], + "selected_reward": -0.006755726411938667 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 vast, cavernous hall within a futuristic base on an alien world?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the interior designed in a stark, dark, and minimalist style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the floor consist of a seamless, high-gloss black material that reflects the structure like a black mirror?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do glowing white cracks spread rapidly across the floor in jagged, unpredictable patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot initially static, low-angle, and wide-angle before the scene begins to tremble violently?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0015193596482276917 + ], + "opposite_logprob_tokens": [ + -6.501519203186035 + ], + "logprob_reward": -0.0015193596482276917, + "yes_no_margin_reward": 6.5, + "token_credit_reward": -0.0015193596482276917, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06397446990013123, + 0.029795784503221512, + 0.08333496749401093, + 0.06507416069507599, + 0.42527398467063904, + 0.036707110702991486, + 0.05916919931769371, + 0.14472948014736176, + 0.05318956449627876, + 0.03875134885311127 + ], + "credit_map": [ + 0.010624506510794163, + 0.008226095698773861, + 0.00686285924166441, + 0.0067684231325984, + 0.007269098423421383, + 0.008556326851248741, + 0.008006811141967773, + 0.007660340052098036, + 0.0037612104788422585, + 0.0030961434822529554, + 0.004550104960799217, + 0.0057440553791821, + 0.004053983371704817, + 0.0022443390917032957, + 0.003289356827735901, + 0.003056587651371956, + 0.012582587078213692, + 0.007013176567852497, + 0.010403666645288467, + 0.007420932408422232, + 0.008405645377933979, + 0.010308468714356422, + 0.012461148202419281, + 0.014739331789314747, + 0.005773554556071758, + 0.006217764690518379, + 0.006971965078264475, + 0.006963933818042278, + 0.007779514882713556, + 0.009181871078908443, + 0.007766190450638533, + 0.014419355429708958, + 0.0152844013646245, + 0.021885370835661888, + 0.043958522379398346, + 0.013895655050873756, + 0.02412266843020916, + 0.020411694422364235, + 0.14021728932857513, + 0.1454983353614807, + 0.004414152819663286, + 0.0047175949439406395, + 0.005886633414775133, + 0.004656653851270676, + 0.0027809860184788704, + 0.005311500281095505, + 0.004287627059966326, + 0.004651958122849464, + 0.006831569131463766, + 0.005205663852393627, + 0.006655512843281031, + 0.0032467932906001806, + 0.011075475253164768, + 0.011058741249144077, + 0.006993166171014309, + 0.008102270774543285, + 0.015240846201777458, + 0.016752637922763824, + 0.02107437513768673, + 0.024142777547240257, + 0.014470105990767479, + 0.013378052972257137, + 0.02390119805932045, + 0.015769468620419502, + 0.005037345457822084, + 0.015551583841443062, + 0.007872622460126877, + 0.003780801547691226, + 0.003990603145211935, + 0.007756911218166351, + 0.0040350365452468395, + 0.005164653994143009, + 0.003411805024370551, + 0.005425885785371065, + 0.005549474619328976, + 0.005351366009563208, + 0.0038162157870829105, + 0.004464662168174982, + 0.006482094991952181, + 0.004249840043485165 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 7, + 2 + ], + "selected_reward": -0.0015193596482276917 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 vast, cavernous hall within a futuristic base on an alien world?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the interior designed in a stark, dark, and minimalist style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the floor consist of a seamless, high-gloss black material that reflects the structure like a black mirror?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do glowing white cracks spread rapidly across the floor in jagged, unpredictable patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot initially static, low-angle, and wide-angle before the scene begins to tremble violently?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018188007175922394 + ], + "opposite_logprob_tokens": [ + -4.018187999725342 + ], + "logprob_reward": -0.018188007175922394, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018188007175922394, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0798487663269043, + 0.19285327196121216, + 0.13558848202228546, + 0.08554860949516296, + 0.23567335307598114, + 0.05226109176874161, + 0.036694999784231186, + 0.07998953014612198, + 0.07269518822431564, + 0.028846731409430504 + ], + "credit_map": [ + 0.011744150891900063, + 0.012178255245089531, + 0.010778146795928478, + 0.00942132156342268, + 0.0073595172725617886, + 0.00987697672098875, + 0.00828602071851492, + 0.010204375721514225, + 0.012139257974922657, + 0.010167507454752922, + 0.06630229949951172, + 0.05715377256274223, + 0.01654251478612423, + 0.008729138411581516, + 0.01102469302713871, + 0.010794112458825111, + 0.015540078282356262, + 0.012034622021019459, + 0.014920239336788654, + 0.01282956637442112, + 0.017233695834875107, + 0.015726923942565918, + 0.021906530484557152, + 0.02539682760834694, + 0.010438717901706696, + 0.008351334370672703, + 0.009164449758827686, + 0.010276790708303452, + 0.01095032598823309, + 0.012467843480408192, + 0.015698565170168877, + 0.008200583979487419, + 0.007890298962593079, + 0.010696151293814182, + 0.027669725939631462, + 0.006936931516975164, + 0.01307083759456873, + 0.009441983886063099, + 0.0844789519906044, + 0.07548847049474716, + 0.005306111183017492, + 0.004812244325876236, + 0.011901764199137688, + 0.003507538465783, + 0.003985338378697634, + 0.00897940993309021, + 0.006580621004104614, + 0.007188064511865377, + 0.0037069942336529493, + 0.0028474496211856604, + 0.0033160413149744272, + 0.0021460989955812693, + 0.0052723996341228485, + 0.008754285983741283, + 0.005021541379392147, + 0.005630190949887037, + 0.006531096994876862, + 0.009437210857868195, + 0.013867564499378204, + 0.013193699531257153, + 0.007573354989290237, + 0.006396781653165817, + 0.014398416504263878, + 0.00859140232205391, + 0.010678296908736229, + 0.019842663779854774, + 0.009958057664334774, + 0.005049070809036493, + 0.005593165755271912, + 0.010490828193724155, + 0.005165691953152418, + 0.005917411297559738, + 0.004472022410482168, + 0.004042769316583872, + 0.0037430988159030676, + 0.0039388942532241344, + 0.0029862753581255674, + 0.003437226405367255, + 0.003598698414862156, + 0.002627747133374214 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 1, + 2 + ], + "selected_reward": -0.018188007175922394 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An initially static, low-angle, wide-angle shot begins to tremble violently, capturing the interior of a vast, cavernous hall within a futuristic base on an alien world, designed in a stark, dark, and minimalist style. The floor, a seamless expanse of polished, high-gloss black material that reflects the entire structure like a black mirror, suddenly begins to crack at its center. A glowing spiderweb of cracks emanating from pure white light erupts across the surface, spreading at a terrifying pace and branching out in jagged, unpredictable patterns. As the violence throughout the structure intensifies, the polished floor begins to bulge and tilt. The only illumination comes from emergency lights, which flicker erratically, casting the scene in a disorienting, stroboscopic glare, creating long, dancing shadows and moments of absolute darkness, establishing an atmosphere of chaotic and utterly urgent catastrophic structural failure.", + "seed": 4451, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000156/group_11/index_5590_seed_4451_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": -0.015760584268718958, + "vlm_logprob": -0.015760584268718958, + "vlm_yes_no_margin": 4.55, + "vlm_token_credit": -0.015760584268718958, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 vast, cavernous hall within a futuristic base on an alien world?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the interior designed in a stark, dark, and minimalist style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the floor consist of a seamless, high-gloss black material that reflects the structure like a black mirror?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do glowing white cracks spread rapidly across the floor in jagged, unpredictable patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot initially static, low-angle, and wide-angle before the scene begins to tremble violently?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.030433597043156624 + ], + "opposite_logprob_tokens": [ + -3.6554336547851562 + ], + "logprob_reward": -0.030433597043156624, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.030433597043156624, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1868252456188202, + 0.07366912066936493, + 0.06030429154634476, + 0.04809381440281868, + 0.15309862792491913, + 0.03570424020290375, + 0.08082625269889832, + 0.0824626088142395, + 0.19881443679332733, + 0.08020132780075073 + ], + "credit_map": [ + 0.013462524861097336, + 0.01583291031420231, + 0.028419625014066696, + 0.033172670751810074, + 0.013344543986022472, + 0.022868890315294266, + 0.03657471388578415, + 0.023149386048316956, + 0.006104596424847841, + 0.0069442144595086575, + 0.00569281866773963, + 0.017999429255723953, + 0.006347222253680229, + 0.006450086832046509, + 0.015373785980045795, + 0.008756962604820728, + 0.007094454485923052, + 0.007160068489611149, + 0.006902673747390509, + 0.005594521760940552, + 0.006316139828413725, + 0.007086772006005049, + 0.012983995489776134, + 0.007165668066591024, + 0.004732705187052488, + 0.005504791159182787, + 0.007156591862440109, + 0.008548279292881489, + 0.009130574762821198, + 0.0055982936173677444, + 0.0035146926529705524, + 0.003907887265086174, + 0.018923712894320488, + 0.011730720289051533, + 0.020701434463262558, + 0.022444775328040123, + 0.02504134178161621, + 0.022726157680153847, + 0.015482245944440365, + 0.016048237681388855, + 0.005027382634580135, + 0.004421429242938757, + 0.0032967827282845974, + 0.0034163224045187235, + 0.0043289694003760815, + 0.005716419778764248, + 0.004357054829597473, + 0.005139879882335663, + 0.009196783415973186, + 0.009853134863078594, + 0.0070575024001300335, + 0.006285103037953377, + 0.007501167245209217, + 0.025969145819544792, + 0.005370969418436289, + 0.009592448361217976, + 0.01750180870294571, + 0.01374286413192749, + 0.004464791622012854, + 0.010977489873766899, + 0.007374272681772709, + 0.009805523790419102, + 0.005614855792373419, + 0.012981005944311619, + 0.02564675733447075, + 0.04308735951781273, + 0.017213044688105583, + 0.03323676064610481, + 0.029388127848505974, + 0.02315972000360489, + 0.015054265037178993, + 0.012028391472995281, + 0.00857024546712637, + 0.009130156598985195, + 0.008351710624992847, + 0.007528123911470175, + 0.013615485280752182, + 0.013576813042163849, + 0.0077207558788359165, + 0.0117080332711339 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 4 + ], + "selected_reward": -0.030433597043156624 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 vast, cavernous hall within a futuristic base on an alien world?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the interior designed in a stark, dark, and minimalist style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the floor consist of a seamless, high-gloss black material that reflects the structure like a black mirror?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do glowing white cracks spread rapidly across the floor in jagged, unpredictable patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot initially static, low-angle, and wide-angle before the scene begins to tremble violently?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009794275276362896 + ], + "opposite_logprob_tokens": [ + -4.634794235229492 + ], + "logprob_reward": -0.009794275276362896, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009794275276362896, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1380009949207306, + 0.12085220217704773, + 0.0918237641453743, + 0.05915428325533867, + 0.09870026260614395, + 0.036773018538951874, + 0.12645958364009857, + 0.04934140294790268, + 0.20519958436489105, + 0.07369481772184372 + ], + "credit_map": [ + 0.01094844937324524, + 0.010277191177010536, + 0.03523053228855133, + 0.022500650957226753, + 0.007904700003564358, + 0.01359169278293848, + 0.021189318969845772, + 0.016358474269509315, + 0.014340652152895927, + 0.01170302927494049, + 0.008633727207779884, + 0.017228635028004646, + 0.01709062047302723, + 0.01495825033634901, + 0.019209355115890503, + 0.017687946557998657, + 0.01219390332698822, + 0.01026469562202692, + 0.012334775179624557, + 0.01214978564530611, + 0.007574190851300955, + 0.00845268927514553, + 0.016543610021471977, + 0.012310121208429337, + 0.008307651616632938, + 0.008026838302612305, + 0.009053898975253105, + 0.011230561882257462, + 0.0070403083227574825, + 0.005787399597465992, + 0.0035519334487617016, + 0.006155697163194418, + 0.017875872552394867, + 0.005748445633798838, + 0.010373801924288273, + 0.012564044445753098, + 0.01781271956861019, + 0.01237513404339552, + 0.011768451891839504, + 0.010181799530982971, + 0.00409623421728611, + 0.0036272883880883455, + 0.0030813051853328943, + 0.005372594576328993, + 0.0067155384458601475, + 0.004761790856719017, + 0.004284440539777279, + 0.004833828657865524, + 0.012883462011814117, + 0.011587091721594334, + 0.00585547462105751, + 0.005973595194518566, + 0.013202820904552937, + 0.047256194055080414, + 0.00717915827408433, + 0.022521808743476868, + 0.004490671679377556, + 0.005715141538530588, + 0.003051955020055175, + 0.01516425609588623, + 0.0026186483446508646, + 0.004008057992905378, + 0.0033218555618077517, + 0.010970821604132652, + 0.02517414651811123, + 0.0392470620572567, + 0.019196202978491783, + 0.03513893485069275, + 0.028271138668060303, + 0.024580327793955803, + 0.017841193825006485, + 0.01575060561299324, + 0.009168886579573154, + 0.010792902670800686, + 0.011146621778607368, + 0.0065841758623719215, + 0.009651600383222103, + 0.009655945003032684, + 0.00885708536952734, + 0.007837614044547081 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 6 + ], + "selected_reward": -0.009794275276362896 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 vast, cavernous hall within a futuristic base on an alien world?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the interior designed in a stark, dark, and minimalist style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the floor consist of a seamless, high-gloss black material that reflects the structure like a black mirror?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do glowing white cracks spread rapidly across the floor in jagged, unpredictable patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot initially static, low-angle, and wide-angle before the scene begins to tremble violently?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005959954112768173 + ], + "opposite_logprob_tokens": [ + -5.130959987640381 + ], + "logprob_reward": -0.005959954112768173, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005959954112768173, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2144479900598526, + 0.08863727748394012, + 0.12861450016498566, + 0.09971205145120621, + 0.05681344121694565, + 0.062131091952323914, + 0.05970834195613861, + 0.07904937118291855, + 0.11236568540334702, + 0.09852016717195511 + ], + "credit_map": [ + 0.014580164104700089, + 0.018521441146731377, + 0.03833571821451187, + 0.03654840588569641, + 0.013846307061612606, + 0.022708473727107048, + 0.0502377487719059, + 0.019669724628329277, + 0.007419044151902199, + 0.010090637020766735, + 0.008285542950034142, + 0.013149370439350605, + 0.010208291001617908, + 0.010392648167908192, + 0.011740942485630512, + 0.017350798472762108, + 0.01602361723780632, + 0.013581881299614906, + 0.023831408470869064, + 0.01970180682837963, + 0.009885882027447224, + 0.014209254644811153, + 0.0139768673107028, + 0.017403781414031982, + 0.007680611684918404, + 0.009524502791464329, + 0.013332204893231392, + 0.020694656297564507, + 0.01419041957706213, + 0.013953572139143944, + 0.009554056450724602, + 0.01078202947974205, + 0.008466103114187717, + 0.004751278553158045, + 0.006071553099900484, + 0.009101210162043571, + 0.008204877376556396, + 0.008794188499450684, + 0.005927944090217352, + 0.0054962849244475365, + 0.0070471870712935925, + 0.007685841992497444, + 0.005957967601716518, + 0.007631412707269192, + 0.01004120334982872, + 0.010482280515134335, + 0.006086468230932951, + 0.0071987309493124485, + 0.006935476325452328, + 0.006921570748090744, + 0.006397029850631952, + 0.008159988559782505, + 0.007792254909873009, + 0.00692228227853775, + 0.008839045651257038, + 0.007740698289126158, + 0.008385803550481796, + 0.009702482260763645, + 0.004773006308823824, + 0.023427121341228485, + 0.004735806491225958, + 0.006826464552432299, + 0.008137423545122147, + 0.01306126918643713, + 0.014930200763046741, + 0.02640535682439804, + 0.009051215834915638, + 0.014129781164228916, + 0.018656475469470024, + 0.013676859438419342, + 0.009073353372514248, + 0.006442445330321789, + 0.01242840476334095, + 0.015597970224916935, + 0.01889505237340927, + 0.01445035357028246, + 0.011714670807123184, + 0.009979544207453728, + 0.009855953976511955, + 0.00559822004288435 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 8 + ], + "selected_reward": -0.005959954112768173 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 vast, cavernous hall within a futuristic base on an alien world?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the interior designed in a stark, dark, and minimalist style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the floor consist of a seamless, high-gloss black material that reflects the structure like a black mirror?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do glowing white cracks spread rapidly across the floor in jagged, unpredictable patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot initially static, low-angle, and wide-angle before the scene begins to tremble violently?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0028216331265866756 + ], + "opposite_logprob_tokens": [ + -5.877821445465088 + ], + "logprob_reward": -0.0028216331265866756, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.0028216331265866756, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.178848996758461, + 0.06519705057144165, + 0.03924824669957161, + 0.05938189476728439, + 0.0448661670088768, + 0.06883595883846283, + 0.07089728116989136, + 0.05487534776329994, + 0.2918802499771118, + 0.12596872448921204 + ], + "credit_map": [ + 0.012531849555671215, + 0.017464283853769302, + 0.028595900163054466, + 0.029065603390336037, + 0.015488315373659134, + 0.024023476988077164, + 0.035423874855041504, + 0.01625569351017475, + 0.006288396194577217, + 0.005047193728387356, + 0.004375442396849394, + 0.00818109791725874, + 0.006621701177209616, + 0.007226237095892429, + 0.010503311641514301, + 0.016953671351075172, + 0.004929682239890099, + 0.005522832740098238, + 0.004052181262522936, + 0.0023661924060434103, + 0.004615670070052147, + 0.004552145954221487, + 0.00753688532859087, + 0.005672658793628216, + 0.007748676463961601, + 0.012280010618269444, + 0.008751912973821163, + 0.00793552864342928, + 0.009382780641317368, + 0.007128492463380098, + 0.003012822475284338, + 0.0031416730489581823, + 0.007079491391777992, + 0.0031434460543096066, + 0.004754896275699139, + 0.007147427648305893, + 0.005245234817266464, + 0.007383228279650211, + 0.005171333905309439, + 0.004941113293170929, + 0.009347311221063137, + 0.009822600521147251, + 0.005730472970753908, + 0.011875413358211517, + 0.004497898742556572, + 0.012033099308609962, + 0.008646058849990368, + 0.006883102469146252, + 0.008082814514636993, + 0.009719808585941792, + 0.005290583707392216, + 0.011163244023919106, + 0.009914455004036427, + 0.008577979169785976, + 0.008983155712485313, + 0.009165238589048386, + 0.0100493673235178, + 0.00785371009260416, + 0.003019522875547409, + 0.008329564705491066, + 0.003961998037993908, + 0.005332570988684893, + 0.007216809317469597, + 0.009111805818974972, + 0.0392950139939785, + 0.06791062653064728, + 0.023739034309983253, + 0.03722807765007019, + 0.048354584723711014, + 0.03359353542327881, + 0.021290425211191177, + 0.020468980073928833, + 0.015044711530208588, + 0.013101840391755104, + 0.020419003441929817, + 0.012880999594926834, + 0.013836441561579704, + 0.01785125397145748, + 0.012951838783919811, + 0.01988263800740242 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 9 + ], + "selected_reward": -0.0028216331265866756 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 vast, cavernous hall within a futuristic base on an alien world?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the interior designed in a stark, dark, and minimalist style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the floor consist of a seamless, high-gloss black material that reflects the structure like a black mirror?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do glowing white cracks spread rapidly across the floor in jagged, unpredictable patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot initially static, low-angle, and wide-angle before the scene begins to tremble violently?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02979346178472042 + ], + "opposite_logprob_tokens": [ + -3.5297935009002686 + ], + "logprob_reward": -0.02979346178472042, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.02979346178472042, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19518671929836273, + 0.1392570436000824, + 0.03370825946331024, + 0.04680251330137253, + 0.05834499001502991, + 0.041856054216623306, + 0.04716723412275314, + 0.18049252033233643, + 0.17689688503742218, + 0.08028785139322281 + ], + "credit_map": [ + 0.01515436451882124, + 0.019570395350456238, + 0.02816842310130596, + 0.02970856800675392, + 0.01620742306113243, + 0.028009267523884773, + 0.038826607167720795, + 0.019541624933481216, + 0.016008146107196808, + 0.011347068473696709, + 0.017465554177761078, + 0.01985982246696949, + 0.01878468133509159, + 0.013491738587617874, + 0.02292383648455143, + 0.019376195967197418, + 0.003883244702592492, + 0.004617705475538969, + 0.005177645944058895, + 0.002538678701967001, + 0.0036349825095385313, + 0.004254200961440802, + 0.0058856927789747715, + 0.003716106992214918, + 0.0043147555552423, + 0.004979276563972235, + 0.006807791069149971, + 0.005315476097166538, + 0.008764070458710194, + 0.011273528449237347, + 0.0024596676230430603, + 0.00288794725202024, + 0.012013144791126251, + 0.0034371791407465935, + 0.00609243381768465, + 0.007771466393023729, + 0.008144669234752655, + 0.0075362431816756725, + 0.007190363015979528, + 0.006159486249089241, + 0.004787012003362179, + 0.0047119492664933205, + 0.0036362777464091778, + 0.004616669379174709, + 0.00950705911964178, + 0.0064104232005774975, + 0.004026590846478939, + 0.004160070791840553, + 0.004364094231277704, + 0.0059935590252280235, + 0.007462773006409407, + 0.004581864923238754, + 0.005734150297939777, + 0.010596610605716705, + 0.004498456604778767, + 0.0039357212372124195, + 0.023963674902915955, + 0.023195961490273476, + 0.01056631002575159, + 0.05243680253624916, + 0.012712386436760426, + 0.014897051267325878, + 0.012227203696966171, + 0.030493123456835747, + 0.022279098629951477, + 0.0396592877805233, + 0.01482091099023819, + 0.03109990991652012, + 0.027181115001440048, + 0.017847266048192978, + 0.013480696827173233, + 0.010528579354286194, + 0.009519982151687145, + 0.009165520779788494, + 0.015306969173252583, + 0.01046421192586422, + 0.007454022765159607, + 0.010240844450891018, + 0.00838556233793497, + 0.009750726632773876 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 8 + ], + "selected_reward": -0.02979346178472042 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An initially static, low-angle, wide-angle shot begins to tremble violently, capturing the interior of a vast, cavernous hall within a futuristic base on an alien world, designed in a stark, dark, and minimalist style. The floor, a seamless expanse of polished, high-gloss black material that reflects the entire structure like a black mirror, suddenly begins to crack at its center. A glowing spiderweb of cracks emanating from pure white light erupts across the surface, spreading at a terrifying pace and branching out in jagged, unpredictable patterns. As the violence throughout the structure intensifies, the polished floor begins to bulge and tilt. The only illumination comes from emergency lights, which flicker erratically, casting the scene in a disorienting, stroboscopic glare, creating long, dancing shadows and moments of absolute darkness, establishing an atmosphere of chaotic and utterly urgent catastrophic structural failure.", + "seed": 4452, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000156/group_11/index_5590_seed_4452_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": -0.012436216697096825, + "vlm_logprob": -0.012436216697096825, + "vlm_yes_no_margin": 4.6, + "vlm_token_credit": -0.012436216697096825, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 vast, cavernous hall within a futuristic base on an alien world?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the interior designed in a stark, dark, and minimalist style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the floor consist of a seamless, high-gloss black material that reflects the structure like a black mirror?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do glowing white cracks spread rapidly across the floor in jagged, unpredictable patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot initially static, low-angle, and wide-angle before the scene begins to tremble violently?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023772219195961952 + ], + "opposite_logprob_tokens": [ + -3.8987722396850586 + ], + "logprob_reward": -0.023772219195961952, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.023772219195961952, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.33678102493286133, + 0.07675155252218246, + 0.12974853813648224, + 0.12205179035663605, + 0.0632113590836525, + 0.03968852385878563, + 0.0312965102493763, + 0.0325842946767807, + 0.09132388234138489, + 0.07656250894069672 + ], + "credit_map": [ + 0.027451978996396065, + 0.035435546189546585, + 0.04090779647231102, + 0.029194161295890808, + 0.05555746704339981, + 0.06146402284502983, + 0.020413389429450035, + 0.06635666638612747, + 0.006916387006640434, + 0.01002196129411459, + 0.009103321470320225, + 0.0073345438577234745, + 0.01792951114475727, + 0.011649738997220993, + 0.007461436092853546, + 0.006334652658551931, + 0.021365944296121597, + 0.013026508502662182, + 0.012442675419151783, + 0.010309417732059956, + 0.01734510064125061, + 0.0269151758402586, + 0.017510466277599335, + 0.01083325408399105, + 0.011542527936398983, + 0.009708618745207787, + 0.012149459682404995, + 0.009602963924407959, + 0.030110983178019524, + 0.02920890785753727, + 0.00863666832447052, + 0.011091664433479309, + 0.010053365491330624, + 0.008698834106326103, + 0.007368084043264389, + 0.005518525373190641, + 0.011096058413386345, + 0.0075720385648310184, + 0.004653930198401213, + 0.008250522427260876, + 0.007275667507201433, + 0.007076956331729889, + 0.003817523131147027, + 0.002665863139554858, + 0.008259721100330353, + 0.005148129537701607, + 0.0026118659880012274, + 0.002832794329151511, + 0.005917131435126066, + 0.005269067361950874, + 0.0024090101942420006, + 0.001609027269296348, + 0.008249790407717228, + 0.003763314802199602, + 0.002581842942163348, + 0.001497327466495335, + 0.00881174486130476, + 0.007911928929388523, + 0.002264935290440917, + 0.001376064377836883, + 0.0030186946969479322, + 0.005913218017667532, + 0.0016798748401924968, + 0.001607837388291955, + 0.026023952290415764, + 0.015712451189756393, + 0.006155030336230993, + 0.005247869528830051, + 0.011938205920159817, + 0.013997092843055725, + 0.006856196094304323, + 0.005393084138631821, + 0.008867795579135418, + 0.01065587904304266, + 0.006310216151177883, + 0.005444886162877083, + 0.011042939499020576, + 0.016332430765032768, + 0.010527588427066803, + 0.0073807681910693645 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 3 + ], + "selected_reward": -0.023772219195961952 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 vast, cavernous hall within a futuristic base on an alien world?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the interior designed in a stark, dark, and minimalist style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the floor consist of a seamless, high-gloss black material that reflects the structure like a black mirror?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do glowing white cracks spread rapidly across the floor in jagged, unpredictable patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot initially static, low-angle, and wide-angle before the scene begins to tremble violently?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011084645055234432 + ], + "opposite_logprob_tokens": [ + -4.51108455657959 + ], + "logprob_reward": -0.011084645055234432, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011084645055234432, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18208886682987213, + 0.11440926790237427, + 0.2458353489637375, + 0.1732601821422577, + 0.09762856364250183, + 0.04823070019483566, + 0.020344577729701996, + 0.015356655232608318, + 0.017903635278344154, + 0.08494224399328232 + ], + "credit_map": [ + 0.013469841331243515, + 0.02258080057799816, + 0.027080168947577477, + 0.020795069634914398, + 0.037598177790641785, + 0.021225571632385254, + 0.013860953040421009, + 0.025478284806013107, + 0.01714358478784561, + 0.017066946253180504, + 0.010991385206580162, + 0.006724708713591099, + 0.03279663622379303, + 0.015571394003927708, + 0.006552655715495348, + 0.007561950944364071, + 0.03180306777358055, + 0.025449339300394058, + 0.025573842227458954, + 0.019402755424380302, + 0.03282475471496582, + 0.05730488896369934, + 0.033416736871004105, + 0.020059965550899506, + 0.021954834461212158, + 0.014128432609140873, + 0.013295253738760948, + 0.014410086907446384, + 0.03804147616028786, + 0.02787492796778679, + 0.015845663845539093, + 0.02770950272679329, + 0.018255485221743584, + 0.013156847096979618, + 0.009238868951797485, + 0.007365926168859005, + 0.02129797264933586, + 0.010944771580398083, + 0.0074966768734157085, + 0.0098720146343112, + 0.00512955104932189, + 0.007841805927455425, + 0.005409709177911282, + 0.004361649043858051, + 0.00936939474195242, + 0.005960484035313129, + 0.0036988905631005764, + 0.006459211930632591, + 0.003957450855523348, + 0.003457183251157403, + 0.0017941772239282727, + 0.0010010036639869213, + 0.0046627456322312355, + 0.0031815150287002325, + 0.0012158058816567063, + 0.0010746951447799802, + 0.00396541366353631, + 0.0020700665190815926, + 0.001528494292870164, + 0.0009407829493284225, + 0.0024181774351745844, + 0.0022526218090206385, + 0.0012822981225326657, + 0.0008988013141788542, + 0.0029203025624156, + 0.0028048029635101557, + 0.0018311175517737865, + 0.0016029321122914553, + 0.0027828046586364508, + 0.0023323732893913984, + 0.0021459581330418587, + 0.0014833429595455527, + 0.008025197312235832, + 0.011675742454826832, + 0.008356445468962193, + 0.0076857758685946465, + 0.013485386036336422, + 0.013938385993242264, + 0.013176464475691319, + 0.008598847314715385 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 3 + ], + "selected_reward": -0.011084645055234432 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 vast, cavernous hall within a futuristic base on an alien world?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the interior designed in a stark, dark, and minimalist style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the floor consist of a seamless, high-gloss black material that reflects the structure like a black mirror?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do glowing white cracks spread rapidly across the floor in jagged, unpredictable patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot initially static, low-angle, and wide-angle before the scene begins to tremble violently?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00764006981626153 + ], + "opposite_logprob_tokens": [ + -4.8826398849487305 + ], + "logprob_reward": -0.00764006981626153, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.00764006981626153, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15073582530021667, + 0.13186241686344147, + 0.23607531189918518, + 0.07111021876335144, + 0.09575407207012177, + 0.07196644693613052, + 0.04520533233880997, + 0.021130232140421867, + 0.050688304007053375, + 0.12547177076339722 + ], + "credit_map": [ + 0.016897166147828102, + 0.017450811341404915, + 0.021620525047183037, + 0.019200071692466736, + 0.02328101545572281, + 0.016713961958885193, + 0.013964381068944931, + 0.02160790003836155, + 0.013015128672122955, + 0.01710011437535286, + 0.014124413952231407, + 0.014436991885304451, + 0.030716078355908394, + 0.01770002208650112, + 0.009630346670746803, + 0.0151393236592412, + 0.03839526325464249, + 0.022957494482398033, + 0.02693481557071209, + 0.021184442564845085, + 0.029694199562072754, + 0.03767727315425873, + 0.03682375326752663, + 0.02240808866918087, + 0.008968831971287727, + 0.00606943853199482, + 0.0035807087551802397, + 0.004946669097989798, + 0.018889423459768295, + 0.015148614533245564, + 0.006044103763997555, + 0.0074624307453632355, + 0.014962048269808292, + 0.013856508769094944, + 0.009472585283219814, + 0.007422429043799639, + 0.01961101032793522, + 0.013693069107830524, + 0.007575938943773508, + 0.009160497225821018, + 0.011959303170442581, + 0.011957285925745964, + 0.007281295955181122, + 0.0051064989529550076, + 0.013577775098383427, + 0.010998371057212353, + 0.004737362265586853, + 0.006348560098558664, + 0.007774479687213898, + 0.007844245061278343, + 0.0043617128394544125, + 0.0019174267072230577, + 0.010832265950739384, + 0.006205444689840078, + 0.0029049094300717115, + 0.003364854957908392, + 0.004268274642527103, + 0.003312892047688365, + 0.0023202234879136086, + 0.0012311823666095734, + 0.00450689485296607, + 0.0028578434139490128, + 0.0015130615793168545, + 0.0011198613792657852, + 0.012831865809857845, + 0.007939200848340988, + 0.0038697647396475077, + 0.003554530907422304, + 0.007393483072519302, + 0.007277028169482946, + 0.004426593892276287, + 0.00339584075845778, + 0.01525514479726553, + 0.017436595633625984, + 0.014639120548963547, + 0.011906836181879044, + 0.017395449802279472, + 0.023173101246356964, + 0.011035151779651642, + 0.01463038194924593 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 1 + ], + "selected_reward": -0.00764006981626153 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 vast, cavernous hall within a futuristic base on an alien world?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the interior designed in a stark, dark, and minimalist style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the floor consist of a seamless, high-gloss black material that reflects the structure like a black mirror?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do glowing white cracks spread rapidly across the floor in jagged, unpredictable patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot initially static, low-angle, and wide-angle before the scene begins to tremble violently?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0036183618940413 + ], + "opposite_logprob_tokens": [ + -5.628618240356445 + ], + "logprob_reward": -0.0036183618940413, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.0036183618940413, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1669415533542633, + 0.06445731222629547, + 0.18097594380378723, + 0.11995365470647812, + 0.047403279691934586, + 0.04783441871404648, + 0.11617033928632736, + 0.04696524888277054, + 0.11626742780208588, + 0.09303078055381775 + ], + "credit_map": [ + 0.015672195702791214, + 0.018207823857665062, + 0.028011996299028397, + 0.018944522365927696, + 0.02792319469153881, + 0.01977222040295601, + 0.01234951801598072, + 0.026060087606310844, + 0.005057518370449543, + 0.008295968174934387, + 0.008154721930623055, + 0.005751670803874731, + 0.01482472289353609, + 0.008345535956323147, + 0.0074385544285178185, + 0.006588625255972147, + 0.0363725945353508, + 0.017102014273405075, + 0.01528991386294365, + 0.012676660902798176, + 0.027615629136562347, + 0.03475170582532883, + 0.02314016968011856, + 0.014027278870344162, + 0.015505039133131504, + 0.010867493227124214, + 0.01006659958511591, + 0.010246246121823788, + 0.026145976036787033, + 0.0246159378439188, + 0.010815524496138096, + 0.011690848506987095, + 0.004227905534207821, + 0.0067531936801970005, + 0.004421874415129423, + 0.003476366400718689, + 0.013520598411560059, + 0.008092493750154972, + 0.0029971112962812185, + 0.003913740161806345, + 0.010011251084506512, + 0.008131402544677258, + 0.0046203709207475185, + 0.003563247388228774, + 0.008005993440747261, + 0.006247426848858595, + 0.0036551090888679028, + 0.003599621821194887, + 0.025833221152424812, + 0.01886148564517498, + 0.010936424136161804, + 0.008631523698568344, + 0.021091952919960022, + 0.015037917532026768, + 0.007724266964942217, + 0.008053544908761978, + 0.00971175916492939, + 0.011017351411283016, + 0.003929086495190859, + 0.0016548980493098497, + 0.007868457585573196, + 0.00726166320964694, + 0.003117212327197194, + 0.0024048243649303913, + 0.034735534340143204, + 0.019425205886363983, + 0.007702508009970188, + 0.006825696211308241, + 0.014831369742751122, + 0.016454948112368584, + 0.008771170862019062, + 0.00752100208774209, + 0.009389681741595268, + 0.010961653664708138, + 0.011471531353890896, + 0.008039060980081558, + 0.015068369917571545, + 0.01369559857994318, + 0.009752274490892887, + 0.014652621001005173 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 3 + ], + "selected_reward": -0.0036183618940413 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set inside a vast, cavernous hall within a futuristic base on an alien world?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the interior designed in a stark, dark, and minimalist style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the floor consist of a seamless, high-gloss black material that reflects the structure like a black mirror?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do glowing white cracks spread rapidly across the floor in jagged, unpredictable patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot initially static, low-angle, and wide-angle before the scene begins to tremble violently?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01606578752398491 + ], + "opposite_logprob_tokens": [ + -4.14106559753418 + ], + "logprob_reward": -0.01606578752398491, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.01606578752398491, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2941751182079315, + 0.11155357211828232, + 0.07594079524278641, + 0.08713608235120773, + 0.15214470028877258, + 0.04817052185535431, + 0.04769972339272499, + 0.07220570743083954, + 0.04441303759813309, + 0.0665607675909996 + ], + "credit_map": [ + 0.027532637119293213, + 0.03059493750333786, + 0.04263579845428467, + 0.030660944059491158, + 0.038156893104314804, + 0.04750937223434448, + 0.02247168868780136, + 0.05461284890770912, + 0.009509158320724964, + 0.0145673006772995, + 0.013932264409959316, + 0.010933822020888329, + 0.022114906460046768, + 0.016008971258997917, + 0.012179790064692497, + 0.012307366356253624, + 0.014097178354859352, + 0.010716764256358147, + 0.007584635633975267, + 0.005438898224383593, + 0.011976934038102627, + 0.011713211424648762, + 0.008441586047410965, + 0.005971583072096109, + 0.015533777885138988, + 0.00846958439797163, + 0.008789192885160446, + 0.007810467388480902, + 0.01805206947028637, + 0.011434024199843407, + 0.006627643946558237, + 0.01041931752115488, + 0.013086154125630856, + 0.02008429914712906, + 0.014873528853058815, + 0.013351445086300373, + 0.029839592054486275, + 0.02368321642279625, + 0.012231490574777126, + 0.024994971230626106, + 0.012038270011544228, + 0.008887954987585545, + 0.003225165419280529, + 0.0030819878447800875, + 0.00916451495140791, + 0.005781346466392279, + 0.0027675433084368706, + 0.003223738633096218, + 0.004769369028508663, + 0.008423828519880772, + 0.0033764862455427647, + 0.0030906565953046083, + 0.012483161874115467, + 0.009111247956752777, + 0.0034770420752465725, + 0.00296793133020401, + 0.016025669872760773, + 0.01430119015276432, + 0.0055853999219834805, + 0.002207500860095024, + 0.015215045772492886, + 0.011371379718184471, + 0.003112347098067403, + 0.004387181252241135, + 0.009392095729708672, + 0.00721692806109786, + 0.003625253215432167, + 0.003625112120062113, + 0.006762164644896984, + 0.006636205594986677, + 0.004030650015920401, + 0.003124626586213708, + 0.00869391206651926, + 0.008157717995345592, + 0.007214006967842579, + 0.0056081912480294704, + 0.012399066239595413, + 0.011669875122606754, + 0.006603588815778494, + 0.006214401684701443 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 1 + ], + "selected_reward": -0.01606578752398491 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An initially static, low-angle, wide-angle shot begins to tremble violently, capturing the interior of a vast, cavernous hall within a futuristic base on an alien world, designed in a stark, dark, and minimalist style. The floor, a seamless expanse of polished, high-gloss black material that reflects the entire structure like a black mirror, suddenly begins to crack at its center. A glowing spiderweb of cracks emanating from pure white light erupts across the surface, spreading at a terrifying pace and branching out in jagged, unpredictable patterns. As the violence throughout the structure intensifies, the polished floor begins to bulge and tilt. The only illumination comes from emergency lights, which flicker erratically, casting the scene in a disorienting, stroboscopic glare, creating long, dancing shadows and moments of absolute darkness, establishing an atmosphere of chaotic and utterly urgent catastrophic structural failure.", + "seed": 4453, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000156/group_11/index_5590_seed_4453_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": -0.013654270418919623, + "vlm_logprob": -0.013654270418919623, + "vlm_yes_no_margin": 4.8249999523162845, + "vlm_token_credit": -0.013654270418919623, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 vast, cavernous hall within a futuristic base on an alien world?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the interior designed in a stark, dark, and minimalist style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the floor consist of a seamless, high-gloss black material that reflects the structure like a black mirror?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do glowing white cracks spread rapidly across the floor in jagged, unpredictable patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot initially static, low-angle, and wide-angle before the scene begins to tremble violently?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03299015760421753 + ], + "opposite_logprob_tokens": [ + -3.5329902172088623 + ], + "logprob_reward": -0.03299015760421753, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.03299015760421753, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3053635358810425, + 0.07292399555444717, + 0.23135247826576233, + 0.12651513516902924, + 0.06256157159805298, + 0.0326046459376812, + 0.01927988789975643, + 0.0868559181690216, + 0.032851334661245346, + 0.029691554605960846 + ], + "credit_map": [ + 0.0324820838868618, + 0.025665946304798126, + 0.032024119049310684, + 0.05920517072081566, + 0.05085347220301628, + 0.04040617495775223, + 0.029685400426387787, + 0.03504114970564842, + 0.00795300118625164, + 0.017004162073135376, + 0.008640100248157978, + 0.00815083459019661, + 0.011565574444830418, + 0.007817966863512993, + 0.006782733369618654, + 0.005009619984775782, + 0.015780551359057426, + 0.03492534160614014, + 0.01747124269604683, + 0.019474219530820847, + 0.03619327023625374, + 0.03302791342139244, + 0.0264129601418972, + 0.048066966235637665, + 0.006659605074673891, + 0.02430807612836361, + 0.01309594139456749, + 0.036927517503499985, + 0.012615571729838848, + 0.0075815338641405106, + 0.011034035123884678, + 0.014292857609689236, + 0.007019470911473036, + 0.013870039954781532, + 0.00686322245746851, + 0.006221863441169262, + 0.0045271678827703, + 0.0060586207546293736, + 0.00725548667833209, + 0.0107456985861063, + 0.005317846778780222, + 0.004690846893936396, + 0.004013859201222658, + 0.002306412672623992, + 0.0034526700619608164, + 0.003886929014697671, + 0.005102155264467001, + 0.0038339251186698675, + 0.0031645374838262796, + 0.0022973953746259212, + 0.0022934682201594114, + 0.0017834004247561097, + 0.0017587208421900868, + 0.0014433346223086119, + 0.0029115851502865553, + 0.00362744415178895, + 0.0070106820203363895, + 0.005163165275007486, + 0.033553607761859894, + 0.0034536917228251696, + 0.0028258615639060736, + 0.0064410921186208725, + 0.0035277903079986572, + 0.024880021810531616, + 0.004902566317468882, + 0.003859421471133828, + 0.004830279387533665, + 0.002207413548603654, + 0.002754782559350133, + 0.007011539302766323, + 0.002949447138234973, + 0.004335882607847452, + 0.0037395586259663105, + 0.004541388712823391, + 0.003622169140726328, + 0.0024462479632347822, + 0.0029449407011270523, + 0.0042193131521344185, + 0.002902083797380328, + 0.005275852512568235 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 3 + ], + "selected_reward": -0.03299015760421753 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 vast, cavernous hall within a futuristic base on an alien world?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the interior designed in a stark, dark, and minimalist style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the floor consist of a seamless, high-gloss black material that reflects the structure like a black mirror?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do glowing white cracks spread rapidly across the floor in jagged, unpredictable patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot initially static, low-angle, and wide-angle before the scene begins to tremble violently?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012552893720567226 + ], + "opposite_logprob_tokens": [ + -4.387552738189697 + ], + "logprob_reward": -0.012552893720567226, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012552893720567226, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.26444876194000244, + 0.10784728825092316, + 0.15597514808177948, + 0.07596521824598312, + 0.09061595797538757, + 0.03340838477015495, + 0.030671797692775726, + 0.1484793722629547, + 0.048616137355566025, + 0.04397190734744072 + ], + "credit_map": [ + 0.02283765748143196, + 0.023095089942216873, + 0.02292659506201744, + 0.06186705827713013, + 0.030132899060845375, + 0.03236127644777298, + 0.03751435875892639, + 0.03371385484933853, + 0.008846516720950603, + 0.013345572166144848, + 0.011975347064435482, + 0.01392462570220232, + 0.015153218060731888, + 0.01058411318808794, + 0.01662941463291645, + 0.017388485372066498, + 0.01831331104040146, + 0.020283449441194534, + 0.012612470425665379, + 0.021314918994903564, + 0.020381910726428032, + 0.017778020352125168, + 0.0185210220515728, + 0.026770062744617462, + 0.007872417569160461, + 0.010100851766765118, + 0.009205225855112076, + 0.014313663356006145, + 0.011961999349296093, + 0.006050935946404934, + 0.008844875730574131, + 0.0076152547262609005, + 0.007011034060269594, + 0.01689576730132103, + 0.011417948640882969, + 0.012005995959043503, + 0.008385962806642056, + 0.007245589513331652, + 0.01183331198990345, + 0.01582036353647709, + 0.0034046447835862637, + 0.0061164554208517075, + 0.004769742954522371, + 0.0022665539290755987, + 0.0057870447635650635, + 0.0035403314977884293, + 0.004456221126019955, + 0.003067394718527794, + 0.004122960846871138, + 0.0041390350088477135, + 0.004424490965902805, + 0.004008763935416937, + 0.0033982039894908667, + 0.002893089549615979, + 0.005213840398937464, + 0.0024714162573218346, + 0.012351230718195438, + 0.009142781607806683, + 0.05913255736231804, + 0.005016872193664312, + 0.0039171078242361546, + 0.010847615078091621, + 0.004743773024529219, + 0.043327447026968, + 0.008686419576406479, + 0.006585625000298023, + 0.006579890381544828, + 0.003220929531380534, + 0.0046128323301672935, + 0.004546483512967825, + 0.0066731772385537624, + 0.0077107870019972324, + 0.0065277209505438805, + 0.004590238444507122, + 0.004569791257381439, + 0.003130052937194705, + 0.005730743054300547, + 0.0062147146090865135, + 0.004201123956590891, + 0.009007528424263 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 7 + ], + "selected_reward": -0.012552893720567226 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 vast, cavernous hall within a futuristic base on an alien world?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the interior designed in a stark, dark, and minimalist style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the floor consist of a seamless, high-gloss black material that reflects the structure like a black mirror?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do glowing white cracks spread rapidly across the floor in jagged, unpredictable patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot initially static, low-angle, and wide-angle before the scene begins to tremble violently?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00282947882078588 + ], + "opposite_logprob_tokens": [ + -5.877829551696777 + ], + "logprob_reward": -0.00282947882078588, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.00282947882078588, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18888327479362488, + 0.11112388968467712, + 0.28637057542800903, + 0.09004764258861542, + 0.09472297132015228, + 0.029771588742733, + 0.047875575721263885, + 0.057306304574012756, + 0.051863111555576324, + 0.04203503951430321 + ], + "credit_map": [ + 0.018996594473719597, + 0.016985230147838593, + 0.014666106551885605, + 0.03596552833914757, + 0.01710584945976734, + 0.026495331898331642, + 0.02807586081326008, + 0.030592771247029305, + 0.010018955916166306, + 0.024907050654292107, + 0.014418268576264381, + 0.01667238213121891, + 0.013918946497142315, + 0.011732412502169609, + 0.011175865307450294, + 0.008280008099973202, + 0.014381740242242813, + 0.04592398554086685, + 0.016293544322252274, + 0.022019650787115097, + 0.07028219103813171, + 0.038907770067453384, + 0.022651294246315956, + 0.05591040104627609, + 0.008035498671233654, + 0.014579151757061481, + 0.0082402303814888, + 0.014652595855295658, + 0.01692202500998974, + 0.008051753044128418, + 0.010900631546974182, + 0.008665758185088634, + 0.009586123749613762, + 0.015307001769542694, + 0.011737764813005924, + 0.01099069882184267, + 0.009632767178118229, + 0.006530539132654667, + 0.014379569329321384, + 0.016558509320020676, + 0.005235214717686176, + 0.004157614894211292, + 0.003547414904460311, + 0.0021552215330302715, + 0.004328426439315081, + 0.0036120221484452486, + 0.003203152446076274, + 0.00353252119384706, + 0.007338652387261391, + 0.004945789463818073, + 0.006824206560850143, + 0.007304138038307428, + 0.004020223394036293, + 0.004085793159902096, + 0.007359231356531382, + 0.005997543688863516, + 0.0057989745400846004, + 0.0037832646630704403, + 0.017115816473960876, + 0.003462711116299033, + 0.0036199791356921196, + 0.004159735050052404, + 0.006320857908576727, + 0.013044965453445911, + 0.008352010510861874, + 0.007384863682091236, + 0.006955623626708984, + 0.004050605930387974, + 0.0054100677371025085, + 0.00581114599481225, + 0.008027981966733932, + 0.005870811641216278, + 0.0068498821929097176, + 0.0072115338407456875, + 0.005236238241195679, + 0.0035272431559860706, + 0.0059889331459999084, + 0.005078192334622145, + 0.0033219207543879747, + 0.004821095149964094 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 1 + ], + "selected_reward": -0.00282947882078588 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 vast, cavernous hall within a futuristic base on an alien world?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the interior designed in a stark, dark, and minimalist style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the floor consist of a seamless, high-gloss black material that reflects the structure like a black mirror?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do glowing white cracks spread rapidly across the floor in jagged, unpredictable patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot initially static, low-angle, and wide-angle before the scene begins to tremble violently?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00171538011636585 + ], + "opposite_logprob_tokens": [ + -6.376715183258057 + ], + "logprob_reward": -0.00171538011636585, + "yes_no_margin_reward": 6.375, + "token_credit_reward": -0.00171538011636585, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1916455179452896, + 0.053507763892412186, + 0.17871518433094025, + 0.055437639355659485, + 0.04736189544200897, + 0.030601952224969864, + 0.05935604125261307, + 0.24759455025196075, + 0.08537741750478745, + 0.05040205642580986 + ], + "credit_map": [ + 0.022571906447410583, + 0.017398670315742493, + 0.0254248958081007, + 0.019253550097346306, + 0.025594107806682587, + 0.02567414380609989, + 0.01865548826754093, + 0.03707275912165642, + 0.005575237330049276, + 0.006800318136811256, + 0.008960054256021976, + 0.007059504743665457, + 0.006291619036346674, + 0.005591669585555792, + 0.007026741746813059, + 0.006202613934874535, + 0.009292359463870525, + 0.02816997654736042, + 0.011103668250143528, + 0.012968199327588081, + 0.03920125216245651, + 0.025256691500544548, + 0.016745416447520256, + 0.03597760573029518, + 0.0036812331527471542, + 0.009529265575110912, + 0.006137239746749401, + 0.010604149661958218, + 0.005498725455254316, + 0.006333880592137575, + 0.007237385958433151, + 0.006415762007236481, + 0.004415853414684534, + 0.008268001489341259, + 0.005829707719385624, + 0.005970868747681379, + 0.004128329921513796, + 0.003407613839954138, + 0.007771163713186979, + 0.00757035193964839, + 0.004822807386517525, + 0.00518400501459837, + 0.0038436143659055233, + 0.0022886425722390413, + 0.004412948153913021, + 0.003529107430949807, + 0.0034431007225066423, + 0.003077726811170578, + 0.005419505760073662, + 0.006477378774434328, + 0.008417290635406971, + 0.007204826455563307, + 0.0074751852080225945, + 0.006999783217906952, + 0.009224951267242432, + 0.008137116208672523, + 0.014350402168929577, + 0.013875390402972698, + 0.09989944845438004, + 0.007583321537822485, + 0.006592713762074709, + 0.019832298159599304, + 0.009315269067883492, + 0.0761457234621048, + 0.010839157737791538, + 0.008431349880993366, + 0.015698330476880074, + 0.005304922815412283, + 0.00984865427017212, + 0.009384813718497753, + 0.010391820222139359, + 0.01547836884856224, + 0.008062287233769894, + 0.009683369658887386, + 0.008117558434605598, + 0.004348915535956621, + 0.004634066019207239, + 0.005348771810531616, + 0.004283464513719082, + 0.005923624616116285 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 2 + ], + "selected_reward": -0.00171538011636585 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 vast, cavernous hall within a futuristic base on an alien world?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the interior designed in a stark, dark, and minimalist style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the floor consist of a seamless, high-gloss black material that reflects the structure like a black mirror?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do glowing white cracks spread rapidly across the floor in jagged, unpredictable patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot initially static, low-angle, and wide-angle before the scene begins to tremble violently?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01818344183266163 + ], + "opposite_logprob_tokens": [ + -4.01818323135376 + ], + "logprob_reward": -0.01818344183266163, + "yes_no_margin_reward": 3.999999761581421, + "token_credit_reward": -0.01818344183266163, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2753942310810089, + 0.07951987534761429, + 0.13763655722141266, + 0.06719327718019485, + 0.05994018539786339, + 0.06706000864505768, + 0.08129524439573288, + 0.13974270224571228, + 0.03681927174329758, + 0.05539870634675026 + ], + "credit_map": [ + 0.027260182425379753, + 0.02609352208673954, + 0.034645043313503265, + 0.038494404405355453, + 0.04120275750756264, + 0.02801894024014473, + 0.03425709530711174, + 0.045422282069921494, + 0.010576639324426651, + 0.014369777403771877, + 0.015946298837661743, + 0.00783502496778965, + 0.008101536892354488, + 0.010711189359426498, + 0.00613596523180604, + 0.005843444727361202, + 0.00836767815053463, + 0.019916687160730362, + 0.00948776863515377, + 0.01230723224580288, + 0.027696717530488968, + 0.016998013481497765, + 0.01587079092860222, + 0.026991678401827812, + 0.0077443961054086685, + 0.00972706452012062, + 0.008458998054265976, + 0.00905291736125946, + 0.007400857750326395, + 0.008985158987343311, + 0.007541994098573923, + 0.008281883783638477, + 0.010480419732630253, + 0.009909234009683132, + 0.007892133668065071, + 0.007253839634358883, + 0.004274850245565176, + 0.005024569109082222, + 0.0061668772250413895, + 0.008938259445130825, + 0.006722685415297747, + 0.011495867744088173, + 0.0107203284278512, + 0.004316163249313831, + 0.010493677109479904, + 0.0073190187104046345, + 0.009317944757640362, + 0.006674316246062517, + 0.016025710850954056, + 0.010188479907810688, + 0.010407709516584873, + 0.00860148947685957, + 0.007499602157622576, + 0.008478253148496151, + 0.011815721169114113, + 0.00827828049659729, + 0.012560570612549782, + 0.007287472486495972, + 0.053339917212724686, + 0.004665383603423834, + 0.004578758962452412, + 0.010320644825696945, + 0.005160016007721424, + 0.04182993993163109, + 0.0052194008603692055, + 0.004477861803025007, + 0.005510756745934486, + 0.0022499244660139084, + 0.005065246485173702, + 0.004612063989043236, + 0.0036450086627155542, + 0.0060390084981918335, + 0.010731412097811699, + 0.012113829143345356, + 0.00974681880325079, + 0.004284477327018976, + 0.0042449538595974445, + 0.004173284862190485, + 0.004642828833311796, + 0.005461100488901138 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 2 + ], + "selected_reward": -0.01818344183266163 + } + ], + "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/0000156/group_12/index_tensor([54, 54])_group_12_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000156/group_12/index_tensor([54, 54])_group_12_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..bed502048d0b629746989c5ec1e8d14d15625a7a --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000156/group_12/index_tensor([54, 54])_group_12_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "static shot from a horizontal angle, rendered in a minimalist graphic animation style, clearly illustrates the principles of sound wave propagation. Against a pure, deep, dark blue background, a shiny silver metal tuning fork stands vertically fixed to the left of the frame, its prongs vibrating vigorously at an extremely high frequency, with a subtle motion blur at the edges. As the fork vibrates, a series of parallel, visible longitudinal waves propagate steadily and rhythmically from the fork into the air to the right, smoothly crossing the entire frame. These waves vividly demonstrate the alternating process of compression and rarefaction: in compressed areas, the visual elements representing air particles are tightly packed together, forming darker, denser, luminous bands; this is followed by rarefaction areas, where the spacing between particles increases significantly, resulting in lighter, nearly transparent bands. The entire scene is illuminated by an even, soft, cool light, highlighting the dynamic structure of the waves and contrasting sharply with the dark background, creating a visual atmosphere of scientific precision and tranquil concentration.", + "seed": 17046, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000156/group_12/index_54_seed_17046_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": -0.03285998278297484, + "vlm_logprob": -0.03285998278297484, + "vlm_yes_no_margin": 4.225, + "vlm_token_credit": -0.03285998278297484, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot as a static shot from a horizontal angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a pure, deep, dark blue color?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a shiny silver metal tuning fork visible standing vertically fixed to the left of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are parallel longitudinal waves propagating from the fork to the right, showing alternating compressed (dark/dense) and rarefied (light/transparent) bands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by an even, soft, cool 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": [ + -0.07779894769191742 + ], + "opposite_logprob_tokens": [ + -2.952798843383789 + ], + "logprob_reward": -0.07779894769191742, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.07779894769191742, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.415752112865448, + 0.1824478805065155, + 0.12862002849578857, + 0.06641341745853424, + 0.0426986888051033, + 0.03835226222872734, + 0.050287339836359024, + 0.023312484845519066, + 0.025153420865535736, + 0.02696233056485653 + ], + "credit_map": [ + 0.04737824574112892, + 0.08107282966375351, + 0.07692915946245193, + 0.05331182852387428, + 0.04319889843463898, + 0.020092880353331566, + 0.05849163234233856, + 0.035276684910058975, + 0.01566927880048752, + 0.04134621098637581, + 0.03365354239940643, + 0.018560100346803665, + 0.01771637797355652, + 0.013081589713692665, + 0.024553706869482994, + 0.017867086455225945, + 0.012034662067890167, + 0.010765179991722107, + 0.01776375062763691, + 0.012350264936685562, + 0.015100364573299885, + 0.017264483496546745, + 0.016663026064634323, + 0.026678310707211494, + 0.006657349411398172, + 0.00679225055500865, + 0.008636563085019588, + 0.0085811298340559, + 0.00997453834861517, + 0.0048825908452272415, + 0.011370268650352955, + 0.009518736973404884, + 0.004770717583596706, + 0.010212856344878674, + 0.006339093204587698, + 0.007414039224386215, + 0.003770116250962019, + 0.0028848694637417793, + 0.0033294649329036474, + 0.003977533895522356, + 0.006316129583865404, + 0.004710968118160963, + 0.007207256741821766, + 0.004541605710983276, + 0.0030282691586762667, + 0.0036396803334355354, + 0.003516558324918151, + 0.0053918007761240005, + 0.006431219633668661, + 0.015367217361927032, + 0.0073142582550644875, + 0.008584225550293922, + 0.0031728961039334536, + 0.002579001011326909, + 0.0026524581480771303, + 0.004186069592833519, + 0.0026453877799212933, + 0.004062168765813112, + 0.003296739887446165, + 0.004356207326054573, + 0.0015307283028960228, + 0.0017207732889801264, + 0.0021181690972298384, + 0.003582312725484371, + 0.003981929738074541, + 0.004434692673385143, + 0.0033715213648974895, + 0.002702686470001936, + 0.0025302551221102476, + 0.0021698034834116697, + 0.003092658007517457, + 0.0028698763344436884, + 0.0033276593312621117, + 0.005063025280833244, + 0.003661727299913764, + 0.006142000202089548, + 0.0021269454155117273, + 0.0018591650296002626, + 0.0024863386061042547, + 0.002295473124831915 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.07779894769191742 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot as a static shot from a horizontal angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a pure, deep, dark blue color?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a shiny silver metal tuning fork visible standing vertically fixed to the left of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are parallel longitudinal waves propagating from the fork to the right, showing alternating compressed (dark/dense) and rarefied (light/transparent) bands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by an even, soft, cool 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": [ + -0.043089091777801514 + ], + "opposite_logprob_tokens": [ + -3.1680891513824463 + ], + "logprob_reward": -0.043089091777801514, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.043089091777801514, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.39344751834869385, + 0.18291036784648895, + 0.068759985268116, + 0.04266524314880371, + 0.04888548329472542, + 0.024873441085219383, + 0.06072146072983742, + 0.0789765864610672, + 0.06747076660394669, + 0.03128911554813385 + ], + "credit_map": [ + 0.043210361152887344, + 0.03176223486661911, + 0.1307380199432373, + 0.020174730569124222, + 0.08361048251390457, + 0.02207682654261589, + 0.03835359215736389, + 0.023521313443779945, + 0.02052777074277401, + 0.029058966785669327, + 0.03935715928673744, + 0.01839713752269745, + 0.0182954054325819, + 0.013059301301836967, + 0.02309313230216503, + 0.021121492609381676, + 0.007832766510546207, + 0.009468610398471355, + 0.008982202969491482, + 0.010930508375167847, + 0.008363889530301094, + 0.005569393280893564, + 0.008069390431046486, + 0.009543219581246376, + 0.004825799260288477, + 0.004685805644840002, + 0.0054173399694263935, + 0.005870339926332235, + 0.005744625348597765, + 0.004229533951729536, + 0.006413770839571953, + 0.0054780263453722, + 0.0066383215598762035, + 0.004779784474521875, + 0.008283457718789577, + 0.005605343729257584, + 0.0045602163299918175, + 0.005763757973909378, + 0.006040447391569614, + 0.007214154116809368, + 0.003671115729957819, + 0.002577928127720952, + 0.0038176393136382103, + 0.002956409240141511, + 0.0023402138613164425, + 0.0030543762259185314, + 0.0028144423849880695, + 0.003641316667199135, + 0.007227154914289713, + 0.009327583014965057, + 0.009927526116371155, + 0.0066916439682245255, + 0.005435540806502104, + 0.005228366237133741, + 0.0052956179715693, + 0.011588028632104397, + 0.012210667133331299, + 0.017345959320664406, + 0.014985341578722, + 0.014524904079735279, + 0.004124780651181936, + 0.004323718603700399, + 0.005054250825196505, + 0.006406963802874088, + 0.009548460133373737, + 0.008758076466619968, + 0.01073785126209259, + 0.010600056499242783, + 0.0058745029382407665, + 0.005430680233985186, + 0.008213757537305355, + 0.008307388052344322, + 0.0035464002285152674, + 0.006045483518391848, + 0.0046926396898925304, + 0.0072136446833610535, + 0.002243863884359598, + 0.001776232966221869, + 0.003122918540611863, + 0.0026479342486709356 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 7 + ], + "selected_reward": -0.043089091777801514 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot as a static shot from a horizontal angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a pure, deep, dark blue color?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a shiny silver metal tuning fork visible standing vertically fixed to the left of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are parallel longitudinal waves propagating from the fork to the right, showing alternating compressed (dark/dense) and rarefied (light/transparent) bands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by an even, soft, cool 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": [ + -0.0024947968777269125 + ], + "opposite_logprob_tokens": [ + -6.002494812011719 + ], + "logprob_reward": -0.0024947968777269125, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.0024947968777269125, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2102232277393341, + 0.25786906480789185, + 0.034772638231515884, + 0.05559442192316055, + 0.07069918513298035, + 0.06299351155757904, + 0.07199510931968689, + 0.10092896968126297, + 0.07444228231906891, + 0.06048159673810005 + ], + "credit_map": [ + 0.01810762844979763, + 0.04648157209157944, + 0.041504956781864166, + 0.013579330407083035, + 0.04860541969537735, + 0.009457102976739407, + 0.021874649450182915, + 0.01061257440596819, + 0.025125036016106606, + 0.06009065732359886, + 0.04521992430090904, + 0.01748124323785305, + 0.023523228242993355, + 0.016562649980187416, + 0.03847035393118858, + 0.031395960599184036, + 0.003909564111381769, + 0.0037128685507923365, + 0.00596516951918602, + 0.006257364526391029, + 0.00417060824111104, + 0.002461150521412492, + 0.004922797437757254, + 0.0033731143921613693, + 0.0053214821964502335, + 0.0068426537327468395, + 0.008028456009924412, + 0.011085445061326027, + 0.004694312810897827, + 0.005131883546710014, + 0.00661109434440732, + 0.007879096083343029, + 0.009607406333088875, + 0.010265493765473366, + 0.011803347617387772, + 0.013292282819747925, + 0.0048651499673724174, + 0.005381873808801174, + 0.008201786316931248, + 0.007281846832484007, + 0.009662722237408161, + 0.014487260021269321, + 0.009435603395104408, + 0.010959260165691376, + 0.00414512911811471, + 0.003657330758869648, + 0.005207659211009741, + 0.005438551772385836, + 0.008645519614219666, + 0.018261265009641647, + 0.01080267783254385, + 0.013614603318274021, + 0.004823030438274145, + 0.004219172988086939, + 0.005040447693318129, + 0.00658839475363493, + 0.013934667222201824, + 0.025253240019083023, + 0.018153833225369453, + 0.02321023680269718, + 0.00431440444663167, + 0.004309200216084719, + 0.0060594710521399975, + 0.0056939204223454, + 0.009616417810320854, + 0.018505405634641647, + 0.012256341055035591, + 0.013833141885697842, + 0.004919990431517363, + 0.0036829407326877117, + 0.006614268757402897, + 0.0050137764774262905, + 0.010108058340847492, + 0.01283307559788227, + 0.009367571212351322, + 0.010864201933145523, + 0.0044972896575927734, + 0.0027383891865611076, + 0.005275125149637461, + 0.00479788426309824 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 7 + ], + "selected_reward": -0.0024947968777269125 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot as a static shot from a horizontal angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a pure, deep, dark blue color?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a shiny silver metal tuning fork visible standing vertically fixed to the left of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are parallel longitudinal waves propagating from the fork to the right, showing alternating compressed (dark/dense) and rarefied (light/transparent) bands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by an even, soft, cool 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": [ + -0.0028177103959023952 + ], + "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": [ + 0.2983165383338928, + 0.18933065235614777, + 0.034808337688446045, + 0.10060425847768784, + 0.05479390174150467, + 0.06186885014176369, + 0.06751388311386108, + 0.037129320204257965, + 0.08922143280506134, + 0.0664127916097641 + ], + "credit_map": [ + 0.032877206802368164, + 0.02729013003408909, + 0.0799851268529892, + 0.03215768188238144, + 0.02796384133398533, + 0.013676192611455917, + 0.060212936252355576, + 0.024153415113687515, + 0.01750124990940094, + 0.038650039583444595, + 0.03227097541093826, + 0.018150510266423225, + 0.017564600333571434, + 0.01761223003268242, + 0.018883783370256424, + 0.028697263449430466, + 0.003931637853384018, + 0.003590776352211833, + 0.0061319321393966675, + 0.0045882235281169415, + 0.003936333116143942, + 0.003222404746338725, + 0.004304757807403803, + 0.005102272145450115, + 0.010175628587603569, + 0.011437533423304558, + 0.014152676798403263, + 0.012136348523199558, + 0.011014887131750584, + 0.013443771749734879, + 0.013247267343103886, + 0.014996148645877838, + 0.0054980055429041386, + 0.004918646067380905, + 0.006380168255418539, + 0.006167203653603792, + 0.005501160863786936, + 0.007118373177945614, + 0.00803191214799881, + 0.011178430169820786, + 0.010788941755890846, + 0.005647754762321711, + 0.013293273746967316, + 0.007910545915365219, + 0.004536905791610479, + 0.0074250842444598675, + 0.004931547679007053, + 0.007334793917834759, + 0.008787330240011215, + 0.01753026619553566, + 0.009685354307293892, + 0.011211707256734371, + 0.00448652496561408, + 0.003661951282992959, + 0.0052480073645710945, + 0.006902739871293306, + 0.004735893104225397, + 0.006536063738167286, + 0.006185217760503292, + 0.008263738825917244, + 0.002172676846385002, + 0.0022229005116969347, + 0.0030002121347934008, + 0.004012617282569408, + 0.010373464785516262, + 0.0179930180311203, + 0.012699952349066734, + 0.011395684443414211, + 0.008767493069171906, + 0.008093372918665409, + 0.00937803927809, + 0.010520407930016518, + 0.00853065773844719, + 0.00730027724057436, + 0.009639309719204903, + 0.008988723158836365, + 0.005549142137169838, + 0.006531799212098122, + 0.008238296955823898, + 0.011634587310254574 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 3 + ], + "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 video shot as a static shot from a horizontal angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a pure, deep, dark blue color?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a shiny silver metal tuning fork visible standing vertically fixed to the left of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are parallel longitudinal waves propagating from the fork to the right, showing alternating compressed (dark/dense) and rarefied (light/transparent) bands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by an even, soft, cool 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": [ + -0.038099367171525955 + ], + "opposite_logprob_tokens": [ + -3.2880992889404297 + ], + "logprob_reward": -0.038099367171525955, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.038099367171525955, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4615787863731384, + 0.21938368678092957, + 0.06962200254201889, + 0.053046438843011856, + 0.043927356600761414, + 0.024524414911866188, + 0.033115312457084656, + 0.028338171541690826, + 0.041112031787633896, + 0.025351757183670998 + ], + "credit_map": [ + 0.05115056037902832, + 0.03582936152815819, + 0.15701386332511902, + 0.028986526653170586, + 0.0719277560710907, + 0.022207343950867653, + 0.06304916739463806, + 0.03141423687338829, + 0.02008412405848503, + 0.04409504309296608, + 0.0417594350874424, + 0.01981426030397415, + 0.018634362146258354, + 0.016509337350726128, + 0.028292469680309296, + 0.030194664373993874, + 0.007184906397014856, + 0.005927531514316797, + 0.010745777748525143, + 0.011659651063382626, + 0.00945668388158083, + 0.005749659147113562, + 0.008548897691071033, + 0.010348904877901077, + 0.005367846693843603, + 0.006984012667089701, + 0.006391754373908043, + 0.01135396771132946, + 0.004448700230568647, + 0.004881232511252165, + 0.006085657048970461, + 0.007533275987952948, + 0.004386963322758675, + 0.0058893789537250996, + 0.004968023393303156, + 0.006478814873844385, + 0.0045377230271697044, + 0.005902716889977455, + 0.005026654340326786, + 0.0067370873875916, + 0.003666533390060067, + 0.003655083477497101, + 0.004191522020846605, + 0.0038711517117917538, + 0.0019518989138305187, + 0.002076175296679139, + 0.002321513369679451, + 0.0027905390597879887, + 0.004402044229209423, + 0.0036053319927304983, + 0.005282146390527487, + 0.004876243881881237, + 0.0027446707244962454, + 0.003207406960427761, + 0.003734676167368889, + 0.005262797698378563, + 0.003209116170182824, + 0.00616688234731555, + 0.004655485972762108, + 0.006426338572055101, + 0.001716621220111847, + 0.0016302544390782714, + 0.002256283536553383, + 0.0022771931253373623, + 0.005284969229251146, + 0.007157727610319853, + 0.006923067849129438, + 0.007479940075427294, + 0.0030174751300364733, + 0.0024263435043394566, + 0.005190211348235607, + 0.0036323033273220062, + 0.0035950366873294115, + 0.004058752674609423, + 0.004031960852444172, + 0.005819092039018869, + 0.0016812897520139813, + 0.0017845305847004056, + 0.002579201478511095, + 0.0018018968403339386 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.038099367171525955 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static shot from a horizontal angle, rendered in a minimalist graphic animation style, clearly illustrates the principles of sound wave propagation. Against a pure, deep, dark blue background, a shiny silver metal tuning fork stands vertically fixed to the left of the frame, its prongs vibrating vigorously at an extremely high frequency, with a subtle motion blur at the edges. As the fork vibrates, a series of parallel, visible longitudinal waves propagate steadily and rhythmically from the fork into the air to the right, smoothly crossing the entire frame. These waves vividly demonstrate the alternating process of compression and rarefaction: in compressed areas, the visual elements representing air particles are tightly packed together, forming darker, denser, luminous bands; this is followed by rarefaction areas, where the spacing between particles increases significantly, resulting in lighter, nearly transparent bands. The entire scene is illuminated by an even, soft, cool light, highlighting the dynamic structure of the waves and contrasting sharply with the dark background, creating a visual atmosphere of scientific precision and tranquil concentration.", + "seed": 17047, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000156/group_12/index_54_seed_17047_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": -0.06334674526005983, + "vlm_logprob": -0.06334674526005983, + "vlm_yes_no_margin": 3.45, + "vlm_token_credit": -0.06334674526005983, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot as a static shot from a horizontal angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a pure, deep, dark blue color?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a shiny silver metal tuning fork visible standing vertically fixed to the left of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are parallel longitudinal waves propagating from the fork to the right, showing alternating compressed (dark/dense) and rarefied (light/transparent) bands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by an even, soft, cool 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": [ + -0.06726500391960144 + ], + "opposite_logprob_tokens": [ + -2.942265033721924 + ], + "logprob_reward": -0.06726500391960144, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.06726500391960144, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3048083782196045, + 0.11532653123140335, + 0.07087975740432739, + 0.04150497168302536, + 0.14395731687545776, + 0.029075700789690018, + 0.16268189251422882, + 0.056422751396894455, + 0.05186896398663521, + 0.023473771288990974 + ], + "credit_map": [ + 0.03795119374990463, + 0.03926461189985275, + 0.028060652315616608, + 0.02608574740588665, + 0.05567594990134239, + 0.04596894979476929, + 0.026670627295970917, + 0.045130595564842224, + 0.016990574076771736, + 0.01806776039302349, + 0.012730264104902744, + 0.011680269613862038, + 0.01432422548532486, + 0.016182074323296547, + 0.0116531727835536, + 0.013698192313313484, + 0.010357984341681004, + 0.008898417465388775, + 0.011597890406847, + 0.009174994193017483, + 0.0070084682665765285, + 0.0055742692202329636, + 0.008219528943300247, + 0.01004820130765438, + 0.006830520462244749, + 0.005133559927344322, + 0.007863285951316357, + 0.005331683438271284, + 0.004181141033768654, + 0.0031342683359980583, + 0.004788338206708431, + 0.0042421757243573666, + 0.03343931585550308, + 0.013793864287436008, + 0.045000556856393814, + 0.01402620505541563, + 0.01031951792538166, + 0.006769770756363869, + 0.007682520896196365, + 0.012925571762025356, + 0.0036404321435838938, + 0.004989697132259607, + 0.004285812843590975, + 0.004214264918118715, + 0.003541697282344103, + 0.002687884261831641, + 0.002420463366433978, + 0.00329545047134161, + 0.017387373372912407, + 0.01498444750905037, + 0.017090028151869774, + 0.012411550618708134, + 0.030871350318193436, + 0.04357967525720596, + 0.012415409088134766, + 0.01394206378608942, + 0.0109034338966012, + 0.007960494607686996, + 0.007847631350159645, + 0.004039784427732229, + 0.006009284406900406, + 0.008642601780593395, + 0.005220966413617134, + 0.005798551253974438, + 0.008012519218027592, + 0.005816436838358641, + 0.006640098989009857, + 0.005014474969357252, + 0.006005040369927883, + 0.00756071787327528, + 0.005260295234620571, + 0.007559381425380707, + 0.002832135884091258, + 0.0030619243625551462, + 0.0026948023587465286, + 0.002420483622699976, + 0.0028861030004918575, + 0.002692056354135275, + 0.003884589532390237, + 0.0030016780365258455 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 4 + ], + "selected_reward": -0.06726500391960144 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot as a static shot from a horizontal angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a pure, deep, dark blue color?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a shiny silver metal tuning fork visible standing vertically fixed to the left of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are parallel longitudinal waves propagating from the fork to the right, showing alternating compressed (dark/dense) and rarefied (light/transparent) bands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by an even, soft, cool 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": [ + -0.17991843819618225 + ], + "opposite_logprob_tokens": [ + -1.8049184083938599 + ], + "logprob_reward": -0.17991843819618225, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.17991843819618225, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22956758737564087, + 0.15234458446502686, + 0.08956652134656906, + 0.03171743452548981, + 0.05785973370075226, + 0.03177155554294586, + 0.21986208856105804, + 0.13423968851566315, + 0.02378265932202339, + 0.02928815595805645 + ], + "credit_map": [ + 0.016374735161662102, + 0.015711987391114235, + 0.017784424126148224, + 0.010741489008069038, + 0.05018992722034454, + 0.06739895790815353, + 0.024080540984869003, + 0.02728552743792534, + 0.031182972714304924, + 0.022877320647239685, + 0.018284019082784653, + 0.013673175126314163, + 0.019231390208005905, + 0.018127504736185074, + 0.01212441548705101, + 0.01684379205107689, + 0.020793529227375984, + 0.006768499035388231, + 0.03024592623114586, + 0.00803214218467474, + 0.00654754089191556, + 0.004155059810727835, + 0.00423715403303504, + 0.008786674588918686, + 0.006193589419126511, + 0.004326808266341686, + 0.004758787341415882, + 0.0038846246898174286, + 0.0036390500608831644, + 0.0028770139906555414, + 0.0031113920267671347, + 0.0029261671006679535, + 0.01236447598785162, + 0.004618612118065357, + 0.01677747629582882, + 0.004632546100765467, + 0.004443817306309938, + 0.004162574652582407, + 0.003513507079333067, + 0.007346721366047859, + 0.005842136684805155, + 0.003427676623687148, + 0.008462049998342991, + 0.0036803872790187597, + 0.002882320899516344, + 0.001990225166082382, + 0.0027870936319231987, + 0.0026996659580618143, + 0.013694755733013153, + 0.016066113486886024, + 0.016853293403983116, + 0.010019050911068916, + 0.055181510746479034, + 0.06960955262184143, + 0.016146723181009293, + 0.022291086614131927, + 0.012847931124269962, + 0.014488003216683865, + 0.012118314392864704, + 0.008019454777240753, + 0.026720397174358368, + 0.03003736585378647, + 0.017045922577381134, + 0.012962291948497295, + 0.003873279085382819, + 0.00243490026332438, + 0.0035373372957110405, + 0.002667150227352977, + 0.002006564987823367, + 0.0027559755835682154, + 0.0028876811265945435, + 0.003619769122451544, + 0.004349512048065662, + 0.003407182637602091, + 0.003186566289514303, + 0.0018944806652143598, + 0.004773999564349651, + 0.0053742025047540665, + 0.00254383054561913, + 0.003758380888029933 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 1 + ], + "selected_reward": -0.17991843819618225 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot as a static shot from a horizontal angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a pure, deep, dark blue color?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a shiny silver metal tuning fork visible standing vertically fixed to the left of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are parallel longitudinal waves propagating from the fork to the right, showing alternating compressed (dark/dense) and rarefied (light/transparent) bands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by an even, soft, cool 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": [ + -0.00595355499535799 + ], + "opposite_logprob_tokens": [ + -5.130953788757324 + ], + "logprob_reward": -0.00595355499535799, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.00595355499535799, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20813173055648804, + 0.14552432298660278, + 0.16365565359592438, + 0.061435721814632416, + 0.1925494521856308, + 0.06760808825492859, + 0.07242269814014435, + 0.036994416266679764, + 0.029023276641964912, + 0.022654687985777855 + ], + "credit_map": [ + 0.016109930351376534, + 0.02178056538105011, + 0.015553007833659649, + 0.01342158205807209, + 0.04436443746089935, + 0.05826385319232941, + 0.01462478470057249, + 0.02401353046298027, + 0.028642384335398674, + 0.03081858716905117, + 0.01895356923341751, + 0.014127901755273342, + 0.009572290815412998, + 0.01075538620352745, + 0.011162320151925087, + 0.02149186283349991, + 0.03788838908076286, + 0.01169088389724493, + 0.05560975521802902, + 0.013366944156587124, + 0.007072218228131533, + 0.007112153340131044, + 0.009015803225338459, + 0.02189948409795761, + 0.011631009168922901, + 0.008183939382433891, + 0.01124425046145916, + 0.008676216937601566, + 0.005674956366419792, + 0.004616422578692436, + 0.005398990586400032, + 0.006009925622493029, + 0.04285866394639015, + 0.023862402886152267, + 0.05822814628481865, + 0.024995921179652214, + 0.010639813728630543, + 0.007116401568055153, + 0.010692973621189594, + 0.014155080541968346, + 0.011007251217961311, + 0.01083195861428976, + 0.015656383708119392, + 0.010672708041965961, + 0.004351915791630745, + 0.00350326974876225, + 0.005250067450106144, + 0.006334524601697922, + 0.013459743931889534, + 0.011261452920734882, + 0.008672849275171757, + 0.005960863549262285, + 0.009104978293180466, + 0.012080489657819271, + 0.005641316995024681, + 0.006240993272513151, + 0.005092978943139315, + 0.005636161658912897, + 0.004970584064722061, + 0.003253913251683116, + 0.004823074210435152, + 0.0047476887702941895, + 0.0036089930217713118, + 0.00486101396381855, + 0.0033475568052381277, + 0.003001208184286952, + 0.004552221857011318, + 0.0025440144818276167, + 0.0031085072550922632, + 0.004166949540376663, + 0.0031736113596707582, + 0.005129202269017696, + 0.0026261035818606615, + 0.003083119634538889, + 0.002782118273898959, + 0.0022749858908355236, + 0.003115706378594041, + 0.0035180721897631884, + 0.0031597339548170567, + 0.0020948434248566628 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 2 + ], + "selected_reward": -0.00595355499535799 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot as a static shot from a horizontal angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a pure, deep, dark blue color?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a shiny silver metal tuning fork visible standing vertically fixed to the left of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are parallel longitudinal waves propagating from the fork to the right, showing alternating compressed (dark/dense) and rarefied (light/transparent) bands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by an even, soft, cool 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": [ + -0.008637804538011551 + ], + "opposite_logprob_tokens": [ + -4.75863790512085 + ], + "logprob_reward": -0.008637804538011551, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008637804538011551, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3034193515777588, + 0.061672382056713104, + 0.172730952501297, + 0.011100055649876595, + 0.053185850381851196, + 0.022577419877052307, + 0.24604371190071106, + 0.07082948088645935, + 0.03469010069966316, + 0.023750660941004753 + ], + "credit_map": [ + 0.018950406461954117, + 0.026942389085888863, + 0.022711748257279396, + 0.019288096576929092, + 0.07044986635446548, + 0.08482451736927032, + 0.026299942284822464, + 0.03395237773656845, + 0.008197114802896976, + 0.009160029701888561, + 0.006467071827501059, + 0.005521500017493963, + 0.007683376781642437, + 0.010935051366686821, + 0.006937168072909117, + 0.006771070417016745, + 0.039536282420158386, + 0.012316732667386532, + 0.05854346975684166, + 0.012675861828029156, + 0.00906637404114008, + 0.008628908544778824, + 0.009473993442952633, + 0.022489318624138832, + 0.0012069132644683123, + 0.0014095526421442628, + 0.0015862835571169853, + 0.0017400099895894527, + 0.0012896176194772124, + 0.0008951768395490944, + 0.0018810795154422522, + 0.0010914222802966833, + 0.011383223347365856, + 0.004518707282841206, + 0.01589961349964142, + 0.004998360760509968, + 0.0033980899024754763, + 0.002509597921743989, + 0.0033565617632120848, + 0.007121698930859566, + 0.003781407605856657, + 0.0033752124290913343, + 0.004601417575031519, + 0.003626951016485691, + 0.0019842544570565224, + 0.00131466262973845, + 0.0019795382395386696, + 0.0019139773212373257, + 0.010900105349719524, + 0.016512589529156685, + 0.018085023388266563, + 0.011070521548390388, + 0.06327066570520401, + 0.0807015523314476, + 0.02035214751958847, + 0.02515111118555069, + 0.005801209714263678, + 0.004851861856877804, + 0.006117352284491062, + 0.003627676283940673, + 0.018693149089813232, + 0.017476197332143784, + 0.0064908829517662525, + 0.007771149277687073, + 0.0058102658949792385, + 0.0035323642659932375, + 0.004458766896277666, + 0.003175522433593869, + 0.0053101591765880585, + 0.005478082224726677, + 0.0029858523048460484, + 0.003939086105674505, + 0.003265130566433072, + 0.0022885496728122234, + 0.0025994505267590284, + 0.002171851461753249, + 0.0032170754857361317, + 0.0037128222174942493, + 0.0034554137382656336, + 0.0030403672717511654 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 2 + ], + "selected_reward": -0.008637804538011551 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot as a static shot from a horizontal angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a pure, deep, dark blue color?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a shiny silver metal tuning fork visible standing vertically fixed to the left of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are parallel longitudinal waves propagating from the fork to the right, showing alternating compressed (dark/dense) and rarefied (light/transparent) bands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by an even, soft, cool 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": [ + -0.054958924651145935 + ], + "opposite_logprob_tokens": [ + -2.9299588203430176 + ], + "logprob_reward": -0.054958924651145935, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.054958924651145935, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2631622552871704, + 0.14638185501098633, + 0.13981521129608154, + 0.03655976057052612, + 0.17436310648918152, + 0.019213180989027023, + 0.11138934642076492, + 0.04272545129060745, + 0.03848003223538399, + 0.027909735217690468 + ], + "credit_map": [ + 0.047075629234313965, + 0.03966231271624565, + 0.032111987471580505, + 0.021703192964196205, + 0.020306702703237534, + 0.027398819103837013, + 0.022664541378617287, + 0.052239082753658295, + 0.03198813274502754, + 0.024689586833119392, + 0.017640693113207817, + 0.015578286722302437, + 0.012998906895518303, + 0.01279249507933855, + 0.01149240043014288, + 0.019201353192329407, + 0.031315360218286514, + 0.011570156551897526, + 0.045723337680101395, + 0.01396715734153986, + 0.009338573552668095, + 0.006827627308666706, + 0.009178517386317253, + 0.011894481256604195, + 0.007680689916014671, + 0.00559539720416069, + 0.007757928688079119, + 0.004084812477231026, + 0.0029558681417256594, + 0.0018379203975200653, + 0.003029651939868927, + 0.0036174943670630455, + 0.041512925177812576, + 0.013976825401186943, + 0.059671297669410706, + 0.01659756526350975, + 0.00967229250818491, + 0.006811064667999744, + 0.00821723137050867, + 0.01790390908718109, + 0.002736289519816637, + 0.003195737488567829, + 0.002791843144223094, + 0.003146787639707327, + 0.001633002539165318, + 0.0012503848411142826, + 0.0023897483479231596, + 0.0020693871192634106, + 0.008425286039710045, + 0.00897219218313694, + 0.009583435952663422, + 0.0070835379883646965, + 0.026168404147028923, + 0.032965533435344696, + 0.008773697540163994, + 0.009417264722287655, + 0.0073251365683972836, + 0.006305253133177757, + 0.0052628726698458195, + 0.00348541303537786, + 0.005193597637116909, + 0.005380888469517231, + 0.005123501643538475, + 0.00464879022911191, + 0.004591795150190592, + 0.003469154704362154, + 0.005233078729361296, + 0.0028226959984749556, + 0.007895825430750847, + 0.006280272267758846, + 0.004147866740822792, + 0.004039343912154436, + 0.005944555159658194, + 0.004195277579128742, + 0.003617189358919859, + 0.002656357130035758, + 0.0026257033459842205, + 0.0035031645093113184, + 0.0024505394976586103, + 0.002916949335485697 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 1 + ], + "selected_reward": -0.054958924651145935 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static shot from a horizontal angle, rendered in a minimalist graphic animation style, clearly illustrates the principles of sound wave propagation. Against a pure, deep, dark blue background, a shiny silver metal tuning fork stands vertically fixed to the left of the frame, its prongs vibrating vigorously at an extremely high frequency, with a subtle motion blur at the edges. As the fork vibrates, a series of parallel, visible longitudinal waves propagate steadily and rhythmically from the fork into the air to the right, smoothly crossing the entire frame. These waves vividly demonstrate the alternating process of compression and rarefaction: in compressed areas, the visual elements representing air particles are tightly packed together, forming darker, denser, luminous bands; this is followed by rarefaction areas, where the spacing between particles increases significantly, resulting in lighter, nearly transparent bands. The entire scene is illuminated by an even, soft, cool light, highlighting the dynamic structure of the waves and contrasting sharply with the dark background, creating a visual atmosphere of scientific precision and tranquil concentration.", + "seed": 17048, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000156/group_12/index_54_seed_17048_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": -0.04107062509283423, + "vlm_logprob": -0.04107062509283423, + "vlm_yes_no_margin": 3.825, + "vlm_token_credit": -0.04107062509283423, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot as a static shot from a horizontal angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a pure, deep, dark blue color?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a shiny silver metal tuning fork visible standing vertically fixed to the left of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are parallel longitudinal waves propagating from the fork to the right, showing alternating compressed (dark/dense) and rarefied (light/transparent) bands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by an even, soft, cool 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": [ + -0.02283433824777603 + ], + "opposite_logprob_tokens": [ + -4.272834300994873 + ], + "logprob_reward": -0.02283433824777603, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.02283433824777603, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2356805056333542, + 0.042462561279535294, + 0.032507434487342834, + 0.05358930677175522, + 0.17361986637115479, + 0.34181615710258484, + 0.024592235684394836, + 0.029738696292042732, + 0.043952710926532745, + 0.02204042673110962 + ], + "credit_map": [ + 0.03546396270394325, + 0.014751533977687359, + 0.0711030513048172, + 0.022150998935103416, + 0.019186077639460564, + 0.020031273365020752, + 0.018037447705864906, + 0.0349561832845211, + 0.005808237940073013, + 0.003248116234317422, + 0.006035349331796169, + 0.0057216971181333065, + 0.005020992364734411, + 0.002948162378743291, + 0.008898475207388401, + 0.00478153582662344, + 0.006627069786190987, + 0.002128571504727006, + 0.008507641032338142, + 0.00269017880782485, + 0.002116748131811619, + 0.002426150953397155, + 0.0023036522325128317, + 0.0057074264623224735, + 0.009807687252759933, + 0.007376219145953655, + 0.010228527709841728, + 0.005332092754542828, + 0.004657618701457977, + 0.004868250340223312, + 0.006689125206321478, + 0.004629792179912329, + 0.022455180063843727, + 0.04050041735172272, + 0.049093373119831085, + 0.013687116093933582, + 0.011594187468290329, + 0.00808740220963955, + 0.010737305507063866, + 0.01746489480137825, + 0.051492203027009964, + 0.01859300024807453, + 0.11273546516895294, + 0.02674717642366886, + 0.02053377404808998, + 0.019986312836408615, + 0.01988825388252735, + 0.07184001803398132, + 0.004149996675550938, + 0.0016792418900877237, + 0.00628822622820735, + 0.0018147081136703491, + 0.001967112999409437, + 0.0016077805776149035, + 0.0018097686115652323, + 0.005275403615087271, + 0.003370781661942601, + 0.0021509758662432432, + 0.006569506134837866, + 0.004545397590845823, + 0.0018074893159791827, + 0.00190652662422508, + 0.003959468100219965, + 0.005428552161902189, + 0.006619453430175781, + 0.0052924491465091705, + 0.005141987930983305, + 0.007149622309952974, + 0.00521365599706769, + 0.002556704916059971, + 0.004771021660417318, + 0.007207820657640696, + 0.003121771151199937, + 0.0013100059004500508, + 0.00521737290546298, + 0.0024538154248148203, + 0.0013139774091541767, + 0.0013860751641914248, + 0.003550253575667739, + 0.0036871579941362143 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 4 + ], + "selected_reward": -0.02283433824777603 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot as a static shot from a horizontal angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a pure, deep, dark blue color?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a shiny silver metal tuning fork visible standing vertically fixed to the left of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are parallel longitudinal waves propagating from the fork to the right, showing alternating compressed (dark/dense) and rarefied (light/transparent) bands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by an even, soft, cool 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": [ + -0.11296695470809937 + ], + "opposite_logprob_tokens": [ + -2.237967014312744 + ], + "logprob_reward": -0.11296695470809937, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11296695470809937, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06490996479988098, + 0.08187408745288849, + 0.03677072003483772, + 0.0916246771812439, + 0.053166713565588, + 0.4076191782951355, + 0.13266152143478394, + 0.0930001363158226, + 0.02575882337987423, + 0.012614178471267223 + ], + "credit_map": [ + 0.010426094755530357, + 0.005154000595211983, + 0.014732587151229382, + 0.009742951951920986, + 0.004688777960836887, + 0.005345175042748451, + 0.007886577397584915, + 0.0069337948225438595, + 0.012426014058291912, + 0.00457765581086278, + 0.028066640719771385, + 0.007167824078351259, + 0.00544660072773695, + 0.005144736263900995, + 0.004093181807547808, + 0.014951434917747974, + 0.005157542414963245, + 0.003320736112073064, + 0.008738650940358639, + 0.0035199562553316355, + 0.003030568826943636, + 0.004700030665844679, + 0.003752916119992733, + 0.004550317768007517, + 0.014842323027551174, + 0.007604706566780806, + 0.029904630035161972, + 0.009427704848349094, + 0.005108434706926346, + 0.005955546163022518, + 0.005862564779818058, + 0.01291877031326294, + 0.0063886563293635845, + 0.009693785570561886, + 0.005775414872914553, + 0.005821388214826584, + 0.005081081762909889, + 0.0035086877178400755, + 0.005588288884609938, + 0.011309410445392132, + 0.06363003700971603, + 0.020746584981679916, + 0.14060437679290771, + 0.03392103686928749, + 0.022247206419706345, + 0.0213632769882679, + 0.020536351948976517, + 0.08457029610872269, + 0.021154362708330154, + 0.005773705895990133, + 0.04342997446656227, + 0.009042828343808651, + 0.006156289484351873, + 0.007357920054346323, + 0.007410969585180283, + 0.03233547881245613, + 0.014111113734543324, + 0.004818725865334272, + 0.02821086347103119, + 0.007242645137012005, + 0.004348693881183863, + 0.007276635151356459, + 0.0058143786154687405, + 0.021177073940634727, + 0.002495987806469202, + 0.0034849003423005342, + 0.0031945453956723213, + 0.0026013304013758898, + 0.004770766478031874, + 0.002199605107307434, + 0.0027475429233163595, + 0.0042641437612473965, + 0.002594335237517953, + 0.0010130321606993675, + 0.0030595094431191683, + 0.0008632515091449022, + 0.0007740260916762054, + 0.0007455800077877939, + 0.001544143189676106, + 0.002020301064476371 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 6, + 7 + ], + "selected_reward": -0.11296695470809937 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot as a static shot from a horizontal angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a pure, deep, dark blue color?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a shiny silver metal tuning fork visible standing vertically fixed to the left of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are parallel longitudinal waves propagating from the fork to the right, showing alternating compressed (dark/dense) and rarefied (light/transparent) bands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by an even, soft, cool 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": [ + -0.008644422516226768 + ], + "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": [ + 0.12947028875350952, + 0.14710430800914764, + 0.01744789630174637, + 0.023973777890205383, + 0.3025760352611542, + 0.12724436819553375, + 0.0346909835934639, + 0.1551532745361328, + 0.03123832307755947, + 0.03110065497457981 + ], + "credit_map": [ + 0.0218412633985281, + 0.009069997817277908, + 0.03337378799915314, + 0.014202157035470009, + 0.011079021729528904, + 0.010897355154156685, + 0.01264780294150114, + 0.01635892316699028, + 0.021371200680732727, + 0.007023127283900976, + 0.05025649070739746, + 0.012554937042295933, + 0.007914762943983078, + 0.0075584654696285725, + 0.009510243311524391, + 0.030915118753910065, + 0.003879619762301445, + 0.0014862720854580402, + 0.002782307332381606, + 0.0014483509585261345, + 0.0013211506884545088, + 0.0022156783379614353, + 0.001968327909708023, + 0.002346192253753543, + 0.003159109503030777, + 0.0025188459549099207, + 0.00640840595588088, + 0.0027930582873523235, + 0.001989685697481036, + 0.0019990524742752314, + 0.0022604011464864016, + 0.0028452237602323294, + 0.04553695023059845, + 0.06586810201406479, + 0.09651504456996918, + 0.02305041253566742, + 0.0186108760535717, + 0.014761642552912235, + 0.010334194637835026, + 0.027898885309696198, + 0.020082766190171242, + 0.0071110171265900135, + 0.04392726346850395, + 0.010763193480670452, + 0.0061601148918271065, + 0.006403195671737194, + 0.006267467048019171, + 0.02652936987578869, + 0.004921099171042442, + 0.002236219123005867, + 0.00999226700514555, + 0.0030400168616324663, + 0.0019407810177654028, + 0.002196179935708642, + 0.002201839117333293, + 0.008162587881088257, + 0.023902539163827896, + 0.005937580484896898, + 0.05137229710817337, + 0.011397328227758408, + 0.004791449289768934, + 0.007367364596575499, + 0.00776095362380147, + 0.04262378066778183, + 0.004010135307908058, + 0.00418676994740963, + 0.0046474989503622055, + 0.0035896822810173035, + 0.004392579197883606, + 0.002125225495547056, + 0.0026682685129344463, + 0.005618170835077763, + 0.00491048488765955, + 0.0017303749918937683, + 0.007556349039077759, + 0.0030656519811600447, + 0.0019497137982398272, + 0.002089331392198801, + 0.004224847070872784, + 0.005573906470090151 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 7, + 1 + ], + "selected_reward": -0.008644422516226768 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot as a static shot from a horizontal angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a pure, deep, dark blue color?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a shiny silver metal tuning fork visible standing vertically fixed to the left of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are parallel longitudinal waves propagating from the fork to the right, showing alternating compressed (dark/dense) and rarefied (light/transparent) bands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by an even, soft, cool 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": [ + -0.005949051585048437 + ], + "opposite_logprob_tokens": [ + -5.130949020385742 + ], + "logprob_reward": -0.005949051585048437, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005949051585048437, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17778654396533966, + 0.0700560063123703, + 0.09004655480384827, + 0.053703103214502335, + 0.1743871420621872, + 0.2131892293691635, + 0.0695682168006897, + 0.0642617791891098, + 0.05604797229170799, + 0.030953409150242805 + ], + "credit_map": [ + 0.02730659767985344, + 0.016209887340664864, + 0.027879862114787102, + 0.025676017627120018, + 0.01721116527915001, + 0.01631346344947815, + 0.020817887037992477, + 0.026371659711003304, + 0.009192286059260368, + 0.005038347095251083, + 0.01719481684267521, + 0.008213028311729431, + 0.006426672451198101, + 0.004255746025592089, + 0.007859250530600548, + 0.011875857599079609, + 0.00938435923308134, + 0.009998227469623089, + 0.017267217859625816, + 0.013873137533664703, + 0.008372331038117409, + 0.008468592539429665, + 0.008381356485188007, + 0.014301333576440811, + 0.007449491415172815, + 0.006779135204851627, + 0.013752677477896214, + 0.005266536958515644, + 0.004164496902376413, + 0.004331828560680151, + 0.003799597267061472, + 0.008159340359270573, + 0.015802692621946335, + 0.0472111701965332, + 0.015952877700328827, + 0.014776356518268585, + 0.020104851573705673, + 0.01257525384426117, + 0.014779648743569851, + 0.033184293657541275, + 0.03280771151185036, + 0.011469187214970589, + 0.07192861288785934, + 0.01757402904331684, + 0.014441486448049545, + 0.011806855909526348, + 0.011123455129563808, + 0.04203789308667183, + 0.010989520698785782, + 0.003374773543328047, + 0.02211832068860531, + 0.004685296677052975, + 0.003264379221946001, + 0.004136044532060623, + 0.004948203917592764, + 0.016051676124334335, + 0.006877186708152294, + 0.004699429962784052, + 0.015529016964137554, + 0.010493543930351734, + 0.0033294910099357367, + 0.0031974702142179012, + 0.009427332319319248, + 0.01070830225944519, + 0.006796375848352909, + 0.008629664778709412, + 0.006762354634702206, + 0.005435350816696882, + 0.009998172521591187, + 0.004353234078735113, + 0.004874477628618479, + 0.009198342449963093, + 0.004703981336206198, + 0.003172690747305751, + 0.005374356172978878, + 0.0036227512173354626, + 0.002934872405603528, + 0.0035722791217267513, + 0.004433826077729464, + 0.003138651605695486 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 4 + ], + "selected_reward": -0.005949051585048437 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot as a static shot from a horizontal angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a pure, deep, dark blue color?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a shiny silver metal tuning fork visible standing vertically fixed to the left of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are parallel longitudinal waves propagating from the fork to the right, showing alternating compressed (dark/dense) and rarefied (light/transparent) bands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by an even, soft, cool 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": [ + -0.05495835840702057 + ], + "opposite_logprob_tokens": [ + -2.9299583435058594 + ], + "logprob_reward": -0.05495835840702057, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05495835840702057, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2723890542984009, + 0.11072671413421631, + 0.04117485508322716, + 0.13112875819206238, + 0.1401514708995819, + 0.0911988914012909, + 0.017287123948335648, + 0.09234341979026794, + 0.04349495843052864, + 0.060104794800281525 + ], + "credit_map": [ + 0.0456610843539238, + 0.016364365816116333, + 0.0781010091304779, + 0.024391962215304375, + 0.020686963573098183, + 0.025592895224690437, + 0.018109159544110298, + 0.04348160699009895, + 0.016695013269782066, + 0.00814882293343544, + 0.01331543829292059, + 0.014216102659702301, + 0.013813162222504616, + 0.010337013751268387, + 0.021108457818627357, + 0.013092698529362679, + 0.005647908430546522, + 0.004995686002075672, + 0.005476919002830982, + 0.00860895961523056, + 0.003495968645438552, + 0.003772645490244031, + 0.004475749097764492, + 0.004701016470789909, + 0.02076152339577675, + 0.011460610665380955, + 0.04069124534726143, + 0.01373987179249525, + 0.00801926851272583, + 0.008128422312438488, + 0.011268887668848038, + 0.017058925703167915, + 0.01769823208451271, + 0.038061823695898056, + 0.03605809807777405, + 0.012466243468225002, + 0.01154625229537487, + 0.006026847753673792, + 0.008640061132609844, + 0.009653901681303978, + 0.01425190083682537, + 0.00497768959030509, + 0.03130348026752472, + 0.007522008381783962, + 0.0049608927220106125, + 0.004989835433661938, + 0.00430767796933651, + 0.018885400146245956, + 0.0021014504600316286, + 0.0021681231446564198, + 0.0027915164828300476, + 0.0023123002611100674, + 0.001243406324647367, + 0.0015770995523780584, + 0.002491176128387451, + 0.0026020491495728493, + 0.014654440805315971, + 0.004415188916027546, + 0.02845768816769123, + 0.0077185602858662605, + 0.004046638961881399, + 0.005677233915776014, + 0.005513130221515894, + 0.021860530599951744, + 0.004897798877209425, + 0.0066821821965277195, + 0.004877102095633745, + 0.006763503886759281, + 0.007785731460899115, + 0.00339182885363698, + 0.005564204882830381, + 0.003532597329467535, + 0.00979097094386816, + 0.0031827660277485847, + 0.01574726030230522, + 0.0045595234259963036, + 0.003210261696949601, + 0.0036710004787892103, + 0.008244087919592857, + 0.011698916554450989 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 3 + ], + "selected_reward": -0.05495835840702057 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static shot from a horizontal angle, rendered in a minimalist graphic animation style, clearly illustrates the principles of sound wave propagation. Against a pure, deep, dark blue background, a shiny silver metal tuning fork stands vertically fixed to the left of the frame, its prongs vibrating vigorously at an extremely high frequency, with a subtle motion blur at the edges. As the fork vibrates, a series of parallel, visible longitudinal waves propagate steadily and rhythmically from the fork into the air to the right, smoothly crossing the entire frame. These waves vividly demonstrate the alternating process of compression and rarefaction: in compressed areas, the visual elements representing air particles are tightly packed together, forming darker, denser, luminous bands; this is followed by rarefaction areas, where the spacing between particles increases significantly, resulting in lighter, nearly transparent bands. The entire scene is illuminated by an even, soft, cool light, highlighting the dynamic structure of the waves and contrasting sharply with the dark background, creating a visual atmosphere of scientific precision and tranquil concentration.", + "seed": 17049, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000156/group_12/index_54_seed_17049_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": -0.046893347054719925, + "vlm_logprob": -0.046893347054719925, + "vlm_yes_no_margin": 3.55, + "vlm_token_credit": -0.046893347054719925, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot as a static shot from a horizontal angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a pure, deep, dark blue color?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a shiny silver metal tuning fork visible standing vertically fixed to the left of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are parallel longitudinal waves propagating from the fork to the right, showing alternating compressed (dark/dense) and rarefied (light/transparent) bands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by an even, soft, cool 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": [ + -0.0945979431271553 + ], + "opposite_logprob_tokens": [ + -2.5945980548858643 + ], + "logprob_reward": -0.0945979431271553, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.0945979431271553, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03762230649590492, + 0.08074206113815308, + 0.3439522683620453, + 0.04985357075929642, + 0.05751395970582962, + 0.019706714898347855, + 0.21533945202827454, + 0.015308892354369164, + 0.05672513693571091, + 0.12323564291000366 + ], + "credit_map": [ + 0.004499034024775028, + 0.004820601548999548, + 0.0035091321915388107, + 0.003768598660826683, + 0.004217846784740686, + 0.005338320042937994, + 0.004979546181857586, + 0.006489223800599575, + 0.006608059164136648, + 0.00425077136605978, + 0.009763622656464577, + 0.014948347583413124, + 0.005692440550774336, + 0.004029839765280485, + 0.029040126129984856, + 0.006408851593732834, + 0.02946026623249054, + 0.029086647555232048, + 0.0198967382311821, + 0.024610919877886772, + 0.06898093968629837, + 0.0905998945236206, + 0.038882479071617126, + 0.04243434965610504, + 0.00547399278730154, + 0.006768602412194014, + 0.006830899510532618, + 0.006066323257982731, + 0.008173389360308647, + 0.004114651586860418, + 0.007211052812635899, + 0.005214657634496689, + 0.004576358012855053, + 0.004492236766964197, + 0.005031673237681389, + 0.007763001136481762, + 0.005703549366444349, + 0.006451421417295933, + 0.017388267442584038, + 0.006107444874942303, + 0.002173048909753561, + 0.0022949925623834133, + 0.002045517321676016, + 0.00204888335429132, + 0.002257645595818758, + 0.0015871133655309677, + 0.0048343725502491, + 0.002465140074491501, + 0.008242937736213207, + 0.008738284930586815, + 0.013666409067809582, + 0.04198280721902847, + 0.012377248145639896, + 0.011367875151336193, + 0.0897269919514656, + 0.02923690155148506, + 0.0014346691314131021, + 0.002038243692368269, + 0.001245866878889501, + 0.0016930330311879516, + 0.001664689858444035, + 0.0014700350584462285, + 0.0029655061662197113, + 0.002796847140416503, + 0.00402367627248168, + 0.0038041186053305864, + 0.003929181955754757, + 0.02131441980600357, + 0.0019588954746723175, + 0.0017365446547046304, + 0.0176643505692482, + 0.0022939436603337526, + 0.010334385558962822, + 0.010619156993925571, + 0.008357532322406769, + 0.008938062004745007, + 0.024129457771778107, + 0.03236672654747963, + 0.014904746785759926, + 0.013585560955107212 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 6, + 9 + ], + "selected_reward": -0.0945979431271553 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot as a static shot from a horizontal angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a pure, deep, dark blue color?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a shiny silver metal tuning fork visible standing vertically fixed to the left of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are parallel longitudinal waves propagating from the fork to the right, showing alternating compressed (dark/dense) and rarefied (light/transparent) bands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by an even, soft, cool 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": [ + -0.06208525598049164 + ], + "opposite_logprob_tokens": [ + -2.8120851516723633 + ], + "logprob_reward": -0.06208525598049164, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06208525598049164, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.40764954686164856, + 0.03476128727197647, + 0.08105450123548508, + 0.040777526795864105, + 0.1401635855436325, + 0.19720077514648438, + 0.033845435827970505, + 0.03308124095201492, + 0.02491982840001583, + 0.006546251010149717 + ], + "credit_map": [ + 0.04096318036317825, + 0.01762579195201397, + 0.036899685859680176, + 0.03611158952116966, + 0.032290372997522354, + 0.022544024512171745, + 0.16001102328300476, + 0.06120391562581062, + 0.0032327426597476006, + 0.003421291708946228, + 0.003760433290153742, + 0.005552410613745451, + 0.00273280031979084, + 0.0020714940037578344, + 0.01106006558984518, + 0.002930048620328307, + 0.006841776426881552, + 0.0085730766877532, + 0.006187560968101025, + 0.007131813559681177, + 0.015497332438826561, + 0.01873699016869068, + 0.009375340305268764, + 0.008710610680282116, + 0.004385859705507755, + 0.0053460365161299706, + 0.004002491477876902, + 0.003541506826877594, + 0.0029003804083913565, + 0.002329636598005891, + 0.01452112291008234, + 0.003750495845451951, + 0.012290668673813343, + 0.008395496755838394, + 0.012643286027014256, + 0.021195553243160248, + 0.01096611563116312, + 0.009367556311190128, + 0.05190521106123924, + 0.013399697840213776, + 0.025799941271543503, + 0.021728262305259705, + 0.02343239262700081, + 0.022516684606671333, + 0.01378199178725481, + 0.017626918852329254, + 0.029212912544608116, + 0.043101660907268524, + 0.0013481430942192674, + 0.0012803255813196301, + 0.002140570431947708, + 0.006413990631699562, + 0.0018771319882944226, + 0.0016001567710191011, + 0.014298571273684502, + 0.0048865461722016335, + 0.003553227987140417, + 0.0033938384149223566, + 0.0035423673689365387, + 0.0030095898546278477, + 0.0031522365752607584, + 0.0029409921262413263, + 0.0082062017172575, + 0.0052827876061201096, + 0.0029590551275759935, + 0.002265029586851597, + 0.0054375347681343555, + 0.005346410442143679, + 0.001175104407593608, + 0.0011730464175343513, + 0.004486652556806803, + 0.002076993929222226, + 0.0005501883570104837, + 0.0005490203038789332, + 0.0005725285154767334, + 0.0005681375041604042, + 0.001189399161376059, + 0.0014687915099784732, + 0.0008221724419854581, + 0.0008260130416601896 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 4 + ], + "selected_reward": -0.06208525598049164 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot as a static shot from a horizontal angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a pure, deep, dark blue color?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a shiny silver metal tuning fork visible standing vertically fixed to the left of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are parallel longitudinal waves propagating from the fork to the right, showing alternating compressed (dark/dense) and rarefied (light/transparent) bands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by an even, soft, cool 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": [ + -0.009788372553884983 + ], + "opposite_logprob_tokens": [ + -4.634788513183594 + ], + "logprob_reward": -0.009788372553884983, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009788372553884983, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0970987007021904, + 0.06381590664386749, + 0.05565016344189644, + 0.0869402140378952, + 0.09312067180871964, + 0.04768603295087814, + 0.41337069869041443, + 0.029332894831895828, + 0.06760329753160477, + 0.04538141191005707 + ], + "credit_map": [ + 0.010516003705561161, + 0.010340758599340916, + 0.0076402295380830765, + 0.01125209592282772, + 0.011639532633125782, + 0.012672154232859612, + 0.015530094504356384, + 0.01750783994793892, + 0.007388066966086626, + 0.006178558338433504, + 0.010451065376400948, + 0.009335663169622421, + 0.005276905372738838, + 0.005469846539199352, + 0.008195017464458942, + 0.01152078527957201, + 0.006128286011517048, + 0.004613265860825777, + 0.004022521432489157, + 0.006239472422748804, + 0.007710624020546675, + 0.009498430415987968, + 0.01081452053040266, + 0.00662304786965251, + 0.009273112751543522, + 0.010191520676016808, + 0.009788519702851772, + 0.008596948347985744, + 0.006932850461453199, + 0.005148538388311863, + 0.027927760034799576, + 0.0090809790417552, + 0.007019843440502882, + 0.0049507953226566315, + 0.009047114290297031, + 0.013704550452530384, + 0.008416710421442986, + 0.006717395968735218, + 0.03459862247109413, + 0.00866564828902483, + 0.0064317695796489716, + 0.004654739983379841, + 0.0055922032333910465, + 0.005166890099644661, + 0.003937870729714632, + 0.0044359564781188965, + 0.008079958148300648, + 0.00938664935529232, + 0.018461858853697777, + 0.015216921456158161, + 0.028201837092638016, + 0.08153564482927322, + 0.023649534210562706, + 0.013574969954788685, + 0.17853789031505585, + 0.05419207364320755, + 0.002752965549007058, + 0.0029915457125753164, + 0.003380174981430173, + 0.0027840654365718365, + 0.00255386158823967, + 0.0026755062863230705, + 0.008296112529933453, + 0.003898666240274906, + 0.004594224970787764, + 0.004600458778440952, + 0.008825602009892464, + 0.027521394193172455, + 0.0023768162354826927, + 0.0021258085034787655, + 0.014673142693936825, + 0.0028858573641628027, + 0.0037480476312339306, + 0.004045544192194939, + 0.0032787052914500237, + 0.0033966670744121075, + 0.008205500431358814, + 0.011757864616811275, + 0.005681519862264395, + 0.005267567001283169 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 4 + ], + "selected_reward": -0.009788372553884983 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot as a static shot from a horizontal angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a pure, deep, dark blue color?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a shiny silver metal tuning fork visible standing vertically fixed to the left of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are parallel longitudinal waves propagating from the fork to the right, showing alternating compressed (dark/dense) and rarefied (light/transparent) bands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by an even, soft, cool 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": [ + -0.005948222242295742 + ], + "opposite_logprob_tokens": [ + -5.130948066711426 + ], + "logprob_reward": -0.005948222242295742, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005948222242295742, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14577433466911316, + 0.09194345027208328, + 0.2104795128107071, + 0.029592247679829597, + 0.14148227870464325, + 0.10002157837152481, + 0.13065370917320251, + 0.04210112988948822, + 0.08237587660551071, + 0.025575853884220123 + ], + "credit_map": [ + 0.016051212325692177, + 0.010707265697419643, + 0.013073785230517387, + 0.013241532258689404, + 0.01649412326514721, + 0.018084703013300896, + 0.03787114471197128, + 0.020250577479600906, + 0.009859481826424599, + 0.008361713029444218, + 0.00838560238480568, + 0.008905477821826935, + 0.01010762806981802, + 0.01078181341290474, + 0.015561597421765327, + 0.019980136305093765, + 0.019966650754213333, + 0.017660096287727356, + 0.011969917453825474, + 0.015153298154473305, + 0.03955962881445885, + 0.05393284559249878, + 0.02577609196305275, + 0.026460980996489525, + 0.003797977464273572, + 0.003515348071232438, + 0.003540044417604804, + 0.0026624018792062998, + 0.0050351289100945, + 0.0032299428712576628, + 0.0042494856752455235, + 0.003561915596947074, + 0.013852554373443127, + 0.012472212314605713, + 0.012281452305614948, + 0.01766512356698513, + 0.013521871529519558, + 0.0178319551050663, + 0.03710712492465973, + 0.01674998551607132, + 0.01293597649782896, + 0.011015628464519978, + 0.010445622727274895, + 0.011204480193555355, + 0.007385140284895897, + 0.00916298944503069, + 0.019241124391555786, + 0.0186306145042181, + 0.0063839079812169075, + 0.005720088724046946, + 0.00787015538662672, + 0.02554641105234623, + 0.008345691487193108, + 0.00579823087900877, + 0.053237948566675186, + 0.01775127463042736, + 0.004249597899615765, + 0.004166375380009413, + 0.0035572091583162546, + 0.0033558609429746866, + 0.00393005833029747, + 0.006662803702056408, + 0.008683588355779648, + 0.0074956389144063, + 0.005889663938432932, + 0.005343123804777861, + 0.011997365392744541, + 0.021893903613090515, + 0.002914544427767396, + 0.0032416447065770626, + 0.02556547522544861, + 0.005530154332518578, + 0.0022994112223386765, + 0.002128210384398699, + 0.0022917762398719788, + 0.002213444095104933, + 0.003918943926692009, + 0.004318614024668932, + 0.0035292422398924828, + 0.004876210819929838 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 4 + ], + "selected_reward": -0.005948222242295742 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot as a static shot from a horizontal angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a pure, deep, dark blue color?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a shiny silver metal tuning fork visible standing vertically fixed to the left of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are parallel longitudinal waves propagating from the fork to the right, showing alternating compressed (dark/dense) and rarefied (light/transparent) bands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by an even, soft, cool 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": [ + -0.06204694136977196 + ], + "opposite_logprob_tokens": [ + -2.812047004699707 + ], + "logprob_reward": -0.06204694136977196, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06204694136977196, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11441616714000702, + 0.058174144476652145, + 0.1263759583234787, + 0.05296469107270241, + 0.045904386788606644, + 0.06217431649565697, + 0.3255353271961212, + 0.09373770654201508, + 0.05699821189045906, + 0.06371907144784927 + ], + "credit_map": [ + 0.01152767427265644, + 0.007868333719670773, + 0.011563771404325962, + 0.010282988660037518, + 0.008847742341458797, + 0.01028234139084816, + 0.030260754749178886, + 0.023782556876540184, + 0.003621466923505068, + 0.005449933465570211, + 0.006663757376372814, + 0.010584178380668163, + 0.0038897942285984755, + 0.0021266532130539417, + 0.021695872768759727, + 0.004142488352954388, + 0.011716540902853012, + 0.012253032065927982, + 0.007659061346203089, + 0.010193484835326672, + 0.023094763979315758, + 0.028641272336244583, + 0.015200523659586906, + 0.017617279663681984, + 0.005830229260027409, + 0.0066082715056836605, + 0.006377990823239088, + 0.0053806304931640625, + 0.004396314732730389, + 0.0029599035624414682, + 0.016846004873514175, + 0.004565345123410225, + 0.0038524565752595663, + 0.0026265529450029135, + 0.0035654972307384014, + 0.00660706777125597, + 0.004394251387566328, + 0.0049025313928723335, + 0.015263080596923828, + 0.004692948888987303, + 0.007971732877194881, + 0.010465518571436405, + 0.006285749841481447, + 0.00776989571750164, + 0.004702375736087561, + 0.00557282380759716, + 0.008974269032478333, + 0.01043194904923439, + 0.013519561849534512, + 0.010850277729332447, + 0.0213499553501606, + 0.06534206122159958, + 0.018101194873452187, + 0.011877337470650673, + 0.141419917345047, + 0.04307505115866661, + 0.008718670345842838, + 0.01779242977499962, + 0.011298952624201775, + 0.01511278748512268, + 0.009502388536930084, + 0.007591120898723602, + 0.01362806186079979, + 0.010093293152749538, + 0.003781283274292946, + 0.003455977886915207, + 0.004236983601003885, + 0.02109198085963726, + 0.0019048969261348248, + 0.0018127901712432504, + 0.018620045855641365, + 0.0020942548289895058, + 0.005152211990207434, + 0.005393106024712324, + 0.004231378436088562, + 0.004576635546982288, + 0.011921294964849949, + 0.01650416851043701, + 0.008471202105283737, + 0.007469079457223415 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 2, + 0 + ], + "selected_reward": -0.06204694136977196 + } + ], + "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/0000156/group_13/index_tensor([3371, 3371])_group_13_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000156/group_13/index_tensor([3371, 3371])_group_13_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..58098d1280e7d5ea4cc982181a8f83b27c515ef5 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000156/group_13/index_tensor([3371, 3371])_group_13_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "low-angle, static medium shot captures the immense force of a yellow backhoe arm as it drives into the ground. Painted in a lightly weathered industrial yellow paint, the heavy steel arm shows signs of wear, with oil stains and light scratches, yet possesses a mechanical precision. Its toothed metal bucket bites deeply into the earth, and with a deliberate, powerful curl, it scoops up a vast mound of rich, dark soil, intermixed with jagged gray rocks of various sizes. The entire scene is illuminated by the harsh, direct light of the midday sun, set against a vast, cloudless, brilliant blue sky. This brilliant light creates sharp, well-defined shadows and brilliant glints on the arm's metal surface and the gleaming facets of unearthed stone. As the arm begins its steady upward lift, straining slightly under its weight, the atmosphere is one of concentrated, raw industrial power, showcasing the brute force of a machine over the natural landscape.", + "seed": 8822, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000156/group_13/index_3371_seed_8822_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": -0.02835198864340782, + "vlm_logprob": -0.02835198864340782, + "vlm_yes_no_margin": 4.1, + "vlm_token_credit": -0.02835198864340782, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 static medium shot of a yellow backhoe arm?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the backhoe arm show signs of wear including oil stains, light scratches, and weathered paint?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the toothed metal bucket scoop up a mound of dark soil mixed with jagged gray rocks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by harsh midday sun creating sharp shadows and glints on the metal?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the background consist of a vast, cloudless, brilliant blue sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05028241500258446 + ], + "opposite_logprob_tokens": [ + -3.9252824783325195 + ], + "logprob_reward": -0.05028241500258446, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.05028241500258446, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.38989368081092834, + 0.12411585450172424, + 0.04497478902339935, + 0.04682355374097824, + 0.03990056738257408, + 0.04285874962806702, + 0.06940481066703796, + 0.08342213928699493, + 0.08904597908258438, + 0.06955984979867935 + ], + "credit_map": [ + 0.12843216955661774, + 0.13905413448810577, + 0.022127332165837288, + 0.025934625416994095, + 0.02918063849210739, + 0.014732686802744865, + 0.020710496231913567, + 0.009721612557768822, + 0.012632343918085098, + 0.019458336755633354, + 0.009722456336021423, + 0.014405968599021435, + 0.017807738855481148, + 0.01668630540370941, + 0.023408373817801476, + 0.009994336403906345, + 0.005612004082649946, + 0.008171181194484234, + 0.0043758489191532135, + 0.007291257847100496, + 0.00546541390940547, + 0.004851819947361946, + 0.005827559623867273, + 0.003379703965038061, + 0.0051560658030211926, + 0.006325739435851574, + 0.00631418963894248, + 0.007703834678977728, + 0.005908302031457424, + 0.004267136100679636, + 0.007478810381144285, + 0.0036694761365652084, + 0.0062229516915977, + 0.005561597179621458, + 0.004410392139106989, + 0.007013392634689808, + 0.005721754860132933, + 0.0033858695533126593, + 0.004716883413493633, + 0.0028677298687398434, + 0.004869369324296713, + 0.007171830628067255, + 0.005878045689314604, + 0.006084323860704899, + 0.00516308331862092, + 0.004331966862082481, + 0.005840649362653494, + 0.003519482212141156, + 0.007300228346139193, + 0.008824892342090607, + 0.0072294664569199085, + 0.007057834416627884, + 0.011360378004610538, + 0.010988723486661911, + 0.011483671143651009, + 0.0051596206612885, + 0.009833774529397488, + 0.009554298594594002, + 0.0056710010394454, + 0.0087805250659585, + 0.014343038201332092, + 0.00977447722107172, + 0.019853226840496063, + 0.005611801519989967, + 0.0072967964224517345, + 0.009499358013272285, + 0.008600904606282711, + 0.009621940553188324, + 0.015731869265437126, + 0.008785571902990341, + 0.016491567716002464, + 0.013017982244491577, + 0.008826692588627338, + 0.008678232319653034, + 0.005389092490077019, + 0.006788613274693489, + 0.014338618144392967, + 0.006075912620872259, + 0.010475151240825653, + 0.00898754969239235 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 8 + ], + "selected_reward": -0.05028241500258446 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 static medium shot of a yellow backhoe arm?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the backhoe arm show signs of wear including oil stains, light scratches, and weathered paint?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the toothed metal bucket scoop up a mound of dark soil mixed with jagged gray rocks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by harsh midday sun creating sharp shadows and glints on the metal?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the background consist of a vast, cloudless, brilliant blue sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.054998189210891724 + ], + "opposite_logprob_tokens": [ + -2.9299981594085693 + ], + "logprob_reward": -0.054998189210891724, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.054998189210891724, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3426811397075653, + 0.0850566104054451, + 0.04492061212658882, + 0.05034566670656204, + 0.04735155776143074, + 0.048573702573776245, + 0.07452251017093658, + 0.05888022854924202, + 0.11125985532999039, + 0.13640812039375305 + ], + "credit_map": [ + 0.10293411463499069, + 0.10052449256181717, + 0.026546895503997803, + 0.03545939177274704, + 0.02412089891731739, + 0.017162146046757698, + 0.02314610965549946, + 0.01278708130121231, + 0.010665264911949635, + 0.015265203081071377, + 0.007321298122406006, + 0.013024791143834591, + 0.010940862819552422, + 0.0077585321851074696, + 0.014928663149476051, + 0.005151993129402399, + 0.005347150843590498, + 0.0068590776063501835, + 0.006103529129177332, + 0.006506504490971565, + 0.005265337415039539, + 0.004608041141182184, + 0.005798390135169029, + 0.00443258136510849, + 0.007020345889031887, + 0.007400615140795708, + 0.00852250400930643, + 0.006286476273089647, + 0.005345305427908897, + 0.004647493828088045, + 0.0053903572261333466, + 0.00573256378993392, + 0.006096874829381704, + 0.006120826583355665, + 0.007755726110190153, + 0.0069722202606499195, + 0.0064048185013234615, + 0.004848912823945284, + 0.006004576571285725, + 0.0031475997529923916, + 0.007303342223167419, + 0.006795885507017374, + 0.006348039023578167, + 0.004793812055140734, + 0.006214476656168699, + 0.0059339022263884544, + 0.006055326666682959, + 0.005128915887326002, + 0.00813179649412632, + 0.00944084208458662, + 0.008848731406033039, + 0.0074731516651809216, + 0.012292067520320415, + 0.009596000425517559, + 0.012911363504827023, + 0.00582854961976409, + 0.006232391111552715, + 0.0067077986896038055, + 0.006726133171468973, + 0.00747276283800602, + 0.00947300624102354, + 0.006112596485763788, + 0.010697604157030582, + 0.005457931198179722, + 0.008844459429383278, + 0.012821896933019161, + 0.012332973070442677, + 0.01123677659779787, + 0.021012237295508385, + 0.01535111665725708, + 0.015913816168904305, + 0.013746571727097034, + 0.017152536660432816, + 0.016782525926828384, + 0.014142303727567196, + 0.017662040889263153, + 0.025639237836003304, + 0.013104431331157684, + 0.015627648681402206, + 0.01629737950861454 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 8 + ], + "selected_reward": -0.054998189210891724 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 static medium shot of a yellow backhoe arm?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the backhoe arm show signs of wear including oil stains, light scratches, and weathered paint?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the toothed metal bucket scoop up a mound of dark soil mixed with jagged gray rocks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by harsh midday sun creating sharp shadows and glints on the metal?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the background consist of a vast, cloudless, brilliant blue sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018191635608673096 + ], + "opposite_logprob_tokens": [ + -4.018191814422607 + ], + "logprob_reward": -0.018191635608673096, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018191635608673096, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06967789679765701, + 0.12617534399032593, + 0.05522822588682175, + 0.039109017699956894, + 0.04244053736329079, + 0.029687562957406044, + 0.23518742620944977, + 0.1021072268486023, + 0.24763639271259308, + 0.05275041237473488 + ], + "credit_map": [ + 0.018216868862509727, + 0.01920258440077305, + 0.0043592085130512714, + 0.005300846882164478, + 0.007951941341161728, + 0.0061876499094069, + 0.005145310889929533, + 0.003313485998660326, + 0.014378565363585949, + 0.018323931843042374, + 0.01624140702188015, + 0.011920337565243244, + 0.016929395496845245, + 0.01503179781138897, + 0.022679362446069717, + 0.010670537129044533, + 0.005697553511708975, + 0.008938614279031754, + 0.005758754909038544, + 0.005658271256834269, + 0.007370454259216785, + 0.008659111335873604, + 0.008494263514876366, + 0.004651203285902739, + 0.0037250646855682135, + 0.0054537090472877026, + 0.004843490663915873, + 0.006714967079460621, + 0.004990380723029375, + 0.005369743797928095, + 0.0055712503381073475, + 0.002440410666167736, + 0.0039170878008008, + 0.006372212432324886, + 0.003754484234377742, + 0.00617896718904376, + 0.009135503321886063, + 0.006302619818598032, + 0.00436496501788497, + 0.002414696617051959, + 0.004408678505569696, + 0.004243494942784309, + 0.0035157769452780485, + 0.003192238276824355, + 0.004446445964276791, + 0.003999471198767424, + 0.0033271382562816143, + 0.0025543184019625187, + 0.02214079350233078, + 0.030266616493463516, + 0.01854688487946987, + 0.02156558260321617, + 0.043585047125816345, + 0.04654375836253166, + 0.03571641817688942, + 0.016822323203086853, + 0.01066717691719532, + 0.0130870146676898, + 0.014600629918277264, + 0.010095336474478245, + 0.018699537962675095, + 0.012943820096552372, + 0.014676982536911964, + 0.007336731534451246, + 0.020093606784939766, + 0.03004976361989975, + 0.02156159095466137, + 0.016704168170690536, + 0.05544370040297508, + 0.03391016274690628, + 0.027268746867775917, + 0.04260464757680893, + 0.0058539751917123795, + 0.006943487096577883, + 0.005551557522267103, + 0.005406652577221394, + 0.00869530439376831, + 0.005714637227356434, + 0.008949032053351402, + 0.005635768175125122 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 6, + 1 + ], + "selected_reward": -0.018191635608673096 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 static medium shot of a yellow backhoe arm?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the backhoe arm show signs of wear including oil stains, light scratches, and weathered paint?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the toothed metal bucket scoop up a mound of dark soil mixed with jagged gray rocks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by harsh midday sun creating sharp shadows and glints on the metal?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the background consist of a vast, cloudless, brilliant blue sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014196357689797878 + ], + "opposite_logprob_tokens": [ + -4.264196395874023 + ], + "logprob_reward": -0.014196357689797878, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014196357689797878, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.29601532220840454, + 0.05988546460866928, + 0.1051863357424736, + 0.1012001484632492, + 0.026766661554574966, + 0.045395877212285995, + 0.07495428621768951, + 0.1105576679110527, + 0.11708427220582962, + 0.06295399367809296 + ], + "credit_map": [ + 0.07804186642169952, + 0.07636398077011108, + 0.03044893778860569, + 0.032099246978759766, + 0.025439923629164696, + 0.015434738248586655, + 0.02457606792449951, + 0.01361052505671978, + 0.009969357401132584, + 0.009330886416137218, + 0.010134560987353325, + 0.008555877953767776, + 0.0050209201872348785, + 0.005011634901165962, + 0.006294701248407364, + 0.005567518528550863, + 0.01916670612990856, + 0.01855561137199402, + 0.014193552546203136, + 0.012939696200191975, + 0.010324404574930668, + 0.010922504588961601, + 0.011295871809124947, + 0.007787973154336214, + 0.014820439741015434, + 0.016462432220578194, + 0.013066096231341362, + 0.024266062304377556, + 0.010458915494382381, + 0.006869264412671328, + 0.009466287679970264, + 0.005790639203041792, + 0.0035327570512890816, + 0.0034759859554469585, + 0.004080325830727816, + 0.005032482091337442, + 0.003411359153687954, + 0.0024538973812013865, + 0.002904079621657729, + 0.0018757705111056566, + 0.007727502845227718, + 0.005922011099755764, + 0.008578913286328316, + 0.005333959124982357, + 0.0060227843932807446, + 0.0039004727732390165, + 0.003754719626158476, + 0.00415550684556365, + 0.008232327178120613, + 0.00999187957495451, + 0.006125219166278839, + 0.007530387956649065, + 0.011962693184614182, + 0.010546782054007053, + 0.01427987590432167, + 0.006285109557211399, + 0.016317563131451607, + 0.015235481783747673, + 0.010763420723378658, + 0.013759177178144455, + 0.01781426928937435, + 0.011164438910782337, + 0.015717435628175735, + 0.009785866364836693, + 0.012195354327559471, + 0.014058497734367847, + 0.00866902619600296, + 0.009788393042981625, + 0.0258535947650671, + 0.01851755939424038, + 0.013337498530745506, + 0.014664335176348686, + 0.009777952916920185, + 0.00841919332742691, + 0.00815922673791647, + 0.008317755535244942, + 0.009915185160934925, + 0.0058256820775568485, + 0.006430312525480986, + 0.006108682602643967 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 7 + ], + "selected_reward": -0.014196357689797878 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 static medium shot of a yellow backhoe arm?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the backhoe arm show signs of wear including oil stains, light scratches, and weathered paint?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the toothed metal bucket scoop up a mound of dark soil mixed with jagged gray rocks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by harsh midday sun creating sharp shadows and glints on the metal?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the background consist of a vast, cloudless, brilliant blue sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004091345705091953 + ], + "opposite_logprob_tokens": [ + -5.504091262817383 + ], + "logprob_reward": -0.004091345705091953, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004091345705091953, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11441195011138916, + 0.053048025816679, + 0.044850077480077744, + 0.0927155539393425, + 0.029052037745714188, + 0.07288443297147751, + 0.09515346586704254, + 0.3212282955646515, + 0.0916401594877243, + 0.08501604199409485 + ], + "credit_map": [ + 0.02236437052488327, + 0.020716723054647446, + 0.011710857041180134, + 0.015602733008563519, + 0.015550049021840096, + 0.010378493927419186, + 0.012010435573756695, + 0.006078280042856932, + 0.008272572420537472, + 0.007111296057701111, + 0.008800559677183628, + 0.008734057657420635, + 0.004041742533445358, + 0.005182739347219467, + 0.007544492371380329, + 0.0033605648204684258, + 0.006809864193201065, + 0.006984037347137928, + 0.006187410559505224, + 0.008642754517495632, + 0.005119859706610441, + 0.0034527373500168324, + 0.004337158054113388, + 0.003316254122182727, + 0.01751113124191761, + 0.016431093215942383, + 0.010256846435368061, + 0.01643170230090618, + 0.010643662884831429, + 0.005951770581305027, + 0.007797835860401392, + 0.007691503502428532, + 0.005106361582875252, + 0.0038375896401703358, + 0.0044045234099030495, + 0.004639604594558477, + 0.00319193908944726, + 0.0024459650740027428, + 0.0033357262145727873, + 0.002090326277539134, + 0.008889838121831417, + 0.009427607990801334, + 0.012802045792341232, + 0.009861528873443604, + 0.009751325473189354, + 0.008546795696020126, + 0.006457931827753782, + 0.007147349417209625, + 0.012483169324696064, + 0.013794461265206337, + 0.014609303325414658, + 0.014311740174889565, + 0.013821490108966827, + 0.00911853276193142, + 0.010075599886476994, + 0.006939160171896219, + 0.039963312447071075, + 0.042870547622442245, + 0.03313601016998291, + 0.033784426748752594, + 0.055047862231731415, + 0.03493260592222214, + 0.05366596579551697, + 0.027827542275190353, + 0.011371191591024399, + 0.010146250016987324, + 0.013728013262152672, + 0.01506195031106472, + 0.011922413483262062, + 0.011801483109593391, + 0.011166943237185478, + 0.006441902834922075, + 0.01659473404288292, + 0.01129947416484356, + 0.014780247583985329, + 0.01416120957583189, + 0.007586906664073467, + 0.0059742433950304985, + 0.005632692016661167, + 0.008986531756818295 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 6 + ], + "selected_reward": -0.004091345705091953 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle, static medium shot captures the immense force of a yellow backhoe arm as it drives into the ground. Painted in a lightly weathered industrial yellow paint, the heavy steel arm shows signs of wear, with oil stains and light scratches, yet possesses a mechanical precision. Its toothed metal bucket bites deeply into the earth, and with a deliberate, powerful curl, it scoops up a vast mound of rich, dark soil, intermixed with jagged gray rocks of various sizes. The entire scene is illuminated by the harsh, direct light of the midday sun, set against a vast, cloudless, brilliant blue sky. This brilliant light creates sharp, well-defined shadows and brilliant glints on the arm's metal surface and the gleaming facets of unearthed stone. As the arm begins its steady upward lift, straining slightly under its weight, the atmosphere is one of concentrated, raw industrial power, showcasing the brute force of a machine over the natural landscape.", + "seed": 8823, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000156/group_13/index_3371_seed_8823_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": -0.03784892968833446, + "vlm_logprob": -0.03784892968833446, + "vlm_yes_no_margin": 3.425, + "vlm_token_credit": -0.03784892968833446, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 static medium shot of a yellow backhoe arm?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the backhoe arm show signs of wear including oil stains, light scratches, and weathered paint?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the toothed metal bucket scoop up a mound of dark soil mixed with jagged gray rocks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by harsh midday sun creating sharp shadows and glints on the metal?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the background consist of a vast, cloudless, brilliant blue sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.043779194355010986 + ], + "opposite_logprob_tokens": [ + -3.793779134750366 + ], + "logprob_reward": -0.043779194355010986, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.043779194355010986, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0887753814458847, + 0.03436737135052681, + 0.04332086816430092, + 0.06572115421295166, + 0.11925482004880905, + 0.4098062217235565, + 0.08992399275302887, + 0.06988207250833511, + 0.03232451528310776, + 0.04662366956472397 + ], + "credit_map": [ + 0.011482764966785908, + 0.009915425442159176, + 0.008586745709180832, + 0.011279747821390629, + 0.009070259518921375, + 0.015247981995344162, + 0.013132335618138313, + 0.010060114786028862, + 0.0031128949485719204, + 0.005521799903362989, + 0.004051353316754103, + 0.003194727236405015, + 0.00377855752594769, + 0.0041487617418169975, + 0.005239772144705057, + 0.005319503135979176, + 0.004137853626161814, + 0.003928113263100386, + 0.006402937229722738, + 0.0037970731500536203, + 0.004412340000271797, + 0.00782542023807764, + 0.0075041125528514385, + 0.005313011817634106, + 0.004984248895198107, + 0.007488610688596964, + 0.004459346178919077, + 0.019611932337284088, + 0.004470673855394125, + 0.011574932374060154, + 0.00747704366222024, + 0.005654357839375734, + 0.020484071224927902, + 0.008535068482160568, + 0.011578736826777458, + 0.00971279013901949, + 0.008991462178528309, + 0.0333319753408432, + 0.012341674417257309, + 0.014279033057391644, + 0.12401402741670609, + 0.013182669878005981, + 0.020350055769085884, + 0.0115346135571599, + 0.02036295272409916, + 0.17879854142665863, + 0.015943672508001328, + 0.025619663298130035, + 0.007302524521946907, + 0.007846018299460411, + 0.005029084626585245, + 0.00859154760837555, + 0.006584780756384134, + 0.014769628643989563, + 0.01919363997876644, + 0.020606763660907745, + 0.004958203062415123, + 0.010112582705914974, + 0.00428302027285099, + 0.003606190672144294, + 0.004821870941668749, + 0.004388472530990839, + 0.021036261692643166, + 0.016675470396876335, + 0.005143948830664158, + 0.004388776607811451, + 0.0034270700998604298, + 0.0028036278672516346, + 0.0037909718230366707, + 0.0059668333269655704, + 0.004039810039103031, + 0.002763474127277732, + 0.002943438943475485, + 0.005487889051437378, + 0.004383397288620472, + 0.015236947685480118, + 0.0033921436406672, + 0.0028046618681401014, + 0.009007064625620842, + 0.003368123434484005 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 4, + 6 + ], + "selected_reward": -0.043779194355010986 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 static medium shot of a yellow backhoe arm?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the backhoe arm show signs of wear including oil stains, light scratches, and weathered paint?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the toothed metal bucket scoop up a mound of dark soil mixed with jagged gray rocks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by harsh midday sun creating sharp shadows and glints on the metal?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the background consist of a vast, cloudless, brilliant blue sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04870099201798439 + ], + "opposite_logprob_tokens": [ + -3.048701047897339 + ], + "logprob_reward": -0.04870099201798439, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04870099201798439, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07185567915439606, + 0.04399873688817024, + 0.05452025309205055, + 0.029564175754785538, + 0.08577697724103928, + 0.3945290744304657, + 0.13461630046367645, + 0.04683540388941765, + 0.09855253249406815, + 0.03975089639425278 + ], + "credit_map": [ + 0.006741238292306662, + 0.008168576285243034, + 0.009080063551664352, + 0.007765717338770628, + 0.007010537665337324, + 0.009433011524379253, + 0.012491045519709587, + 0.011165489442646503, + 0.00406579440459609, + 0.005775787401944399, + 0.005835175514221191, + 0.006143060512840748, + 0.0032982658594846725, + 0.0041176010854542255, + 0.008568910881876945, + 0.006194140762090683, + 0.0055600800551474094, + 0.0075690289959311485, + 0.0069602979347109795, + 0.005531031172722578, + 0.004326837137341499, + 0.008645148947834969, + 0.008659740909934044, + 0.0072680870071053505, + 0.0028071950655430555, + 0.0026082287076860666, + 0.0028843842446804047, + 0.005790695548057556, + 0.0018871426582336426, + 0.004130655899643898, + 0.0057296124286949635, + 0.003726259106770158, + 0.010697735473513603, + 0.009994380176067352, + 0.008759749121963978, + 0.008341081440448761, + 0.010623027570545673, + 0.015972420573234558, + 0.012712528929114342, + 0.008676060475409031, + 0.11928249895572662, + 0.012636455707252026, + 0.01931818760931492, + 0.010828317143023014, + 0.018814489245414734, + 0.17493227124214172, + 0.013800256885588169, + 0.024916617199778557, + 0.015318033285439014, + 0.007352966349571943, + 0.0061393119394779205, + 0.008017138577997684, + 0.009242641739547253, + 0.04516807571053505, + 0.012607770971953869, + 0.030770374462008476, + 0.003792023519054055, + 0.006199211347848177, + 0.005425688344985247, + 0.004122267477214336, + 0.005560735240578651, + 0.006884443573653698, + 0.00712777441367507, + 0.00772325973957777, + 0.007442138623446226, + 0.005337012931704521, + 0.005437352228909731, + 0.007920259609818459, + 0.022612059488892555, + 0.006019726395606995, + 0.0352461151778698, + 0.008537865243852139, + 0.0038086450658738613, + 0.006907708942890167, + 0.004175565205514431, + 0.010723812505602837, + 0.0038522693794220686, + 0.0026338263414800167, + 0.005398732144385576, + 0.002250338438898325 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 6, + 8 + ], + "selected_reward": -0.04870099201798439 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 static medium shot of a yellow backhoe arm?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the backhoe arm show signs of wear including oil stains, light scratches, and weathered paint?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the toothed metal bucket scoop up a mound of dark soil mixed with jagged gray rocks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by harsh midday sun creating sharp shadows and glints on the metal?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the background consist of a vast, cloudless, brilliant blue sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03809947893023491 + ], + "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": [ + 0.043289508670568466, + 0.0335647389292717, + 0.03487955778837204, + 0.09425456076860428, + 0.26488548517227173, + 0.2528723180294037, + 0.070877306163311, + 0.08806996792554855, + 0.04017110541462898, + 0.07713547348976135 + ], + "credit_map": [ + 0.004894063342362642, + 0.004962145816534758, + 0.005990361794829369, + 0.004985142964869738, + 0.0033809267915785313, + 0.004087400157004595, + 0.007702726870775223, + 0.007286738138645887, + 0.0035565709695219994, + 0.003394695231691003, + 0.002996135037392378, + 0.003190741641446948, + 0.004671588074415922, + 0.008526033721864223, + 0.0037183910608291626, + 0.00351058435626328, + 0.0035003037191927433, + 0.0029998922254890203, + 0.007345380261540413, + 0.004309916868805885, + 0.003392862854525447, + 0.0037896689027547836, + 0.005365405231714249, + 0.0041761258617043495, + 0.012554413639008999, + 0.01130334846675396, + 0.008526095189154148, + 0.01448184996843338, + 0.009335650131106377, + 0.02236553281545639, + 0.008662950247526169, + 0.007024722173810005, + 0.024752773344516754, + 0.04323187470436096, + 0.03083007037639618, + 0.02804054319858551, + 0.03641822189092636, + 0.0335405208170414, + 0.036968082189559937, + 0.03110339865088463, + 0.07549203932285309, + 0.007626512553542852, + 0.013473068363964558, + 0.007492660079151392, + 0.011543731205165386, + 0.10993503034114838, + 0.011340618133544922, + 0.015968652442097664, + 0.008005345240235329, + 0.0032269468065351248, + 0.003116974839940667, + 0.00448621017858386, + 0.005360423121601343, + 0.021182164549827576, + 0.010653151199221611, + 0.01484608929604292, + 0.010308018885552883, + 0.02138477750122547, + 0.008293083868920803, + 0.0072662425227463245, + 0.008792079985141754, + 0.008883747272193432, + 0.013500283472239971, + 0.00964172650128603, + 0.004192893858999014, + 0.004902903456240892, + 0.0047499337233603, + 0.0038448020350188017, + 0.006380343344062567, + 0.00436732592061162, + 0.007744180504232645, + 0.0039887214079499245, + 0.005365144461393356, + 0.007601879071444273, + 0.0060894060879945755, + 0.021902337670326233, + 0.006242056377232075, + 0.005034542642533779, + 0.01980157010257244, + 0.005098546389490366 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 5, + 3 + ], + "selected_reward": -0.03809947893023491 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 static medium shot of a yellow backhoe arm?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the backhoe arm show signs of wear including oil stains, light scratches, and weathered paint?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the toothed metal bucket scoop up a mound of dark soil mixed with jagged gray rocks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by harsh midday sun creating sharp shadows and glints on the metal?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the background consist of a vast, cloudless, brilliant blue sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02057950384914875 + ], + "opposite_logprob_tokens": [ + -3.8955795764923096 + ], + "logprob_reward": -0.02057950384914875, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.02057950384914875, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17164576053619385, + 0.07651612907648087, + 0.06586241722106934, + 0.06869404762983322, + 0.11886797845363617, + 0.27078673243522644, + 0.07937045395374298, + 0.03939899802207947, + 0.04730154946446419, + 0.0615558996796608 + ], + "credit_map": [ + 0.015116250142455101, + 0.015964822843670845, + 0.024848027154803276, + 0.0279752928763628, + 0.016467992216348648, + 0.021442554891109467, + 0.024563388898968697, + 0.025267425924539566, + 0.005909536965191364, + 0.011569507420063019, + 0.007300520781427622, + 0.006090064067393541, + 0.008942357264459133, + 0.01681593805551529, + 0.009363534860312939, + 0.010524673387408257, + 0.007011733949184418, + 0.007451969664543867, + 0.00980334635823965, + 0.006703717168420553, + 0.00591699406504631, + 0.008315067738294601, + 0.011465867049992085, + 0.00919372495263815, + 0.00775276031345129, + 0.004470227286219597, + 0.005029110237956047, + 0.022693227976560593, + 0.004185187630355358, + 0.0056264824233949184, + 0.012098992243409157, + 0.006838055793195963, + 0.01203719712793827, + 0.011613436043262482, + 0.014489857479929924, + 0.011860588565468788, + 0.016716551035642624, + 0.015906158834695816, + 0.01954326033592224, + 0.016700930893421173, + 0.07783534377813339, + 0.010595444589853287, + 0.01545529905706644, + 0.009839478880167007, + 0.013421356678009033, + 0.11047161370515823, + 0.013243294321000576, + 0.019924862310290337, + 0.00604553148150444, + 0.00478129368275404, + 0.003961994778364897, + 0.005658947862684727, + 0.0072340876795351505, + 0.011046743020415306, + 0.01752987876534462, + 0.023111972957849503, + 0.0037494520656764507, + 0.0074902973137795925, + 0.0039439029060304165, + 0.0029348433017730713, + 0.004707367159426212, + 0.005841331090778112, + 0.005841468460857868, + 0.004890334326773882, + 0.006759791634976864, + 0.0065711066126823425, + 0.0045646317303180695, + 0.004175099078565836, + 0.008064107969403267, + 0.006797613110393286, + 0.006227259989827871, + 0.00414193794131279, + 0.005637757945805788, + 0.010652005672454834, + 0.0062459721229970455, + 0.007992984727025032, + 0.00935987289994955, + 0.0089140385389328, + 0.008065483532845974, + 0.004687782842665911 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 4 + ], + "selected_reward": -0.02057950384914875 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 static medium shot of a yellow backhoe arm?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the backhoe arm show signs of wear including oil stains, light scratches, and weathered paint?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the toothed metal bucket scoop up a mound of dark soil mixed with jagged gray rocks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by harsh midday sun creating sharp shadows and glints on the metal?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the background consist of a vast, cloudless, brilliant blue sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03808547928929329 + ], + "opposite_logprob_tokens": [ + -3.288085460662842 + ], + "logprob_reward": -0.03808547928929329, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03808547928929329, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16707947850227356, + 0.07870625704526901, + 0.045581262558698654, + 0.04044381156563759, + 0.0734744742512703, + 0.2855457067489624, + 0.17590656876564026, + 0.06648099422454834, + 0.0344453826546669, + 0.032336026430130005 + ], + "credit_map": [ + 0.014849802479147911, + 0.008984612300992012, + 0.021383749321103096, + 0.020958703011274338, + 0.01656181551516056, + 0.02408841997385025, + 0.037399984896183014, + 0.022852405905723572, + 0.007669754326343536, + 0.010692005045711994, + 0.010044741444289684, + 0.010706375353038311, + 0.007622182369232178, + 0.00965090375393629, + 0.010726820677518845, + 0.011593476869165897, + 0.004245993681252003, + 0.006754378322511911, + 0.005394152831286192, + 0.005423133261501789, + 0.0038482879754155874, + 0.006505900528281927, + 0.00771842198446393, + 0.005690994206815958, + 0.00511045241728425, + 0.003703191177919507, + 0.0039024685975164175, + 0.006074743811041117, + 0.003954300656914711, + 0.009297731332480907, + 0.004296232480555773, + 0.004104692488908768, + 0.010634872131049633, + 0.009979789145290852, + 0.008452421985566616, + 0.007327309809625149, + 0.005722943693399429, + 0.01332919206470251, + 0.009779766201972961, + 0.008248180150985718, + 0.08559154719114304, + 0.009895818307995796, + 0.01565198414027691, + 0.00882964488118887, + 0.013573105446994305, + 0.12307552248239517, + 0.011563523672521114, + 0.01736459508538246, + 0.01926354505121708, + 0.014533483423292637, + 0.008064165711402893, + 0.010289878584444523, + 0.014724234119057655, + 0.04187555983662605, + 0.04023490473628044, + 0.02692081220448017, + 0.0058666677214205265, + 0.00473776925355196, + 0.005838675890117884, + 0.005679304711520672, + 0.006143166683614254, + 0.0051348102279007435, + 0.02088259905576706, + 0.012198001146316528, + 0.005575438495725393, + 0.003513485658913851, + 0.0032649391796439886, + 0.002328377915546298, + 0.005943010561168194, + 0.006606260314583778, + 0.003851731074973941, + 0.003362140618264675, + 0.0035259767901152372, + 0.004897316452115774, + 0.0030160865280777216, + 0.007033741567283869, + 0.003515292191877961, + 0.004142938181757927, + 0.003784595290198922, + 0.0024200796615332365 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 6, + 0 + ], + "selected_reward": -0.03808547928929329 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle, static medium shot captures the immense force of a yellow backhoe arm as it drives into the ground. Painted in a lightly weathered industrial yellow paint, the heavy steel arm shows signs of wear, with oil stains and light scratches, yet possesses a mechanical precision. Its toothed metal bucket bites deeply into the earth, and with a deliberate, powerful curl, it scoops up a vast mound of rich, dark soil, intermixed with jagged gray rocks of various sizes. The entire scene is illuminated by the harsh, direct light of the midday sun, set against a vast, cloudless, brilliant blue sky. This brilliant light creates sharp, well-defined shadows and brilliant glints on the arm's metal surface and the gleaming facets of unearthed stone. As the arm begins its steady upward lift, straining slightly under its weight, the atmosphere is one of concentrated, raw industrial power, showcasing the brute force of a machine over the natural landscape.", + "seed": 8824, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000156/group_13/index_3371_seed_8824_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": -0.0458706671372056, + "vlm_logprob": -0.0458706671372056, + "vlm_yes_no_margin": 3.725, + "vlm_token_credit": -0.0458706671372056, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 static medium shot of a yellow backhoe arm?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the backhoe arm show signs of wear including oil stains, light scratches, and weathered paint?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the toothed metal bucket scoop up a mound of dark soil mixed with jagged gray rocks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by harsh midday sun creating sharp shadows and glints on the metal?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the background consist of a vast, cloudless, brilliant blue sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05215862765908241 + ], + "opposite_logprob_tokens": [ + -4.302158832550049 + ], + "logprob_reward": -0.05215862765908241, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.05215862765908241, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07201310992240906, + 0.035382289439439774, + 0.024657338857650757, + 0.038112517446279526, + 0.0423082560300827, + 0.1846798062324524, + 0.11071000248193741, + 0.18760760128498077, + 0.15821224451065063, + 0.14631681144237518 + ], + "credit_map": [ + 0.008327990770339966, + 0.013345812447369099, + 0.006397265009582043, + 0.007273236755281687, + 0.012912977486848831, + 0.010593254119157791, + 0.006191416177898645, + 0.006971160415560007, + 0.0029121406842023134, + 0.003322367323562503, + 0.0024766186252236366, + 0.002659593941643834, + 0.009560853242874146, + 0.00711699016392231, + 0.003578656353056431, + 0.0037550677079707384, + 0.003117114072665572, + 0.003285651560872793, + 0.001994171878322959, + 0.0019534460734575987, + 0.005131965037435293, + 0.003991716541349888, + 0.002728160237893462, + 0.002455115085467696, + 0.005297008901834488, + 0.004459735006093979, + 0.0032848345581442118, + 0.003909680526703596, + 0.00631455285474658, + 0.005517608020454645, + 0.004056697711348534, + 0.0052724010311067104, + 0.005308237392455339, + 0.005912004970014095, + 0.003965412732213736, + 0.004052046220749617, + 0.008977199904620647, + 0.00789934303611517, + 0.0031248941086232662, + 0.0030691183637827635, + 0.005693530198186636, + 0.0076508051715791225, + 0.004326165188103914, + 0.00591917522251606, + 0.07650138437747955, + 0.07626104354858398, + 0.0036316753830760717, + 0.004696019925177097, + 0.015353086404502392, + 0.020650582388043404, + 0.009242851287126541, + 0.011531751602888107, + 0.021893871948122978, + 0.019860897213220596, + 0.007207551971077919, + 0.004969416651874781, + 0.016047704964876175, + 0.03016567789018154, + 0.011344599537551403, + 0.028486885130405426, + 0.03653603047132492, + 0.029400141909718513, + 0.02305888757109642, + 0.012567663565278053, + 0.01922023482620716, + 0.01954854279756546, + 0.01149730384349823, + 0.03448915481567383, + 0.02264859341084957, + 0.02135615609586239, + 0.01718798838555813, + 0.01226427685469389, + 0.02110167220234871, + 0.023605884984135628, + 0.0136268250644207, + 0.017049888148903847, + 0.022394375875592232, + 0.0200868658721447, + 0.014960060827434063, + 0.01349124126136303 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 5, + 8 + ], + "selected_reward": -0.05215862765908241 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 static medium shot of a yellow backhoe arm?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the backhoe arm show signs of wear including oil stains, light scratches, and weathered paint?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the toothed metal bucket scoop up a mound of dark soil mixed with jagged gray rocks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by harsh midday sun creating sharp shadows and glints on the metal?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the background consist of a vast, cloudless, brilliant blue sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.127162903547287 + ], + "opposite_logprob_tokens": [ + -2.1271629333496094 + ], + "logprob_reward": -0.127162903547287, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.127162903547287, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06861092150211334, + 0.043657951056957245, + 0.04322706535458565, + 0.023447005078196526, + 0.052047085016965866, + 0.23314893245697021, + 0.0789860337972641, + 0.24764904379844666, + 0.08967844396829605, + 0.1195475310087204 + ], + "credit_map": [ + 0.006959992460906506, + 0.008766569197177887, + 0.004461228847503662, + 0.0065034544095396996, + 0.012663786299526691, + 0.010520819574594498, + 0.009188970550894737, + 0.009546102955937386, + 0.004030096810311079, + 0.004275554325431585, + 0.005306578241288662, + 0.004795820452272892, + 0.008290843106806278, + 0.006441664416342974, + 0.005096017383038998, + 0.005421379581093788, + 0.007349527440965176, + 0.0075261094607412815, + 0.0035283691249787807, + 0.003870795015245676, + 0.00661814957857132, + 0.00558654498308897, + 0.0042915986850857735, + 0.004455976653844118, + 0.0038723221514374018, + 0.0029064251575618982, + 0.0020587944891303778, + 0.0024874324444681406, + 0.004473478067666292, + 0.003630994586274028, + 0.0018698663916438818, + 0.002147694118320942, + 0.007109536789357662, + 0.008469464257359505, + 0.004350421018898487, + 0.00578291667625308, + 0.009613949805498123, + 0.009209993295371532, + 0.00398007920011878, + 0.0035307249054312706, + 0.006416424177587032, + 0.010663515888154507, + 0.005478151608258486, + 0.005496588535606861, + 0.09756031632423401, + 0.09744081646203995, + 0.004591356962919235, + 0.005501784849911928, + 0.009472930803894997, + 0.016800520941615105, + 0.0052357385866343975, + 0.007006639149039984, + 0.016260966658592224, + 0.015383751131594181, + 0.0051362235099077225, + 0.0036892700009047985, + 0.02711227722465992, + 0.044222526252269745, + 0.015881096944212914, + 0.04512423276901245, + 0.037503667175769806, + 0.03556192293763161, + 0.02669735625386238, + 0.015545979142189026, + 0.012508104555308819, + 0.009869925677776337, + 0.007050506304949522, + 0.02014552801847458, + 0.013520915061235428, + 0.011979232542216778, + 0.00770576111972332, + 0.006898479536175728, + 0.014232084155082703, + 0.021528460085392, + 0.011817156337201595, + 0.014210227876901627, + 0.01793859153985977, + 0.015759466215968132, + 0.01326025277376175, + 0.010801296681165695 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 5, + 9 + ], + "selected_reward": -0.127162903547287 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 static medium shot of a yellow backhoe arm?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the backhoe arm show signs of wear including oil stains, light scratches, and weathered paint?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the toothed metal bucket scoop up a mound of dark soil mixed with jagged gray rocks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by harsh midday sun creating sharp shadows and glints on the metal?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the background consist of a vast, cloudless, brilliant blue sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023297274485230446 + ], + "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": [ + 0.10667914152145386, + 0.15617337822914124, + 0.03545212373137474, + 0.060051750391721725, + 0.0731707513332367, + 0.09577926993370056, + 0.11588999629020691, + 0.18372398614883423, + 0.08882264792919159, + 0.08425698429346085 + ], + "credit_map": [ + 0.011323334649205208, + 0.012356364168226719, + 0.008244608528912067, + 0.012123397551476955, + 0.01602691039443016, + 0.014371424913406372, + 0.014413357712328434, + 0.017819736152887344, + 0.014475703239440918, + 0.017492515966296196, + 0.009931648150086403, + 0.012084311805665493, + 0.046407733112573624, + 0.032923370599746704, + 0.012785525061190128, + 0.010072565637528896, + 0.0030622296035289764, + 0.003948997240513563, + 0.0028339705895632505, + 0.0023384292144328356, + 0.009929266758263111, + 0.006535347085446119, + 0.003292911686003208, + 0.003510973183438182, + 0.006327154114842415, + 0.005724720656871796, + 0.003912199288606644, + 0.005039528012275696, + 0.015396813862025738, + 0.011239616200327873, + 0.004700331948697567, + 0.00771138072013855, + 0.007301132194697857, + 0.0056261639110744, + 0.0051744552329182625, + 0.007019766140729189, + 0.01732068881392479, + 0.016864918172359467, + 0.006984535604715347, + 0.0068790847435593605, + 0.008521812967956066, + 0.013983642682433128, + 0.007414414081722498, + 0.008905849419534206, + 0.02261664904654026, + 0.02409779466688633, + 0.005271990317851305, + 0.004967114422470331, + 0.011807476170361042, + 0.014047101140022278, + 0.007104987278580666, + 0.009939495474100113, + 0.030879313126206398, + 0.025608954951167107, + 0.010114936158061028, + 0.006387724541127682, + 0.023313162848353386, + 0.036434970796108246, + 0.013578477315604687, + 0.028829025104641914, + 0.031458429992198944, + 0.02567179501056671, + 0.013488282449543476, + 0.010949826799333096, + 0.009170329198241234, + 0.011401494964957237, + 0.007231511641293764, + 0.021973466500639915, + 0.010780151933431625, + 0.014419245533645153, + 0.006633164826780558, + 0.0072132800705730915, + 0.009341639466583729, + 0.013292507268488407, + 0.008709202520549297, + 0.011823481880128384, + 0.010373975150287151, + 0.011825754307210445, + 0.008634474128484726, + 0.01025594025850296 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 1, + 6 + ], + "selected_reward": -0.023297274485230446 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 static medium shot of a yellow backhoe arm?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the backhoe arm show signs of wear including oil stains, light scratches, and weathered paint?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the toothed metal bucket scoop up a mound of dark soil mixed with jagged gray rocks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by harsh midday sun creating sharp shadows and glints on the metal?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the background consist of a vast, cloudless, brilliant blue sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014200471341609955 + ], + "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": [ + 0.10724025219678879, + 0.04292026907205582, + 0.12231496721506119, + 0.1944839358329773, + 0.1455257683992386, + 0.03922799602150917, + 0.06912936270236969, + 0.10995227098464966, + 0.07562685012817383, + 0.0935782864689827 + ], + "credit_map": [ + 0.008497296832501888, + 0.01115291565656662, + 0.00822635181248188, + 0.008762783370912075, + 0.02592938207089901, + 0.01928793638944626, + 0.011908902786672115, + 0.013474689796566963, + 0.0033381779212504625, + 0.00429625203832984, + 0.004397760145366192, + 0.004595006350427866, + 0.005784905049949884, + 0.006541769485920668, + 0.0064195431768894196, + 0.007546857465058565, + 0.009048564359545708, + 0.010617479681968689, + 0.007846622727811337, + 0.007297522388398647, + 0.03307603672146797, + 0.023812847211956978, + 0.012723630294203758, + 0.017892269417643547, + 0.01652579940855503, + 0.019315065816044807, + 0.011602364480495453, + 0.01449250616133213, + 0.051037855446338654, + 0.03811538964509964, + 0.01744554564356804, + 0.02594943344593048, + 0.006905618589371443, + 0.008920918218791485, + 0.006059546489268541, + 0.00771730300039053, + 0.04839985817670822, + 0.04726410284638405, + 0.010615052655339241, + 0.009643368422985077, + 0.002022830070927739, + 0.002338261576369405, + 0.0021692002192139626, + 0.0032983894925564528, + 0.012008041143417358, + 0.011601953767240047, + 0.002667247084900737, + 0.003122074529528618, + 0.005442694295197725, + 0.008352259173989296, + 0.003026731079444289, + 0.005055463872849941, + 0.02238154225051403, + 0.016907421872019768, + 0.005090600810945034, + 0.002872650045901537, + 0.015598047524690628, + 0.015573547221720219, + 0.007663892116397619, + 0.014866609126329422, + 0.021146215498447418, + 0.01769045740365982, + 0.010368788614869118, + 0.007044721860438585, + 0.008500459603965282, + 0.007489131763577461, + 0.0072595807723701, + 0.01780126430094242, + 0.010096881538629532, + 0.011120937764644623, + 0.006958384532481432, + 0.0064002107828855515, + 0.010650752112269402, + 0.0158114954829216, + 0.011606364510953426, + 0.009778167121112347, + 0.014275429770350456, + 0.012268592603504658, + 0.011011101305484772, + 0.008176382631063461 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 4, + 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 shot a low-angle static medium shot of a yellow backhoe arm?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the backhoe arm show signs of wear including oil stains, light scratches, and weathered paint?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the toothed metal bucket scoop up a mound of dark soil mixed with jagged gray rocks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by harsh midday sun creating sharp shadows and glints on the metal?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the background consist of a vast, cloudless, brilliant blue sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012534058652818203 + ], + "opposite_logprob_tokens": [ + -4.387534141540527 + ], + "logprob_reward": -0.012534058652818203, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012534058652818203, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05577870458364487, + 0.059493061155080795, + 0.10979726165533066, + 0.13918760418891907, + 0.2203906774520874, + 0.1987229585647583, + 0.10792723298072815, + 0.0669025182723999, + 0.015901455655694008, + 0.025898519903421402 + ], + "credit_map": [ + 0.0055671678856015205, + 0.00549723394215107, + 0.004349720664322376, + 0.00547444075345993, + 0.008629320189356804, + 0.008994009345769882, + 0.0078001804649829865, + 0.009466630406677723, + 0.004772012121975422, + 0.005880399141460657, + 0.005468966439366341, + 0.00580865191295743, + 0.009181544184684753, + 0.009981281124055386, + 0.008086197078227997, + 0.01031401101499796, + 0.00782078504562378, + 0.009976248256862164, + 0.0070743681862950325, + 0.006271681282669306, + 0.029476657509803772, + 0.021212400868535042, + 0.01134627964347601, + 0.016618842259049416, + 0.011235982179641724, + 0.013050185516476631, + 0.007987023331224918, + 0.009578092023730278, + 0.03961576521396637, + 0.028907541185617447, + 0.011167263612151146, + 0.017645740881562233, + 0.012883509509265423, + 0.013693230226635933, + 0.011274708434939384, + 0.016141312196850777, + 0.0676567405462265, + 0.06402159482240677, + 0.017949825152754784, + 0.016769761219620705, + 0.005187210626900196, + 0.00783271249383688, + 0.0031860850285738707, + 0.004189514555037022, + 0.08652282506227493, + 0.08670379221439362, + 0.0025840813759714365, + 0.0025167365092784166, + 0.005994539242237806, + 0.008850581012666225, + 0.004312748089432716, + 0.010175428353250027, + 0.03678447753190994, + 0.02895708940923214, + 0.008446140214800835, + 0.004406226798892021, + 0.008304187096655369, + 0.011254079639911652, + 0.006546515040099621, + 0.011325825937092304, + 0.009688656777143478, + 0.009060901589691639, + 0.006145945284515619, + 0.004576404113322496, + 0.0013750807847827673, + 0.0018419240368530154, + 0.001181535655632615, + 0.0033681087661534548, + 0.002468631835654378, + 0.0025088151451200247, + 0.0016818286385387182, + 0.0014755309093743563, + 0.0036127662751823664, + 0.00513165770098567, + 0.0025826215278357267, + 0.0026816395111382008, + 0.003957181237637997, + 0.0035026944242417812, + 0.0022247282322496176, + 0.0022052305284887552 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 5, + 3 + ], + "selected_reward": -0.012534058652818203 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle, static medium shot captures the immense force of a yellow backhoe arm as it drives into the ground. Painted in a lightly weathered industrial yellow paint, the heavy steel arm shows signs of wear, with oil stains and light scratches, yet possesses a mechanical precision. Its toothed metal bucket bites deeply into the earth, and with a deliberate, powerful curl, it scoops up a vast mound of rich, dark soil, intermixed with jagged gray rocks of various sizes. The entire scene is illuminated by the harsh, direct light of the midday sun, set against a vast, cloudless, brilliant blue sky. This brilliant light creates sharp, well-defined shadows and brilliant glints on the arm's metal surface and the gleaming facets of unearthed stone. As the arm begins its steady upward lift, straining slightly under its weight, the atmosphere is one of concentrated, raw industrial power, showcasing the brute force of a machine over the natural landscape.", + "seed": 8825, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000156/group_13/index_3371_seed_8825_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": -0.030359969474375248, + "vlm_logprob": -0.030359969474375248, + "vlm_yes_no_margin": 3.8, + "vlm_token_credit": -0.030359969474375248, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 static medium shot of a yellow backhoe arm?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the backhoe arm show signs of wear including oil stains, light scratches, and weathered paint?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the toothed metal bucket scoop up a mound of dark soil mixed with jagged gray rocks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by harsh midday sun creating sharp shadows and glints on the metal?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the background consist of a vast, cloudless, brilliant blue sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04968367516994476 + ], + "opposite_logprob_tokens": [ + -3.4246835708618164 + ], + "logprob_reward": -0.04968367516994476, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.04968367516994476, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11792778968811035, + 0.19350461661815643, + 0.12204183638095856, + 0.086885467171669, + 0.032215531915426254, + 0.022462155669927597, + 0.09382230043411255, + 0.07936684787273407, + 0.15037815272808075, + 0.10139523446559906 + ], + "credit_map": [ + 0.014883549883961678, + 0.024953432381153107, + 0.012244750745594501, + 0.01144866831600666, + 0.01731143333017826, + 0.010552660562098026, + 0.015972858294844627, + 0.010560447350144386, + 0.046352919191122055, + 0.024671947583556175, + 0.01615222729742527, + 0.014721622690558434, + 0.017160193994641304, + 0.04944285377860069, + 0.012714248150587082, + 0.012288621626794338, + 0.02337874099612236, + 0.013524995185434818, + 0.011548939161002636, + 0.009730901569128036, + 0.01725851185619831, + 0.023074524477124214, + 0.011758952401578426, + 0.011766289360821247, + 0.009379000402987003, + 0.01228283904492855, + 0.011217590421438217, + 0.011797348037362099, + 0.015044860541820526, + 0.009006098844110966, + 0.008496751077473164, + 0.00966098252683878, + 0.0035713359247893095, + 0.003514583222568035, + 0.00568587938323617, + 0.003460196079686284, + 0.006752850487828255, + 0.003270896850153804, + 0.003050976898521185, + 0.0029088177252560854, + 0.0025053115095943213, + 0.0027044997550547123, + 0.0021827630698680878, + 0.0031214612536132336, + 0.006262266077101231, + 0.0017486342694610357, + 0.0017925892025232315, + 0.002144632861018181, + 0.013852576725184917, + 0.011326505802571774, + 0.0067838807590305805, + 0.009024541825056076, + 0.016705496236681938, + 0.007167289964854717, + 0.014917203225195408, + 0.014044815674424171, + 0.010221848264336586, + 0.010281049646437168, + 0.009095458313822746, + 0.0068241930566728115, + 0.011316961608827114, + 0.007477649487555027, + 0.015433642081916332, + 0.008716056123375893, + 0.025284942239522934, + 0.02107582800090313, + 0.007746996358036995, + 0.00969740841537714, + 0.02964186668395996, + 0.01766597479581833, + 0.018793610855937004, + 0.020471546798944473, + 0.008825433440506458, + 0.011001256294548512, + 0.011950979940593243, + 0.013309832662343979, + 0.010453592985868454, + 0.011806297115981579, + 0.016705336049199104, + 0.017342515289783478 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 8, + 2 + ], + "selected_reward": -0.04968367516994476 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 static medium shot of a yellow backhoe arm?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the backhoe arm show signs of wear including oil stains, light scratches, and weathered paint?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the toothed metal bucket scoop up a mound of dark soil mixed with jagged gray rocks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by harsh midday sun creating sharp shadows and glints on the metal?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the background consist of a vast, cloudless, brilliant blue sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0487048514187336 + ], + "opposite_logprob_tokens": [ + -3.0487048625946045 + ], + "logprob_reward": -0.0487048514187336, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.0487048514187336, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09280908852815628, + 0.19582264125347137, + 0.07987482100725174, + 0.1348729133605957, + 0.04510323330760002, + 0.021091308444738388, + 0.10781773924827576, + 0.12512989342212677, + 0.07215666025876999, + 0.12532170116901398 + ], + "credit_map": [ + 0.008943098597228527, + 0.01951378397643566, + 0.009316523559391499, + 0.010959909297525883, + 0.012794909067451954, + 0.007024480029940605, + 0.01375261414796114, + 0.010503765195608139, + 0.04885078966617584, + 0.02353621833026409, + 0.014057393185794353, + 0.01881183497607708, + 0.02189522236585617, + 0.050414640456438065, + 0.010041325353085995, + 0.008215222507715225, + 0.0164675023406744, + 0.009182985872030258, + 0.006243958603590727, + 0.007529465481638908, + 0.011507447808980942, + 0.01675705797970295, + 0.007131948601454496, + 0.00505444360896945, + 0.01448015682399273, + 0.019648799672722816, + 0.0198647640645504, + 0.0195314921438694, + 0.0180494524538517, + 0.012412053532898426, + 0.014251218177378178, + 0.01663496531546116, + 0.005444060079753399, + 0.005405541975051165, + 0.006970759481191635, + 0.006827961187809706, + 0.005150758661329746, + 0.0034027588553726673, + 0.005312894470989704, + 0.006588499993085861, + 0.0021874383091926575, + 0.0021142989862710238, + 0.0026053700130432844, + 0.005217378959059715, + 0.0027862461283802986, + 0.0016889589605852962, + 0.002558878157287836, + 0.0019327366026118398, + 0.014434700831770897, + 0.011282457038760185, + 0.012568383477628231, + 0.019066743552684784, + 0.012835905887186527, + 0.007101578637957573, + 0.016533872112631798, + 0.013994083739817142, + 0.018509240821003914, + 0.017120879143476486, + 0.0121062733232975, + 0.0120698232203722, + 0.021664945408701897, + 0.011155360378324986, + 0.021643301472067833, + 0.010860062204301357, + 0.009781280532479286, + 0.00787777453660965, + 0.005885127000510693, + 0.005938190035521984, + 0.010694495402276516, + 0.011840958148241043, + 0.009498788975179195, + 0.010640034452080727, + 0.011248900555074215, + 0.010284545831382275, + 0.01838798262178898, + 0.018095800653100014, + 0.00900631956756115, + 0.010970698669552803, + 0.02560414746403694, + 0.0217233058065176 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 3, + 9 + ], + "selected_reward": -0.0487048514187336 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 static medium shot of a yellow backhoe arm?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the backhoe arm show signs of wear including oil stains, light scratches, and weathered paint?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the toothed metal bucket scoop up a mound of dark soil mixed with jagged gray rocks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by harsh midday sun creating sharp shadows and glints on the metal?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the background consist of a vast, cloudless, brilliant blue sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.033704210072755814 + ], + "opposite_logprob_tokens": [ + -3.4087042808532715 + ], + "logprob_reward": -0.033704210072755814, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.033704210072755814, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1754375845193863, + 0.09563986957073212, + 0.20374076068401337, + 0.0714133083820343, + 0.050493933260440826, + 0.025517411530017853, + 0.0426936037838459, + 0.06553905457258224, + 0.1853189468383789, + 0.0842055156826973 + ], + "credit_map": [ + 0.01771608553826809, + 0.03321634605526924, + 0.016141653060913086, + 0.020021263509988785, + 0.03535313159227371, + 0.015520104207098484, + 0.020322155207395554, + 0.017146846279501915, + 0.020583834499120712, + 0.012518268078565598, + 0.009418177418410778, + 0.007814641110599041, + 0.010244595818221569, + 0.021987048909068108, + 0.006223879288882017, + 0.006849427707493305, + 0.06249724328517914, + 0.010705188848078251, + 0.012331144884228706, + 0.013171590864658356, + 0.014220086857676506, + 0.06652260571718216, + 0.010804162360727787, + 0.013488740660250187, + 0.007817608304321766, + 0.010508875362575054, + 0.009162870235741138, + 0.011189458891749382, + 0.009885840117931366, + 0.005819452926516533, + 0.007219281513243914, + 0.009809925220906734, + 0.0057913074269890785, + 0.004764303099364042, + 0.005965397693216801, + 0.00941036269068718, + 0.007613514084368944, + 0.004633869510143995, + 0.005886313039809465, + 0.006428868044167757, + 0.002873143181204796, + 0.002662911545485258, + 0.00415808567777276, + 0.0035044767428189516, + 0.004578805062919855, + 0.0020516104996204376, + 0.0034310317132622004, + 0.002257349668070674, + 0.004656135570257902, + 0.00448418827727437, + 0.0035854135639965534, + 0.004878389183431864, + 0.005772837437689304, + 0.002535028150305152, + 0.01139829121530056, + 0.005383325740695, + 0.006762888748198748, + 0.005502521526068449, + 0.006727633532136679, + 0.0065194363705813885, + 0.010733778588473797, + 0.00575641356408596, + 0.016226187348365784, + 0.007310202345252037, + 0.02472587302327156, + 0.020764179527759552, + 0.010153059847652912, + 0.01308689545840025, + 0.033839769661426544, + 0.029553022235631943, + 0.026048826053738594, + 0.027147343382239342, + 0.006906587164849043, + 0.007739079650491476, + 0.009597549214959145, + 0.014487303793430328, + 0.008628811687231064, + 0.010223455727100372, + 0.012077906168997288, + 0.01454482413828373 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 8, + 0 + ], + "selected_reward": -0.033704210072755814 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 static medium shot of a yellow backhoe arm?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the backhoe arm show signs of wear including oil stains, light scratches, and weathered paint?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the toothed metal bucket scoop up a mound of dark soil mixed with jagged gray rocks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by harsh midday sun creating sharp shadows and glints on the metal?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the background consist of a vast, cloudless, brilliant blue sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01107415184378624 + ], + "opposite_logprob_tokens": [ + -4.511074066162109 + ], + "logprob_reward": -0.01107415184378624, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.01107415184378624, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10369190573692322, + 0.08199073374271393, + 0.0766148641705513, + 0.048894770443439484, + 0.05817737802863121, + 0.020111603662371635, + 0.40233996510505676, + 0.04487829655408859, + 0.09857834130525589, + 0.06472214311361313 + ], + "credit_map": [ + 0.013556857593357563, + 0.017434895038604736, + 0.010322523303329945, + 0.012491056695580482, + 0.013627796433866024, + 0.009216837584972382, + 0.017371004447340965, + 0.009670930914580822, + 0.021361218765378, + 0.006168670486658812, + 0.00621167803183198, + 0.007396051660180092, + 0.005681519862264395, + 0.023249339312314987, + 0.006151022855192423, + 0.00577123137190938, + 0.016583096235990524, + 0.0051333485171198845, + 0.008396257646381855, + 0.004233159590512514, + 0.009191999211907387, + 0.017281709238886833, + 0.0065779974684119225, + 0.009217297658324242, + 0.00524521991610527, + 0.007414279971271753, + 0.007079550996422768, + 0.00746925501152873, + 0.006213514134287834, + 0.004588189069181681, + 0.0054346611723303795, + 0.0054501015692949295, + 0.007607521489262581, + 0.011827753856778145, + 0.00598191050812602, + 0.012935674749314785, + 0.004721213597804308, + 0.004990814253687859, + 0.004791656043380499, + 0.005320835392922163, + 0.002714195754379034, + 0.002798567758873105, + 0.0030678214970976114, + 0.0041101970709860325, + 0.0019473161082714796, + 0.0016642194241285324, + 0.001985597424209118, + 0.0018236886244267225, + 0.05483861267566681, + 0.03944329172372818, + 0.04415074735879898, + 0.07168710231781006, + 0.048220399767160416, + 0.02639988623559475, + 0.06061488762497902, + 0.056985046714544296, + 0.006301240064203739, + 0.0052928077057003975, + 0.004847246687859297, + 0.0052535803988575935, + 0.007177173160016537, + 0.0039607505314052105, + 0.007608475629240274, + 0.004437020979821682, + 0.009604298509657383, + 0.008250962011516094, + 0.009676624089479446, + 0.007668276317417622, + 0.016035353764891624, + 0.012822188436985016, + 0.018994348123669624, + 0.015526290982961655, + 0.005729398224502802, + 0.005012061912566423, + 0.009163177572190762, + 0.008966288529336452, + 0.005681096110492945, + 0.006346398498862982, + 0.011366366408765316, + 0.012457361444830894 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 8 + ], + "selected_reward": -0.01107415184378624 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 static medium shot of a yellow backhoe arm?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the backhoe arm show signs of wear including oil stains, light scratches, and weathered paint?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the toothed metal bucket scoop up a mound of dark soil mixed with jagged gray rocks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by harsh midday sun creating sharp shadows and glints on the metal?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the background consist of a vast, cloudless, brilliant blue sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008632958866655827 + ], + "opposite_logprob_tokens": [ + -4.758633136749268 + ], + "logprob_reward": -0.008632958866655827, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008632958866655827, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04940040782094002, + 0.044439539313316345, + 0.10874667763710022, + 0.05011725425720215, + 0.03806964308023453, + 0.02287992462515831, + 0.25885745882987976, + 0.10210536420345306, + 0.24623721837997437, + 0.07914646714925766 + ], + "credit_map": [ + 0.006118461024016142, + 0.00791166815906763, + 0.004708395805209875, + 0.006408768706023693, + 0.00784074142575264, + 0.004393276292830706, + 0.006586727220565081, + 0.005432366859167814, + 0.014645108953118324, + 0.0026527754962444305, + 0.0028716286178678274, + 0.0019450701074674726, + 0.0024944634642452, + 0.015779007226228714, + 0.0019904363434761763, + 0.002061048988252878, + 0.03215951845049858, + 0.007247498258948326, + 0.006525844801217318, + 0.006587665528059006, + 0.009087632410228252, + 0.0347454659640789, + 0.00560361985117197, + 0.006789432838559151, + 0.005844344384968281, + 0.008350452408194542, + 0.005201201420277357, + 0.008259686641395092, + 0.006669233553111553, + 0.003943779971450567, + 0.004825585521757603, + 0.007022967096418142, + 0.004198675975203514, + 0.005429653916507959, + 0.006053660064935684, + 0.005690139252692461, + 0.006558801978826523, + 0.00231112539768219, + 0.003408096730709076, + 0.004419488832354546, + 0.0030806642025709152, + 0.0035266648046672344, + 0.0037419982254505157, + 0.00312421633861959, + 0.004106342792510986, + 0.0013336819829419255, + 0.0024979973677545786, + 0.0014683578629046679, + 0.04453333839774132, + 0.03387965261936188, + 0.02603508159518242, + 0.036523304879665375, + 0.0296326931566, + 0.01742173172533512, + 0.03633491322398186, + 0.0344967283308506, + 0.018911486491560936, + 0.01808266155421734, + 0.012165266089141369, + 0.011858318001031876, + 0.01345478743314743, + 0.007555393036454916, + 0.013104695826768875, + 0.0069727604277431965, + 0.03160766884684563, + 0.024742981418967247, + 0.02310563623905182, + 0.017372841015458107, + 0.047340936958789825, + 0.03252734616398811, + 0.03015594743192196, + 0.03938385099172592, + 0.007715508341789246, + 0.0068944962695240974, + 0.009852833114564419, + 0.009682142175734043, + 0.008111925795674324, + 0.009859079495072365, + 0.012861678376793861, + 0.014168807305395603 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 8, + 2 + ], + "selected_reward": -0.008632958866655827 + } + ], + "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/0000156/group_14/index_tensor([483, 483])_group_14_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000156/group_14/index_tensor([483, 483])_group_14_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..cb9c91ef11efad800095834febb75e8eba6224f0 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000156/group_14/index_tensor([483, 483])_group_14_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "static, first-person perspective shot overlooks the polished surface of an ornate, dark-grained wooden dressing table, the entire scene rendered in the style of a whimsical fairytale illustration. A soft, ethereal light, like a ray of morning sun, permeates the frame, capturing dust particles in a gentle glow and casting a warm glow over the setting. At its center lies a delicate hairpin, its body meticulously crafted from gleaming, polished silver wire, twisted into the shape of a blooming flower, its center resting on a smooth, pale green jade stone. Initially motionless, the hairpin begins to sway with a life of its own; it begins as a faint, almost imperceptible tremor, then escalates into a visible vibration, causing its silver and jade components to gently rattle against the wood. This magical animation builds as the hairpin lifts ever so slightly, in a delicate, enchanting dance across the short distance of the vanity, creating an atmosphere of quiet wonder and palpable unfolding magic.", + "seed": 15989, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000156/group_14/index_483_seed_15989_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": -0.03130339868366718, + "vlm_logprob": -0.03130339868366718, + "vlm_yes_no_margin": 3.799999952316284, + "vlm_token_credit": -0.03130339868366718, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 static, first-person perspective overlooking a polished wooden dressing table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the table dark-grained and ornate, with a delicate hairpin made of polished silver wire and a pale green jade stone at its center?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene rendered in the style of a whimsical fairytale illustration with soft, ethereal light and visible dust particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the hairpin begin motionless before swaying, vibrating, and rattling against the wood?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the hairpin lift slightly and dance across the surface of the vanity?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04954947903752327 + ], + "opposite_logprob_tokens": [ + -4.174549579620361 + ], + "logprob_reward": -0.04954947903752327, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.04954947903752327, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21291592717170715, + 0.058236733078956604, + 0.20405977964401245, + 0.08934010565280914, + 0.03574937954545021, + 0.043361637741327286, + 0.07405143231153488, + 0.0815267264842987, + 0.09494474530220032, + 0.10581357032060623 + ], + "credit_map": [ + 0.046260107308626175, + 0.042206428945064545, + 0.022579912096261978, + 0.016822006553411484, + 0.020440058782696724, + 0.02197543904185295, + 0.016001904383301735, + 0.026630058884620667, + 0.011724238283932209, + 0.008607883006334305, + 0.006832157261669636, + 0.005754814948886633, + 0.0047562383115291595, + 0.005673230160027742, + 0.0058179679326713085, + 0.009070204570889473, + 0.034923672676086426, + 0.04814700782299042, + 0.020238928496837616, + 0.021378088742494583, + 0.016999322921037674, + 0.020392587408423424, + 0.022150760516524315, + 0.0198293998837471, + 0.011955007910728455, + 0.02030712552368641, + 0.007712110877037048, + 0.009903894737362862, + 0.011964664794504642, + 0.01087120734155178, + 0.007942312397062778, + 0.008683783002197742, + 0.0044228695333004, + 0.006288209464401007, + 0.003734036348760128, + 0.004699749406427145, + 0.00477079302072525, + 0.0039962660521268845, + 0.004312742967158556, + 0.0035247139167040586, + 0.006482400465756655, + 0.008269185200333595, + 0.004427681211382151, + 0.005153580568730831, + 0.005744545720517635, + 0.004883741959929466, + 0.004498749505728483, + 0.003901754505932331, + 0.008419341407716274, + 0.013629784807562828, + 0.008258101530373096, + 0.008676942437887192, + 0.0080691147595644, + 0.00673122052103281, + 0.013032065704464912, + 0.007234863005578518, + 0.010518013499677181, + 0.018375296145677567, + 0.006964110303670168, + 0.008695453405380249, + 0.011144720017910004, + 0.009796170517802238, + 0.009346434846520424, + 0.0066865189000964165, + 0.00994513276964426, + 0.010864648036658764, + 0.008799702860414982, + 0.013403663411736488, + 0.019056910648941994, + 0.011220932938158512, + 0.014260109513998032, + 0.007393650244921446, + 0.014054926112294197, + 0.01504281535744667, + 0.0116776954382658, + 0.019741028547286987, + 0.014854829758405685, + 0.010648688301444054, + 0.011491169221699238, + 0.008302418515086174 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 9 + ], + "selected_reward": -0.04954947903752327 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 static, first-person perspective overlooking a polished wooden dressing table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the table dark-grained and ornate, with a delicate hairpin made of polished silver wire and a pale green jade stone at its center?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene rendered in the style of a whimsical fairytale illustration with soft, ethereal light and visible dust particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the hairpin begin motionless before swaying, vibrating, and rattling against the wood?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the hairpin lift slightly and dance across the surface of the vanity?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03813252970576286 + ], + "opposite_logprob_tokens": [ + -3.288132429122925 + ], + "logprob_reward": -0.03813252970576286, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03813252970576286, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07752952724695206, + 0.05800025537610054, + 0.46369069814682007, + 0.18747873604297638, + 0.0444180928170681, + 0.03239784017205238, + 0.02488243393599987, + 0.05102475732564926, + 0.016195842996239662, + 0.04438178613781929 + ], + "credit_map": [ + 0.011219391599297523, + 0.01872965507209301, + 0.0070909676142036915, + 0.006730435881763697, + 0.0071249352768063545, + 0.010196040384471416, + 0.006511444225907326, + 0.009926659055054188, + 0.010564428754150867, + 0.010995831340551376, + 0.006063373293727636, + 0.004931824281811714, + 0.0061001842841506, + 0.0060984352603554726, + 0.006016813684254885, + 0.007229369133710861, + 0.07328905165195465, + 0.11231260001659393, + 0.04293382540345192, + 0.04917322099208832, + 0.041388217359781265, + 0.04667405039072037, + 0.04834786057472229, + 0.049571890383958817, + 0.02683761902153492, + 0.040161553770303726, + 0.021471146494150162, + 0.020228130742907524, + 0.021197926253080368, + 0.02342386357486248, + 0.013330895453691483, + 0.020827602595090866, + 0.008533068932592869, + 0.007456916384398937, + 0.004874101839959621, + 0.004632393829524517, + 0.0039239730685949326, + 0.004008013755083084, + 0.005009337328374386, + 0.005980286281555891, + 0.004346499685198069, + 0.007455896120518446, + 0.0029943713452667, + 0.0033864667639136314, + 0.0037629290018230677, + 0.003370011458173394, + 0.004061589948832989, + 0.0030200767796486616, + 0.0038357784505933523, + 0.004493530374020338, + 0.0026829440612345934, + 0.0030005055014044046, + 0.0026166443713009357, + 0.002340484643355012, + 0.003449066076427698, + 0.0024634802248328924, + 0.007983713410794735, + 0.010746229439973831, + 0.00445773359388113, + 0.0059214625507593155, + 0.006311818491667509, + 0.006058249622583389, + 0.005401485599577427, + 0.004144064150750637, + 0.0018163967179134488, + 0.0023056052159518003, + 0.001423700712621212, + 0.002290577394887805, + 0.0030060838907957077, + 0.0021202843636274338, + 0.0016558808274567127, + 0.0015773142222315073, + 0.005672370549291372, + 0.007279434707015753, + 0.003358942922204733, + 0.005267123691737652, + 0.009374132379889488, + 0.00524876918643713, + 0.004216561559587717, + 0.003964452538639307 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 3, + 0 + ], + "selected_reward": -0.03813252970576286 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 static, first-person perspective overlooking a polished wooden dressing table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the table dark-grained and ornate, with a delicate hairpin made of polished silver wire and a pale green jade stone at its center?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene rendered in the style of a whimsical fairytale illustration with soft, ethereal light and visible dust particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the hairpin begin motionless before swaying, vibrating, and rattling against the wood?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the hairpin lift slightly and dance across the surface of the vanity?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01819819025695324 + ], + "opposite_logprob_tokens": [ + -4.018198013305664 + ], + "logprob_reward": -0.01819819025695324, + "yes_no_margin_reward": 3.999999761581421, + "token_credit_reward": -0.01819819025695324, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.175597682595253, + 0.07489485293626785, + 0.3264927864074707, + 0.031478215008974075, + 0.045222360640764236, + 0.08235123753547668, + 0.019464673474431038, + 0.023736007511615753, + 0.13687191903591156, + 0.08389025181531906 + ], + "credit_map": [ + 0.034633658826351166, + 0.03817639499902725, + 0.01812383532524109, + 0.01613457128405571, + 0.016950199380517006, + 0.016173802316188812, + 0.015549352392554283, + 0.01985587552189827, + 0.01431160606443882, + 0.017266128212213516, + 0.008035147562623024, + 0.006213217508047819, + 0.0071990713477134705, + 0.0072363209910690784, + 0.006902390159666538, + 0.0077309696935117245, + 0.050811685621738434, + 0.0810575932264328, + 0.03002675250172615, + 0.034290771931409836, + 0.02914939448237419, + 0.03442052751779556, + 0.03130683675408363, + 0.0354292206466198, + 0.004946188069880009, + 0.006813534535467625, + 0.0033921508584171534, + 0.003059736220166087, + 0.003833673195913434, + 0.003528845962136984, + 0.0024895931128412485, + 0.0034144942183047533, + 0.005582475569099188, + 0.007893498986959457, + 0.0041139074601233006, + 0.005892344284802675, + 0.00727335037663579, + 0.005250278860330582, + 0.00552761135622859, + 0.0036888918839395046, + 0.012488294392824173, + 0.02063746564090252, + 0.007586261723190546, + 0.0072355614975094795, + 0.007388269063085318, + 0.00880651455372572, + 0.00902334600687027, + 0.009185530245304108, + 0.002267013769596815, + 0.0029630507342517376, + 0.0020357363391667604, + 0.0028054136782884598, + 0.0028540920466184616, + 0.0023138613905757666, + 0.0026756557635962963, + 0.0015498497523367405, + 0.0031666066497564316, + 0.004391772672533989, + 0.0023894920013844967, + 0.0034059155732393265, + 0.002785064745694399, + 0.0024517627898603678, + 0.0033624316565692425, + 0.0017829622374847531, + 0.016778044402599335, + 0.019754664972424507, + 0.012246979400515556, + 0.018252350389957428, + 0.022579962387681007, + 0.014951042830944061, + 0.02254176139831543, + 0.009767111390829086, + 0.011094660498201847, + 0.013026202097535133, + 0.007868886925280094, + 0.01320689171552658, + 0.014447344467043877, + 0.008849994279444218, + 0.008704107254743576, + 0.006692163180559874 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 8 + ], + "selected_reward": -0.01819819025695324 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 static, first-person perspective overlooking a polished wooden dressing table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the table dark-grained and ornate, with a delicate hairpin made of polished silver wire and a pale green jade stone at its center?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene rendered in the style of a whimsical fairytale illustration with soft, ethereal light and visible dust particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the hairpin begin motionless before swaying, vibrating, and rattling against the wood?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the hairpin lift slightly and dance across the surface of the vanity?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012535706162452698 + ], + "opposite_logprob_tokens": [ + -4.387535572052002 + ], + "logprob_reward": -0.012535706162452698, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012535706162452698, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06751230359077454, + 0.05183907225728035, + 0.4099370837211609, + 0.0816585049033165, + 0.026441017165780067, + 0.0630892962217331, + 0.029156267642974854, + 0.04613213986158371, + 0.10535283386707306, + 0.11888148635625839 + ], + "credit_map": [ + 0.010921558365225792, + 0.016485782340168953, + 0.005951923318207264, + 0.006699304096400738, + 0.006593064405024052, + 0.007483797147870064, + 0.005321338772773743, + 0.008055539801716805, + 0.007138176821172237, + 0.012454796582460403, + 0.003956300672143698, + 0.004051949363201857, + 0.007795811165124178, + 0.006335856392979622, + 0.004723844118416309, + 0.005382339004427195, + 0.06424816697835922, + 0.09497472643852234, + 0.037757012993097305, + 0.044446174055337906, + 0.038656871765851974, + 0.03965781256556511, + 0.04545283317565918, + 0.04474349692463875, + 0.01117696426808834, + 0.016760103404521942, + 0.008045489899814129, + 0.009198064915835857, + 0.010724307969212532, + 0.009784978814423084, + 0.008526705205440521, + 0.007441896945238113, + 0.0028778179548680782, + 0.004777058959007263, + 0.0021091375965625048, + 0.0037893373519182205, + 0.004365533124655485, + 0.0031056476291269064, + 0.0029987182933837175, + 0.0024177669547498226, + 0.008858461864292622, + 0.012394526042044163, + 0.005998648703098297, + 0.006241180468350649, + 0.007646541576832533, + 0.006678096950054169, + 0.008568798191845417, + 0.006703044753521681, + 0.00304289348423481, + 0.003982288297265768, + 0.0027458539698272943, + 0.004017003811895847, + 0.005158043932169676, + 0.0032576434314250946, + 0.0041509270668029785, + 0.002801614347845316, + 0.006527591962367296, + 0.007988804951310158, + 0.0042929379269480705, + 0.0056217447854578495, + 0.006036294158548117, + 0.005169975105673075, + 0.0063832001760602, + 0.004111590329557657, + 0.01672133058309555, + 0.015045776031911373, + 0.00936826504766941, + 0.013410575687885284, + 0.019340291619300842, + 0.011898180469870567, + 0.011467973701655865, + 0.008100444450974464, + 0.01190156675875187, + 0.015045728534460068, + 0.010085832327604294, + 0.019350450485944748, + 0.02030463144183159, + 0.011716585606336594, + 0.022029323503375053, + 0.008447371423244476 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 9, + 8 + ], + "selected_reward": -0.012535706162452698 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 static, first-person perspective overlooking a polished wooden dressing table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the table dark-grained and ornate, with a delicate hairpin made of polished silver wire and a pale green jade stone at its center?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene rendered in the style of a whimsical fairytale illustration with soft, ethereal light and visible dust particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the hairpin begin motionless before swaying, vibrating, and rattling against the wood?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the hairpin lift slightly and dance across the surface of the vanity?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.038101088255643845 + ], + "opposite_logprob_tokens": [ + -3.2881011962890625 + ], + "logprob_reward": -0.038101088255643845, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.038101088255643845, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07755527645349503, + 0.03478928655385971, + 0.47076869010925293, + 0.15521277487277985, + 0.023424627259373665, + 0.01724488101899624, + 0.030275141820311546, + 0.032228950411081314, + 0.08528848737478256, + 0.07321187853813171 + ], + "credit_map": [ + 0.013802497647702694, + 0.01278915349394083, + 0.008044417947530746, + 0.007628343999385834, + 0.009549696929752827, + 0.00801259558647871, + 0.008723585866391659, + 0.009004976600408554, + 0.006761443335562944, + 0.0067244176752865314, + 0.0037832180969417095, + 0.0029423334635794163, + 0.003905751509591937, + 0.0033877671230584383, + 0.0035243602469563484, + 0.0037599937058985233, + 0.07325268536806107, + 0.11070601642131805, + 0.04301721602678299, + 0.050610169768333435, + 0.04439590126276016, + 0.046398747712373734, + 0.05139446258544922, + 0.05099347233772278, + 0.02194567583501339, + 0.03306252136826515, + 0.017363078892230988, + 0.015583939850330353, + 0.017726201564073563, + 0.019418051466345787, + 0.01158086210489273, + 0.018532443791627884, + 0.003223973559215665, + 0.004002692177891731, + 0.002348846523091197, + 0.0030857338570058346, + 0.0025545579846948385, + 0.0021529069636017084, + 0.0035863127559423447, + 0.00246960180811584, + 0.0027398692909628153, + 0.003304590005427599, + 0.0015577564481645823, + 0.00190575304441154, + 0.0025062416680157185, + 0.0017350097186863422, + 0.0020219660364091396, + 0.0014736931771039963, + 0.0028193348553031683, + 0.004279310815036297, + 0.0031501739285886288, + 0.004087215755134821, + 0.004809196572750807, + 0.0033241931814700365, + 0.005040379706770182, + 0.0027653337456285954, + 0.002781791379675269, + 0.004643853288143873, + 0.0031732439529150724, + 0.005746933165937662, + 0.0047144098207354546, + 0.0033760482911020517, + 0.0055595035664737225, + 0.002233167178928852, + 0.012060625478625298, + 0.011075306683778763, + 0.007326560094952583, + 0.011095329187810421, + 0.014725237153470516, + 0.009521369822323322, + 0.011197417974472046, + 0.008286634460091591, + 0.007835006341338158, + 0.008944833651185036, + 0.006209622137248516, + 0.010280628688633442, + 0.01431181374937296, + 0.008257661946117878, + 0.012736812233924866, + 0.004635501652956009 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 3, + 8 + ], + "selected_reward": -0.038101088255643845 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, first-person perspective shot overlooks the polished surface of an ornate, dark-grained wooden dressing table, the entire scene rendered in the style of a whimsical fairytale illustration. A soft, ethereal light, like a ray of morning sun, permeates the frame, capturing dust particles in a gentle glow and casting a warm glow over the setting. At its center lies a delicate hairpin, its body meticulously crafted from gleaming, polished silver wire, twisted into the shape of a blooming flower, its center resting on a smooth, pale green jade stone. Initially motionless, the hairpin begins to sway with a life of its own; it begins as a faint, almost imperceptible tremor, then escalates into a visible vibration, causing its silver and jade components to gently rattle against the wood. This magical animation builds as the hairpin lifts ever so slightly, in a delicate, enchanting dance across the short distance of the vanity, creating an atmosphere of quiet wonder and palpable unfolding magic.", + "seed": 15990, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000156/group_14/index_483_seed_15990_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": -0.0611617598682642, + "vlm_logprob": -0.0611617598682642, + "vlm_yes_no_margin": 3.4999999523162844, + "vlm_token_credit": -0.0611617598682642, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 static, first-person perspective overlooking a polished wooden dressing table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the table dark-grained and ornate, with a delicate hairpin made of polished silver wire and a pale green jade stone at its center?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene rendered in the style of a whimsical fairytale illustration with soft, ethereal light and visible dust particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the hairpin begin motionless before swaying, vibrating, and rattling against the wood?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the hairpin lift slightly and dance across the surface of the vanity?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04132525622844696 + ], + "opposite_logprob_tokens": [ + -4.041325092315674 + ], + "logprob_reward": -0.04132525622844696, + "yes_no_margin_reward": 3.999999761581421, + "token_credit_reward": -0.04132525622844696, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.34308236837387085, + 0.06755854934453964, + 0.05019451677799225, + 0.03139219433069229, + 0.03447635471820831, + 0.06787125021219254, + 0.11701905727386475, + 0.09925947338342667, + 0.08622892946004868, + 0.10291732102632523 + ], + "credit_map": [ + 0.03916996717453003, + 0.07265664637088776, + 0.04355575516819954, + 0.027779534459114075, + 0.02154846489429474, + 0.05758928880095482, + 0.019738543778657913, + 0.061044178903102875, + 0.007670221850275993, + 0.013812306337058544, + 0.007538869511336088, + 0.005136475898325443, + 0.004207659512758255, + 0.016460943967103958, + 0.004141290672123432, + 0.008590780198574066, + 0.005345324520021677, + 0.010922207497060299, + 0.006177815143018961, + 0.0036155746784061193, + 0.0033288674894720316, + 0.011759007349610329, + 0.001955262152478099, + 0.00709045585244894, + 0.0047758473083376884, + 0.003761146916076541, + 0.006100212689489126, + 0.0036095953546464443, + 0.002412623493000865, + 0.0024140796158462763, + 0.0023915087804198265, + 0.005927179008722305, + 0.003457317128777504, + 0.008281170390546322, + 0.004175993148237467, + 0.0031367342453449965, + 0.002628580667078495, + 0.005757422186434269, + 0.002654659328982234, + 0.00438448041677475, + 0.011029168963432312, + 0.00730214174836874, + 0.009149054996669292, + 0.008617332205176353, + 0.007018888369202614, + 0.008353764191269875, + 0.006175306160002947, + 0.010225591249763966, + 0.012282811105251312, + 0.013744079507887363, + 0.008772710338234901, + 0.01105007715523243, + 0.028295066207647324, + 0.014211487956345081, + 0.017192086204886436, + 0.011470739729702473, + 0.013017090037465096, + 0.013509911485016346, + 0.009722886607050896, + 0.01782400533556938, + 0.010381601750850677, + 0.010448532178997993, + 0.014469049870967865, + 0.009886397048830986, + 0.012680529616773129, + 0.008706768043339252, + 0.007265027146786451, + 0.012768671847879887, + 0.011576895602047443, + 0.015499461442232132, + 0.009402817115187645, + 0.0083287563174963, + 0.014025827869772911, + 0.01235109195113182, + 0.018008101731538773, + 0.011588318273425102, + 0.009390573017299175, + 0.01082231942564249, + 0.0097673824056983, + 0.016963714733719826 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 9 + ], + "selected_reward": -0.04132525622844696 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 static, first-person perspective overlooking a polished wooden dressing table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the table dark-grained and ornate, with a delicate hairpin made of polished silver wire and a pale green jade stone at its center?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene rendered in the style of a whimsical fairytale illustration with soft, ethereal light and visible dust particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the hairpin begin motionless before swaying, vibrating, and rattling against the wood?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the hairpin lift slightly and dance across the surface of the vanity?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008653167635202408 + ], + "opposite_logprob_tokens": [ + -4.758653163909912 + ], + "logprob_reward": -0.008653167635202408, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008653167635202408, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16451120376586914, + 0.13344819843769073, + 0.0781220942735672, + 0.0478171780705452, + 0.07311415672302246, + 0.09266480803489685, + 0.14569786190986633, + 0.08559948951005936, + 0.07875663042068481, + 0.10026829689741135 + ], + "credit_map": [ + 0.022583864629268646, + 0.021308831870555878, + 0.020789578557014465, + 0.027846187353134155, + 0.01371137984097004, + 0.022768327966332436, + 0.015356713905930519, + 0.020146330818533897, + 0.013132665306329727, + 0.019914668053388596, + 0.009998016990721226, + 0.012124288827180862, + 0.011381982825696468, + 0.03525521606206894, + 0.010986014269292355, + 0.020655348896980286, + 0.007639276795089245, + 0.013909450732171535, + 0.005787951871752739, + 0.007480221334844828, + 0.006502559874206781, + 0.01942034251987934, + 0.0067550744861364365, + 0.010627221316099167, + 0.004353196360170841, + 0.014304754324257374, + 0.003451265161857009, + 0.003636628622189164, + 0.003515496850013733, + 0.011006954126060009, + 0.002835564548149705, + 0.004713322967290878, + 0.006631959229707718, + 0.021339794620871544, + 0.005914529785513878, + 0.006480047944933176, + 0.005333079490810633, + 0.01773076131939888, + 0.003963143564760685, + 0.00572084728628397, + 0.014369367621839046, + 0.010869292542338371, + 0.00968683697283268, + 0.015578651800751686, + 0.008845333941280842, + 0.01152867916971445, + 0.008207865059375763, + 0.013578793033957481, + 0.015173288062214851, + 0.018679337576031685, + 0.013171566650271416, + 0.030220594257116318, + 0.017729731276631355, + 0.012639309279620647, + 0.019454268738627434, + 0.01862979494035244, + 0.01162694487720728, + 0.010271548293530941, + 0.007515902165323496, + 0.014710077084600925, + 0.010145073756575584, + 0.00942754466086626, + 0.013614293187856674, + 0.008288117125630379, + 0.01004370953887701, + 0.014779276214540005, + 0.006472920998930931, + 0.009139597415924072, + 0.011366492137312889, + 0.007035382557660341, + 0.01334163174033165, + 0.006577625405043364, + 0.01245846226811409, + 0.0126519575715065, + 0.01199998427182436, + 0.011132432147860527, + 0.01471028383821249, + 0.011109165847301483, + 0.01494131051003933, + 0.01126471534371376 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 1 + ], + "selected_reward": -0.008653167635202408 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 static, first-person perspective overlooking a polished wooden dressing table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the table dark-grained and ornate, with a delicate hairpin made of polished silver wire and a pale green jade stone at its center?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene rendered in the style of a whimsical fairytale illustration with soft, ethereal light and visible dust particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the hairpin begin motionless before swaying, vibrating, and rattling against the wood?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the hairpin lift slightly and dance across the surface of the vanity?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03370582312345505 + ], + "opposite_logprob_tokens": [ + -3.408705711364746 + ], + "logprob_reward": -0.03370582312345505, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03370582312345505, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16522550582885742, + 0.03212319314479828, + 0.022509519010782242, + 0.017221417278051376, + 0.02746201865375042, + 0.142296701669693, + 0.09226706624031067, + 0.07398488372564316, + 0.046309422701597214, + 0.38060030341148376 + ], + "credit_map": [ + 0.02360970340669155, + 0.018871719017624855, + 0.02298903465270996, + 0.01903236284852028, + 0.011506465263664722, + 0.015837274491786957, + 0.012845480814576149, + 0.04053346812725067, + 0.003991716541349888, + 0.0038850915152579546, + 0.003197903512045741, + 0.004382578656077385, + 0.0027697146870195866, + 0.005123014561831951, + 0.0026542271953076124, + 0.0061189476400613785, + 0.0021801828406751156, + 0.0044471947476267815, + 0.0026709723751991987, + 0.002259590430185199, + 0.001717928797006607, + 0.004695986397564411, + 0.0017057225340977311, + 0.002831941470503807, + 0.0022379558067768812, + 0.0027096818666905165, + 0.0019069084664806724, + 0.0021733760368078947, + 0.0011711055412888527, + 0.002888616407290101, + 0.0013596267672255635, + 0.002774145919829607, + 0.004058518912643194, + 0.003179156221449375, + 0.0031331046484410763, + 0.0037382792215794325, + 0.002288584131747484, + 0.003987804986536503, + 0.0019978764466941357, + 0.0050786943174898624, + 0.017594318836927414, + 0.018175484612584114, + 0.016436276957392693, + 0.01554135326296091, + 0.017128175124526024, + 0.023408446460962296, + 0.01301069837063551, + 0.021001942455768585, + 0.009075803682208061, + 0.013339175842702389, + 0.009554723277688026, + 0.017974216490983963, + 0.00966093223541975, + 0.007976530119776726, + 0.012494901195168495, + 0.012190783396363258, + 0.007282670121639967, + 0.012073655612766743, + 0.006230509839951992, + 0.012183419428765774, + 0.008791022934019566, + 0.01099107414484024, + 0.009529681876301765, + 0.006902850698679686, + 0.006088943220674992, + 0.011202055960893631, + 0.005021169316023588, + 0.005442773923277855, + 0.00515095004811883, + 0.0038266561459749937, + 0.006297317799180746, + 0.0032795583829283714, + 0.03680731728672981, + 0.08479990810155869, + 0.04230572283267975, + 0.02477629855275154, + 0.03226276487112045, + 0.09645126760005951, + 0.02351648546755314, + 0.03968055546283722 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 5 + ], + "selected_reward": -0.03370582312345505 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 static, first-person perspective overlooking a polished wooden dressing table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the table dark-grained and ornate, with a delicate hairpin made of polished silver wire and a pale green jade stone at its center?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene rendered in the style of a whimsical fairytale illustration with soft, ethereal light and visible dust particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the hairpin begin motionless before swaying, vibrating, and rattling against the wood?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the hairpin lift slightly and dance across the surface of the vanity?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020572146400809288 + ], + "opposite_logprob_tokens": [ + -3.8955721855163574 + ], + "logprob_reward": -0.020572146400809288, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020572146400809288, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18601524829864502, + 0.046188872307538986, + 0.045767009258270264, + 0.026376832276582718, + 0.03285890445113182, + 0.09837767481803894, + 0.0975552350282669, + 0.11417233198881149, + 0.09734226763248444, + 0.2553456127643585 + ], + "credit_map": [ + 0.018292702734470367, + 0.03991977497935295, + 0.01741367019712925, + 0.018198389559984207, + 0.016605045646429062, + 0.039380885660648346, + 0.014790849760174751, + 0.021413912996649742, + 0.006545332260429859, + 0.006456261500716209, + 0.004276189021766186, + 0.005703764967620373, + 0.005265065003186464, + 0.006544738542288542, + 0.00411281269043684, + 0.007284705527126789, + 0.003358204383403063, + 0.007051391992717981, + 0.004244845826178789, + 0.004482879303395748, + 0.005788596346974373, + 0.010575716383755207, + 0.005368176382035017, + 0.004897198174148798, + 0.0035972807090729475, + 0.004198594484478235, + 0.0023450793232768774, + 0.0027634024154394865, + 0.002985789906233549, + 0.002572642406448722, + 0.002755870809778571, + 0.00515817292034626, + 0.004759687930345535, + 0.0036298930644989014, + 0.0028771671932190657, + 0.002959062345325947, + 0.002551394049078226, + 0.00593205401673913, + 0.002105242107063532, + 0.008044402115046978, + 0.009292659349739552, + 0.011449089273810387, + 0.008954828605055809, + 0.011686662212014198, + 0.014170780777931213, + 0.014836969785392284, + 0.012300277128815651, + 0.015686411410570145, + 0.009001926518976688, + 0.010776488110423088, + 0.009293682873249054, + 0.01724272035062313, + 0.013788756914436817, + 0.011380821466445923, + 0.012299551628530025, + 0.013771296478807926, + 0.01262490451335907, + 0.01435044128447771, + 0.009888621047139168, + 0.012402382679283619, + 0.017497286200523376, + 0.016653427854180336, + 0.013437856920063496, + 0.01731742173433304, + 0.009139824658632278, + 0.028413180261850357, + 0.007577789481729269, + 0.011187950149178505, + 0.01055852323770523, + 0.01262346189469099, + 0.010961158201098442, + 0.006880383938550949, + 0.023656431585550308, + 0.03930991515517235, + 0.027618952095508575, + 0.027801506221294403, + 0.03590807318687439, + 0.03624711185693741, + 0.036423999816179276, + 0.02837962470948696 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 7 + ], + "selected_reward": -0.020572146400809288 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 static, first-person perspective overlooking a polished wooden dressing table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the table dark-grained and ornate, with a delicate hairpin made of polished silver wire and a pale green jade stone at its center?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene rendered in the style of a whimsical fairytale illustration with soft, ethereal light and visible dust particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the hairpin begin motionless before swaying, vibrating, and rattling against the wood?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the hairpin lift slightly and dance across the surface of the vanity?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2015524059534073 + ], + "opposite_logprob_tokens": [ + -1.701552391052246 + ], + "logprob_reward": -0.2015524059534073, + "yes_no_margin_reward": 1.5, + "token_credit_reward": -0.2015524059534073, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18720443546772003, + 0.05906493961811066, + 0.04458490014076233, + 0.027007807046175003, + 0.03801514208316803, + 0.06909231841564178, + 0.08213115483522415, + 0.10678938776254654, + 0.1270170360803604, + 0.2590928375720978 + ], + "credit_map": [ + 0.0231829434633255, + 0.023310212418437004, + 0.026497915387153625, + 0.026263348758220673, + 0.014703020453453064, + 0.02147897705435753, + 0.01541154645383358, + 0.03635647892951965, + 0.006566766649484634, + 0.00974868331104517, + 0.008091133087873459, + 0.006745923776179552, + 0.0059498208574950695, + 0.008362521417438984, + 0.005719039123505354, + 0.00788105744868517, + 0.004006814677268267, + 0.008654994890093803, + 0.004604985471814871, + 0.003454320365563035, + 0.004299635533243418, + 0.012050269171595573, + 0.0033670885022729635, + 0.004146793391555548, + 0.003255074378103018, + 0.0051615298725664616, + 0.0045716529712080956, + 0.003738171886652708, + 0.0025666537694633007, + 0.002388822613283992, + 0.0025947019457817078, + 0.0027312021702528, + 0.004205591510981321, + 0.006947897840291262, + 0.0031791802030056715, + 0.002438206458464265, + 0.0030879920814186335, + 0.009050101973116398, + 0.002147659892216325, + 0.0069585153833031654, + 0.007032901048660278, + 0.009114060550928116, + 0.006268603261560202, + 0.008252331055700779, + 0.010144910775125027, + 0.010722869075834751, + 0.008375694043934345, + 0.009180951863527298, + 0.008865906856954098, + 0.012193510308861732, + 0.00741977384313941, + 0.012654013931751251, + 0.011662652716040611, + 0.009034701623022556, + 0.012345554307103157, + 0.007955053821206093, + 0.009518355131149292, + 0.013162591494619846, + 0.009087296202778816, + 0.01373481284826994, + 0.020395535975694656, + 0.010469677858054638, + 0.01969710737466812, + 0.010724017396569252, + 0.010974504984915257, + 0.03134283795952797, + 0.012201154604554176, + 0.015519077889621258, + 0.009797453880310059, + 0.017130671069025993, + 0.010107415728271008, + 0.019943926483392715, + 0.02612554468214512, + 0.03231565281748772, + 0.026440365239977837, + 0.02715112827718258, + 0.0455823689699173, + 0.02897498942911625, + 0.047512561082839966, + 0.024990255013108253 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 8 + ], + "selected_reward": -0.2015524059534073 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, first-person perspective shot overlooks the polished surface of an ornate, dark-grained wooden dressing table, the entire scene rendered in the style of a whimsical fairytale illustration. A soft, ethereal light, like a ray of morning sun, permeates the frame, capturing dust particles in a gentle glow and casting a warm glow over the setting. At its center lies a delicate hairpin, its body meticulously crafted from gleaming, polished silver wire, twisted into the shape of a blooming flower, its center resting on a smooth, pale green jade stone. Initially motionless, the hairpin begins to sway with a life of its own; it begins as a faint, almost imperceptible tremor, then escalates into a visible vibration, causing its silver and jade components to gently rattle against the wood. This magical animation builds as the hairpin lifts ever so slightly, in a delicate, enchanting dance across the short distance of the vanity, creating an atmosphere of quiet wonder and palpable unfolding magic.", + "seed": 15991, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000156/group_14/index_483_seed_15991_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": -0.029212336987257004, + "vlm_logprob": -0.029212336987257004, + "vlm_yes_no_margin": 3.95, + "vlm_token_credit": -0.029212336987257004, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 static, first-person perspective overlooking a polished wooden dressing table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the table dark-grained and ornate, with a delicate hairpin made of polished silver wire and a pale green jade stone at its center?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene rendered in the style of a whimsical fairytale illustration with soft, ethereal light and visible dust particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the hairpin begin motionless before swaying, vibrating, and rattling against the wood?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the hairpin lift slightly and dance across the surface of the vanity?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04938507825136185 + ], + "opposite_logprob_tokens": [ + -3.7993850708007812 + ], + "logprob_reward": -0.04938507825136185, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.04938507825136185, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.35866695642471313, + 0.08494941890239716, + 0.047024160623550415, + 0.06396616995334625, + 0.02701815776526928, + 0.07399103790521622, + 0.05150184407830238, + 0.07328588515520096, + 0.18452729284763336, + 0.035069044679403305 + ], + "credit_map": [ + 0.06061970070004463, + 0.02928736060857773, + 0.07257600873708725, + 0.027591360732913017, + 0.022697869688272476, + 0.04002450034022331, + 0.04866480827331543, + 0.057205356657505035, + 0.011591603048145771, + 0.01193563174456358, + 0.013020696118474007, + 0.010864011012017727, + 0.006966083310544491, + 0.008440099656581879, + 0.012901702895760536, + 0.009229592978954315, + 0.0095764035359025, + 0.0038268365897238255, + 0.007961579598486423, + 0.005244013387709856, + 0.003411457408219576, + 0.005308794789016247, + 0.004639283753931522, + 0.007055792026221752, + 0.008738800883293152, + 0.017078109085559845, + 0.007023020647466183, + 0.009585680440068245, + 0.006469706539064646, + 0.0054174913093447685, + 0.005189360119402409, + 0.004464001394808292, + 0.0028161092195659876, + 0.004045436158776283, + 0.0033537407871335745, + 0.0035152812488377094, + 0.0030371916946023703, + 0.003146386006847024, + 0.003458054270595312, + 0.0036459583789110184, + 0.011171571910381317, + 0.011683502234518528, + 0.011975451372563839, + 0.008586576208472252, + 0.007943118922412395, + 0.007200706750154495, + 0.007991766557097435, + 0.007438340224325657, + 0.006560040172189474, + 0.008868740871548653, + 0.009713328443467617, + 0.005918503273278475, + 0.0036414354108273983, + 0.005162976216524839, + 0.004783086944371462, + 0.006853729952126741, + 0.0056442576460540295, + 0.008372627198696136, + 0.00826586876064539, + 0.008266719989478588, + 0.011498311534523964, + 0.009692007675766945, + 0.009937576949596405, + 0.011608509346842766, + 0.0190056711435318, + 0.028125455603003502, + 0.02467229589819908, + 0.022634822875261307, + 0.019841285422444344, + 0.021625587716698647, + 0.02413065731525421, + 0.02449151873588562, + 0.003392979968339205, + 0.0042682308703660965, + 0.004509742371737957, + 0.0047373282723128796, + 0.005142464302480221, + 0.004276648163795471, + 0.005165862385183573, + 0.0035757904406636953 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 1 + ], + "selected_reward": -0.04938507825136185 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 static, first-person perspective overlooking a polished wooden dressing table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the table dark-grained and ornate, with a delicate hairpin made of polished silver wire and a pale green jade stone at its center?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene rendered in the style of a whimsical fairytale illustration with soft, ethereal light and visible dust particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the hairpin begin motionless before swaying, vibrating, and rattling against the wood?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the hairpin lift slightly and dance across the surface of the vanity?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008654704317450523 + ], + "opposite_logprob_tokens": [ + -4.758654594421387 + ], + "logprob_reward": -0.008654704317450523, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008654704317450523, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18730854988098145, + 0.1327681541442871, + 0.04828450828790665, + 0.09203969687223434, + 0.027031371369957924, + 0.03731824457645416, + 0.09422177821397781, + 0.31673792004585266, + 0.03503885120153427, + 0.029250994324684143 + ], + "credit_map": [ + 0.013149803504347801, + 0.007218488026410341, + 0.01564079336822033, + 0.008791031315922737, + 0.00735363457351923, + 0.06081584841012955, + 0.04986587166786194, + 0.02447306178510189, + 0.012329913675785065, + 0.008225672878324986, + 0.04225071147084236, + 0.010395714081823826, + 0.0069169956259429455, + 0.02245856449007988, + 0.012293502688407898, + 0.01789706200361252, + 0.007388226687908173, + 0.004681477788835764, + 0.005402219481766224, + 0.00509560015052557, + 0.004609634634107351, + 0.006334594450891018, + 0.0059413425624370575, + 0.008831406012177467, + 0.011720186099410057, + 0.016013391315937042, + 0.011867539957165718, + 0.008795051835477352, + 0.010856428183615208, + 0.007934917695820332, + 0.008538637310266495, + 0.016313541680574417, + 0.002586639253422618, + 0.0026891431771218777, + 0.004297588020563126, + 0.004012485034763813, + 0.0034178607165813446, + 0.002494002692401409, + 0.0033792986068874598, + 0.004154351074248552, + 0.006638760212808847, + 0.006031204480677843, + 0.006360804196447134, + 0.0038030303549021482, + 0.0027833585627377033, + 0.004381200764328241, + 0.0031544307712465525, + 0.0041654519736766815, + 0.007644779048860073, + 0.014306486584246159, + 0.015429061837494373, + 0.013517457991838455, + 0.010254026390612125, + 0.008145363070070744, + 0.012529868632555008, + 0.01239472720772028, + 0.02781720459461212, + 0.04509580507874489, + 0.03492879867553711, + 0.037897586822509766, + 0.040287084877491, + 0.04063272848725319, + 0.044152144342660904, + 0.045926522463560104, + 0.0032155951485037804, + 0.004112417809665203, + 0.004728074185550213, + 0.004099167883396149, + 0.002974053146317601, + 0.003953787498176098, + 0.007000803016126156, + 0.004954944830387831, + 0.002077497774735093, + 0.004264564719051123, + 0.003757899859920144, + 0.0051625859923660755, + 0.003766816109418869, + 0.0029646612238138914, + 0.004286722280085087, + 0.002970242640003562 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 1 + ], + "selected_reward": -0.008654704317450523 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 static, first-person perspective overlooking a polished wooden dressing table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the table dark-grained and ornate, with a delicate hairpin made of polished silver wire and a pale green jade stone at its center?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene rendered in the style of a whimsical fairytale illustration with soft, ethereal light and visible dust particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the hairpin begin motionless before swaying, vibrating, and rattling against the wood?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the hairpin lift slightly and dance across the surface of the vanity?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009788254275918007 + ], + "opposite_logprob_tokens": [ + -4.6347880363464355 + ], + "logprob_reward": -0.009788254275918007, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009788254275918007, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14453116059303284, + 0.08511964231729507, + 0.03792644664645195, + 0.0552249439060688, + 0.049816347658634186, + 0.07585719972848892, + 0.06819465756416321, + 0.21834412217140198, + 0.19782422482967377, + 0.06716129183769226 + ], + "credit_map": [ + 0.01139681227505207, + 0.007236307021230459, + 0.016799617558717728, + 0.009119635447859764, + 0.008985308930277824, + 0.03724358603358269, + 0.031761687248945236, + 0.02198820933699608, + 0.00808004941791296, + 0.005536148324608803, + 0.025936691090464592, + 0.0053228228352963924, + 0.0031732171773910522, + 0.005611862521618605, + 0.0057031987234950066, + 0.02575564756989479, + 0.00563304778188467, + 0.0029939853120595217, + 0.004916342906653881, + 0.003712819889187813, + 0.0034453249536454678, + 0.004771104082465172, + 0.005337010603398085, + 0.007116809021681547, + 0.005968409590423107, + 0.010949840769171715, + 0.007125684525817633, + 0.006380484439432621, + 0.006568410433828831, + 0.004551950376480818, + 0.005243896506726742, + 0.008436263538897038, + 0.008066119626164436, + 0.005790457595139742, + 0.00792620051652193, + 0.006046968512237072, + 0.005551621317863464, + 0.005697494838386774, + 0.005051192361861467, + 0.005686290096491575, + 0.009801672771573067, + 0.010988911613821983, + 0.013304363936185837, + 0.007514531724154949, + 0.009409753605723381, + 0.00784336868673563, + 0.009014725685119629, + 0.007979868911206722, + 0.00756117282435298, + 0.012285633943974972, + 0.01565014198422432, + 0.007171278353780508, + 0.004658423364162445, + 0.006283171474933624, + 0.005634270142763853, + 0.00895056314766407, + 0.01815667189657688, + 0.032271988689899445, + 0.022810863330960274, + 0.028171909973025322, + 0.026392821222543716, + 0.024446267634630203, + 0.03354933112859726, + 0.032544251531362534, + 0.01867797039449215, + 0.03337975963950157, + 0.02728021703660488, + 0.022362975403666496, + 0.01607886143028736, + 0.024605771526694298, + 0.01948499120771885, + 0.03595364838838577, + 0.007960123009979725, + 0.011435138992965221, + 0.008311416953802109, + 0.006762205623090267, + 0.00635183323174715, + 0.011102174408733845, + 0.006720093544572592, + 0.00851830281317234 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 8, + 0 + ], + "selected_reward": -0.009788254275918007 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 static, first-person perspective overlooking a polished wooden dressing table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the table dark-grained and ornate, with a delicate hairpin made of polished silver wire and a pale green jade stone at its center?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene rendered in the style of a whimsical fairytale illustration with soft, ethereal light and visible dust particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the hairpin begin motionless before swaying, vibrating, and rattling against the wood?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the hairpin lift slightly and dance across the surface of the vanity?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023279689252376556 + ], + "opposite_logprob_tokens": [ + -3.7732796669006348 + ], + "logprob_reward": -0.023279689252376556, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023279689252376556, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10569790750741959, + 0.10860524326562881, + 0.04322041943669319, + 0.06772224605083466, + 0.09517662227153778, + 0.07804067432880402, + 0.14198654890060425, + 0.0725206732749939, + 0.2512527406215668, + 0.03577689453959465 + ], + "credit_map": [ + 0.01334345992654562, + 0.006802062503993511, + 0.007873876951634884, + 0.008210239000618458, + 0.006535169202834368, + 0.023900002241134644, + 0.030301038175821304, + 0.008732059970498085, + 0.009913432411849499, + 0.009053545072674751, + 0.031636793166399, + 0.009442316368222237, + 0.0067026447504758835, + 0.0083449212834239, + 0.007332917768508196, + 0.026178672909736633, + 0.008543134666979313, + 0.0037057786248624325, + 0.006750436965376139, + 0.004964184015989304, + 0.004613272380083799, + 0.004839298781007528, + 0.0043658362701535225, + 0.005438475403934717, + 0.009043842554092407, + 0.016010133549571037, + 0.007665842305868864, + 0.008163326419889927, + 0.008629104122519493, + 0.004814473446458578, + 0.006586314179003239, + 0.0068092141300439835, + 0.015036779455840588, + 0.012570193037390709, + 0.014581842347979546, + 0.010212356224656105, + 0.010434407740831375, + 0.011472584679722786, + 0.009368867613375187, + 0.011499590240418911, + 0.009319623000919819, + 0.009972875006496906, + 0.012097806669771671, + 0.007677462417632341, + 0.011823763139545918, + 0.00817668717354536, + 0.01070993673056364, + 0.008262516930699348, + 0.014089065603911877, + 0.024970494210720062, + 0.02971019595861435, + 0.015362833626568317, + 0.009479614906013012, + 0.013410714454948902, + 0.014540305361151695, + 0.020423322916030884, + 0.007617393974214792, + 0.01244718674570322, + 0.008290683850646019, + 0.007902930490672588, + 0.008535617962479591, + 0.007037450093775988, + 0.011068232357501984, + 0.009621184319257736, + 0.018392277881503105, + 0.026712166145443916, + 0.0340379998087883, + 0.041259631514549255, + 0.02798112854361534, + 0.03702634945511818, + 0.044018980115652084, + 0.02182423323392868, + 0.0036523796152323484, + 0.0037689702585339546, + 0.004981364123523235, + 0.00382597791031003, + 0.004403325263410807, + 0.004924813751131296, + 0.004570791497826576, + 0.005649275612086058 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 6, + 1 + ], + "selected_reward": -0.023279689252376556 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a static, first-person perspective overlooking a polished wooden dressing table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the table dark-grained and ornate, with a delicate hairpin made of polished silver wire and a pale green jade stone at its center?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene rendered in the style of a whimsical fairytale illustration with soft, ethereal light and visible dust particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the hairpin begin motionless before swaying, vibrating, and rattling against the wood?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the hairpin lift slightly and dance across the surface of the vanity?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.054953958839178085 + ], + "opposite_logprob_tokens": [ + -2.9299540519714355 + ], + "logprob_reward": -0.054953958839178085, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.054953958839178085, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20407356321811676, + 0.09002144634723663, + 0.022141460329294205, + 0.04651343822479248, + 0.06051045283675194, + 0.07719326764345169, + 0.07152417302131653, + 0.0842655673623085, + 0.2766401469707489, + 0.06711646914482117 + ], + "credit_map": [ + 0.0167660191655159, + 0.006330499891191721, + 0.011353564448654652, + 0.007741687353700399, + 0.008104502223432064, + 0.07251817733049393, + 0.04312380030751228, + 0.038135308772325516, + 0.010961386375129223, + 0.007386631797999144, + 0.017561253160238266, + 0.010634232312440872, + 0.008336692117154598, + 0.012069325894117355, + 0.010957024991512299, + 0.01211489923298359, + 0.0026663432363420725, + 0.00163790094666183, + 0.003647897392511368, + 0.0032263253815472126, + 0.0024671151768416166, + 0.0028381068259477615, + 0.0022011161781847477, + 0.0034566563554108143, + 0.00620719650760293, + 0.006770393345504999, + 0.005677537061274052, + 0.006051675416529179, + 0.0065543921664357185, + 0.004852548241615295, + 0.005450245924293995, + 0.004949450958520174, + 0.00861912127584219, + 0.0075966049917042255, + 0.011715788394212723, + 0.0052721574902534485, + 0.00524376193061471, + 0.007550687063485384, + 0.004687676206231117, + 0.009824655950069427, + 0.011336744762957096, + 0.012785377912223339, + 0.014113268814980984, + 0.00678121717646718, + 0.006724300794303417, + 0.00844547338783741, + 0.007826428860425949, + 0.009180459193885326, + 0.007526603527367115, + 0.00984023418277502, + 0.013569822534918785, + 0.010791989043354988, + 0.006052425131201744, + 0.006684884894639254, + 0.008313222788274288, + 0.00874499510973692, + 0.009356452152132988, + 0.012466873042285442, + 0.011228788644075394, + 0.0110716437920928, + 0.008508030325174332, + 0.009562091901898384, + 0.013212621212005615, + 0.008859066292643547, + 0.022722458466887474, + 0.03011283092200756, + 0.033147502690553665, + 0.040743328630924225, + 0.03603559359908104, + 0.04122942313551903, + 0.04627656936645508, + 0.026372451335191727, + 0.006071342620998621, + 0.008415678516030312, + 0.01014922745525837, + 0.0074635641649365425, + 0.00918498169630766, + 0.007764002773910761, + 0.01029964443296194, + 0.0077680260874331 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 1 + ], + "selected_reward": -0.054953958839178085 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, first-person perspective shot overlooks the polished surface of an ornate, dark-grained wooden dressing table, the entire scene rendered in the style of a whimsical fairytale illustration. A soft, ethereal light, like a ray of morning sun, permeates the frame, capturing dust particles in a gentle glow and casting a warm glow over the setting. At its center lies a delicate hairpin, its body meticulously crafted from gleaming, polished silver wire, twisted into the shape of a blooming flower, its center resting on a smooth, pale green jade stone. Initially motionless, the hairpin begins to sway with a life of its own; it begins as a faint, almost imperceptible tremor, then escalates into a visible vibration, causing its silver and jade components to gently rattle against the wood. This magical animation builds as the hairpin lifts ever so slightly, in a delicate, enchanting dance across the short distance of the vanity, creating an atmosphere of quiet wonder and palpable unfolding magic.", + "seed": 15992, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000156/group_14/index_483_seed_15992_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": -0.029450251162052153, + "vlm_logprob": -0.029450251162052153, + "vlm_yes_no_margin": 3.875, + "vlm_token_credit": -0.029450251162052153, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 static, first-person perspective overlooking a polished wooden dressing table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the table dark-grained and ornate, with a delicate hairpin made of polished silver wire and a pale green jade stone at its center?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene rendered in the style of a whimsical fairytale illustration with soft, ethereal light and visible dust particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the hairpin begin motionless before swaying, vibrating, and rattling against the wood?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the hairpin lift slightly and dance across the surface of the vanity?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.042158305644989014 + ], + "opposite_logprob_tokens": [ + -4.417158126831055 + ], + "logprob_reward": -0.042158305644989014, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.042158305644989014, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09877936542034149, + 0.024915724992752075, + 0.03054388239979744, + 0.4827934801578522, + 0.03892712667584419, + 0.04751165211200714, + 0.05282203108072281, + 0.06076209619641304, + 0.11042337864637375, + 0.0525212399661541 + ], + "credit_map": [ + 0.0072072590701282024, + 0.008329445496201515, + 0.012626159004867077, + 0.01725662127137184, + 0.00986640527844429, + 0.012751453556120396, + 0.01304871216416359, + 0.017693310976028442, + 0.0025304537266492844, + 0.0026526176370680332, + 0.0027803669217973948, + 0.0033047301694750786, + 0.0027506384067237377, + 0.0034655011259019375, + 0.0036495875101536512, + 0.0037818308919668198, + 0.002630964620038867, + 0.003123678034171462, + 0.004526953212916851, + 0.004692760296165943, + 0.004099591169506311, + 0.004731074906885624, + 0.0036219407338649035, + 0.003116917796432972, + 0.02089184708893299, + 0.013949496671557426, + 0.11565638333559036, + 0.05485999211668968, + 0.01688963919878006, + 0.08447016030550003, + 0.013745282776653767, + 0.16233068704605103, + 0.004708359483629465, + 0.007222179789096117, + 0.004500982817262411, + 0.0041860188357532024, + 0.004150340333580971, + 0.005681775510311127, + 0.004715572111308575, + 0.003761898260563612, + 0.005377903115004301, + 0.006250352133065462, + 0.007369665894657373, + 0.006909441202878952, + 0.005461282096803188, + 0.006942620035260916, + 0.004843086004257202, + 0.004357300233095884, + 0.00672762980684638, + 0.010527476668357849, + 0.004548563621938229, + 0.005245443433523178, + 0.005048664752393961, + 0.008345803245902061, + 0.006052189972251654, + 0.006326264701783657, + 0.005783182568848133, + 0.007606081664562225, + 0.00636369688436389, + 0.008614325895905495, + 0.006999385077506304, + 0.007650450803339481, + 0.008777177892625332, + 0.008967798203229904, + 0.010587369091808796, + 0.02056996524333954, + 0.011635863222181797, + 0.014747570268809795, + 0.011120895855128765, + 0.018932532519102097, + 0.0123372171074152, + 0.010491959750652313, + 0.007193267811089754, + 0.005606159567832947, + 0.006834158208221197, + 0.006193836219608784, + 0.00730079086497426, + 0.00729491887614131, + 0.006875205785036087, + 0.005222903098911047 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 8, + 0 + ], + "selected_reward": -0.042158305644989014 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 static, first-person perspective overlooking a polished wooden dressing table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the table dark-grained and ornate, with a delicate hairpin made of polished silver wire and a pale green jade stone at its center?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene rendered in the style of a whimsical fairytale illustration with soft, ethereal light and visible dust particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the hairpin begin motionless before swaying, vibrating, and rattling against the wood?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the hairpin lift slightly and dance across the surface of the vanity?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020603327080607414 + ], + "opposite_logprob_tokens": [ + -3.8956034183502197 + ], + "logprob_reward": -0.020603327080607414, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020603327080607414, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12222606688737869, + 0.05485312640666962, + 0.6455409526824951, + 0.044284336268901825, + 0.022890396416187286, + 0.014831001870334148, + 0.010180789045989513, + 0.02141883783042431, + 0.0381711982190609, + 0.025603272020816803 + ], + "credit_map": [ + 0.011433870531618595, + 0.018696988001465797, + 0.013533557765185833, + 0.015095205046236515, + 0.012203088030219078, + 0.019996384158730507, + 0.013462433591485023, + 0.017804542556405067, + 0.006138981785625219, + 0.007326915394514799, + 0.005994502920657396, + 0.006451999768614769, + 0.0063888514414429665, + 0.008165454491972923, + 0.006666839122772217, + 0.007719588465988636, + 0.033992912620306015, + 0.04658687859773636, + 0.16363680362701416, + 0.10006498545408249, + 0.07478418946266174, + 0.17684699594974518, + 0.023363998159766197, + 0.026264214888215065, + 0.0026489507872611284, + 0.002653385978192091, + 0.011815940029919147, + 0.006147938780486584, + 0.003718829248100519, + 0.004348697140812874, + 0.0031556987669318914, + 0.009794899262487888, + 0.0018986827926710248, + 0.002293655415996909, + 0.002406870014965534, + 0.002583591267466545, + 0.0028765371534973383, + 0.005594586953520775, + 0.002619612729176879, + 0.0026168625336140394, + 0.0015798149397596717, + 0.001489926129579544, + 0.0022345660254359245, + 0.0021520317532122135, + 0.0015527537325397134, + 0.0018764331471174955, + 0.002506868215277791, + 0.0014386093243956566, + 0.001132903853431344, + 0.0013672268250957131, + 0.0011978675611317158, + 0.0012916979612782598, + 0.0012498905416578054, + 0.001557459938339889, + 0.0012770213652402163, + 0.0011067227460443974, + 0.002484304131940007, + 0.004027706570923328, + 0.0027770488522946835, + 0.002405301434919238, + 0.002021221211180091, + 0.00316593237221241, + 0.002159293508157134, + 0.002378032309934497, + 0.0037971106357872486, + 0.005873863119632006, + 0.0038178961258381605, + 0.004069165792316198, + 0.007448282558470964, + 0.005952803883701563, + 0.003923269920051098, + 0.0032888096757233143, + 0.0025744542945176363, + 0.0020022860262542963, + 0.0031275872606784105, + 0.003630675608292222, + 0.0036855742800980806, + 0.0031931758858263493, + 0.004236373584717512, + 0.003153145546093583 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 1 + ], + "selected_reward": -0.020603327080607414 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 static, first-person perspective overlooking a polished wooden dressing table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the table dark-grained and ornate, with a delicate hairpin made of polished silver wire and a pale green jade stone at its center?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene rendered in the style of a whimsical fairytale illustration with soft, ethereal light and visible dust particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the hairpin begin motionless before swaying, vibrating, and rattling against the wood?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the hairpin lift slightly and dance across the surface of the vanity?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02330007031559944 + ], + "opposite_logprob_tokens": [ + -3.7733001708984375 + ], + "logprob_reward": -0.02330007031559944, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.02330007031559944, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05083848908543587, + 0.010929922573268414, + 0.7840184569358826, + 0.014340762048959732, + 0.014754229225218296, + 0.006446176674216986, + 0.03598586842417717, + 0.012313837185502052, + 0.02813488431274891, + 0.04223731532692909 + ], + "credit_map": [ + 0.0049501219764351845, + 0.0048022144474089146, + 0.005245963577181101, + 0.00893485452979803, + 0.006423432379961014, + 0.006208815146237612, + 0.007920240983366966, + 0.00635285209864378, + 0.001099683460779488, + 0.0011414597975090146, + 0.0013156969798728824, + 0.0013838212471455336, + 0.0012391643831506371, + 0.0020799345802515745, + 0.0013854976277798414, + 0.0012846665922552347, + 0.040504761040210724, + 0.05729761719703674, + 0.1958063691854477, + 0.12058303505182266, + 0.09108839184045792, + 0.21729721128940582, + 0.029289938509464264, + 0.032151203602552414, + 0.000862296496052295, + 0.0009484434849582613, + 0.0028671417385339737, + 0.0018724986584857106, + 0.00165500957518816, + 0.00417686253786087, + 0.0006573802093043923, + 0.0013011315604671836, + 0.001608059974387288, + 0.0022650989703834057, + 0.0022619643714278936, + 0.0019482445204630494, + 0.0015286700800061226, + 0.0024600226897746325, + 0.0013110305881127715, + 0.0013711407082155347, + 0.0007323449826799333, + 0.0011266417568549514, + 0.0006306286668404937, + 0.0008707576780579984, + 0.0008939028484746814, + 0.0008299107430502772, + 0.0007467087707482278, + 0.0006152821588329971, + 0.004159391392022371, + 0.0069570960476994514, + 0.004559810273349285, + 0.003717069048434496, + 0.0031382625456899405, + 0.005836468189954758, + 0.003449973650276661, + 0.004167801700532436, + 0.0012487906496971846, + 0.002002112567424774, + 0.0014061969704926014, + 0.0014617227716371417, + 0.0014932284830138087, + 0.0017430292209610343, + 0.0015071475645527244, + 0.0014516107039526105, + 0.0028614909388124943, + 0.005217464175075293, + 0.0029862134251743555, + 0.003580135991796851, + 0.0026365553494542837, + 0.005117922089993954, + 0.0026495675556361675, + 0.0030855387449264526, + 0.0036953457165509462, + 0.003970971796661615, + 0.006147262640297413, + 0.0057316436432302, + 0.005721937865018845, + 0.004954570904374123, + 0.006501344032585621, + 0.005514242686331272 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 9 + ], + "selected_reward": -0.02330007031559944 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 static, first-person perspective overlooking a polished wooden dressing table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the table dark-grained and ornate, with a delicate hairpin made of polished silver wire and a pale green jade stone at its center?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene rendered in the style of a whimsical fairytale illustration with soft, ethereal light and visible dust particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the hairpin begin motionless before swaying, vibrating, and rattling against the wood?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the hairpin lift slightly and dance across the surface of the vanity?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01253500021994114 + ], + "opposite_logprob_tokens": [ + -4.387535095214844 + ], + "logprob_reward": -0.01253500021994114, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.01253500021994114, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.056226033717393875, + 0.035535987466573715, + 0.2510223984718323, + 0.3482261896133423, + 0.0218769870698452, + 0.03827609866857529, + 0.033969566226005554, + 0.05945947766304016, + 0.0958380252122879, + 0.059569161385297775 + ], + "credit_map": [ + 0.0046527874656021595, + 0.007971267215907574, + 0.005097162444144487, + 0.007373925764113665, + 0.006054176948964596, + 0.00878682266920805, + 0.008520681411027908, + 0.007769216783344746, + 0.0029962758999317884, + 0.0030863152351230383, + 0.0033611825201660395, + 0.004063286818563938, + 0.005709235090762377, + 0.006920721847563982, + 0.005004476755857468, + 0.004394495859742165, + 0.01376065518707037, + 0.0174960158765316, + 0.06534742563962936, + 0.03955995291471481, + 0.028581008315086365, + 0.0659206435084343, + 0.009056736715137959, + 0.011299998499453068, + 0.014671019278466702, + 0.009188083000481129, + 0.08026176691055298, + 0.03649260848760605, + 0.011690672487020493, + 0.06605175137519836, + 0.009854567237198353, + 0.12001575529575348, + 0.0021739834919571877, + 0.0029069825541228056, + 0.00206713262014091, + 0.002574199577793479, + 0.002590718911960721, + 0.00356182549148798, + 0.003367004217579961, + 0.002635142533108592, + 0.00350868864916265, + 0.003963122144341469, + 0.0052198064513504505, + 0.005383678711950779, + 0.00517967389896512, + 0.004926637280732393, + 0.00594089413061738, + 0.004153600428253412, + 0.0035317912697792053, + 0.005796097684651613, + 0.0030677649192512035, + 0.0030930903740227222, + 0.0061216154135763645, + 0.004811896476894617, + 0.00423995079472661, + 0.00330736325122416, + 0.004596879705786705, + 0.0053822994232177734, + 0.006700454279780388, + 0.008366001769900322, + 0.00925953034311533, + 0.00784984976053238, + 0.011717098765075207, + 0.005587366409599781, + 0.00887036882340908, + 0.011985654942691326, + 0.012484122067689896, + 0.015121116302907467, + 0.011942167766392231, + 0.01268762070685625, + 0.013585290871560574, + 0.009161688387393951, + 0.005574206355959177, + 0.006248082499951124, + 0.008223405107855797, + 0.007728970143944025, + 0.0070593287236988544, + 0.0075314780697226524, + 0.009519451297819614, + 0.007684242911636829 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 2, + 8 + ], + "selected_reward": -0.01253500021994114 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 static, first-person perspective overlooking a polished wooden dressing table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the table dark-grained and ornate, with a delicate hairpin made of polished silver wire and a pale green jade stone at its center?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene rendered in the style of a whimsical fairytale illustration with soft, ethereal light and visible dust particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the hairpin begin motionless before swaying, vibrating, and rattling against the wood?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the hairpin lift slightly and dance across the surface of the vanity?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.048654552549123764 + ], + "opposite_logprob_tokens": [ + -3.048654556274414 + ], + "logprob_reward": -0.048654552549123764, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.048654552549123764, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.02450323849916458, + 0.014261445961892605, + 0.4640796482563019, + 0.2807328402996063, + 0.013551164418458939, + 0.023826470598578453, + 0.023804860189557076, + 0.06253159046173096, + 0.04992448911070824, + 0.04278421401977539 + ], + "credit_map": [ + 0.0024582850746810436, + 0.002878638682886958, + 0.002626864006742835, + 0.002977474359795451, + 0.00260122399777174, + 0.003892304142937064, + 0.0033584563061594963, + 0.0037099928595125675, + 0.0012147721135988832, + 0.0012049473589286208, + 0.0015830483753234148, + 0.0020466777496039867, + 0.002203024225309491, + 0.001958161825314164, + 0.002397763542830944, + 0.0016530497232452035, + 0.02384052239358425, + 0.03390101343393326, + 0.11541138589382172, + 0.07077988237142563, + 0.05410877242684364, + 0.1294623762369156, + 0.017399797216057777, + 0.019175883382558823, + 0.011985627934336662, + 0.00779982702806592, + 0.06525949388742447, + 0.029600780457258224, + 0.009202752262353897, + 0.052338093519210815, + 0.008016429841518402, + 0.09652986377477646, + 0.0011458462104201317, + 0.0012141034239903092, + 0.001523840008303523, + 0.00225052866153419, + 0.0019035596633329988, + 0.002453031949698925, + 0.001824591658078134, + 0.0012356635415926576, + 0.002129641594365239, + 0.002320821862667799, + 0.0027157573495060205, + 0.0029722023755311966, + 0.003991343080997467, + 0.003672537626698613, + 0.0036796305794268847, + 0.002344536129385233, + 0.0027094732504338026, + 0.00457072863355279, + 0.0015861362917348742, + 0.0022153512109071016, + 0.0032210240606218576, + 0.003532246919348836, + 0.0031919633038342, + 0.0027779345400631428, + 0.004810551647096872, + 0.0065595549531280994, + 0.006871752440929413, + 0.008236391469836235, + 0.010214786045253277, + 0.0069389245472848415, + 0.011098144575953484, + 0.007801488041877747, + 0.004673993214964867, + 0.007823790423572063, + 0.004469068255275488, + 0.006523314863443375, + 0.005494647193700075, + 0.008765272796154022, + 0.0067774467170238495, + 0.005396957043558359, + 0.0039782156236469746, + 0.004523875191807747, + 0.005903341341763735, + 0.006014358252286911, + 0.004954324569553137, + 0.0053069014102220535, + 0.007167196832597256, + 0.004935998935252428 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 3, + 7 + ], + "selected_reward": -0.048654552549123764 + } + ], + "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/0000156/group_15/index_tensor([751, 751])_group_15_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000156/group_15/index_tensor([751, 751])_group_15_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..7324a6e1f72edd76638ffae493ff7607a0f173d4 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000156/group_15/index_tensor([751, 751])_group_15_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "static, high-angle, medium-close-up captures a young Southeast Asian man with thick, wavy black hair, absorbed in quiet reflection. He sits at a richly textured wooden table, leaning slightly forward, maintaining a posture of intense concentration. He wears a soft, cream-colored linen shirt, its texture visible in the warm light. Golden hour sunlight streams through an unseen window to his right, bathing one side of his face in brilliant light and casting long, soft shadows across the scene, illuminating the dancing dust particles in the air. His dark eyes stare at the pages of a worn, leather-bound journal, open before him. His left hand steadying the cream-colored, textured paper, his right hand carefully guiding a classic fountain pen with a polished dark barrel and silver nib. The pen glides smoothly and deliberately across the page, leaving a trail of dark ink. His pensive expression—a slightly furrowed brow and soft lips—conveys a profound inner dialogue. The background is softly blurred, suggesting a cozy, book-lined study. This, combined with the warm color palette and focused silence of the action, creates an atmosphere of profound tranquility and introspection.", + "seed": 33217, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000156/group_15/index_751_seed_33217_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": -0.06421124339103698, + "vlm_logprob": -0.06421124339103698, + "vlm_yes_no_margin": 3.05, + "vlm_token_credit": -0.06421124339103698, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject a young Southeast Asian man with thick, wavy black hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a soft, cream-colored linen shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man holding a worn, leather-bound journal and writing with a classic fountain pen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by golden hour sunlight streaming from the right, casting long soft shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera positioned at a static, high-angle medium-close-up?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.14514654874801636 + ], + "opposite_logprob_tokens": [ + -2.020146608352661 + ], + "logprob_reward": -0.14514654874801636, + "yes_no_margin_reward": 1.875, + "token_credit_reward": -0.14514654874801636, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2652991712093353, + 0.08617375791072845, + 0.06792251020669937, + 0.08937761187553406, + 0.12115704268217087, + 0.048618197441101074, + 0.03939245641231537, + 0.15036270022392273, + 0.08390466123819351, + 0.04779189079999924 + ], + "credit_map": [ + 0.05911615863442421, + 0.03195507824420929, + 0.028946809470653534, + 0.0182941947132349, + 0.022684214636683464, + 0.07179722934961319, + 0.0174375269562006, + 0.015067948028445244, + 0.02596425823867321, + 0.007799972780048847, + 0.007131451740860939, + 0.006526497658342123, + 0.009403618983924389, + 0.02008747309446335, + 0.004762416705489159, + 0.004498067311942577, + 0.02297169901430607, + 0.005563592072576284, + 0.006232436280697584, + 0.004056592471897602, + 0.00440890621393919, + 0.01786435954272747, + 0.003422611393034458, + 0.0034023094922304153, + 0.022942036390304565, + 0.008443263359367847, + 0.010380093939602375, + 0.0062877945601940155, + 0.0077835917472839355, + 0.017644857987761497, + 0.009747939184308052, + 0.0061480277217924595, + 0.027631843462586403, + 0.020073048770427704, + 0.011227460578083992, + 0.009237457998096943, + 0.014520131051540375, + 0.022945716977119446, + 0.008397663943469524, + 0.007123715244233608, + 0.011148637160658836, + 0.008105915039777756, + 0.0049372403882443905, + 0.0032521209213882685, + 0.004673912655562162, + 0.009670080617070198, + 0.003851378569379449, + 0.0029789134860038757, + 0.005721653811633587, + 0.004725084640085697, + 0.005316329654306173, + 0.0053894734010100365, + 0.00523722218349576, + 0.0052552297711372375, + 0.0042376285418868065, + 0.003509832313284278, + 0.024455396458506584, + 0.017855845391750336, + 0.012680161744356155, + 0.009815090335905552, + 0.02359694615006447, + 0.04044816642999649, + 0.009779131039977074, + 0.011731951497495174, + 0.012383787892758846, + 0.011614962480962276, + 0.007909528911113739, + 0.00837432499974966, + 0.012697954662144184, + 0.01668909378349781, + 0.008064914494752884, + 0.006170096807181835, + 0.0069560944102704525, + 0.005063524004071951, + 0.004926025401800871, + 0.0046861134469509125, + 0.007525409571826458, + 0.011399887502193451, + 0.003769794711843133, + 0.0034650382585823536 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 4 + ], + "selected_reward": -0.14514654874801636 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject a young Southeast Asian man with thick, wavy black hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a soft, cream-colored linen shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man holding a worn, leather-bound journal and writing with a classic fountain pen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by golden hour sunlight streaming from the right, casting long soft shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera positioned at a static, high-angle medium-close-up?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02332068420946598 + ], + "opposite_logprob_tokens": [ + -3.7733206748962402 + ], + "logprob_reward": -0.02332068420946598, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.02332068420946598, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3829193711280823, + 0.20492558181285858, + 0.08507297933101654, + 0.07779483497142792, + 0.04792593792080879, + 0.030786974355578423, + 0.08118760585784912, + 0.039898429065942764, + 0.02606966160237789, + 0.023418670520186424 + ], + "credit_map": [ + 0.08496028929948807, + 0.04699331149458885, + 0.03705267608165741, + 0.028378013521432877, + 0.037531640380620956, + 0.09846802800893784, + 0.03007652424275875, + 0.01945888064801693, + 0.054753970354795456, + 0.01737639680504799, + 0.02824474684894085, + 0.014487064443528652, + 0.019480574876070023, + 0.04258519411087036, + 0.017122574150562286, + 0.010875042527914047, + 0.022260472178459167, + 0.007496223319321871, + 0.011385462246835232, + 0.008256040513515472, + 0.008488434366881847, + 0.015559641644358635, + 0.0067116981372237206, + 0.004915006458759308, + 0.013880359940230846, + 0.01011926494538784, + 0.008389072492718697, + 0.01062993984669447, + 0.00861771684139967, + 0.012422698549926281, + 0.007657680194824934, + 0.006078100763261318, + 0.008891524747014046, + 0.0063859084621071815, + 0.004419364500790834, + 0.00521458126604557, + 0.0056586191058158875, + 0.008442851714789867, + 0.005069998558610678, + 0.0038430874701589346, + 0.005957015324383974, + 0.004087500739842653, + 0.005342144053429365, + 0.0036506939213722944, + 0.0015515958657488227, + 0.005431060679256916, + 0.0028106162790209055, + 0.001956348540261388, + 0.029429005458950996, + 0.007081491407006979, + 0.006804360542446375, + 0.0029278933070600033, + 0.005724938586354256, + 0.021487221121788025, + 0.003934096544981003, + 0.0037986033130437136, + 0.006300806999206543, + 0.005546176806092262, + 0.0050386786460876465, + 0.0038413957227021456, + 0.007323443423956633, + 0.006482744123786688, + 0.002807278884574771, + 0.002557907020673156, + 0.003773710923269391, + 0.0036304390523582697, + 0.003374390071257949, + 0.0018079968867823482, + 0.0038432239089161158, + 0.005426390096545219, + 0.0021933538373559713, + 0.002020155545324087, + 0.003211676375940442, + 0.003356850938871503, + 0.0031160314101725817, + 0.002213231520727277, + 0.0030623183120042086, + 0.004170178901404142, + 0.002481296891346574, + 0.001807084889151156 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.02332068420946598 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject a young Southeast Asian man with thick, wavy black hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a soft, cream-colored linen shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man holding a worn, leather-bound journal and writing with a classic fountain pen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by golden hour sunlight streaming from the right, casting long soft shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera positioned at a static, high-angle medium-close-up?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012553011998534203 + ], + "opposite_logprob_tokens": [ + -4.3875532150268555 + ], + "logprob_reward": -0.012553011998534203, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012553011998534203, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.26058363914489746, + 0.17303438484668732, + 0.1531715989112854, + 0.06898020207881927, + 0.04489583149552345, + 0.06995086371898651, + 0.08810380101203918, + 0.04468169063329697, + 0.034044139087200165, + 0.06255388259887695 + ], + "credit_map": [ + 0.07864965498447418, + 0.02413894049823284, + 0.028677118942141533, + 0.01318265963345766, + 0.01392301544547081, + 0.0768977627158165, + 0.01471804454922676, + 0.010396435856819153, + 0.05352671071887016, + 0.01799720898270607, + 0.021906815469264984, + 0.009727203287184238, + 0.009825609624385834, + 0.0435187928378582, + 0.010168919339776039, + 0.0063631245866417885, + 0.05695243179798126, + 0.013272767886519432, + 0.014873450621962547, + 0.007184941321611404, + 0.006353634409606457, + 0.04267289489507675, + 0.004982223734259605, + 0.006879250053316355, + 0.014513727277517319, + 0.006436120253056288, + 0.011032401584088802, + 0.007882696576416492, + 0.008016823790967464, + 0.009825402870774269, + 0.004922828171402216, + 0.006350204814225435, + 0.010560297407209873, + 0.007732138969004154, + 0.005798663012683392, + 0.004241133574396372, + 0.0035128421150147915, + 0.006211277563124895, + 0.0036333652678877115, + 0.0032061124220490456, + 0.015167939476668835, + 0.008887622505426407, + 0.013785501942038536, + 0.006313900463283062, + 0.004411354195326567, + 0.01143732387572527, + 0.005562268663197756, + 0.004384953062981367, + 0.029033422470092773, + 0.008105623535811901, + 0.009502552449703217, + 0.0042428551241755486, + 0.0052022384479641914, + 0.02087545394897461, + 0.006564058363437653, + 0.004577599931508303, + 0.00703772297129035, + 0.004831097088754177, + 0.004782813601195812, + 0.003474470227956772, + 0.008053823374211788, + 0.010563661344349384, + 0.003371309954673052, + 0.0025667930021882057, + 0.005426457617431879, + 0.003886442631483078, + 0.005500676576048136, + 0.0030231194104999304, + 0.00523108197376132, + 0.005152547266334295, + 0.0031158856581896544, + 0.002707927953451872, + 0.011524003930389881, + 0.008093596436083317, + 0.010197547264397144, + 0.0057936119846999645, + 0.008273020386695862, + 0.009315905161201954, + 0.005049008876085281, + 0.0043071866966784 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.012553011998534203 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject a young Southeast Asian man with thick, wavy black hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a soft, cream-colored linen shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man holding a worn, leather-bound journal and writing with a classic fountain pen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by golden hour sunlight streaming from the right, casting long soft shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera positioned at a static, high-angle medium-close-up?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0700308233499527 + ], + "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": [ + 0.3601589798927307, + 0.16270631551742554, + 0.18692713975906372, + 0.042459312826395035, + 0.057287998497486115, + 0.03959866240620613, + 0.03188072144985199, + 0.036020390689373016, + 0.033155083656311035, + 0.04980538785457611 + ], + "credit_map": [ + 0.08874053508043289, + 0.053703851997852325, + 0.03638116270303726, + 0.023623425513505936, + 0.023726969957351685, + 0.10036438703536987, + 0.01807933859527111, + 0.015539291314780712, + 0.026301143690943718, + 0.024567464366555214, + 0.033906884491443634, + 0.023296557366847992, + 0.016794424504041672, + 0.016595249995589256, + 0.013995405286550522, + 0.007249181158840656, + 0.06659629195928574, + 0.015673793852329254, + 0.017899446189403534, + 0.007982977665960789, + 0.01082648802548647, + 0.05250925570726395, + 0.0070093669928610325, + 0.00842952448874712, + 0.007397436071187258, + 0.005656019318848848, + 0.0065291267819702625, + 0.003954028245061636, + 0.00570050859823823, + 0.005645920988172293, + 0.00457498524338007, + 0.0030012892093509436, + 0.008492719382047653, + 0.011786031536757946, + 0.009276588447391987, + 0.007283545099198818, + 0.00457637757062912, + 0.007494289893656969, + 0.004887929651886225, + 0.0034905190113931894, + 0.005205943714827299, + 0.005756914149969816, + 0.004577300511300564, + 0.004251402337104082, + 0.004487204365432262, + 0.006725449115037918, + 0.004667988978326321, + 0.003926458302885294, + 0.0058168768882751465, + 0.0035058725625276566, + 0.004535620100796223, + 0.0029799246694892645, + 0.0032806096132844687, + 0.0056329467333853245, + 0.003969704732298851, + 0.002159167779609561, + 0.005796687677502632, + 0.003905697027221322, + 0.0026883280370384455, + 0.002919895341619849, + 0.006501212250441313, + 0.007871130481362343, + 0.0032461448572576046, + 0.003091295249760151, + 0.005367596168071032, + 0.003238001372665167, + 0.007287987042218447, + 0.00327503914013505, + 0.0039598108269274235, + 0.0042989193461835384, + 0.0027832435443997383, + 0.0029444864485412836, + 0.0073061599396169186, + 0.007403289433568716, + 0.004726614337414503, + 0.004694359842687845, + 0.007408185862004757, + 0.007485838141292334, + 0.006048089358955622, + 0.004732854664325714 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 1 + ], + "selected_reward": -0.0700308233499527 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject a young Southeast Asian man with thick, wavy black hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a soft, cream-colored linen shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man holding a worn, leather-bound journal and writing with a classic fountain pen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by golden hour sunlight streaming from the right, casting long soft shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera positioned at a static, high-angle medium-close-up?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0700051486492157 + ], + "opposite_logprob_tokens": [ + -2.695005178451538 + ], + "logprob_reward": -0.0700051486492157, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.0700051486492157, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3074464499950409, + 0.11373373121023178, + 0.08580523729324341, + 0.0792732760310173, + 0.06022018566727638, + 0.055426549166440964, + 0.08550283312797546, + 0.07250478863716125, + 0.04605253040790558, + 0.09403439611196518 + ], + "credit_map": [ + 0.07286936044692993, + 0.03188854083418846, + 0.06674383580684662, + 0.02293686382472515, + 0.022843800485134125, + 0.05426585301756859, + 0.021819472312927246, + 0.014078762382268906, + 0.0267943125218153, + 0.010664093308150768, + 0.014209779910743237, + 0.00734316511079669, + 0.012029668316245079, + 0.026485208421945572, + 0.009672904387116432, + 0.0065346164628863335, + 0.01675707846879959, + 0.009904964827001095, + 0.010014550760388374, + 0.010331192053854465, + 0.011229843832552433, + 0.01375830639153719, + 0.00883940514177084, + 0.004969905596226454, + 0.023908566683530807, + 0.008596843108534813, + 0.007743244059383869, + 0.005280232988297939, + 0.006797394249588251, + 0.016446592286229134, + 0.006215292494744062, + 0.004285119939595461, + 0.008622415363788605, + 0.010164868086576462, + 0.007677061017602682, + 0.007527752313762903, + 0.007047425489872694, + 0.009026634506881237, + 0.005603920668363571, + 0.004550111014395952, + 0.006826584693044424, + 0.007624994497746229, + 0.010567843914031982, + 0.007336808368563652, + 0.006670636124908924, + 0.008026659488677979, + 0.004904128611087799, + 0.003468894399702549, + 0.03052719682455063, + 0.007078446447849274, + 0.00801126565784216, + 0.0035086674615740776, + 0.005234483629465103, + 0.022477351129055023, + 0.004869754426181316, + 0.0037956819869577885, + 0.010829155333340168, + 0.009499761275947094, + 0.010093463584780693, + 0.0052034915424883366, + 0.008907263167202473, + 0.01610778272151947, + 0.00501332338899374, + 0.006850559264421463, + 0.0076603456400334835, + 0.006744213402271271, + 0.005516534671187401, + 0.0034891741815954447, + 0.007077330723404884, + 0.007628829218447208, + 0.004425087943673134, + 0.003511016024276614, + 0.014930639415979385, + 0.012092441320419312, + 0.006636462640017271, + 0.006521565839648247, + 0.015902794897556305, + 0.019612573087215424, + 0.009998269379138947, + 0.0083396565169096 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 9 + ], + "selected_reward": -0.0700051486492157 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, high-angle, medium-close-up captures a young Southeast Asian man with thick, wavy black hair, absorbed in quiet reflection. He sits at a richly textured wooden table, leaning slightly forward, maintaining a posture of intense concentration. He wears a soft, cream-colored linen shirt, its texture visible in the warm light. Golden hour sunlight streams through an unseen window to his right, bathing one side of his face in brilliant light and casting long, soft shadows across the scene, illuminating the dancing dust particles in the air. His dark eyes stare at the pages of a worn, leather-bound journal, open before him. His left hand steadying the cream-colored, textured paper, his right hand carefully guiding a classic fountain pen with a polished dark barrel and silver nib. The pen glides smoothly and deliberately across the page, leaving a trail of dark ink. His pensive expression—a slightly furrowed brow and soft lips—conveys a profound inner dialogue. The background is softly blurred, suggesting a cozy, book-lined study. This, combined with the warm color palette and focused silence of the action, creates an atmosphere of profound tranquility and introspection.", + "seed": 33218, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000156/group_15/index_751_seed_33218_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": -0.03169086994603276, + "vlm_logprob": -0.03169086994603276, + "vlm_yes_no_margin": 3.8, + "vlm_token_credit": -0.03169086994603276, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject a young Southeast Asian man with thick, wavy black hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a soft, cream-colored linen shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man holding a worn, leather-bound journal and writing with a classic fountain pen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by golden hour sunlight streaming from the right, casting long soft shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera positioned at a static, high-angle medium-close-up?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03687673062086105 + ], + "opposite_logprob_tokens": [ + -3.411876678466797 + ], + "logprob_reward": -0.03687673062086105, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03687673062086105, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20798152685165405, + 0.09121300280094147, + 0.09457894414663315, + 0.09498213231563568, + 0.07720572501420975, + 0.052770283073186874, + 0.10154656320810318, + 0.1372300684452057, + 0.08798682689666748, + 0.05450490489602089 + ], + "credit_map": [ + 0.04127638041973114, + 0.04617885500192642, + 0.015971796587109566, + 0.025047721341252327, + 0.028494320809841156, + 0.028976740315556526, + 0.0128355473279953, + 0.00920016411691904, + 0.013581641018390656, + 0.012371084652841091, + 0.015704616904258728, + 0.013708130456507206, + 0.011670178733766079, + 0.00849065836519003, + 0.007260098122060299, + 0.00842659268528223, + 0.0146468011662364, + 0.015925902873277664, + 0.020304180681705475, + 0.007822324521839619, + 0.011468366719782352, + 0.00852292776107788, + 0.004995741415768862, + 0.010892705991864204, + 0.01687944494187832, + 0.010212247259914875, + 0.011375159956514835, + 0.006871132180094719, + 0.01502401102334261, + 0.018533015623688698, + 0.0048535410314798355, + 0.011233585886657238, + 0.01200182270258665, + 0.010627882555127144, + 0.01576942577958107, + 0.006395530421286821, + 0.008012183010578156, + 0.012075073085725307, + 0.004429113585501909, + 0.007894692942500114, + 0.008467759937047958, + 0.006841425783932209, + 0.006475152913480997, + 0.006973755080252886, + 0.0072440048679709435, + 0.0055435337126255035, + 0.006023244000971317, + 0.005201404914259911, + 0.031030120328068733, + 0.01193536352366209, + 0.011016501113772392, + 0.0067769200541079044, + 0.010534976609051228, + 0.019389677792787552, + 0.005386334843933582, + 0.005476670805364847, + 0.018685370683670044, + 0.018729357048869133, + 0.01856180652976036, + 0.0200209803879261, + 0.019604675471782684, + 0.017090803012251854, + 0.01110562589019537, + 0.013431454077363014, + 0.013703346252441406, + 0.011821669526398182, + 0.007944975048303604, + 0.012003784067928791, + 0.023131195455789566, + 0.009115996770560741, + 0.005086159333586693, + 0.0051797013729810715, + 0.010409807786345482, + 0.006901504471898079, + 0.010762241668999195, + 0.005827983375638723, + 0.008445273153483868, + 0.0060869138687849045, + 0.0029601366259157658, + 0.003111042082309723 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 6 + ], + "selected_reward": -0.03687673062086105 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject a young Southeast Asian man with thick, wavy black hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a soft, cream-colored linen shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man holding a worn, leather-bound journal and writing with a classic fountain pen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by golden hour sunlight streaming from the right, casting long soft shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera positioned at a static, high-angle medium-close-up?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016089484095573425 + ], + "opposite_logprob_tokens": [ + -4.14108943939209 + ], + "logprob_reward": -0.016089484095573425, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016089484095573425, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1671416014432907, + 0.13883572816848755, + 0.102174773812294, + 0.07161445915699005, + 0.09280916303396225, + 0.114771269261837, + 0.08939748257398605, + 0.08835243433713913, + 0.04437517002224922, + 0.09052792936563492 + ], + "credit_map": [ + 0.032418087124824524, + 0.019481493160128593, + 0.021809983998537064, + 0.02165343426167965, + 0.02856305055320263, + 0.018806904554367065, + 0.013009285554289818, + 0.011399352923035622, + 0.01836174912750721, + 0.015714384615421295, + 0.025642510503530502, + 0.02021222747862339, + 0.021795984357595444, + 0.01430940255522728, + 0.01136204693466425, + 0.011437426321208477, + 0.014419385232031345, + 0.014490429311990738, + 0.02003512904047966, + 0.010329987853765488, + 0.012074296362698078, + 0.01406356692314148, + 0.006016294006258249, + 0.010745679028332233, + 0.016865739598870277, + 0.007447429466992617, + 0.008106067776679993, + 0.005290112923830748, + 0.00914617720991373, + 0.014552622102200985, + 0.0034702711272984743, + 0.006736036855727434, + 0.019132912158966064, + 0.012461594305932522, + 0.015652205795049667, + 0.006764151155948639, + 0.009322019293904305, + 0.013769440352916718, + 0.006004089955240488, + 0.009702753275632858, + 0.03575121611356735, + 0.015619911253452301, + 0.008829888887703419, + 0.008883395232260227, + 0.009239692240953445, + 0.024096889421343803, + 0.005323863122612238, + 0.007026414852589369, + 0.023785142228007317, + 0.00972544401884079, + 0.009575076401233673, + 0.005826241802424192, + 0.007970087230205536, + 0.02227378450334072, + 0.004110921639949083, + 0.0061307866126298904, + 0.01372319832444191, + 0.009988422505557537, + 0.01564505696296692, + 0.01497002225369215, + 0.013943192549049854, + 0.0074319979175925255, + 0.005876539740711451, + 0.006774003151804209, + 0.008259913884103298, + 0.006229988764971495, + 0.007132400292903185, + 0.006103333551436663, + 0.00609507504850626, + 0.0040282029658555984, + 0.002627978567034006, + 0.0038982767146080732, + 0.013449872843921185, + 0.01044243574142456, + 0.018704092130064964, + 0.008431456051766872, + 0.02220820263028145, + 0.008354363031685352, + 0.004755341447889805, + 0.004182170145213604 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 5 + ], + "selected_reward": -0.016089484095573425 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject a young Southeast Asian man with thick, wavy black hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a soft, cream-colored linen shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man holding a worn, leather-bound journal and writing with a classic fountain pen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by golden hour sunlight streaming from the right, casting long soft shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera positioned at a static, high-angle medium-close-up?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005954266060143709 + ], + "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": [ + 0.17335724830627441, + 0.06774783879518509, + 0.040608249604701996, + 0.04131419584155083, + 0.06492417305707932, + 0.0915050283074379, + 0.10577889531850815, + 0.20864218473434448, + 0.1124187707901001, + 0.09370341151952744 + ], + "credit_map": [ + 0.03856354206800461, + 0.018370045349001884, + 0.017398830503225327, + 0.012356474064290524, + 0.03595951199531555, + 0.021134616807103157, + 0.01443241536617279, + 0.015141802839934826, + 0.01024239044636488, + 0.006682445295155048, + 0.006993616931140423, + 0.008840319700539112, + 0.013892678543925285, + 0.007673503365367651, + 0.007089389022439718, + 0.006333492696285248, + 0.007040776312351227, + 0.006523908581584692, + 0.007809992879629135, + 0.003462048713117838, + 0.005729729309678078, + 0.004976121708750725, + 0.002218936337158084, + 0.002846732269972563, + 0.006113896146416664, + 0.0060125915333628654, + 0.00536212557926774, + 0.004271996207535267, + 0.008424063213169575, + 0.005217862315475941, + 0.0029522518161684275, + 0.0029594064690172672, + 0.012442330829799175, + 0.00809037871658802, + 0.011660671792924404, + 0.005640196148306131, + 0.008386459201574326, + 0.007846903055906296, + 0.004575036000460386, + 0.006282193120568991, + 0.018729114904999733, + 0.011836211197078228, + 0.007933435030281544, + 0.009141978807747364, + 0.019130781292915344, + 0.01221736241132021, + 0.0061706374399363995, + 0.006345509085804224, + 0.02369714342057705, + 0.011368626728653908, + 0.012562500312924385, + 0.007243693806231022, + 0.016154112294316292, + 0.02284548431634903, + 0.005510248243808746, + 0.006397079210728407, + 0.025561394169926643, + 0.028392760083079338, + 0.04127034544944763, + 0.030028831213712692, + 0.02886064536869526, + 0.01893889345228672, + 0.01649647392332554, + 0.019092831760644913, + 0.02050269953906536, + 0.019930966198444366, + 0.01353740505874157, + 0.014765658415853977, + 0.014313880354166031, + 0.010620614513754845, + 0.007151301950216293, + 0.011596241034567356, + 0.01300533302128315, + 0.011284330859780312, + 0.01879541389644146, + 0.008094718679785728, + 0.02726418897509575, + 0.0071601117961108685, + 0.004726659972220659, + 0.0033726580440998077 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 8 + ], + "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 subject a young Southeast Asian man with thick, wavy black hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a soft, cream-colored linen shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man holding a worn, leather-bound journal and writing with a classic fountain pen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by golden hour sunlight streaming from the right, casting long soft shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera positioned at a static, high-angle medium-close-up?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020582539960741997 + ], + "opposite_logprob_tokens": [ + -3.895582437515259 + ], + "logprob_reward": -0.020582539960741997, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020582539960741997, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.35156261920928955, + 0.15927237272262573, + 0.02901659533381462, + 0.030711477622389793, + 0.1536024510860443, + 0.07748395949602127, + 0.05781218037009239, + 0.057092536240816116, + 0.036932703107595444, + 0.04651310667395592 + ], + "credit_map": [ + 0.07509679347276688, + 0.05163080617785454, + 0.04112827777862549, + 0.032728321850299835, + 0.06798204779624939, + 0.03581307455897331, + 0.02532103657722473, + 0.021862264722585678, + 0.03412903845310211, + 0.018870895728468895, + 0.015550132840871811, + 0.013044576160609722, + 0.028086937963962555, + 0.02496640384197235, + 0.009630773216485977, + 0.014993617311120033, + 0.004467656835913658, + 0.003120302688330412, + 0.0043895733542740345, + 0.002975822426378727, + 0.005358981899917126, + 0.0037233575712889433, + 0.0017075291834771633, + 0.0032733711414039135, + 0.0077313524670898914, + 0.004002051893621683, + 0.0023484956473112106, + 0.002513585612177849, + 0.004122424405068159, + 0.006294566672295332, + 0.0015072185778990388, + 0.002191780135035515, + 0.02816087380051613, + 0.020960455760359764, + 0.03538889437913895, + 0.013521765358746052, + 0.018245289102196693, + 0.014261636883020401, + 0.00884042214602232, + 0.01422311831265688, + 0.023917989805340767, + 0.010909211821854115, + 0.005585098639130592, + 0.0030679628252983093, + 0.009203597903251648, + 0.017603620886802673, + 0.0026601050049066544, + 0.004536375869065523, + 0.014522336423397064, + 0.007899482734501362, + 0.008564363233745098, + 0.0033915145322680473, + 0.008856302127242088, + 0.007740126922726631, + 0.003079761052504182, + 0.0037582917138934135, + 0.006289721932262182, + 0.008147009648382664, + 0.0070277187041938305, + 0.005944124422967434, + 0.015293837524950504, + 0.006935339421033859, + 0.0036616597790271044, + 0.0037931236438453197, + 0.005771313328295946, + 0.005722446367144585, + 0.0030038554687052965, + 0.006900850683450699, + 0.00718203280121088, + 0.0032609906047582626, + 0.0023015153128653765, + 0.0027896983083337545, + 0.005659439601004124, + 0.005366105120629072, + 0.008581726811826229, + 0.003748510731384158, + 0.012150654569268227, + 0.004312087781727314, + 0.0029329280368983746, + 0.0037616577465087175 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 4 + ], + "selected_reward": -0.020582539960741997 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject a young Southeast Asian man with thick, wavy black hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a soft, cream-colored linen shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man holding a worn, leather-bound journal and writing with a classic fountain pen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by golden hour sunlight streaming from the right, casting long soft shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera positioned at a static, high-angle medium-close-up?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07895132899284363 + ], + "opposite_logprob_tokens": [ + -2.578951358795166 + ], + "logprob_reward": -0.07895132899284363, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07895132899284363, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.26672372221946716, + 0.03880900517106056, + 0.06378399580717087, + 0.07218008488416672, + 0.08145004510879517, + 0.04191087931394577, + 0.04490189254283905, + 0.0865403413772583, + 0.24709729850292206, + 0.05660262703895569 + ], + "credit_map": [ + 0.054309312254190445, + 0.034689001739025116, + 0.030785931274294853, + 0.03205252066254616, + 0.05123254284262657, + 0.022395990788936615, + 0.02092033438384533, + 0.020338093861937523, + 0.005757048260420561, + 0.00561269698664546, + 0.006845870520919561, + 0.005604233592748642, + 0.0039643156342208385, + 0.005013319663703442, + 0.003127657575532794, + 0.002883862005546689, + 0.009524449706077576, + 0.009584922343492508, + 0.015322561375796795, + 0.007531489711254835, + 0.007098152302205563, + 0.006183345336467028, + 0.00397952226921916, + 0.004559548571705818, + 0.012187612242996693, + 0.00852768961340189, + 0.010862179100513458, + 0.00722500728443265, + 0.012649301439523697, + 0.008015211671590805, + 0.005332614295184612, + 0.007380473893135786, + 0.015242882072925568, + 0.012768019922077656, + 0.019547127187252045, + 0.006777916569262743, + 0.007926247082650661, + 0.006611987948417664, + 0.004687043838202953, + 0.007888823747634888, + 0.00866667553782463, + 0.0056877341121435165, + 0.00509107019752264, + 0.003520668251439929, + 0.008899364620447159, + 0.0048092808574438095, + 0.002921793609857559, + 0.0023142914287745953, + 0.012786490842700005, + 0.004762435797601938, + 0.00448974221944809, + 0.0038209431804716587, + 0.005816297605633736, + 0.008403041400015354, + 0.002337477868422866, + 0.0024854643270373344, + 0.013783239759504795, + 0.011824442073702812, + 0.011972621083259583, + 0.010265703313052654, + 0.01348832156509161, + 0.00910261645913124, + 0.0074223908595740795, + 0.008681011386215687, + 0.039191968739032745, + 0.03817109018564224, + 0.01845465786755085, + 0.040749549865722656, + 0.05587158724665642, + 0.02282758057117462, + 0.014069479890167713, + 0.01776139810681343, + 0.0090421037748456, + 0.008681484498083591, + 0.0068667675368487835, + 0.0054833791218698025, + 0.01168776024132967, + 0.005948600824922323, + 0.004310370422899723, + 0.004582161083817482 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 7 + ], + "selected_reward": -0.07895132899284363 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, high-angle, medium-close-up captures a young Southeast Asian man with thick, wavy black hair, absorbed in quiet reflection. He sits at a richly textured wooden table, leaning slightly forward, maintaining a posture of intense concentration. He wears a soft, cream-colored linen shirt, its texture visible in the warm light. Golden hour sunlight streams through an unseen window to his right, bathing one side of his face in brilliant light and casting long, soft shadows across the scene, illuminating the dancing dust particles in the air. His dark eyes stare at the pages of a worn, leather-bound journal, open before him. His left hand steadying the cream-colored, textured paper, his right hand carefully guiding a classic fountain pen with a polished dark barrel and silver nib. The pen glides smoothly and deliberately across the page, leaving a trail of dark ink. His pensive expression—a slightly furrowed brow and soft lips—conveys a profound inner dialogue. The background is softly blurred, suggesting a cozy, book-lined study. This, combined with the warm color palette and focused silence of the action, creates an atmosphere of profound tranquility and introspection.", + "seed": 33219, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000156/group_15/index_751_seed_33219_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": -0.024617314897477628, + "vlm_logprob": -0.024617314897477628, + "vlm_yes_no_margin": 4.05, + "vlm_token_credit": -0.024617314897477628, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject a young Southeast Asian man with thick, wavy black hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a soft, cream-colored linen shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man holding a worn, leather-bound journal and writing with a classic fountain pen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by golden hour sunlight streaming from the right, casting long soft shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera positioned at a static, high-angle medium-close-up?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06353749334812164 + ], + "opposite_logprob_tokens": [ + -2.81353759765625 + ], + "logprob_reward": -0.06353749334812164, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06353749334812164, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.128376767039299, + 0.042068637907505035, + 0.02634635753929615, + 0.04114840179681778, + 0.417946457862854, + 0.23586979508399963, + 0.030864497646689415, + 0.01634635403752327, + 0.01133657991886139, + 0.04969619959592819 + ], + "credit_map": [ + 0.019523903727531433, + 0.012696325778961182, + 0.028212781995534897, + 0.014222774654626846, + 0.01084204576909542, + 0.01449417695403099, + 0.011353843845427036, + 0.01703091338276863, + 0.006735071539878845, + 0.008428581990301609, + 0.005842817015945911, + 0.0055144187062978745, + 0.0036127641797065735, + 0.004114392679184675, + 0.0033898567780852318, + 0.004430738277733326, + 0.003700661240145564, + 0.003157620783895254, + 0.0034191804006695747, + 0.0033112054225057364, + 0.0027438905090093613, + 0.0033513440284878016, + 0.003079649992287159, + 0.0035828049294650555, + 0.005372157786041498, + 0.005566757638007402, + 0.0056383959017694, + 0.005041306372731924, + 0.004394328687340021, + 0.006954692304134369, + 0.004289970267564058, + 0.0038907909765839577, + 0.051143158227205276, + 0.050459910184144974, + 0.06465515494346619, + 0.07506419718265533, + 0.03157147765159607, + 0.05567961186170578, + 0.042414870113134384, + 0.04695805534720421, + 0.02866051346063614, + 0.029017753899097443, + 0.04306352138519287, + 0.04173940047621727, + 0.01687464490532875, + 0.031767912209033966, + 0.018010860309004784, + 0.02673518657684326, + 0.0039082784205675125, + 0.004043599124997854, + 0.004233191721141338, + 0.0051256814040243626, + 0.002171541564166546, + 0.004643942229449749, + 0.0028518035542219877, + 0.0038864589296281338, + 0.002335111377760768, + 0.002057983772829175, + 0.0035065494012087584, + 0.0017910921014845371, + 0.0014939620159566402, + 0.0018343550618737936, + 0.0019347213674336672, + 0.0013925785897299647, + 0.001413959194906056, + 0.001352719496935606, + 0.001755217555910349, + 0.0018618650501593947, + 0.0011418882058933377, + 0.001525339437648654, + 0.001209984184242785, + 0.0010756070259958506, + 0.008007613942027092, + 0.008033400401473045, + 0.004488750360906124, + 0.0051998416893184185, + 0.004605697933584452, + 0.007161370944231749, + 0.006137419492006302, + 0.006062104366719723 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 5, + 0 + ], + "selected_reward": -0.06353749334812164 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject a young Southeast Asian man with thick, wavy black hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a soft, cream-colored linen shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man holding a worn, leather-bound journal and writing with a classic fountain pen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by golden hour sunlight streaming from the right, casting long soft shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera positioned at a static, high-angle medium-close-up?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016092533245682716 + ], + "opposite_logprob_tokens": [ + -4.141092300415039 + ], + "logprob_reward": -0.016092533245682716, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016092533245682716, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09029868245124817, + 0.06651121377944946, + 0.039000362157821655, + 0.05014200881123543, + 0.3265393078327179, + 0.2441229224205017, + 0.08413129299879074, + 0.044997744262218475, + 0.019651195034384727, + 0.03460529446601868 + ], + "credit_map": [ + 0.015938613563776016, + 0.015679635107517242, + 0.012361233122646809, + 0.010058015584945679, + 0.009396574459969997, + 0.008225461468100548, + 0.008187937550246716, + 0.010451218113303185, + 0.009411667473614216, + 0.008993014693260193, + 0.008188284002244473, + 0.007528054062277079, + 0.007872756570577621, + 0.009411617182195187, + 0.006283709779381752, + 0.008822103962302208, + 0.005526058375835419, + 0.005368197802454233, + 0.005307720508426428, + 0.004381668288260698, + 0.0036317629273980856, + 0.006473206449300051, + 0.003976959735155106, + 0.004334789700806141, + 0.0064338818192481995, + 0.006585513241589069, + 0.006872891448438168, + 0.0060080261901021, + 0.005862672347575426, + 0.007712701801210642, + 0.005361330229789019, + 0.005304989404976368, + 0.0396699458360672, + 0.03603741526603699, + 0.05266190320253372, + 0.058506131172180176, + 0.025512099266052246, + 0.04428583011031151, + 0.03366754949092865, + 0.03619841858744621, + 0.030097516253590584, + 0.030209388583898544, + 0.04366515204310417, + 0.042134691029787064, + 0.017725883051753044, + 0.03181803226470947, + 0.02106471173465252, + 0.027407538145780563, + 0.010484116151928902, + 0.010761335492134094, + 0.01399329211562872, + 0.014740649610757828, + 0.005665377248078585, + 0.010432641953229904, + 0.008228358812630177, + 0.009825523942708969, + 0.006943189073354006, + 0.006045580375939608, + 0.00789580587297678, + 0.006297246087342501, + 0.0039713201113045216, + 0.004720706958323717, + 0.005362255964428186, + 0.003761640749871731, + 0.002980719320476055, + 0.0019767004996538162, + 0.0067856465466320515, + 0.0016904054209589958, + 0.0012942590983584523, + 0.0015845567686483264, + 0.0016783187165856361, + 0.0016605902928858995, + 0.0050661880522966385, + 0.0036627529188990593, + 0.0067182485945522785, + 0.003570142900571227, + 0.003185256151482463, + 0.003969836980104446, + 0.0042549800127744675, + 0.004177889786660671 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 5, + 0 + ], + "selected_reward": -0.016092533245682716 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject a young Southeast Asian man with thick, wavy black hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a soft, cream-colored linen shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man holding a worn, leather-bound journal and writing with a classic fountain pen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by golden hour sunlight streaming from the right, casting long soft shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera positioned at a static, high-angle medium-close-up?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004100368358194828 + ], + "opposite_logprob_tokens": [ + -5.504100322723389 + ], + "logprob_reward": -0.004100368358194828, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004100368358194828, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2124674916267395, + 0.07976845651865005, + 0.06290114670991898, + 0.018529245629906654, + 0.40912696719169617, + 0.15456938743591309, + 0.012676308862864971, + 0.029624028131365776, + 0.00661137979477644, + 0.013725675642490387 + ], + "credit_map": [ + 0.03040127642452717, + 0.022193633019924164, + 0.04435032233595848, + 0.02674436941742897, + 0.016518766060471535, + 0.02897779829800129, + 0.0217170137912035, + 0.021564291790127754, + 0.011200149543583393, + 0.00839180126786232, + 0.01793070323765278, + 0.009999328292906284, + 0.005710772704333067, + 0.010464213788509369, + 0.008162097074091434, + 0.007909385487437248, + 0.008998189121484756, + 0.0077959513291716576, + 0.011170564219355583, + 0.008583083748817444, + 0.0048179118894040585, + 0.007822884246706963, + 0.007062167394906282, + 0.006650392897427082, + 0.002765397075563669, + 0.00205523194745183, + 0.004672579001635313, + 0.0019455809378996491, + 0.0015538278967142105, + 0.0023447819985449314, + 0.0014140948187559843, + 0.001777751138433814, + 0.050647348165512085, + 0.049082476645708084, + 0.06260626763105392, + 0.0734802857041359, + 0.031108466908335686, + 0.05483632907271385, + 0.040583137422800064, + 0.046782609075307846, + 0.019317347556352615, + 0.020804233849048615, + 0.025914525613188744, + 0.02723567746579647, + 0.011687155812978745, + 0.020033949986100197, + 0.012786115519702435, + 0.016790369525551796, + 0.0019166144775226712, + 0.0014718418242409825, + 0.0039094723761081696, + 0.0012970117386430502, + 0.0009135911823250353, + 0.001226951484568417, + 0.0009856270626187325, + 0.0009551974362693727, + 0.0036729408893734217, + 0.004120451398193836, + 0.004290721844881773, + 0.004325341433286667, + 0.002330176066607237, + 0.0034230442252010107, + 0.0037618433125317097, + 0.00369950826279819, + 0.0010590462479740381, + 0.0011578184785321355, + 0.0011444585397839546, + 0.0008624106994830072, + 0.0004926964757032692, + 0.0006345774163492024, + 0.0006951065734028816, + 0.0005652648396790028, + 0.0018004857702180743, + 0.0017696490976959467, + 0.0037962947972118855, + 0.001338841044344008, + 0.0008941175765357912, + 0.001638863468542695, + 0.0012266122503206134, + 0.0012608097167685628 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 5 + ], + "selected_reward": -0.004100368358194828 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject a young Southeast Asian man with thick, wavy black hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a soft, cream-colored linen shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man holding a worn, leather-bound journal and writing with a classic fountain pen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by golden hour sunlight streaming from the right, casting long soft shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera positioned at a static, high-angle medium-close-up?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023291800171136856 + ], + "opposite_logprob_tokens": [ + -3.773291826248169 + ], + "logprob_reward": -0.023291800171136856, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023291800171136856, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12489882111549377, + 0.05930718034505844, + 0.06841332465410233, + 0.04024799168109894, + 0.5439528822898865, + 0.057407550513744354, + 0.0271537397056818, + 0.0462920144200325, + 0.011501168832182884, + 0.020825311541557312 + ], + "credit_map": [ + 0.020791999995708466, + 0.029156681150197983, + 0.014217935502529144, + 0.012645942158997059, + 0.011638568714261055, + 0.01585802435874939, + 0.008849477395415306, + 0.011740191839635372, + 0.009265630505979061, + 0.00630892813205719, + 0.011449302546679974, + 0.005591193679720163, + 0.005412647034972906, + 0.008346200920641422, + 0.004239145666360855, + 0.00869413185864687, + 0.012097659520804882, + 0.007669647224247456, + 0.0142371766269207, + 0.006163850426673889, + 0.005010609515011311, + 0.010160178877413273, + 0.005365591496229172, + 0.007708614692091942, + 0.005417149513959885, + 0.0050995540805161, + 0.00921575166285038, + 0.004554471932351589, + 0.003239706391468644, + 0.005637275520712137, + 0.002738228067755699, + 0.004345858469605446, + 0.0693291649222374, + 0.06031818687915802, + 0.07715478539466858, + 0.09814838320016861, + 0.04408266395330429, + 0.07431865483522415, + 0.055363234132528305, + 0.06523779779672623, + 0.008413213305175304, + 0.010493223555386066, + 0.012652467004954815, + 0.006567533127963543, + 0.00487722922116518, + 0.005985078401863575, + 0.004196771886199713, + 0.004222038201987743, + 0.0038590666372328997, + 0.0037841126322746277, + 0.006210499908775091, + 0.00302538531832397, + 0.0017753222491592169, + 0.003698919666931033, + 0.0022607501596212387, + 0.002539681037887931, + 0.005900230724364519, + 0.007048425264656544, + 0.00752894626930356, + 0.0071025085635483265, + 0.003318399889394641, + 0.006061372812837362, + 0.004378420766443014, + 0.004953709430992603, + 0.001510809757746756, + 0.0011767293326556683, + 0.0024328394792973995, + 0.00174095225520432, + 0.0009569565881974995, + 0.0013852373231202364, + 0.0013760121073573828, + 0.0009216314647346735, + 0.0029532448388636112, + 0.0023573017679154873, + 0.003715700702741742, + 0.0021112975664436817, + 0.0017794655868783593, + 0.00396303366869688, + 0.0018074484542012215, + 0.0021378195378929377 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 2 + ], + "selected_reward": -0.023291800171136856 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject a young Southeast Asian man with thick, wavy black hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a soft, cream-colored linen shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man holding a worn, leather-bound journal and writing with a classic fountain pen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by golden hour sunlight streaming from the right, casting long soft shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera positioned at a static, high-angle medium-close-up?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01606437936425209 + ], + "opposite_logprob_tokens": [ + -4.141064167022705 + ], + "logprob_reward": -0.01606437936425209, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.01606437936425209, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0858616754412651, + 0.0451347641646862, + 0.03184745833277702, + 0.10317210853099823, + 0.4995642602443695, + 0.1400761902332306, + 0.03396957367658615, + 0.028260940685868263, + 0.015374662354588509, + 0.016738276928663254 + ], + "credit_map": [ + 0.011873179115355015, + 0.009996283799409866, + 0.016732467338442802, + 0.009820559062063694, + 0.007062975317239761, + 0.013111533597111702, + 0.009128975681960583, + 0.008135718293488026, + 0.005524055100977421, + 0.005571951158344746, + 0.005585747305303812, + 0.005155921448022127, + 0.005920375231653452, + 0.007331330794841051, + 0.004075301345437765, + 0.005970087368041277, + 0.004331743810325861, + 0.0038019302301108837, + 0.004445234313607216, + 0.004761677235364914, + 0.003035160945728421, + 0.004699224606156349, + 0.003893753746524453, + 0.0028787371702492237, + 0.012280340306460857, + 0.013632340356707573, + 0.018519839271903038, + 0.016841420903801918, + 0.008823348209261894, + 0.01288433838635683, + 0.011679989285767078, + 0.008510498329997063, + 0.0618181973695755, + 0.06523624807596207, + 0.07402396202087402, + 0.08946065604686737, + 0.03617703914642334, + 0.06653264909982681, + 0.05003948137164116, + 0.05627608671784401, + 0.01707136258482933, + 0.017975015565752983, + 0.019702265039086342, + 0.025202801451086998, + 0.010288695804774761, + 0.022002046927809715, + 0.01177946012467146, + 0.016054561361670494, + 0.00656089698895812, + 0.00477488012984395, + 0.009023516438901424, + 0.003344080876559019, + 0.0018743033288046718, + 0.0033154403790831566, + 0.002378284465521574, + 0.0026981751434504986, + 0.0036668910179287195, + 0.003987011034041643, + 0.004459122195839882, + 0.00412240345031023, + 0.0021812233608216047, + 0.0038498269859701395, + 0.0030034552328288555, + 0.0029910095036029816, + 0.0023426322732120752, + 0.001764596439898014, + 0.005132393911480904, + 0.001435542362742126, + 0.0009319212986156344, + 0.0014020787784829736, + 0.0011868029832839966, + 0.001178696402348578, + 0.0027174942661076784, + 0.001976330066099763, + 0.003598144045099616, + 0.0021110489033162594, + 0.0012559095630422235, + 0.001584705663844943, + 0.0020222137682139874, + 0.001472432166337967 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 5, + 3 + ], + "selected_reward": -0.01606437936425209 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, high-angle, medium-close-up captures a young Southeast Asian man with thick, wavy black hair, absorbed in quiet reflection. He sits at a richly textured wooden table, leaning slightly forward, maintaining a posture of intense concentration. He wears a soft, cream-colored linen shirt, its texture visible in the warm light. Golden hour sunlight streams through an unseen window to his right, bathing one side of his face in brilliant light and casting long, soft shadows across the scene, illuminating the dancing dust particles in the air. His dark eyes stare at the pages of a worn, leather-bound journal, open before him. His left hand steadying the cream-colored, textured paper, his right hand carefully guiding a classic fountain pen with a polished dark barrel and silver nib. The pen glides smoothly and deliberately across the page, leaving a trail of dark ink. His pensive expression—a slightly furrowed brow and soft lips—conveys a profound inner dialogue. The background is softly blurred, suggesting a cozy, book-lined study. This, combined with the warm color palette and focused silence of the action, creates an atmosphere of profound tranquility and introspection.", + "seed": 33220, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000156/group_15/index_751_seed_33220_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": -0.03463519420474768, + "vlm_logprob": -0.03463519420474768, + "vlm_yes_no_margin": 3.675, + "vlm_token_credit": -0.03463519420474768, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject a young Southeast Asian man with thick, wavy black hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a soft, cream-colored linen shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man holding a worn, leather-bound journal and writing with a classic fountain pen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by golden hour sunlight streaming from the right, casting long soft shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera positioned at a static, high-angle medium-close-up?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05773080512881279 + ], + "opposite_logprob_tokens": [ + -2.9327309131622314 + ], + "logprob_reward": -0.05773080512881279, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05773080512881279, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17461265623569489, + 0.07838314771652222, + 0.05433100834488869, + 0.08603434264659882, + 0.0509168766438961, + 0.10070475190877914, + 0.11965571343898773, + 0.09016762673854828, + 0.12721742689609528, + 0.11797643452882767 + ], + "credit_map": [ + 0.019851207733154297, + 0.029755255207419395, + 0.015101471915841103, + 0.013215944170951843, + 0.02564948983490467, + 0.02911781519651413, + 0.017481699585914612, + 0.024439770728349686, + 0.006945624947547913, + 0.008588124066591263, + 0.0071740346029400826, + 0.007990451529622078, + 0.011564867570996284, + 0.011520291678607464, + 0.006611455697566271, + 0.017988303676247597, + 0.007822655141353607, + 0.0066998181864619255, + 0.00493124732747674, + 0.004225320648401976, + 0.009136549197137356, + 0.005961088463664055, + 0.0038149890024214983, + 0.011739338748157024, + 0.010111472569406033, + 0.0078451968729496, + 0.009821671061217785, + 0.005416356958448887, + 0.013397892005741596, + 0.013856574892997742, + 0.007726056035608053, + 0.017859121784567833, + 0.0067344605922698975, + 0.006565824616700411, + 0.004111023619771004, + 0.0033611280377954245, + 0.008002376183867455, + 0.008084811270236969, + 0.004366982728242874, + 0.009690271690487862, + 0.010550737380981445, + 0.018700236454606056, + 0.007252881303429604, + 0.005606220103800297, + 0.013193245977163315, + 0.0198622215539217, + 0.008364657871425152, + 0.017174554988741875, + 0.010636657476425171, + 0.017751365900039673, + 0.011264159344136715, + 0.018357301130890846, + 0.014283686876296997, + 0.017260456457734108, + 0.013115433044731617, + 0.016986660659313202, + 0.010973187163472176, + 0.01732661761343479, + 0.00934012420475483, + 0.004770938772708178, + 0.014292428269982338, + 0.010582870803773403, + 0.007792485877871513, + 0.015088976360857487, + 0.018321581184864044, + 0.014453612267971039, + 0.019692694768309593, + 0.009735682047903538, + 0.01573764719069004, + 0.016796816140413284, + 0.012686585076153278, + 0.019792811945080757, + 0.014235872775316238, + 0.026196761056780815, + 0.011414011009037495, + 0.006239448208361864, + 0.018802350386977196, + 0.018315531313419342, + 0.007100225891917944, + 0.015672236680984497 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 6 + ], + "selected_reward": -0.05773080512881279 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject a young Southeast Asian man with thick, wavy black hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a soft, cream-colored linen shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man holding a worn, leather-bound journal and writing with a classic fountain pen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by golden hour sunlight streaming from the right, casting long soft shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera positioned at a static, high-angle medium-close-up?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009806551970541477 + ], + "opposite_logprob_tokens": [ + -4.6348066329956055 + ], + "logprob_reward": -0.009806551970541477, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009806551970541477, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3215501308441162, + 0.13831055164337158, + 0.048425305634737015, + 0.08240718394517899, + 0.059942103922367096, + 0.0688689798116684, + 0.07107923179864883, + 0.04994676262140274, + 0.08589350432157516, + 0.07357628643512726 + ], + "credit_map": [ + 0.03214416652917862, + 0.038269657641649246, + 0.02187569998204708, + 0.02971167117357254, + 0.04417344927787781, + 0.05601797252893448, + 0.04035555198788643, + 0.05900195986032486, + 0.013765950687229633, + 0.022368650883436203, + 0.010342498309910297, + 0.01287009846419096, + 0.01743709295988083, + 0.02522250823676586, + 0.015283659100532532, + 0.02102009765803814, + 0.005794051103293896, + 0.006411593873053789, + 0.004694752860814333, + 0.003734349040314555, + 0.005791279021650553, + 0.007585157640278339, + 0.005851296242326498, + 0.008562826551496983, + 0.011342945508658886, + 0.01695491187274456, + 0.007687143515795469, + 0.006619654130190611, + 0.010598631575703621, + 0.012959587387740612, + 0.006435565184801817, + 0.009808743372559547, + 0.007783509325236082, + 0.011212476529181004, + 0.0041577634401619434, + 0.005112841259688139, + 0.007829797454178333, + 0.010784248821437359, + 0.005005550105124712, + 0.008055920712649822, + 0.011166045442223549, + 0.008352864533662796, + 0.005012286361306906, + 0.004651809576898813, + 0.010972192510962486, + 0.010547919198870659, + 0.006902801338583231, + 0.011263061314821243, + 0.007225630339235067, + 0.011482806876301765, + 0.005322600249201059, + 0.008027848787605762, + 0.009257212281227112, + 0.013476945459842682, + 0.006455674301832914, + 0.009830513969063759, + 0.006218641996383667, + 0.0068100616335868835, + 0.0052838255651295185, + 0.0028631500899791718, + 0.007249615155160427, + 0.00716834980994463, + 0.004921660292893648, + 0.00943145900964737, + 0.010301865637302399, + 0.010401617735624313, + 0.015136901289224625, + 0.007180530112236738, + 0.010634834878146648, + 0.01051320880651474, + 0.008274499326944351, + 0.013450045138597488, + 0.008614669553935528, + 0.01308058388531208, + 0.010528072714805603, + 0.006572525016963482, + 0.010861962102353573, + 0.008129999972879887, + 0.006133113522082567, + 0.009655360132455826 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 8 + ], + "selected_reward": -0.009806551970541477 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject a young Southeast Asian man with thick, wavy black hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a soft, cream-colored linen shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man holding a worn, leather-bound journal and writing with a classic fountain pen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by golden hour sunlight streaming from the right, casting long soft shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera positioned at a static, high-angle medium-close-up?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007638650014996529 + ], + "opposite_logprob_tokens": [ + -4.882638454437256 + ], + "logprob_reward": -0.007638650014996529, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007638650014996529, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2676237225532532, + 0.09187766164541245, + 0.08819838613271713, + 0.08986779302358627, + 0.05209185555577278, + 0.04207528010010719, + 0.057446908205747604, + 0.07741828262805939, + 0.08650592714548111, + 0.14689421653747559 + ], + "credit_map": [ + 0.039363496005535126, + 0.02468671277165413, + 0.019880088046193123, + 0.028595298528671265, + 0.04328802227973938, + 0.044450223445892334, + 0.024387134239077568, + 0.042972736060619354, + 0.016563767567276955, + 0.009441624395549297, + 0.008114726282656193, + 0.009928238578140736, + 0.014417418278753757, + 0.014318853616714478, + 0.006450498476624489, + 0.01264253444969654, + 0.01218150183558464, + 0.022009769454598427, + 0.007599191274493933, + 0.008493339642882347, + 0.011984270066022873, + 0.006776492577046156, + 0.0069247144274413586, + 0.012229102663695812, + 0.012198119424283504, + 0.015344763174653053, + 0.008676934987306595, + 0.008021892979741096, + 0.01419044192880392, + 0.009269896894693375, + 0.007340035866945982, + 0.014825710095465183, + 0.0065137892961502075, + 0.009069154039025307, + 0.005534782074391842, + 0.006406533997505903, + 0.00704564992338419, + 0.005908933933824301, + 0.0044991676695644855, + 0.007113845087587833, + 0.009374729357659817, + 0.007863281294703484, + 0.0034125091042369604, + 0.0038337258156389, + 0.005154182203114033, + 0.004912217613309622, + 0.0033337431959807873, + 0.004190893843770027, + 0.007247792091220617, + 0.008021813817322254, + 0.005185839720070362, + 0.007169176824390888, + 0.007173115853220224, + 0.008616073988378048, + 0.006567925214767456, + 0.007465176749974489, + 0.012725342065095901, + 0.010595024563372135, + 0.008458721451461315, + 0.009619234129786491, + 0.009412126615643501, + 0.010055659338831902, + 0.007223794236779213, + 0.009328379295766354, + 0.011994512751698494, + 0.009803233668208122, + 0.016339873895049095, + 0.006363170687109232, + 0.009908458217978477, + 0.009673344902694225, + 0.010438939556479454, + 0.011984395794570446, + 0.01786433719098568, + 0.03016328252851963, + 0.02508704923093319, + 0.012737597338855267, + 0.018525464460253716, + 0.01454149466007948, + 0.012035605497658253, + 0.015939390286803246 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 1 + ], + "selected_reward": -0.007638650014996529 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject a young Southeast Asian man with thick, wavy black hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a soft, cream-colored linen shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man holding a worn, leather-bound journal and writing with a classic fountain pen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by golden hour sunlight streaming from the right, casting long soft shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera positioned at a static, high-angle medium-close-up?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04306168854236603 + ], + "opposite_logprob_tokens": [ + -3.1680617332458496 + ], + "logprob_reward": -0.04306168854236603, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04306168854236603, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2886640131473541, + 0.08691924810409546, + 0.07302317023277283, + 0.06610473245382309, + 0.04723299294710159, + 0.051621269434690475, + 0.21730506420135498, + 0.023234941065311432, + 0.0836617723107338, + 0.062232840806245804 + ], + "credit_map": [ + 0.0217716284096241, + 0.038405951112508774, + 0.02399817295372486, + 0.024044301360845566, + 0.048900067806243896, + 0.04007260501384735, + 0.022207099944353104, + 0.06926419585943222, + 0.006467798259109259, + 0.013977074064314365, + 0.007055199239403009, + 0.008186631835997105, + 0.012679274193942547, + 0.012969428673386574, + 0.008470980450510979, + 0.01711285673081875, + 0.007645796984434128, + 0.015791229903697968, + 0.004757191985845566, + 0.006127166096121073, + 0.011143255047500134, + 0.008134415373206139, + 0.005550440400838852, + 0.013873684220016003, + 0.005717643070966005, + 0.009111505933105946, + 0.004928140435367823, + 0.004431778099387884, + 0.010358134284615517, + 0.010992719791829586, + 0.005585612263530493, + 0.01497920136898756, + 0.00622654939070344, + 0.005642581731081009, + 0.004589491989463568, + 0.004830700345337391, + 0.007290185894817114, + 0.006614083889871836, + 0.004497448448091745, + 0.007541951257735491, + 0.006483405828475952, + 0.008511745370924473, + 0.004529026802629232, + 0.0028586231637746096, + 0.007603601552546024, + 0.008888988755643368, + 0.0048154499381780624, + 0.007930424995720387, + 0.023314692080020905, + 0.03438672050833702, + 0.015741005539894104, + 0.02388651855289936, + 0.030208321288228035, + 0.04103962332010269, + 0.019028708338737488, + 0.029699480161070824, + 0.0030303667299449444, + 0.004544962663203478, + 0.0020953966304659843, + 0.0015003369189798832, + 0.0032641852740198374, + 0.0028646185528486967, + 0.002053303876891732, + 0.0038817720487713814, + 0.009726558811962605, + 0.010478585958480835, + 0.013659687712788582, + 0.005596034694463015, + 0.012221143580973148, + 0.010903031565248966, + 0.007695350795984268, + 0.01338137872517109, + 0.007334906607866287, + 0.00840449146926403, + 0.005725793074816465, + 0.0048407260328531265, + 0.010937627404928207, + 0.010560406371951103, + 0.0067810663022100925, + 0.007647825870662928 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 1 + ], + "selected_reward": -0.04306168854236603 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject a young Southeast Asian man with thick, wavy black hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a soft, cream-colored linen shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man holding a worn, leather-bound journal and writing with a classic fountain pen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by golden hour sunlight streaming from the right, casting long soft shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera positioned at a static, high-angle medium-close-up?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05493827536702156 + ], + "opposite_logprob_tokens": [ + -2.929938316345215 + ], + "logprob_reward": -0.05493827536702156, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05493827536702156, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23503360152244568, + 0.10641299933195114, + 0.03936324268579483, + 0.1383393555879593, + 0.0945095494389534, + 0.05483516305685043, + 0.0800599753856659, + 0.1344599425792694, + 0.05563976988196373, + 0.061346348375082016 + ], + "credit_map": [ + 0.03171461075544357, + 0.034748658537864685, + 0.020727332681417465, + 0.021574538201093674, + 0.034124955534935, + 0.0354316309094429, + 0.022725140675902367, + 0.03398672863841057, + 0.010385500267148018, + 0.01502918265759945, + 0.010143875144422054, + 0.013094106689095497, + 0.01587488129734993, + 0.013232018798589706, + 0.012716542929410934, + 0.015936898067593575, + 0.004663701169192791, + 0.00500614196062088, + 0.003119225613772869, + 0.004203713033348322, + 0.005750247742980719, + 0.005204853601753712, + 0.0037231429014354944, + 0.007692215498536825, + 0.0173342265188694, + 0.0362262986600399, + 0.011749946512281895, + 0.013283288106322289, + 0.021290002390742302, + 0.014665994793176651, + 0.009067484177649021, + 0.014722121879458427, + 0.00806078128516674, + 0.026428701356053352, + 0.006700906902551651, + 0.009070299565792084, + 0.014853697270154953, + 0.008694272488355637, + 0.008098915219306946, + 0.012601979076862335, + 0.006665749475359917, + 0.013820142485201359, + 0.004535506945103407, + 0.00391031801700592, + 0.007367035839706659, + 0.00805612001568079, + 0.004128318279981613, + 0.006351970136165619, + 0.010347326286137104, + 0.010893004015088081, + 0.006242609582841396, + 0.008020506240427494, + 0.009455809369683266, + 0.01581558585166931, + 0.0062538436613976955, + 0.013031288981437683, + 0.012965229339897633, + 0.031925786286592484, + 0.01216817181557417, + 0.00965869054198265, + 0.020472481846809387, + 0.01611456274986267, + 0.013466154225170612, + 0.017688877880573273, + 0.008500469848513603, + 0.009786760434508324, + 0.005129130557179451, + 0.004655785858631134, + 0.007444526068866253, + 0.008836913853883743, + 0.004648256581276655, + 0.006637924816459417, + 0.007938344962894917, + 0.011585029773414135, + 0.00519524747505784, + 0.0041716680862009525, + 0.0077896625734865665, + 0.010269811376929283, + 0.00618814816698432, + 0.008208440616726875 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 7 + ], + "selected_reward": -0.05493827536702156 + } + ], + "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/0000156/group_2/index_tensor([4352, 4352])_group_2_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000156/group_2/index_tensor([4352, 4352])_group_2_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..8ea2ebfae6d1cf3cc4231382d55f8474aa740831 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000156/group_2/index_tensor([4352, 4352])_group_2_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "static, extreme close-up with an incredibly shallow depth of field focuses on the mouth and nose of a young woman, her eyes gently closed in deep, serene concentration. Warm, soft light, like that streaming through a nearby window in the late afternoon, sculpts her features, highlighting the smooth texture of her fair skin and the delicate, dark lashes framing her cheeks. A perfect, shimmering drop of thick, golden-amber honey erratically lands on the bow of a cupid's plump upper lip, its viscosity catching the light and causing it to glow. Slowly and deliberately, her tongue emerges from slightly parted lips, tracing a sensual path upward in search of the honey. The camera captures the microscopic textures of her tongue's contact, the way the honey clings and stretches before being delicately licked away. The entire action is unhurried, creating an atmosphere of intense sensual focus and quiet, indulgent pleasure, leaving her upper lip wet and shimmering in a warm glow, while the softly blurred background is a hazy, warm hue.", + "seed": 15794, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000156/group_2/index_4352_seed_15794_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": -0.027576595824211837, + "vlm_logprob": -0.027576595824211837, + "vlm_yes_no_margin": 4.075, + "vlm_token_credit": -0.027576595824211837, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 the mouth and nose of a young woman?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the depth of field incredibly shallow with a softly blurred background in a hazy warm hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a perfect, shimmering drop of thick golden-amber honey visible on the woman's upper lip?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the woman's tongue emerge slowly and deliberately to lick the honey off her lip?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting warm and soft, resembling late afternoon light streaming through a window?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.022060826420783997 + ], + "opposite_logprob_tokens": [ + -4.147060871124268 + ], + "logprob_reward": -0.022060826420783997, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.022060826420783997, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21937628090381622, + 0.07138236612081528, + 0.08392415940761566, + 0.0634964257478714, + 0.3059762716293335, + 0.08439157903194427, + 0.0391964428126812, + 0.032441411167383194, + 0.0539729930460453, + 0.04584207758307457 + ], + "credit_map": [ + 0.03596432879567146, + 0.022519933059811592, + 0.023072663694620132, + 0.018184905871748924, + 0.024592788890004158, + 0.023252753540873528, + 0.045481614768505096, + 0.02630729414522648, + 0.007677248679101467, + 0.008067178539931774, + 0.00842433050274849, + 0.005588633473962545, + 0.010512633249163628, + 0.009162094444036484, + 0.014728802256286144, + 0.007221447303891182, + 0.01163602527230978, + 0.009807164780795574, + 0.009065458551049232, + 0.009146603755652905, + 0.009124832227826118, + 0.0067111351527273655, + 0.020675037056207657, + 0.007757906801998615, + 0.009352467954158783, + 0.00728216115385294, + 0.006498745176941156, + 0.006030107382684946, + 0.006957150995731354, + 0.006953068543225527, + 0.014146038331091404, + 0.006276683881878853, + 0.03092009574174881, + 0.032838184386491776, + 0.03147228807210922, + 0.024622516706585884, + 0.041145745664834976, + 0.02560913935303688, + 0.0930657684803009, + 0.026302531361579895, + 0.008887499570846558, + 0.009455417282879353, + 0.008804034441709518, + 0.008344783447682858, + 0.008750610053539276, + 0.0070115611888468266, + 0.02562725357711315, + 0.007510423194617033, + 0.006409771740436554, + 0.003332143649458885, + 0.004652582108974457, + 0.0034570307470858097, + 0.004941950086504221, + 0.0051122973673045635, + 0.007102603558450937, + 0.00418806541711092, + 0.00435169693082571, + 0.002869379473850131, + 0.00481113838031888, + 0.002670872490853071, + 0.0046155876480042934, + 0.004717608448117971, + 0.004014375153928995, + 0.00439075194299221, + 0.011051525361835957, + 0.005923394579440355, + 0.009421072900295258, + 0.008208703249692917, + 0.004402026534080505, + 0.004807291552424431, + 0.005770498886704445, + 0.004388483706861734, + 0.005113676656037569, + 0.005494475830346346, + 0.0057455128990113735, + 0.004241909831762314, + 0.0055275289341807365, + 0.004722945392131805, + 0.00993134081363678, + 0.00506468303501606 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 5 + ], + "selected_reward": -0.022060826420783997 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 the mouth and nose of a young woman?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the depth of field incredibly shallow with a softly blurred background in a hazy warm hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a perfect, shimmering drop of thick golden-amber honey visible on the woman's upper lip?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the woman's tongue emerge slowly and deliberately to lick the honey off her lip?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting warm and soft, resembling late afternoon light streaming through a window?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006771237589418888 + ], + "opposite_logprob_tokens": [ + -5.006771087646484 + ], + "logprob_reward": -0.006771237589418888, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006771237589418888, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1775674670934677, + 0.09092626720666885, + 0.19039858877658844, + 0.033337194472551346, + 0.24910138547420502, + 0.05338888615369797, + 0.04683452099561691, + 0.12051726877689362, + 0.0233572106808424, + 0.014571276493370533 + ], + "credit_map": [ + 0.028835013508796692, + 0.01914912648499012, + 0.027422215789556503, + 0.01972014456987381, + 0.019390910863876343, + 0.01595822535455227, + 0.02621392160654068, + 0.0208779014647007, + 0.012758324854075909, + 0.009101251140236855, + 0.010594929568469524, + 0.009861357510089874, + 0.008850631304085255, + 0.012666014023125172, + 0.017128460109233856, + 0.009965299628674984, + 0.023896824568510056, + 0.021342994645237923, + 0.02063608355820179, + 0.02012738026678562, + 0.015422247350215912, + 0.032878972589969635, + 0.03084796853363514, + 0.02524610236287117, + 0.00537754176184535, + 0.0033241724595427513, + 0.004126721993088722, + 0.002808663295581937, + 0.004102466162294149, + 0.004408866632729769, + 0.004128385800868273, + 0.005060376599431038, + 0.025318600237369537, + 0.02737724781036377, + 0.026060333475470543, + 0.02004038542509079, + 0.0338730625808239, + 0.01866142451763153, + 0.07583943009376526, + 0.021930890157818794, + 0.009126277640461922, + 0.006631861440837383, + 0.006098515819758177, + 0.005324407480657101, + 0.005775000900030136, + 0.006053335964679718, + 0.00982724130153656, + 0.004552243743091822, + 0.006454933900386095, + 0.004920010454952717, + 0.005649009719491005, + 0.003966426942497492, + 0.00564402574673295, + 0.004283736925572157, + 0.011971667408943176, + 0.003944712225347757, + 0.02634703554213047, + 0.009703760966658592, + 0.02109464444220066, + 0.013059050776064396, + 0.009818132035434246, + 0.016153989359736443, + 0.01077963039278984, + 0.01356102991849184, + 0.003871204797178507, + 0.0021571882534772158, + 0.003438957268372178, + 0.004153252579271793, + 0.001830937690101564, + 0.0021974118426442146, + 0.0033726892434060574, + 0.002335570054128766, + 0.0024023319128900766, + 0.0016886547673493624, + 0.0019357812125235796, + 0.0015951738459989429, + 0.001636919449083507, + 0.0012937525752931833, + 0.0027280908543616533, + 0.001290571759454906 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 2, + 0 + ], + "selected_reward": -0.006771237589418888 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 the mouth and nose of a young woman?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the depth of field incredibly shallow with a softly blurred background in a hazy warm hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a perfect, shimmering drop of thick golden-amber honey visible on the woman's upper lip?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the woman's tongue emerge slowly and deliberately to lick the honey off her lip?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting warm and soft, resembling late afternoon light streaming through a window?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.033722881227731705 + ], + "opposite_logprob_tokens": [ + -3.4087228775024414 + ], + "logprob_reward": -0.033722881227731705, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.033722881227731705, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1067652478814125, + 0.13638721406459808, + 0.3539433479309082, + 0.08694341778755188, + 0.1324673593044281, + 0.048522792756557465, + 0.04155712574720383, + 0.030612513422966003, + 0.017809808254241943, + 0.04499119520187378 + ], + "credit_map": [ + 0.011805293150246143, + 0.012053240090608597, + 0.013016284443438053, + 0.010241972282528877, + 0.012889356352388859, + 0.010343799367547035, + 0.022840559482574463, + 0.013574725016951561, + 0.016346802935004234, + 0.011707458645105362, + 0.01523405872285366, + 0.01325621735304594, + 0.014336301013827324, + 0.011175406165421009, + 0.03963673114776611, + 0.014694235287606716, + 0.03563249856233597, + 0.04222864285111427, + 0.04001026600599289, + 0.037775661796331406, + 0.03781062737107277, + 0.027675451710820198, + 0.10196411609649658, + 0.03084605745971203, + 0.012614233419299126, + 0.01018261257559061, + 0.009875345975160599, + 0.00802670232951641, + 0.006556570064276457, + 0.009217685088515282, + 0.02178182639181614, + 0.008688430301845074, + 0.014475503005087376, + 0.010020734742283821, + 0.014045451767742634, + 0.01113216858357191, + 0.015322118997573853, + 0.01225239597260952, + 0.03877297788858414, + 0.016445988789200783, + 0.004138363059610128, + 0.004405321553349495, + 0.005068256054073572, + 0.003995837643742561, + 0.005851864814758301, + 0.004676471929997206, + 0.014433817937970161, + 0.005952852312475443, + 0.005873978603631258, + 0.003346705110743642, + 0.004323343746364117, + 0.004219903610646725, + 0.004831148311495781, + 0.004620147868990898, + 0.010490400716662407, + 0.0038514987099915743, + 0.007328165229409933, + 0.002886599861085415, + 0.004365603439509869, + 0.003503864398226142, + 0.002957578282803297, + 0.003247790038585663, + 0.004286649636924267, + 0.0020362597424536943, + 0.0024750232696533203, + 0.0024490333162248135, + 0.0022334663663059473, + 0.0021526783239096403, + 0.0021813102066516876, + 0.0016861422918736935, + 0.0028487704694271088, + 0.0017833823803812265, + 0.007396089844405651, + 0.0054292273707687855, + 0.0067474814131855965, + 0.0062914346344769, + 0.004704905208200216, + 0.003711513476446271, + 0.0063967215828597546, + 0.004313813056796789 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 1, + 4 + ], + "selected_reward": -0.033722881227731705 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 the mouth and nose of a young woman?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the depth of field incredibly shallow with a softly blurred background in a hazy warm hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a perfect, shimmering drop of thick golden-amber honey visible on the woman's upper lip?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the woman's tongue emerge slowly and deliberately to lick the honey off her lip?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting warm and soft, resembling late afternoon light streaming through a window?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07005894184112549 + ], + "opposite_logprob_tokens": [ + -2.695058822631836 + ], + "logprob_reward": -0.07005894184112549, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07005894184112549, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0807829201221466, + 0.09451992064714432, + 0.19225835800170898, + 0.07191058248281479, + 0.4632844924926758, + 0.011970617808401585, + 0.0412418507039547, + 0.02021261677145958, + 0.015714379027485847, + 0.008104221895337105 + ], + "credit_map": [ + 0.010360786691308022, + 0.007932817563414574, + 0.009243367239832878, + 0.008428080007433891, + 0.009329107590019703, + 0.006672352086752653, + 0.020624589174985886, + 0.008191828615963459, + 0.009412441402673721, + 0.010377890430390835, + 0.01029811892658472, + 0.007718690671026707, + 0.010569236241281033, + 0.007240395527333021, + 0.030062366276979446, + 0.008840780705213547, + 0.022394396364688873, + 0.021799247711896896, + 0.02208578586578369, + 0.01617925800383091, + 0.02796025015413761, + 0.02079072594642639, + 0.04234842211008072, + 0.01870027743279934, + 0.008334051817655563, + 0.008590171113610268, + 0.0078828614205122, + 0.006345746107399464, + 0.008548825047910213, + 0.00853610597550869, + 0.016520582139492035, + 0.0071522509679198265, + 0.04683460667729378, + 0.04856232926249504, + 0.0480661541223526, + 0.03574361279606819, + 0.06382109224796295, + 0.04310233145952225, + 0.1367689073085785, + 0.04038550704717636, + 0.0016844968777149916, + 0.0015239061322063208, + 0.0012845757883042097, + 0.0010323278838768601, + 0.0014253839617595077, + 0.0017069898312911391, + 0.0022778857965022326, + 0.0010350530501455069, + 0.004720435943454504, + 0.004436459392309189, + 0.004756969399750233, + 0.0034581006038933992, + 0.00530893774703145, + 0.0038091065362095833, + 0.011050851084291935, + 0.003700993489474058, + 0.003438456216827035, + 0.002385263331234455, + 0.0028039393946528435, + 0.0028995394241064787, + 0.0022909122053533792, + 0.0024060900323092937, + 0.002637505065649748, + 0.0013509135460481048, + 0.0024582580663263798, + 0.0019145322730764747, + 0.002292066579684615, + 0.0022245515137910843, + 0.0015580515610054135, + 0.0014694957062602043, + 0.0023217445705085993, + 0.0014756805030629039, + 0.0009556914446875453, + 0.0007202517008408904, + 0.0010905101662501693, + 0.0009674544562585652, + 0.0009312389302067459, + 0.0008913275087252259, + 0.0015747195575386286, + 0.000973029003944248 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 2, + 1 + ], + "selected_reward": -0.07005894184112549 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 the mouth and nose of a young woman?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the depth of field incredibly shallow with a softly blurred background in a hazy warm hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a perfect, shimmering drop of thick golden-amber honey visible on the woman's upper lip?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the woman's tongue emerge slowly and deliberately to lick the honey off her lip?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting warm and soft, resembling late afternoon light streaming through a window?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005269092041999102 + ], + "opposite_logprob_tokens": [ + -5.2552690505981445 + ], + "logprob_reward": -0.005269092041999102, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005269092041999102, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09317050129175186, + 0.1259303241968155, + 0.1819511204957962, + 0.07361705601215363, + 0.15280067920684814, + 0.03247777372598648, + 0.08145102113485336, + 0.18760930001735687, + 0.050997521728277206, + 0.019994670525193214 + ], + "credit_map": [ + 0.01709907315671444, + 0.007794548757374287, + 0.01152376364916563, + 0.009480041451752186, + 0.011320976540446281, + 0.009992861188948154, + 0.01618543080985546, + 0.009773805737495422, + 0.019431831315159798, + 0.014081076718866825, + 0.017824826762080193, + 0.011721969582140446, + 0.00997910462319851, + 0.018549825996160507, + 0.017967013642191887, + 0.01637469418346882, + 0.018446190282702446, + 0.02424340695142746, + 0.02208487130701542, + 0.015100983902812004, + 0.023791169747710228, + 0.017920399084687233, + 0.04363754391670227, + 0.016726555302739143, + 0.011946706101298332, + 0.009196247905492783, + 0.008497072383761406, + 0.008508746512234211, + 0.006345386616885662, + 0.013498938642442226, + 0.007438874803483486, + 0.008185082115232944, + 0.01725965552031994, + 0.014768279157578945, + 0.017918996512889862, + 0.010971448384225368, + 0.02166071906685829, + 0.012710709124803543, + 0.043904174119234085, + 0.013606706634163857, + 0.0066388756968081, + 0.0029766422230750322, + 0.0038498479407280684, + 0.003344205906614661, + 0.0032592425122857094, + 0.0035298443399369717, + 0.004848743323236704, + 0.004030371550470591, + 0.014934808015823364, + 0.00958856102079153, + 0.01218058168888092, + 0.0077829319052398205, + 0.009992905892431736, + 0.010065421462059021, + 0.009853844530880451, + 0.007051966618746519, + 0.04153132066130638, + 0.02105698361992836, + 0.03104146011173725, + 0.023941276594996452, + 0.01927885226905346, + 0.018284128978848457, + 0.020892122760415077, + 0.011583163402974606, + 0.009219054132699966, + 0.0054337033070623875, + 0.008372251875698566, + 0.008201004937291145, + 0.003838587086647749, + 0.004823909141123295, + 0.006570746656507254, + 0.004538265522569418, + 0.0036952069494873285, + 0.001936553162522614, + 0.0025987690314650536, + 0.002473859116435051, + 0.0023255650885403156, + 0.0019618868827819824, + 0.0026761109475046396, + 0.0023267194628715515 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 2, + 4 + ], + "selected_reward": -0.005269092041999102 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, extreme close-up with an incredibly shallow depth of field focuses on the mouth and nose of a young woman, her eyes gently closed in deep, serene concentration. Warm, soft light, like that streaming through a nearby window in the late afternoon, sculpts her features, highlighting the smooth texture of her fair skin and the delicate, dark lashes framing her cheeks. A perfect, shimmering drop of thick, golden-amber honey erratically lands on the bow of a cupid's plump upper lip, its viscosity catching the light and causing it to glow. Slowly and deliberately, her tongue emerges from slightly parted lips, tracing a sensual path upward in search of the honey. The camera captures the microscopic textures of her tongue's contact, the way the honey clings and stretches before being delicately licked away. The entire action is unhurried, creating an atmosphere of intense sensual focus and quiet, indulgent pleasure, leaving her upper lip wet and shimmering in a warm glow, while the softly blurred background is a hazy, warm hue.", + "seed": 15795, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000156/group_2/index_4352_seed_15795_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": -0.01110481177456677, + "vlm_logprob": -0.01110481177456677, + "vlm_yes_no_margin": 4.85, + "vlm_token_credit": -0.01110481177456677, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 the mouth and nose of a young woman?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the depth of field incredibly shallow with a softly blurred background in a hazy warm hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a perfect, shimmering drop of thick golden-amber honey visible on the woman's upper lip?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the woman's tongue emerge slowly and deliberately to lick the honey off her lip?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting warm and soft, resembling late afternoon light streaming through a window?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01743186078965664 + ], + "opposite_logprob_tokens": [ + -4.642431735992432 + ], + "logprob_reward": -0.01743186078965664, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.01743186078965664, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21678844094276428, + 0.2187856286764145, + 0.03882405906915665, + 0.0766221210360527, + 0.0512363538146019, + 0.04851093888282776, + 0.037627462297677994, + 0.05618946626782417, + 0.1461951732635498, + 0.10922032594680786 + ], + "credit_map": [ + 0.020480984821915627, + 0.017382076010107994, + 0.030403008684515953, + 0.020912669599056244, + 0.04442581534385681, + 0.037002772092819214, + 0.02650180645287037, + 0.019679326564073563, + 0.018845602869987488, + 0.01832588016986847, + 0.02989816665649414, + 0.022183241322636604, + 0.022453317418694496, + 0.016040347516536713, + 0.06607761979103088, + 0.02496146224439144, + 0.004300497937947512, + 0.0037375777028501034, + 0.004751289263367653, + 0.004675204399973154, + 0.007868273183703423, + 0.004364809021353722, + 0.005509114358574152, + 0.0036172971595078707, + 0.008010516874492168, + 0.007625814061611891, + 0.010472780093550682, + 0.006746289320290089, + 0.010424355044960976, + 0.008586930111050606, + 0.0155189149081707, + 0.009236518293619156, + 0.0033045264426618814, + 0.004720560275018215, + 0.0056703039444983006, + 0.004018004052340984, + 0.009031709283590317, + 0.006396602839231491, + 0.013165494427084923, + 0.004929158836603165, + 0.004259446170181036, + 0.007516924757510424, + 0.00671324273571372, + 0.004245060030370951, + 0.008779872208833694, + 0.005457928404211998, + 0.007320457138121128, + 0.0042180088348686695, + 0.0034852575045078993, + 0.007119075395166874, + 0.004631345625966787, + 0.0037809391506016254, + 0.004878873471170664, + 0.004340514075011015, + 0.0066752624697983265, + 0.002716196235269308, + 0.0078037274070084095, + 0.007904290221631527, + 0.0069601284340023994, + 0.005011802539229393, + 0.00645429827272892, + 0.005936324596405029, + 0.010353373363614082, + 0.005765525158494711, + 0.01279087271541357, + 0.012758725322782993, + 0.01610073447227478, + 0.012680943123996258, + 0.02597448229789734, + 0.017798811197280884, + 0.03297141566872597, + 0.015119199641048908, + 0.011338631622493267, + 0.009915574453771114, + 0.013186163268983364, + 0.007824893109500408, + 0.018089378252625465, + 0.014401735737919807, + 0.024097871035337448, + 0.010366080328822136 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 8 + ], + "selected_reward": -0.01743186078965664 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 the mouth and nose of a young woman?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the depth of field incredibly shallow with a softly blurred background in a hazy warm hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a perfect, shimmering drop of thick golden-amber honey visible on the woman's upper lip?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the woman's tongue emerge slowly and deliberately to lick the honey off her lip?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting warm and soft, resembling late afternoon light streaming through a window?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0025119201745837927 + ], + "opposite_logprob_tokens": [ + -6.002511978149414 + ], + "logprob_reward": -0.0025119201745837927, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.0025119201745837927, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1656625121831894, + 0.22649088501930237, + 0.11466497927904129, + 0.03622008115053177, + 0.04166730120778084, + 0.05220237001776695, + 0.09522087872028351, + 0.11966084688901901, + 0.08934319764375687, + 0.058866988867521286 + ], + "credit_map": [ + 0.019515786319971085, + 0.017436137422919273, + 0.021183444187045097, + 0.018198398873209953, + 0.026538267731666565, + 0.0254112109541893, + 0.02014145255088806, + 0.017237817868590355, + 0.018683025613427162, + 0.021507924422621727, + 0.030475564301013947, + 0.02237079292535782, + 0.02324414625763893, + 0.015740539878606796, + 0.07087797671556473, + 0.023590929806232452, + 0.01745297573506832, + 0.014361599460244179, + 0.016217365860939026, + 0.015507001429796219, + 0.013843456283211708, + 0.0100706797093153, + 0.013996591791510582, + 0.013215303421020508, + 0.00581045588478446, + 0.002830802695825696, + 0.005764326546341181, + 0.003100320231169462, + 0.003035565372556448, + 0.004186528269201517, + 0.004564409609884024, + 0.006927674636244774, + 0.003431226359680295, + 0.004536419175565243, + 0.0043311938643455505, + 0.0031211620662361383, + 0.007089322432875633, + 0.005107690114527941, + 0.008696351200342178, + 0.0053539336659014225, + 0.004244259092956781, + 0.005552489310503006, + 0.007358034141361713, + 0.004652298055589199, + 0.010211769491434097, + 0.006064560730010271, + 0.009439687244594097, + 0.0046792724169790745, + 0.008501147851347923, + 0.01489218883216381, + 0.01220612321048975, + 0.008144188672304153, + 0.012649147771298885, + 0.015092246234416962, + 0.013843745924532413, + 0.009892089292407036, + 0.009486909955739975, + 0.011669536121189594, + 0.014660903252661228, + 0.009382197633385658, + 0.017590979114174843, + 0.014538519084453583, + 0.02710023708641529, + 0.015231559984385967, + 0.007679078262299299, + 0.00809403695166111, + 0.010108647868037224, + 0.00842141080647707, + 0.012969426810741425, + 0.010421870276331902, + 0.023272527381777763, + 0.008376190438866615, + 0.0057837520726025105, + 0.005037292372435331, + 0.006809033919125795, + 0.0038504181429743767, + 0.008702936582267284, + 0.007129518780857325, + 0.01575753092765808, + 0.0057965051382780075 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 7 + ], + "selected_reward": -0.0025119201745837927 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 the mouth and nose of a young woman?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the depth of field incredibly shallow with a softly blurred background in a hazy warm hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a perfect, shimmering drop of thick golden-amber honey visible on the woman's upper lip?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the woman's tongue emerge slowly and deliberately to lick the honey off her lip?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting warm and soft, resembling late afternoon light streaming through a window?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014223036356270313 + ], + "opposite_logprob_tokens": [ + -4.264223098754883 + ], + "logprob_reward": -0.014223036356270313, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014223036356270313, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.058397307991981506, + 0.1587435007095337, + 0.02704785205423832, + 0.030658956617116928, + 0.09598296880722046, + 0.10548476129770279, + 0.2993755042552948, + 0.10859496146440506, + 0.04514102265238762, + 0.07057313621044159 + ], + "credit_map": [ + 0.005646147765219212, + 0.006117398850619793, + 0.006676459684967995, + 0.006322488654404879, + 0.0095452181994915, + 0.007251259405165911, + 0.00955219566822052, + 0.007286139763891697, + 0.014416687190532684, + 0.011120381765067577, + 0.022082384675741196, + 0.014301649294793606, + 0.01814875937998295, + 0.014523981139063835, + 0.047064486891031265, + 0.017085175961256027, + 0.0023552034981548786, + 0.002919537015259266, + 0.003353617386892438, + 0.0033782843966037035, + 0.004622349515557289, + 0.0024855078663676977, + 0.0047608003951609135, + 0.003172551514580846, + 0.002825483912602067, + 0.002589163603261113, + 0.003764354856684804, + 0.002540977904573083, + 0.004499441012740135, + 0.0025903077330440283, + 0.007996818982064724, + 0.0038524093106389046, + 0.005896304734051228, + 0.009008266031742096, + 0.010590569116175175, + 0.00822488684207201, + 0.01392815075814724, + 0.011068235151469707, + 0.02249208465218544, + 0.014774465002119541, + 0.008647097274661064, + 0.010930183343589306, + 0.014574401080608368, + 0.009093662723898888, + 0.019580502063035965, + 0.012272878549993038, + 0.021487371996045113, + 0.008898668922483921, + 0.030338024720549583, + 0.04064645245671272, + 0.04413731023669243, + 0.029055964201688766, + 0.04145139828324318, + 0.03525201603770256, + 0.04625026509165764, + 0.03224407136440277, + 0.010083121247589588, + 0.013984370976686478, + 0.014218976721167564, + 0.008245970122516155, + 0.011697549372911453, + 0.009984937496483326, + 0.027981489896774292, + 0.012398537248373032, + 0.004973636474460363, + 0.003214497584849596, + 0.005272945389151573, + 0.0029939888045191765, + 0.007916872389614582, + 0.004773771855980158, + 0.011654647067189217, + 0.004340664017945528, + 0.007110674865543842, + 0.00820254161953926, + 0.009267191402614117, + 0.005505830980837345, + 0.007436070591211319, + 0.005964063107967377, + 0.021296391263604164, + 0.005790363997220993 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 1, + 7 + ], + "selected_reward": -0.014223036356270313 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 the mouth and nose of a young woman?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the depth of field incredibly shallow with a softly blurred background in a hazy warm hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a perfect, shimmering drop of thick golden-amber honey visible on the woman's upper lip?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the woman's tongue emerge slowly and deliberately to lick the honey off her lip?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting warm and soft, resembling late afternoon light streaming through a window?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016081271693110466 + ], + "opposite_logprob_tokens": [ + -4.1410813331604 + ], + "logprob_reward": -0.016081271693110466, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016081271693110466, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09627999365329742, + 0.08432819694280624, + 0.026964373886585236, + 0.02516218088567257, + 0.0958276018500328, + 0.04195600375533104, + 0.17790110409259796, + 0.18812118470668793, + 0.18991321325302124, + 0.07354618608951569 + ], + "credit_map": [ + 0.0077508799731731415, + 0.011975924484431744, + 0.012732678093016148, + 0.010626571252942085, + 0.01303025707602501, + 0.01462455652654171, + 0.015203068032860756, + 0.01033604983240366, + 0.0066660079173743725, + 0.007388966623693705, + 0.011565393768250942, + 0.006507307291030884, + 0.01251183357089758, + 0.007256415206938982, + 0.023697499185800552, + 0.008734770119190216, + 0.003489898983389139, + 0.002684362232685089, + 0.0030467109754681587, + 0.0032876390032470226, + 0.004671177361160517, + 0.002698182128369808, + 0.004886392503976822, + 0.0022000072058290243, + 0.0018303345423191786, + 0.0022639743983745575, + 0.002944361185654998, + 0.0016948754200711846, + 0.0047374628484249115, + 0.002799080219119787, + 0.005082528572529554, + 0.0038095603231340647, + 0.005374576896429062, + 0.011044340208172798, + 0.010875360108911991, + 0.00839660968631506, + 0.010788421146571636, + 0.013111520558595657, + 0.022280823439359665, + 0.013955953530967236, + 0.0034949537366628647, + 0.0042737070471048355, + 0.005880434066057205, + 0.004264357965439558, + 0.00723362248390913, + 0.004914918914437294, + 0.007506459020078182, + 0.004387546796351671, + 0.019729629158973694, + 0.027880050241947174, + 0.027782948687672615, + 0.017218763008713722, + 0.02272062562406063, + 0.025072580203413963, + 0.0192144513130188, + 0.01828203909099102, + 0.015444723889231682, + 0.027199476957321167, + 0.022660091519355774, + 0.012476415373384953, + 0.029094811528921127, + 0.0231622364372015, + 0.035865433514118195, + 0.02221798151731491, + 0.01786213368177414, + 0.013981601223349571, + 0.021547481417655945, + 0.013685361482203007, + 0.02683340385556221, + 0.021495306864380836, + 0.05405329912900925, + 0.02045460045337677, + 0.007156815845519304, + 0.008110676892101765, + 0.008333204314112663, + 0.005755332764238119, + 0.00776344258338213, + 0.007224787026643753, + 0.022117413580417633, + 0.007084504701197147 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 6 + ], + "selected_reward": -0.016081271693110466 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 the mouth and nose of a young woman?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the depth of field incredibly shallow with a softly blurred background in a hazy warm hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a perfect, shimmering drop of thick golden-amber honey visible on the woman's upper lip?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the woman's tongue emerge slowly and deliberately to lick the honey off her lip?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting warm and soft, resembling late afternoon light streaming through a window?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005275969859212637 + ], + "opposite_logprob_tokens": [ + -5.255276203155518 + ], + "logprob_reward": -0.005275969859212637, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005275969859212637, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10358519852161407, + 0.10114685446023941, + 0.17610101401805878, + 0.0862017273902893, + 0.05211517587304115, + 0.06645598262548447, + 0.05561866983771324, + 0.1572081446647644, + 0.13812053203582764, + 0.06344670057296753 + ], + "credit_map": [ + 0.0157468244433403, + 0.01142433937638998, + 0.014842074364423752, + 0.014875801280140877, + 0.015284023247659206, + 0.009030620567500591, + 0.011680420488119125, + 0.010701099410653114, + 0.010978286154568195, + 0.009000862948596478, + 0.01540564838796854, + 0.013930200599133968, + 0.01229315809905529, + 0.012407226487994194, + 0.01386666763573885, + 0.013264809735119343, + 0.022449275478720665, + 0.02082391455769539, + 0.029534529894590378, + 0.021646132692694664, + 0.01867765188217163, + 0.016862349584698677, + 0.023977117612957954, + 0.022130033001303673, + 0.014606478624045849, + 0.005504126660525799, + 0.012971960008144379, + 0.008407184854149818, + 0.008192772977054119, + 0.00903195422142744, + 0.013937822543084621, + 0.013549422845244408, + 0.003630442079156637, + 0.005718375090509653, + 0.0059434049762785435, + 0.004276836756616831, + 0.007493619341403246, + 0.007747044786810875, + 0.010181605815887451, + 0.0071238442324101925, + 0.007455185055732727, + 0.007939892821013927, + 0.011762510053813457, + 0.00606429623439908, + 0.010499434545636177, + 0.007181690074503422, + 0.009653099812567234, + 0.005899875890463591, + 0.0051004658453166485, + 0.009167900308966637, + 0.007171014789491892, + 0.005735756363719702, + 0.00794995203614235, + 0.008518733084201813, + 0.006958466023206711, + 0.00501638138666749, + 0.014648398384451866, + 0.01563039980828762, + 0.019009321928024292, + 0.014734708704054356, + 0.015349093824625015, + 0.01979929767549038, + 0.03467321768403053, + 0.02336370199918747, + 0.014486780390143394, + 0.010784883983433247, + 0.017940733581781387, + 0.00913051888346672, + 0.020409230142831802, + 0.015020741149783134, + 0.038628995418548584, + 0.011718654073774815, + 0.008281916379928589, + 0.008429456502199173, + 0.008975289762020111, + 0.005661013536155224, + 0.009541207924485207, + 0.007900996133685112, + 0.008876848965883255, + 0.005779970437288284 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 7, + 8 + ], + "selected_reward": -0.005275969859212637 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, extreme close-up with an incredibly shallow depth of field focuses on the mouth and nose of a young woman, her eyes gently closed in deep, serene concentration. Warm, soft light, like that streaming through a nearby window in the late afternoon, sculpts her features, highlighting the smooth texture of her fair skin and the delicate, dark lashes framing her cheeks. A perfect, shimmering drop of thick, golden-amber honey erratically lands on the bow of a cupid's plump upper lip, its viscosity catching the light and causing it to glow. Slowly and deliberately, her tongue emerges from slightly parted lips, tracing a sensual path upward in search of the honey. The camera captures the microscopic textures of her tongue's contact, the way the honey clings and stretches before being delicately licked away. The entire action is unhurried, creating an atmosphere of intense sensual focus and quiet, indulgent pleasure, leaving her upper lip wet and shimmering in a warm glow, while the softly blurred background is a hazy, warm hue.", + "seed": 15796, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000156/group_2/index_4352_seed_15796_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": -0.011756272008642554, + "vlm_logprob": -0.011756272008642554, + "vlm_yes_no_margin": 4.975, + "vlm_token_credit": -0.011756272008642554, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 the mouth and nose of a young woman?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the depth of field incredibly shallow with a softly blurred background in a hazy warm hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a perfect, shimmering drop of thick golden-amber honey visible on the woman's upper lip?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the woman's tongue emerge slowly and deliberately to lick the honey off her lip?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting warm and soft, resembling late afternoon light streaming through a window?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00847233273088932 + ], + "opposite_logprob_tokens": [ + -5.758472442626953 + ], + "logprob_reward": -0.00847233273088932, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.00847233273088932, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0704277902841568, + 0.049996741116046906, + 0.08776313066482544, + 0.15909960865974426, + 0.06067337840795517, + 0.06861985474824905, + 0.06948292255401611, + 0.36974242329597473, + 0.025881769135594368, + 0.0383123904466629 + ], + "credit_map": [ + 0.010687758214771748, + 0.0074183340184390545, + 0.012439736165106297, + 0.007605523802340031, + 0.006564995273947716, + 0.010045538656413555, + 0.006916580721735954, + 0.008749324828386307, + 0.00785136129707098, + 0.004677480086684227, + 0.008692389354109764, + 0.005183184985071421, + 0.004241787828505039, + 0.00853864848613739, + 0.004159607458859682, + 0.006652283016592264, + 0.011136619374155998, + 0.0037724466528743505, + 0.035250984132289886, + 0.00923803262412548, + 0.0036396009381860495, + 0.012296373024582863, + 0.003945172298699617, + 0.00848389696329832, + 0.006840684916824102, + 0.004664100706577301, + 0.053351618349552155, + 0.00655916566029191, + 0.007085147313773632, + 0.06151339411735535, + 0.010015296749770641, + 0.009070189669728279, + 0.006600659340620041, + 0.0073548173531889915, + 0.009077467024326324, + 0.005529134999960661, + 0.007031469605863094, + 0.008131497539579868, + 0.01156905759125948, + 0.005379270296543837, + 0.0064637428149580956, + 0.008144362829625607, + 0.008006432093679905, + 0.0056193131022155285, + 0.008220026269555092, + 0.007725061383098364, + 0.01722988672554493, + 0.007211025804281235, + 0.005876298062503338, + 0.005540826357901096, + 0.007973150350153446, + 0.010455873794853687, + 0.007263215724378824, + 0.016308700665831566, + 0.008194472640752792, + 0.007870380766689777, + 0.026318952441215515, + 0.017289545387029648, + 0.016383852809667587, + 0.12025110423564911, + 0.010352765209972858, + 0.12283042818307877, + 0.016184723004698753, + 0.04013100638985634, + 0.0037255799397826195, + 0.002057380974292755, + 0.003415736835449934, + 0.0031193646136671305, + 0.003215062664821744, + 0.0032259547151625156, + 0.003503131680190563, + 0.003619556315243244, + 0.005048238206654787, + 0.0034830309450626373, + 0.005343729164451361, + 0.007135508581995964, + 0.0052070459350943565, + 0.0035304976627230644, + 0.00480780890211463, + 0.003756527788937092 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 3, + 2 + ], + "selected_reward": -0.00847233273088932 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 the mouth and nose of a young woman?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the depth of field incredibly shallow with a softly blurred background in a hazy warm hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a perfect, shimmering drop of thick golden-amber honey visible on the woman's upper lip?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the woman's tongue emerge slowly and deliberately to lick the honey off her lip?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting warm and soft, resembling late afternoon light streaming through a window?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003211818402633071 + ], + "opposite_logprob_tokens": [ + -5.753211975097656 + ], + "logprob_reward": -0.003211818402633071, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.003211818402633071, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.045911725610494614, + 0.03097894974052906, + 0.07142090797424316, + 0.29771196842193604, + 0.0489625558257103, + 0.06364999711513519, + 0.02117578312754631, + 0.2698872685432434, + 0.07996818423271179, + 0.07033269852399826 + ], + "credit_map": [ + 0.005761508364230394, + 0.005177071318030357, + 0.006648594979196787, + 0.0048517752438783646, + 0.004896506201475859, + 0.008470679633319378, + 0.005790838040411472, + 0.004314752295613289, + 0.004882002715021372, + 0.0026959835086017847, + 0.00508568761870265, + 0.0030671420972794294, + 0.003426984418183565, + 0.005890822969377041, + 0.0032091978937387466, + 0.0027211280539631844, + 0.0075586591847240925, + 0.003094986081123352, + 0.02942757122218609, + 0.006184110883623362, + 0.0026836437173187733, + 0.01357114128768444, + 0.003026183694601059, + 0.005874614231288433, + 0.013903792947530746, + 0.006789427250623703, + 0.10896550118923187, + 0.011723635718226433, + 0.009059162810444832, + 0.12586286664009094, + 0.006513537373393774, + 0.014894023537635803, + 0.005610805936157703, + 0.006270561367273331, + 0.009700966067612171, + 0.004645209293812513, + 0.00473558297380805, + 0.006286732852458954, + 0.008369393646717072, + 0.0033433008939027786, + 0.004631257615983486, + 0.0060863676480948925, + 0.006925961934030056, + 0.005387577693909407, + 0.007608860731124878, + 0.0082801328971982, + 0.01735912635922432, + 0.0073707131668925285, + 0.0017811368452385068, + 0.002318943152204156, + 0.0024528547655791044, + 0.002636775840073824, + 0.0025348004419356585, + 0.004256292711943388, + 0.002400995697826147, + 0.0027939837891608477, + 0.009766172617673874, + 0.0061750225722789764, + 0.0133581617847085, + 0.021790435537695885, + 0.007169833406805992, + 0.11565305292606354, + 0.006807737518101931, + 0.08916685730218887, + 0.013779646717011929, + 0.006993201095610857, + 0.01808966137468815, + 0.009680339135229588, + 0.005329712759703398, + 0.00858759880065918, + 0.005996563471853733, + 0.011511464603245258, + 0.0119678545743227, + 0.00623348169028759, + 0.013264144770801067, + 0.0099427979439497, + 0.005235782824456692, + 0.00701520312577486, + 0.0073316642083227634, + 0.009341767989099026 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 7, + 8 + ], + "selected_reward": -0.003211818402633071 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 the mouth and nose of a young woman?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the depth of field incredibly shallow with a softly blurred background in a hazy warm hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a perfect, shimmering drop of thick golden-amber honey visible on the woman's upper lip?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the woman's tongue emerge slowly and deliberately to lick the honey off her lip?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting warm and soft, resembling late afternoon light streaming through a window?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020596086978912354 + ], + "opposite_logprob_tokens": [ + -3.8955960273742676 + ], + "logprob_reward": -0.020596086978912354, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020596086978912354, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.026038315147161484, + 0.01660318300127983, + 0.0514671765267849, + 0.3252745568752289, + 0.08259876072406769, + 0.1329634040594101, + 0.03366009518504143, + 0.23217816650867462, + 0.05730358883738518, + 0.04191279038786888 + ], + "credit_map": [ + 0.0034165361430495977, + 0.0027613253332674503, + 0.004265116527676582, + 0.0025364647153764963, + 0.002803258365020156, + 0.0038621067069470882, + 0.0033651445992290974, + 0.003028362989425659, + 0.001842042664065957, + 0.0014829813735559583, + 0.002182705793529749, + 0.0019912985153496265, + 0.0023058943916112185, + 0.0022743484005331993, + 0.0027128392830491066, + 0.0018110745586454868, + 0.006056452635675669, + 0.003182005137205124, + 0.016913246363401413, + 0.004539172630757093, + 0.004029128234833479, + 0.006896884646266699, + 0.004761323798447847, + 0.0050889612175524235, + 0.016082556918263435, + 0.006550215650349855, + 0.11961379647254944, + 0.01268585305660963, + 0.010442616418004036, + 0.13713966310024261, + 0.0074844155460596085, + 0.015275433659553528, + 0.007542054168879986, + 0.010146720334887505, + 0.008175250142812729, + 0.0075498949736356735, + 0.009014560841023922, + 0.009740172885358334, + 0.022234028205275536, + 0.008196075446903706, + 0.009452992118895054, + 0.010768737643957138, + 0.011340499855577946, + 0.011756557039916515, + 0.01761096902191639, + 0.018015600740909576, + 0.03764844685792923, + 0.01636960729956627, + 0.0031811383087188005, + 0.002591925673186779, + 0.0033939785789698362, + 0.004636686760932207, + 0.003980468958616257, + 0.007342831697314978, + 0.004776109009981155, + 0.0037569573614746332, + 0.009389445185661316, + 0.008647831156849861, + 0.00787439476698637, + 0.058893993496894836, + 0.007740736473351717, + 0.09050039201974869, + 0.010276534594595432, + 0.03885482996702194, + 0.005132324527949095, + 0.004776765126734972, + 0.009058188647031784, + 0.0058699906803667545, + 0.008264005184173584, + 0.006581876426935196, + 0.011548422276973724, + 0.006072015035897493, + 0.0066557335667312145, + 0.002719214418902993, + 0.006368937902152538, + 0.009051382541656494, + 0.003642996773123741, + 0.003538019023835659, + 0.005043931305408478, + 0.004892578814178705 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 7, + 5 + ], + "selected_reward": -0.020596086978912354 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 the mouth and nose of a young woman?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the depth of field incredibly shallow with a softly blurred background in a hazy warm hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a perfect, shimmering drop of thick golden-amber honey visible on the woman's upper lip?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the woman's tongue emerge slowly and deliberately to lick the honey off her lip?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting warm and soft, resembling late afternoon light streaming through a window?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023297740146517754 + ], + "opposite_logprob_tokens": [ + -3.7732977867126465 + ], + "logprob_reward": -0.023297740146517754, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023297740146517754, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.036927562206983566, + 0.03605732321739197, + 0.07759048044681549, + 0.20036360621452332, + 0.10118241608142853, + 0.07601356506347656, + 0.13851386308670044, + 0.2737569510936737, + 0.03461235389113426, + 0.02498193457722664 + ], + "credit_map": [ + 0.0052999877370893955, + 0.0038658748380839825, + 0.007002437021583319, + 0.0033614912535995245, + 0.004902262706309557, + 0.004029730800539255, + 0.003976362757384777, + 0.0044894106686115265, + 0.00517929345369339, + 0.0036582360044121742, + 0.006068152841180563, + 0.004418720491230488, + 0.0037952931597828865, + 0.005950074642896652, + 0.00369382300414145, + 0.003293730318546295, + 0.007482045330107212, + 0.003247571177780628, + 0.029222261160612106, + 0.007027859799563885, + 0.004361416678875685, + 0.01197119988501072, + 0.007008879911154509, + 0.00726924417540431, + 0.008863434195518494, + 0.006150804925709963, + 0.07026752084493637, + 0.008075493387877941, + 0.009257689118385315, + 0.07068732380867004, + 0.01626693271100521, + 0.010794389061629772, + 0.008494479581713676, + 0.01154057402163744, + 0.012266615405678749, + 0.007335119415074587, + 0.009609244763851166, + 0.018713202327489853, + 0.023545052856206894, + 0.009678119793534279, + 0.00846150889992714, + 0.005446175113320351, + 0.007026113104075193, + 0.005444048438221216, + 0.016250286251306534, + 0.012062101624906063, + 0.012761619873344898, + 0.00856170803308487, + 0.011138545349240303, + 0.010919934138655663, + 0.01190119981765747, + 0.01565607078373432, + 0.019445179030299187, + 0.03112904168665409, + 0.02149876393377781, + 0.016825100407004356, + 0.014583492651581764, + 0.010655696503818035, + 0.01124438364058733, + 0.07665681093931198, + 0.008714954368770123, + 0.1015891432762146, + 0.010876111686229706, + 0.03943632170557976, + 0.0035469236318022013, + 0.0028181569650769234, + 0.004825684241950512, + 0.0066462052054703236, + 0.004075944423675537, + 0.0035830261185765266, + 0.006210541818290949, + 0.0029058665968477726, + 0.003989004995673895, + 0.0019335317192599177, + 0.003625972429290414, + 0.0027328624855726957, + 0.0033635974396020174, + 0.0021521870512515306, + 0.00491632055491209, + 0.002268455224111676 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 3, + 6 + ], + "selected_reward": -0.023297740146517754 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 the mouth and nose of a young woman?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the depth of field incredibly shallow with a softly blurred background in a hazy warm hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a perfect, shimmering drop of thick golden-amber honey visible on the woman's upper lip?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the woman's tongue emerge slowly and deliberately to lick the honey off her lip?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting warm and soft, resembling late afternoon light streaming through a window?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003203381784260273 + ], + "opposite_logprob_tokens": [ + -5.753203392028809 + ], + "logprob_reward": -0.003203381784260273, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.003203381784260273, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05556650087237358, + 0.11090365052223206, + 0.042538583278656006, + 0.16239455342292786, + 0.030380303040146828, + 0.036896731704473495, + 0.01897413656115532, + 0.2357032150030136, + 0.15745429694652557, + 0.14918795228004456 + ], + "credit_map": [ + 0.009058967232704163, + 0.005118261557072401, + 0.014074437320232391, + 0.005459449253976345, + 0.005010603461414576, + 0.0068122283555567265, + 0.005421032197773457, + 0.004611518699675798, + 0.01926160790026188, + 0.008710569702088833, + 0.031831804662942886, + 0.009869443252682686, + 0.00773903401568532, + 0.015542106702923775, + 0.00850279163569212, + 0.009446287527680397, + 0.004426657687872648, + 0.0017624844331294298, + 0.016978833824396133, + 0.003691965015605092, + 0.0019435053691267967, + 0.008464530110359192, + 0.002072557108476758, + 0.0031980490311980247, + 0.007384373340755701, + 0.003947205375880003, + 0.056842271238565445, + 0.006626334972679615, + 0.005766532849520445, + 0.06788203865289688, + 0.0050383061170578, + 0.008907497860491276, + 0.002766914665699005, + 0.003125967690721154, + 0.0048813470639288425, + 0.0021272143349051476, + 0.0025513984728604555, + 0.006040489766746759, + 0.006190233863890171, + 0.0026967343874275684, + 0.003277492942288518, + 0.004099318291991949, + 0.003879933152347803, + 0.003553741378709674, + 0.004717790987342596, + 0.004376913420855999, + 0.009893377311527729, + 0.0030981649179011583, + 0.002129739848896861, + 0.0025382426101714373, + 0.0024193988647311926, + 0.0020270992536097765, + 0.00225983839482069, + 0.002187678124755621, + 0.003402085741981864, + 0.0020100532565265894, + 0.00833226926624775, + 0.005581348668783903, + 0.016299625858664513, + 0.019833652302622795, + 0.007257656194269657, + 0.100281722843647, + 0.006059616804122925, + 0.0720573216676712, + 0.016582634299993515, + 0.010231590829789639, + 0.02169991284608841, + 0.04698677361011505, + 0.007345191668719053, + 0.012418755330145359, + 0.010350454598665237, + 0.03183898702263832, + 0.027761993929743767, + 0.01442453172057867, + 0.03566999360918999, + 0.02198369987308979, + 0.009683715179562569, + 0.013744735158979893, + 0.012853076681494713, + 0.01306620892137289 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 3, + 8 + ], + "selected_reward": -0.003203381784260273 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, extreme close-up with an incredibly shallow depth of field focuses on the mouth and nose of a young woman, her eyes gently closed in deep, serene concentration. Warm, soft light, like that streaming through a nearby window in the late afternoon, sculpts her features, highlighting the smooth texture of her fair skin and the delicate, dark lashes framing her cheeks. A perfect, shimmering drop of thick, golden-amber honey erratically lands on the bow of a cupid's plump upper lip, its viscosity catching the light and causing it to glow. Slowly and deliberately, her tongue emerges from slightly parted lips, tracing a sensual path upward in search of the honey. The camera captures the microscopic textures of her tongue's contact, the way the honey clings and stretches before being delicately licked away. The entire action is unhurried, creating an atmosphere of intense sensual focus and quiet, indulgent pleasure, leaving her upper lip wet and shimmering in a warm glow, while the softly blurred background is a hazy, warm hue.", + "seed": 15797, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000156/group_2/index_4352_seed_15797_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": -0.030866612028330565, + "vlm_logprob": -0.030866612028330565, + "vlm_yes_no_margin": 4.25, + "vlm_token_credit": -0.030866612028330565, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 the mouth and nose of a young woman?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the depth of field incredibly shallow with a softly blurred background in a hazy warm hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a perfect, shimmering drop of thick golden-amber honey visible on the woman's upper lip?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the woman's tongue emerge slowly and deliberately to lick the honey off her lip?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting warm and soft, resembling late afternoon light streaming through a window?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01574278250336647 + ], + "opposite_logprob_tokens": [ + -4.515742778778076 + ], + "logprob_reward": -0.01574278250336647, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.01574278250336647, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16306261718273163, + 0.15478432178497314, + 0.04572655260562897, + 0.10556326061487198, + 0.17199651896953583, + 0.13178126513957977, + 0.05037515237927437, + 0.03039042092859745, + 0.05374285206198692, + 0.09257706254720688 + ], + "credit_map": [ + 0.01882479526102543, + 0.01819583587348461, + 0.022338949143886566, + 0.02307955175638199, + 0.021869203075766563, + 0.020730121061205864, + 0.024918844923377037, + 0.013105313293635845, + 0.013326911255717278, + 0.009907622821629047, + 0.019711341708898544, + 0.0180380716919899, + 0.019886834546923637, + 0.023191023617982864, + 0.03774695098400116, + 0.012975563295185566, + 0.0044561875984072685, + 0.004613992292433977, + 0.005193899851292372, + 0.0052688391879200935, + 0.005789394956082106, + 0.005130053497850895, + 0.01000172644853592, + 0.0052724555134773254, + 0.017825474962592125, + 0.006180208642035723, + 0.014586660079658031, + 0.008052588440477848, + 0.012926318682730198, + 0.012881070375442505, + 0.013389701023697853, + 0.019721245393157005, + 0.021842874586582184, + 0.01838419958949089, + 0.023687539622187614, + 0.024976510554552078, + 0.02664855308830738, + 0.021616365760564804, + 0.02194574847817421, + 0.012894726358354092, + 0.014795981347560883, + 0.012163951992988586, + 0.02031160145998001, + 0.02148035541176796, + 0.02079145424067974, + 0.016147926449775696, + 0.014780097641050816, + 0.01130988821387291, + 0.009165207855403423, + 0.002899469807744026, + 0.007880090735852718, + 0.004705472849309444, + 0.007985895499587059, + 0.006120845675468445, + 0.005858348682522774, + 0.005759820342063904, + 0.006426509469747543, + 0.002070607617497444, + 0.004750005900859833, + 0.0024899113923311234, + 0.0030461314599961042, + 0.003713530022650957, + 0.0038679218851029873, + 0.00402580201625824, + 0.008814112283289433, + 0.003793330630287528, + 0.010488424450159073, + 0.005858500953763723, + 0.006158136762678623, + 0.005040553864091635, + 0.007924219593405724, + 0.005665576551109552, + 0.011412162333726883, + 0.007002774626016617, + 0.012726527638733387, + 0.007589309476315975, + 0.011477881111204624, + 0.008573666214942932, + 0.02411087043583393, + 0.009683875367045403 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 1 + ], + "selected_reward": -0.01574278250336647 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 the mouth and nose of a young woman?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the depth of field incredibly shallow with a softly blurred background in a hazy warm hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a perfect, shimmering drop of thick golden-amber honey visible on the woman's upper lip?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the woman's tongue emerge slowly and deliberately to lick the honey off her lip?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting warm and soft, resembling late afternoon light streaming through a window?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006757739465683699 + ], + "opposite_logprob_tokens": [ + -5.006757736206055 + ], + "logprob_reward": -0.006757739465683699, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006757739465683699, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12257180362939835, + 0.12541694939136505, + 0.04189032316207886, + 0.08684065192937851, + 0.17396710813045502, + 0.1309375911951065, + 0.06950577348470688, + 0.20245930552482605, + 0.014061097986996174, + 0.03234941512346268 + ], + "credit_map": [ + 0.012812064029276371, + 0.011395877227187157, + 0.016352102160453796, + 0.016270041465759277, + 0.016759661957621574, + 0.012727541849017143, + 0.017619498074054718, + 0.018635014072060585, + 0.013851313851773739, + 0.010323398746550083, + 0.016066206619143486, + 0.013659020885825157, + 0.014447825960814953, + 0.014998436905443668, + 0.032207101583480835, + 0.009863641113042831, + 0.005520953796803951, + 0.004003913141787052, + 0.00570289371535182, + 0.004120546858757734, + 0.005121933296322823, + 0.0040313019417226315, + 0.009460339322686195, + 0.0039284429512917995, + 0.01205342449247837, + 0.0050689829513430595, + 0.011076339520514011, + 0.00821305625140667, + 0.010697687044739723, + 0.010978755541145802, + 0.013089924119412899, + 0.01566246896982193, + 0.0260988287627697, + 0.016437536105513573, + 0.026834385469555855, + 0.02481091022491455, + 0.02475842647254467, + 0.01826479285955429, + 0.0213607344776392, + 0.015401502139866352, + 0.020239800214767456, + 0.016885412856936455, + 0.012673311866819859, + 0.02201974391937256, + 0.020756008103489876, + 0.012868866324424744, + 0.01408739760518074, + 0.011407050304114819, + 0.015422709286212921, + 0.005616278387606144, + 0.012881895527243614, + 0.005945220123976469, + 0.00917739886790514, + 0.007466237060725689, + 0.005991669371724129, + 0.007004363927990198, + 0.042135193943977356, + 0.016848595812916756, + 0.03648756444454193, + 0.017305046319961548, + 0.019430210813879967, + 0.02098805084824562, + 0.028624985367059708, + 0.020639663562178612, + 0.0015854033408686519, + 0.0010455549927428365, + 0.0016216812655329704, + 0.002019566250964999, + 0.001776625053025782, + 0.0012555395951494575, + 0.0026700631715357304, + 0.0020866640843451023, + 0.0028920769691467285, + 0.0026102736592292786, + 0.004278167150914669, + 0.0029740682803094387, + 0.004446101374924183, + 0.0031723796855658293, + 0.007991557009518147, + 0.003984792623668909 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 4, + 5 + ], + "selected_reward": -0.006757739465683699 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 the mouth and nose of a young woman?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the depth of field incredibly shallow with a softly blurred background in a hazy warm hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a perfect, shimmering drop of thick golden-amber honey visible on the woman's upper lip?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the woman's tongue emerge slowly and deliberately to lick the honey off her lip?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting warm and soft, resembling late afternoon light streaming through a window?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"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": [ + 0.04718266427516937, + 0.14852015674114227, + 0.0815901830792427, + 0.06138068065047264, + 0.2656358778476715, + 0.32296624779701233, + 0.030884921550750732, + 0.012376053258776665, + 0.015275460667908192, + 0.014187782071530819 + ], + "credit_map": [ + 0.004845513962209225, + 0.004508042708039284, + 0.006606826093047857, + 0.006347056478261948, + 0.0065421368926763535, + 0.006067104171961546, + 0.006838898174464703, + 0.005427084863185883, + 0.014595761895179749, + 0.012369350530207157, + 0.017030268907546997, + 0.018635524436831474, + 0.017520058900117874, + 0.016027677804231644, + 0.03974561765789986, + 0.012595891952514648, + 0.008497047238051891, + 0.007032531779259443, + 0.008023038506507874, + 0.006654751021414995, + 0.011537044309079647, + 0.009483939968049526, + 0.021609241142868996, + 0.008752592839300632, + 0.007797894533723593, + 0.005011511500924826, + 0.00736964913085103, + 0.007817864418029785, + 0.009016606956720352, + 0.00693928450345993, + 0.011664357967674732, + 0.005763508379459381, + 0.0374220572412014, + 0.02742343209683895, + 0.033082474023103714, + 0.036717548966407776, + 0.04289431497454643, + 0.032465092837810516, + 0.034527406096458435, + 0.021103542298078537, + 0.04442761465907097, + 0.03946305066347122, + 0.035563357174396515, + 0.04854463413357735, + 0.0558585487306118, + 0.0336848646402359, + 0.035418421030044556, + 0.03000575862824917, + 0.006600170861929655, + 0.0020021928939968348, + 0.00555455731227994, + 0.0022462119814008474, + 0.0038476039189845324, + 0.0031261765398085117, + 0.00301920622587204, + 0.004488800186663866, + 0.0017510849284008145, + 0.0009129422251135111, + 0.0016163368709385395, + 0.0012900126166641712, + 0.0011884365230798721, + 0.0012501653982326388, + 0.0028064772486686707, + 0.001560597331263125, + 0.0012518437579274178, + 0.0013238136889413, + 0.0021049289498478174, + 0.0025025804061442614, + 0.0016414395067840815, + 0.001361739938147366, + 0.00343984249047935, + 0.001649271696805954, + 0.0016212223563343287, + 0.001066719414666295, + 0.0019941693171858788, + 0.0012401487911120057, + 0.002069217851385474, + 0.0017098239623010159, + 0.003181109670549631, + 0.0013053705915808678 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 4, + 1 + ], + "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\": \"Is the shot an extreme close-up focusing on the mouth and nose of a young woman?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the depth of field incredibly shallow with a softly blurred background in a hazy warm hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a perfect, shimmering drop of thick golden-amber honey visible on the woman's upper lip?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the woman's tongue emerge slowly and deliberately to lick the honey off her lip?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting warm and soft, resembling late afternoon light streaming through a window?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1129710003733635 + ], + "opposite_logprob_tokens": [ + -2.237971067428589 + ], + "logprob_reward": -0.1129710003733635, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.1129710003733635, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06521382927894592, + 0.09091532975435257, + 0.060833655297756195, + 0.10729644447565079, + 0.3611584007740021, + 0.14117655158042908, + 0.020825417712330818, + 0.03098917193710804, + 0.016332780942320824, + 0.10525843501091003 + ], + "credit_map": [ + 0.006397614721208811, + 0.0068390448577702045, + 0.007602144032716751, + 0.008149184286594391, + 0.0087622394785285, + 0.007271760609000921, + 0.012223856523633003, + 0.007967979647219181, + 0.00874715019017458, + 0.006632851902395487, + 0.011568988673388958, + 0.010116340592503548, + 0.011593937873840332, + 0.011569757014513016, + 0.020969389006495476, + 0.009716921485960484, + 0.0061621107161045074, + 0.005387801211327314, + 0.006683086510747671, + 0.008986363187432289, + 0.00731129664927721, + 0.005946803372353315, + 0.014411627314984798, + 0.005944566335529089, + 0.01558099128305912, + 0.009113145060837269, + 0.011946020647883415, + 0.0110712219029665, + 0.013484390452504158, + 0.011530669406056404, + 0.018407970666885376, + 0.016162032261490822, + 0.04360852763056755, + 0.04391328617930412, + 0.049812886863946915, + 0.06279236823320389, + 0.04743717610836029, + 0.03761918842792511, + 0.050520870834589005, + 0.025454076007008553, + 0.01580924354493618, + 0.01495672669261694, + 0.019351880997419357, + 0.023910071700811386, + 0.015447722747921944, + 0.017147142440080643, + 0.017979150637984276, + 0.016574621200561523, + 0.0041743079200387, + 0.0014284657081589103, + 0.0030995970591902733, + 0.001973601756617427, + 0.0023364489898085594, + 0.002153042471036315, + 0.0020067719742655754, + 0.003653181716799736, + 0.005680580623447895, + 0.0023705768398940563, + 0.005239722318947315, + 0.0027144853956997395, + 0.003342872951179743, + 0.003378043184056878, + 0.004749913699924946, + 0.003512977622449398, + 0.0017810299759730697, + 0.0012151036644354463, + 0.0020477420184761286, + 0.0014802758814767003, + 0.002041898900642991, + 0.0017929347231984138, + 0.0046289581805467606, + 0.0013448380632326007, + 0.00902494601905346, + 0.009074147790670395, + 0.012524348683655262, + 0.007897148840129375, + 0.013375602662563324, + 0.009472411125898361, + 0.03426546975970268, + 0.009624358266592026 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 5, + 3 + ], + "selected_reward": -0.1129710003733635 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 the mouth and nose of a young woman?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the depth of field incredibly shallow with a softly blurred background in a hazy warm hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a perfect, shimmering drop of thick golden-amber honey visible on the woman's upper lip?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the woman's tongue emerge slowly and deliberately to lick the honey off her lip?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting warm and soft, resembling late afternoon light streaming through a window?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"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": [ + 0.11545268446207047, + 0.05937984213232994, + 0.10064569860696793, + 0.07099492847919464, + 0.17246833443641663, + 0.10758446156978607, + 0.2395658791065216, + 0.037584323436021805, + 0.03741326928138733, + 0.058910541236400604 + ], + "credit_map": [ + 0.015447973273694515, + 0.009283346123993397, + 0.022748909890651703, + 0.012367567978799343, + 0.014271982945501804, + 0.01519566960632801, + 0.012322823517024517, + 0.013814413920044899, + 0.00581695418804884, + 0.005451275501400232, + 0.007475983817130327, + 0.007631243672221899, + 0.005992527585476637, + 0.005914522334933281, + 0.015455568209290504, + 0.005641765892505646, + 0.01622490957379341, + 0.00714416429400444, + 0.017254767939448357, + 0.006949126720428467, + 0.014025373384356499, + 0.011067137122154236, + 0.013040678575634956, + 0.014939540065824986, + 0.009386973455548286, + 0.007584049366414547, + 0.01087612472474575, + 0.007934301160275936, + 0.009368372149765491, + 0.008597107604146004, + 0.010187188163399696, + 0.007060820236802101, + 0.031545259058475494, + 0.021033739671111107, + 0.021532870829105377, + 0.0177695881575346, + 0.026126718148589134, + 0.016753671690821648, + 0.018737902864813805, + 0.01896858401596546, + 0.019088465720415115, + 0.013993998058140278, + 0.01532155554741621, + 0.011673915199935436, + 0.013615362346172333, + 0.011617911048233509, + 0.009384523145854473, + 0.012888740748167038, + 0.055947195738554, + 0.01653110794723034, + 0.04856313392519951, + 0.018726548179984093, + 0.03201931342482567, + 0.02436823956668377, + 0.019193315878510475, + 0.024217037484049797, + 0.006216999609023333, + 0.0028992551378905773, + 0.003908514976501465, + 0.004212700296193361, + 0.004552236758172512, + 0.003557702526450157, + 0.008381388150155544, + 0.00385552947409451, + 0.004964573308825493, + 0.0029375681187957525, + 0.004561128094792366, + 0.007849597372114658, + 0.004560690373182297, + 0.0027566554490476847, + 0.00700742844492197, + 0.002775628585368395, + 0.012516875751316547, + 0.004938377533107996, + 0.01029240433126688, + 0.005767899565398693, + 0.005591314285993576, + 0.005387206096202135, + 0.008783003315329552, + 0.005633458960801363 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 4, + 0 + ], + "selected_reward": -0.004650253802537918 + } + ], + "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/0000156/group_3/index_tensor([3603, 3603])_group_3_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000156/group_3/index_tensor([3603, 3603])_group_3_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..e35b177aca38090bc133d88a9f79300e44e1593b --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000156/group_3/index_tensor([3603, 3603])_group_3_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "narrow, documentary-style shot, suspended almost motionlessly in the water, captures a massive nurse shark, resting peacefully on the fine, white sandy seabed, taking center stage. The scene is bathed in a deep, almost monochromatic sapphire blue, creating a tranquil and mysterious atmosphere of the deep sea. The shark's gray-brown skin appears thick against the blue water. Its broad, flat head rests close to the sand, its small, deep-set eyes seemingly closed in repose, and its signature barbels near its mouth sway gently with the barely perceptible current. Sunlight streaming from above penetrates the deep water, constantly refracted and distorted by the ripples on the surface, creating a constantly shifting, dancing web of light. These shifting specks of light and shadow, like a silent narrative, slowly and rhythmically glide across the shark's broad, smooth back and body curves, vividly highlighting the texture of its skin and the contours of its muscles. The image, in its profound stillness, reveals a primal and majestic vitality.", + "seed": 31474, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000156/group_3/index_3603_seed_31474_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": -0.02187858633697033, + "vlm_logprob": -0.02187858633697033, + "vlm_yes_no_margin": 4.025, + "vlm_token_credit": -0.02187858633697033, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"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 massive nurse shark visible resting on a fine, white sandy seabed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene bathed in a deep, monochromatic sapphire blue color?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does sunlight stream from above, creating shifting specks of light and shadow on the shark's back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the shot framed as a narrow, documentary-style view with the shark taking center stage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the shark's signature barbels near its mouth sway gently with the current?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011642840690910816 + ], + "opposite_logprob_tokens": [ + -4.511642932891846 + ], + "logprob_reward": -0.011642840690910816, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011642840690910816, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09024670720100403, + 0.06076416000723839, + 0.1535259336233139, + 0.05362667143344879, + 0.24208152294158936, + 0.050276707857847214, + 0.0972844660282135, + 0.1634567677974701, + 0.04044842720031738, + 0.04828866571187973 + ], + "credit_map": [ + 0.008269891142845154, + 0.011585247702896595, + 0.01012816559523344, + 0.017110437154769897, + 0.01137213222682476, + 0.011284771375358105, + 0.012181715108454227, + 0.008314341306686401, + 0.010520034469664097, + 0.009171041660010815, + 0.007688374258577824, + 0.008030165918171406, + 0.006884430069476366, + 0.006549557205289602, + 0.0069994693621993065, + 0.004921081010252237, + 0.018871597945690155, + 0.01382970530539751, + 0.04345102980732918, + 0.013143843039870262, + 0.026178916916251183, + 0.015255465172231197, + 0.012185978703200817, + 0.010609382763504982, + 0.009221192449331284, + 0.010223409160971642, + 0.006159503478556871, + 0.0059604523703455925, + 0.0068003516644239426, + 0.006678571924567223, + 0.004816452506929636, + 0.0037667290307581425, + 0.06356213241815567, + 0.023814577609300613, + 0.07474566251039505, + 0.022284574806690216, + 0.016037968918681145, + 0.01563805527985096, + 0.013501889072358608, + 0.012496630661189556, + 0.0066449521109461784, + 0.007150707766413689, + 0.006708427332341671, + 0.005864254664629698, + 0.006445770617574453, + 0.006361145526170731, + 0.006777833681553602, + 0.004323610104620457, + 0.018215732648968697, + 0.020411960780620575, + 0.013351332396268845, + 0.01216733269393444, + 0.012790353037416935, + 0.006727301981300116, + 0.007031586021184921, + 0.0065888590179383755, + 0.034083619713783264, + 0.04404262825846672, + 0.012867686338722706, + 0.02531101554632187, + 0.020398259162902832, + 0.007953760214149952, + 0.012021007016301155, + 0.0067787799052894115, + 0.004581891931593418, + 0.005101290065795183, + 0.007849656045436859, + 0.004748477600514889, + 0.0035036238841712475, + 0.005189486779272556, + 0.005188492592424154, + 0.0042855036444962025, + 0.0070418259128928185, + 0.010932432487607002, + 0.005618937313556671, + 0.004544377326965332, + 0.006311400327831507, + 0.00423795310780406, + 0.004465685226023197, + 0.0051360479556024075 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 7, + 2 + ], + "selected_reward": -0.011642840690910816 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"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 massive nurse shark visible resting on a fine, white sandy seabed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene bathed in a deep, monochromatic sapphire blue color?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does sunlight stream from above, creating shifting specks of light and shadow on the shark's back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the shot framed as a narrow, documentary-style view with the shark taking center stage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the shark's signature barbels near its mouth sway gently with the current?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"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": [ + 0.05585678666830063, + 0.10167568922042847, + 0.12391979992389679, + 0.06664546579122543, + 0.12490268051624298, + 0.08629374951124191, + 0.08456703275442123, + 0.231101855635643, + 0.04226905107498169, + 0.08276789635419846 + ], + "credit_map": [ + 0.0060551343485713005, + 0.007961216382682323, + 0.0064104534685611725, + 0.006904487498104572, + 0.0057031563483178616, + 0.009699039161205292, + 0.0057295397855341434, + 0.007393757347017527, + 0.011069698259234428, + 0.010771562345325947, + 0.009267437271773815, + 0.013815678656101227, + 0.012230897322297096, + 0.014378241263329983, + 0.013791936449706554, + 0.01635023020207882, + 0.01668693870306015, + 0.011143452487885952, + 0.041767146438360214, + 0.008721555583178997, + 0.021854441612958908, + 0.01012871041893959, + 0.006615744438022375, + 0.007001812104135752, + 0.009746200405061245, + 0.009538835845887661, + 0.015293973498046398, + 0.0069016688503324986, + 0.006560101173818111, + 0.007687215227633715, + 0.004943661857396364, + 0.005973802413791418, + 0.017772071063518524, + 0.015392627567052841, + 0.01968958042562008, + 0.01202608086168766, + 0.01242782361805439, + 0.017940888181328773, + 0.01954370178282261, + 0.01010991632938385, + 0.011074007488787174, + 0.010664429515600204, + 0.01857019029557705, + 0.008526955731213093, + 0.012279411777853966, + 0.008363042958080769, + 0.010953892022371292, + 0.005861820187419653, + 0.011742175556719303, + 0.01393747329711914, + 0.010474254377186298, + 0.007924675941467285, + 0.013312624767422676, + 0.009008960798382759, + 0.00736608961597085, + 0.010800778865814209, + 0.044381316751241684, + 0.06067828834056854, + 0.015501045621931553, + 0.03205866366624832, + 0.03091771900653839, + 0.014279395341873169, + 0.018565410748124123, + 0.014720014296472073, + 0.007225763984024525, + 0.009239287115633488, + 0.004704556427896023, + 0.004867538809776306, + 0.004894850309938192, + 0.004417369142174721, + 0.003955415450036526, + 0.002964274724945426, + 0.012821478769183159, + 0.021931929513812065, + 0.010707373730838299, + 0.005183226428925991, + 0.011247984133660793, + 0.006583443842828274, + 0.008454607799649239, + 0.005837851669639349 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 4, + 2 + ], + "selected_reward": -0.007637348957359791 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"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 massive nurse shark visible resting on a fine, white sandy seabed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene bathed in a deep, monochromatic sapphire blue color?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does sunlight stream from above, creating shifting specks of light and shadow on the shark's back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the shot framed as a narrow, documentary-style view with the shark taking center stage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the shark's signature barbels near its mouth sway gently with the current?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01818847469985485 + ], + "opposite_logprob_tokens": [ + -4.0181884765625 + ], + "logprob_reward": -0.01818847469985485, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.01818847469985485, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07579085230827332, + 0.06362828612327576, + 0.058765582740306854, + 0.052075549960136414, + 0.3115149140357971, + 0.05952732637524605, + 0.1116621270775795, + 0.14583058655261993, + 0.021886629983782768, + 0.09931813180446625 + ], + "credit_map": [ + 0.010417292825877666, + 0.009939237497746944, + 0.008045072667300701, + 0.014231656678020954, + 0.0083948178216815, + 0.011113879270851612, + 0.007559269666671753, + 0.006089629605412483, + 0.009667053818702698, + 0.0072218007408082485, + 0.00930702593177557, + 0.0064182947389781475, + 0.0072522941045463085, + 0.007716748397797346, + 0.005594939924776554, + 0.010450123809278011, + 0.009385756216943264, + 0.00965031236410141, + 0.01094277948141098, + 0.007228802423924208, + 0.007185768336057663, + 0.006227653939276934, + 0.004488492384552956, + 0.0036560194566845894, + 0.008698059245944023, + 0.011297560296952724, + 0.007285444065928459, + 0.006104691885411739, + 0.0061191716231405735, + 0.006774409674108028, + 0.0031390541698783636, + 0.002657160395756364, + 0.08296366035938263, + 0.028409872204065323, + 0.10555677115917206, + 0.022112509235739708, + 0.025079088285565376, + 0.017887283116579056, + 0.011294304393231869, + 0.01821143925189972, + 0.008424974977970123, + 0.010798121802508831, + 0.012453554198145866, + 0.005495206452906132, + 0.0061570582911372185, + 0.005550698842853308, + 0.00664517330005765, + 0.004002538044005632, + 0.020043140277266502, + 0.02098073810338974, + 0.012196474708616734, + 0.013258728198707104, + 0.015454831533133984, + 0.01089830044656992, + 0.010653479024767876, + 0.00817643478512764, + 0.02889532782137394, + 0.040124814957380295, + 0.009219668805599213, + 0.020330028608441353, + 0.01898408867418766, + 0.008421136997640133, + 0.010408115573227406, + 0.00944740790873766, + 0.0036247919779270887, + 0.0034439449664205313, + 0.004244499374181032, + 0.0018594048451632261, + 0.002216880675405264, + 0.0024426935706287622, + 0.002228023251518607, + 0.0018263922538608313, + 0.012351089157164097, + 0.015097926370799541, + 0.01778634637594223, + 0.00674927793443203, + 0.012110559269785881, + 0.010904129594564438, + 0.010851130820810795, + 0.013467665761709213 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 7, + 6 + ], + "selected_reward": -0.01818847469985485 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"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 massive nurse shark visible resting on a fine, white sandy seabed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene bathed in a deep, monochromatic sapphire blue color?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does sunlight stream from above, creating shifting specks of light and shadow on the shark's back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the shot framed as a narrow, documentary-style view with the shark taking center stage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the shark's signature barbels near its mouth sway gently with the current?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023280387744307518 + ], + "opposite_logprob_tokens": [ + -3.773280382156372 + ], + "logprob_reward": -0.023280387744307518, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023280387744307518, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05841870605945587, + 0.08465828001499176, + 0.05972655862569809, + 0.0453542061150074, + 0.3552429974079132, + 0.057131338864564896, + 0.11238361895084381, + 0.13744260370731354, + 0.024036364629864693, + 0.06560534983873367 + ], + "credit_map": [ + 0.0059639946557581425, + 0.007341485004872084, + 0.00651956582441926, + 0.009060157462954521, + 0.00891029927879572, + 0.007562180515378714, + 0.006013568490743637, + 0.00704745389521122, + 0.008068171329796314, + 0.012059573084115982, + 0.008401105180382729, + 0.006306854542344809, + 0.011313031427562237, + 0.01919754408299923, + 0.01033774483948946, + 0.008974256925284863, + 0.007021043449640274, + 0.0069324360229074955, + 0.012093042954802513, + 0.006425152067095041, + 0.007780600804835558, + 0.007927902974188328, + 0.005769089329987764, + 0.005777294281870127, + 0.005447221454232931, + 0.004724493715912104, + 0.008426194079220295, + 0.004943070467561483, + 0.005586487706750631, + 0.007335105910897255, + 0.004319574683904648, + 0.004572056233882904, + 0.06931805610656738, + 0.03324287757277489, + 0.0987340658903122, + 0.029604846611618996, + 0.03267957270145416, + 0.030934829264879227, + 0.03889303654432297, + 0.021835682913661003, + 0.008696481585502625, + 0.00789518840610981, + 0.010874961502850056, + 0.007171718869358301, + 0.006609579548239708, + 0.005336796399205923, + 0.006985761690884829, + 0.003560855519026518, + 0.016618529334664345, + 0.014937758445739746, + 0.019501779228448868, + 0.011941804550588131, + 0.017389971762895584, + 0.009047283791005611, + 0.01146670151501894, + 0.011479792185127735, + 0.027705546468496323, + 0.038898762315511703, + 0.008793113753199577, + 0.02075301669538021, + 0.01876564882695675, + 0.007446722127497196, + 0.009552929550409317, + 0.005526863969862461, + 0.0032787220552563667, + 0.00383179634809494, + 0.0027128569781780243, + 0.002617489779368043, + 0.0030136210843920708, + 0.003660680493339896, + 0.002213670639321208, + 0.002707528416067362, + 0.010140348225831985, + 0.020636016502976418, + 0.005878581199795008, + 0.004005847033113241, + 0.010172107256948948, + 0.005604363977909088, + 0.004250383470207453, + 0.004917699843645096 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 7, + 6 + ], + "selected_reward": -0.023280387744307518 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"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 massive nurse shark visible resting on a fine, white sandy seabed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene bathed in a deep, monochromatic sapphire blue color?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does sunlight stream from above, creating shifting specks of light and shadow on the shark's back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the shot framed as a narrow, documentary-style view with the shark taking center stage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the shark's signature barbels near its mouth sway gently with the current?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04864387959241867 + ], + "opposite_logprob_tokens": [ + -3.0486438274383545 + ], + "logprob_reward": -0.04864387959241867, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04864387959241867, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08015412092208862, + 0.07151389122009277, + 0.10159635543823242, + 0.09957020729780197, + 0.06245649978518486, + 0.053501177579164505, + 0.06506764143705368, + 0.21684660017490387, + 0.05425311625003815, + 0.19504036009311676 + ], + "credit_map": [ + 0.010537082329392433, + 0.011407963931560516, + 0.008125566877424717, + 0.014913151040673256, + 0.008411898277699947, + 0.0110514797270298, + 0.008340408094227314, + 0.007366564590483904, + 0.010129103437066078, + 0.009927541017532349, + 0.007498793303966522, + 0.008903713896870613, + 0.008925862610340118, + 0.011020394042134285, + 0.008346580900251865, + 0.0067619020119309425, + 0.013247848488390446, + 0.009664621204137802, + 0.03205493092536926, + 0.009674260392785072, + 0.01744488812983036, + 0.008439446799457073, + 0.005978787783533335, + 0.005091574043035507, + 0.018334873020648956, + 0.020222917199134827, + 0.012677285820245743, + 0.01096529420465231, + 0.012185218743979931, + 0.010210234671831131, + 0.008106883615255356, + 0.006867502816021442, + 0.011091434396803379, + 0.007086407858878374, + 0.012835515663027763, + 0.006763579323887825, + 0.006520973052829504, + 0.006386495660990477, + 0.006842968054115772, + 0.0049291253089904785, + 0.006113186478614807, + 0.007068898528814316, + 0.0074132513254880905, + 0.007411693688482046, + 0.007314377464354038, + 0.006141205318272114, + 0.006836330983787775, + 0.005202233791351318, + 0.01176415104418993, + 0.01167305838316679, + 0.007633989676833153, + 0.00904360506683588, + 0.00721272686496377, + 0.0071431235410273075, + 0.006069901864975691, + 0.004527086392045021, + 0.042988065630197525, + 0.05899950861930847, + 0.014863841235637665, + 0.03247684985399246, + 0.027067134156823158, + 0.013573753647506237, + 0.015514552593231201, + 0.011362898163497448, + 0.009850456379354, + 0.010472557507455349, + 0.006904213689267635, + 0.005674323067069054, + 0.006238366477191448, + 0.005291836801916361, + 0.005595364607870579, + 0.004226000979542732, + 0.028087057173252106, + 0.053587280213832855, + 0.027003340423107147, + 0.012198373675346375, + 0.02986481599509716, + 0.016148220747709274, + 0.0142815625295043, + 0.013869710266590118 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 9, + 2 + ], + "selected_reward": -0.04864387959241867 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "narrow, documentary-style shot, suspended almost motionlessly in the water, captures a massive nurse shark, resting peacefully on the fine, white sandy seabed, taking center stage. The scene is bathed in a deep, almost monochromatic sapphire blue, creating a tranquil and mysterious atmosphere of the deep sea. The shark's gray-brown skin appears thick against the blue water. Its broad, flat head rests close to the sand, its small, deep-set eyes seemingly closed in repose, and its signature barbels near its mouth sway gently with the barely perceptible current. Sunlight streaming from above penetrates the deep water, constantly refracted and distorted by the ripples on the surface, creating a constantly shifting, dancing web of light. These shifting specks of light and shadow, like a silent narrative, slowly and rhythmically glide across the shark's broad, smooth back and body curves, vividly highlighting the texture of its skin and the contours of its muscles. The image, in its profound stillness, reveals a primal and majestic vitality.", + "seed": 31475, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000156/group_3/index_3603_seed_31475_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": -0.024312366172671317, + "vlm_logprob": -0.024312366172671317, + "vlm_yes_no_margin": 3.875, + "vlm_token_credit": -0.024312366172671317, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"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 massive nurse shark visible resting on a fine, white sandy seabed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene bathed in a deep, monochromatic sapphire blue color?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does sunlight stream from above, creating shifting specks of light and shadow on the shark's back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the shot framed as a narrow, documentary-style view with the shark taking center stage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the shark's signature barbels near its mouth sway gently with the current?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.021928466856479645 + ], + "opposite_logprob_tokens": [ + -3.896928548812866 + ], + "logprob_reward": -0.021928466856479645, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.021928466856479645, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0839013159275055, + 0.20737776160240173, + 0.14371709525585175, + 0.10706677287817001, + 0.08175113052129745, + 0.046897999942302704, + 0.03820516914129257, + 0.12098760157823563, + 0.107110895216465, + 0.06298433989286423 + ], + "credit_map": [ + 0.014457913115620613, + 0.0126160504296422, + 0.020877458155155182, + 0.007053259294480085, + 0.009562183171510696, + 0.00983048789203167, + 0.004841689486056566, + 0.004662265069782734, + 0.03471265733242035, + 0.018816927447915077, + 0.05808861553668976, + 0.016909122467041016, + 0.022535765543580055, + 0.03048916533589363, + 0.01251806877553463, + 0.013307427987456322, + 0.018874870613217354, + 0.016711391508579254, + 0.02259998954832554, + 0.016502736136317253, + 0.0208307895809412, + 0.02184668369591236, + 0.01264724787324667, + 0.013703370466828346, + 0.014056297950446606, + 0.01184491440653801, + 0.02001618780195713, + 0.0102248415350914, + 0.01558246836066246, + 0.0176406130194664, + 0.009024702943861485, + 0.008676734752953053, + 0.010721932165324688, + 0.00782732106745243, + 0.01278020441532135, + 0.011735650710761547, + 0.006375511642545462, + 0.00887052807956934, + 0.010472843423485756, + 0.012967129237949848, + 0.008383418433368206, + 0.006312054581940174, + 0.011309536173939705, + 0.003926449920982122, + 0.004861270543187857, + 0.004605494439601898, + 0.0036501926369965076, + 0.0038495801854878664, + 0.005313892383128405, + 0.0033594644628465176, + 0.0053452965803444386, + 0.0030044401064515114, + 0.007573682349175215, + 0.007195348851382732, + 0.003618599846959114, + 0.002794439671561122, + 0.022515419870615005, + 0.01033400185406208, + 0.03508822247385979, + 0.015491088852286339, + 0.010441012680530548, + 0.014032045379281044, + 0.00663565332069993, + 0.0064501469023525715, + 0.012662604451179504, + 0.028249146416783333, + 0.013795946724712849, + 0.009623873047530651, + 0.01680176332592964, + 0.011846639215946198, + 0.007552347611635923, + 0.00657855998724699, + 0.008285501040518284, + 0.008039138279855251, + 0.011675700545310974, + 0.008345285430550575, + 0.005500368773937225, + 0.0076559907756745815, + 0.004559264052659273, + 0.008923089131712914 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 2, + 7 + ], + "selected_reward": -0.021928466856479645 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"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 massive nurse shark visible resting on a fine, white sandy seabed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene bathed in a deep, monochromatic sapphire blue color?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does sunlight stream from above, creating shifting specks of light and shadow on the shark's back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the shot framed as a narrow, documentary-style view with the shark taking center stage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the shark's signature barbels near its mouth sway gently with the current?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016079511493444443 + ], + "opposite_logprob_tokens": [ + -4.141079425811768 + ], + "logprob_reward": -0.016079511493444443, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016079511493444443, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08788200467824936, + 0.3489326536655426, + 0.07720820605754852, + 0.03117995895445347, + 0.057016756385564804, + 0.03108980320394039, + 0.03817284479737282, + 0.17799976468086243, + 0.08219143748283386, + 0.06832659989595413 + ], + "credit_map": [ + 0.014770456589758396, + 0.019599543884396553, + 0.008442927151918411, + 0.0061371298506855965, + 0.014226315543055534, + 0.012500575743615627, + 0.005858718417584896, + 0.006346334237605333, + 0.08710893243551254, + 0.022645099088549614, + 0.14895695447921753, + 0.0300188846886158, + 0.015179676003754139, + 0.018309136852622032, + 0.012467661872506142, + 0.014246293343603611, + 0.010732896625995636, + 0.009949594736099243, + 0.013483550399541855, + 0.0056137703359127045, + 0.012230508960783482, + 0.014200839214026928, + 0.004997053649276495, + 0.005999986547976732, + 0.004432430025190115, + 0.0038854898884892464, + 0.004537740256637335, + 0.0030093190725892782, + 0.004231473430991173, + 0.005789718125015497, + 0.0031352825462818146, + 0.0021585046779364347, + 0.007307847961783409, + 0.007045818958431482, + 0.008914921432733536, + 0.009547188878059387, + 0.005961548537015915, + 0.007719603832811117, + 0.005387426353991032, + 0.005132402293384075, + 0.004856759216636419, + 0.004411451518535614, + 0.0042604729533195496, + 0.002652146155014634, + 0.00534031493589282, + 0.004400491714477539, + 0.002547765616327524, + 0.0026203999295830727, + 0.005084648262709379, + 0.003346649929881096, + 0.00383393163792789, + 0.0026264560874551535, + 0.009138727560639381, + 0.006643980275839567, + 0.00402413122355938, + 0.003474318888038397, + 0.037749942392110825, + 0.014616245403885841, + 0.06281211227178574, + 0.019453616812825203, + 0.012610215693712234, + 0.013410424813628197, + 0.009203501977026463, + 0.008143700659275055, + 0.012764985673129559, + 0.025065993890166283, + 0.010347569361329079, + 0.007463141810148954, + 0.012256207875907421, + 0.007217549718916416, + 0.00350078078918159, + 0.0035752099938690662, + 0.008134549483656883, + 0.007748862262815237, + 0.016505982726812363, + 0.012413703836500645, + 0.006218929775059223, + 0.007923658937215805, + 0.0036572145763784647, + 0.005723695270717144 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 7, + 0 + ], + "selected_reward": -0.016079511493444443 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"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 massive nurse shark visible resting on a fine, white sandy seabed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene bathed in a deep, monochromatic sapphire blue color?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does sunlight stream from above, creating shifting specks of light and shadow on the shark's back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the shot framed as a narrow, documentary-style view with the shark taking center stage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the shark's signature barbels near its mouth sway gently with the current?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"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": [ + 0.06781958788633347, + 0.08063657581806183, + 0.07177378982305527, + 0.03622395545244217, + 0.06375780701637268, + 0.02452128753066063, + 0.02580983191728592, + 0.24608492851257324, + 0.2732345163822174, + 0.11013765633106232 + ], + "credit_map": [ + 0.010468809865415096, + 0.012352066114544868, + 0.00919736921787262, + 0.00750814750790596, + 0.009399965405464172, + 0.009114807471632957, + 0.0046334355138242245, + 0.005144994705915451, + 0.013220984488725662, + 0.005907616578042507, + 0.020066071301698685, + 0.00710753770545125, + 0.0093799764290452, + 0.012082170695066452, + 0.006910346448421478, + 0.005961875896900892, + 0.010937084443867207, + 0.007776474114507437, + 0.022093206644058228, + 0.006451931782066822, + 0.005975617095828056, + 0.00887943897396326, + 0.003908244427293539, + 0.005751797929406166, + 0.005131269805133343, + 0.00380178214982152, + 0.008128706365823746, + 0.003864844096824527, + 0.003668770892545581, + 0.004571098834276199, + 0.004018643405288458, + 0.0030388452578336, + 0.009319636970758438, + 0.010916740633547306, + 0.006357691716402769, + 0.006048244424164295, + 0.008326536975800991, + 0.010083639062941074, + 0.007569544482976198, + 0.005135782063007355, + 0.003577197901904583, + 0.002347908215597272, + 0.004007883369922638, + 0.0032942851539701223, + 0.0032621424179524183, + 0.0034869336523115635, + 0.002018342725932598, + 0.0025265971198678017, + 0.0039022238925099373, + 0.0023706424981355667, + 0.0045428466983139515, + 0.0025039208121597767, + 0.0046006920747458935, + 0.003240691963583231, + 0.00230845483019948, + 0.0023403617087751627, + 0.049552202224731445, + 0.02019137144088745, + 0.07881804555654526, + 0.026823949068784714, + 0.020427975803613663, + 0.024587690830230713, + 0.01386068481951952, + 0.011823032982647419, + 0.037184182554483414, + 0.09615601599216461, + 0.02553589642047882, + 0.01921618916094303, + 0.05318081006407738, + 0.021845165640115738, + 0.009540718048810959, + 0.010575567372143269, + 0.015584875829517841, + 0.013990425504744053, + 0.026401914656162262, + 0.015983905643224716, + 0.010766725055873394, + 0.009468418546020985, + 0.008088563568890095, + 0.009852835908532143 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 9 + ], + "selected_reward": -0.01607493683695793 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"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 massive nurse shark visible resting on a fine, white sandy seabed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene bathed in a deep, monochromatic sapphire blue color?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does sunlight stream from above, creating shifting specks of light and shadow on the shark's back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the shot framed as a narrow, documentary-style view with the shark taking center stage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the shark's signature barbels near its mouth sway gently with the current?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01253500021994114 + ], + "opposite_logprob_tokens": [ + -4.387535095214844 + ], + "logprob_reward": -0.01253500021994114, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.01253500021994114, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11341878026723862, + 0.38097018003463745, + 0.0958818569779396, + 0.04715336486697197, + 0.0481274239718914, + 0.022519826889038086, + 0.023213675245642662, + 0.09316840767860413, + 0.11888450384140015, + 0.056661978363990784 + ], + "credit_map": [ + 0.01935557648539543, + 0.021647121757268906, + 0.017037633806467056, + 0.01101305615156889, + 0.01724747195839882, + 0.015178164467215538, + 0.005634088534861803, + 0.0063056680373847485, + 0.09755082428455353, + 0.023468216881155968, + 0.16878283023834229, + 0.034121330827474594, + 0.015585602261126041, + 0.013059820979833603, + 0.01327971089631319, + 0.015121850185096264, + 0.013679476454854012, + 0.014360175468027592, + 0.011697215028107166, + 0.0058019463904201984, + 0.016457950696349144, + 0.01822253316640854, + 0.007532130461186171, + 0.008130423724651337, + 0.0058867731131613255, + 0.00557002704590559, + 0.007211001589894295, + 0.004294226411730051, + 0.007558303885161877, + 0.008097886107861996, + 0.00407192250713706, + 0.004463220480829477, + 0.006465301848948002, + 0.005374549422413111, + 0.005607026629149914, + 0.005763854831457138, + 0.005813357420265675, + 0.007684623822569847, + 0.005364513956010342, + 0.00605419697239995, + 0.0031724784057587385, + 0.0022044985089451075, + 0.004496355541050434, + 0.0018571428954601288, + 0.003579739946871996, + 0.003166191978380084, + 0.002230271464213729, + 0.0018131483811885118, + 0.0034744504373520613, + 0.0018687567207962275, + 0.003306817030534148, + 0.00152806774713099, + 0.006626487243920565, + 0.002769791055470705, + 0.001991220284253359, + 0.0016480847261846066, + 0.016163336113095284, + 0.010094424709677696, + 0.02777552604675293, + 0.011093834415078163, + 0.008305234834551811, + 0.010172312147915363, + 0.004382836166769266, + 0.005180906038731337, + 0.01832975633442402, + 0.04460293427109718, + 0.009132547304034233, + 0.00839877501130104, + 0.02255544252693653, + 0.007512100040912628, + 0.0043079848401248455, + 0.004044965375214815, + 0.007128686178475618, + 0.0069555253721773624, + 0.008121510967612267, + 0.005940812174230814, + 0.008304224349558353, + 0.009077983908355236, + 0.005675133783370256, + 0.005458102561533451 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 8, + 0 + ], + "selected_reward": -0.01253500021994114 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"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 massive nurse shark visible resting on a fine, white sandy seabed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene bathed in a deep, monochromatic sapphire blue color?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does sunlight stream from above, creating shifting specks of light and shadow on the shark's back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the shot framed as a narrow, documentary-style view with the shark taking center stage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the shark's signature barbels near its mouth sway gently with the current?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05494391545653343 + ], + "opposite_logprob_tokens": [ + -2.929943799972534 + ], + "logprob_reward": -0.05494391545653343, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05494391545653343, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10067646950483322, + 0.4475034475326538, + 0.03830273821949959, + 0.030786463990807533, + 0.02983274683356285, + 0.03262161836028099, + 0.01634315401315689, + 0.16693231463432312, + 0.04198778420686722, + 0.09501327574253082 + ], + "credit_map": [ + 0.016521146520972252, + 0.016433414071798325, + 0.018476713448762894, + 0.008258513174951077, + 0.01672118529677391, + 0.01223007496446371, + 0.006358890328556299, + 0.0056765410117805, + 0.10504554212093353, + 0.02718772552907467, + 0.17801660299301147, + 0.03584528714418411, + 0.026055332273244858, + 0.03712235391139984, + 0.01719583198428154, + 0.021034780889749527, + 0.004728554282337427, + 0.005318053998053074, + 0.006569799035787582, + 0.004390583373606205, + 0.005282856523990631, + 0.0056052240543067455, + 0.002559206448495388, + 0.003848462598398328, + 0.004096159245818853, + 0.004690247122198343, + 0.005013301502913237, + 0.0034701060503721237, + 0.0039586289785802364, + 0.003415252547711134, + 0.003852826077491045, + 0.0022899440955370665, + 0.004847186617553234, + 0.0036219432950019836, + 0.005911706481128931, + 0.0032766566146165133, + 0.0036618132144212723, + 0.0030803768895566463, + 0.003466886468231678, + 0.001966178650036454, + 0.0055637541227042675, + 0.0044438280165195465, + 0.005190430674701929, + 0.0027167026419192553, + 0.004757441580295563, + 0.004160040058195591, + 0.0026812858413904905, + 0.0031081384513527155, + 0.001977513777092099, + 0.0015589712420478463, + 0.0016760661965236068, + 0.001150821684859693, + 0.003494239877909422, + 0.003242437494918704, + 0.0019884665962308645, + 0.001254636561498046, + 0.035644423216581345, + 0.012208512052893639, + 0.05943155661225319, + 0.020542223006486893, + 0.01062118262052536, + 0.011532195843756199, + 0.007931116037070751, + 0.00902111642062664, + 0.00539443688467145, + 0.005474962294101715, + 0.004495143424719572, + 0.004238137509673834, + 0.006327608600258827, + 0.007557939272373915, + 0.004214184358716011, + 0.004285372793674469, + 0.01372693944722414, + 0.01264999806880951, + 0.010048214346170425, + 0.010320213623344898, + 0.01608320325613022, + 0.013657969422638416, + 0.009661913849413395, + 0.008864824660122395 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 7, + 0 + ], + "selected_reward": -0.05494391545653343 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "narrow, documentary-style shot, suspended almost motionlessly in the water, captures a massive nurse shark, resting peacefully on the fine, white sandy seabed, taking center stage. The scene is bathed in a deep, almost monochromatic sapphire blue, creating a tranquil and mysterious atmosphere of the deep sea. The shark's gray-brown skin appears thick against the blue water. Its broad, flat head rests close to the sand, its small, deep-set eyes seemingly closed in repose, and its signature barbels near its mouth sway gently with the barely perceptible current. Sunlight streaming from above penetrates the deep water, constantly refracted and distorted by the ripples on the surface, creating a constantly shifting, dancing web of light. These shifting specks of light and shadow, like a silent narrative, slowly and rhythmically glide across the shark's broad, smooth back and body curves, vividly highlighting the texture of its skin and the contours of its muscles. The image, in its profound stillness, reveals a primal and majestic vitality.", + "seed": 31476, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000156/group_3/index_3603_seed_31476_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": -0.048844918981194495, + "vlm_logprob": -0.048844918981194495, + "vlm_yes_no_margin": 3.375, + "vlm_token_credit": -0.048844918981194495, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"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 massive nurse shark visible resting on a fine, white sandy seabed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene bathed in a deep, monochromatic sapphire blue color?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does sunlight stream from above, creating shifting specks of light and shadow on the shark's back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the shot framed as a narrow, documentary-style view with the shark taking center stage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the shark's signature barbels near its mouth sway gently with the current?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.13020484149456024 + ], + "opposite_logprob_tokens": [ + -2.130204916000366 + ], + "logprob_reward": -0.13020484149456024, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.13020484149456024, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0963643491268158, + 0.07115145027637482, + 0.040431421250104904, + 0.08314415067434311, + 0.032272208482027054, + 0.04700404033064842, + 0.07014747709035873, + 0.33247110247612, + 0.08484514057636261, + 0.14216862618923187 + ], + "credit_map": [ + 0.013355272822082043, + 0.011565647087991238, + 0.017270317301154137, + 0.008054747246205807, + 0.013500570319592953, + 0.012440398335456848, + 0.00653320737183094, + 0.013644194230437279, + 0.010431995615363121, + 0.005777021870017052, + 0.015652354806661606, + 0.008138847537338734, + 0.008336052298545837, + 0.009392756037414074, + 0.004807407967746258, + 0.008615012280642986, + 0.005062067415565252, + 0.005234085489064455, + 0.009302601218223572, + 0.005098484922200441, + 0.004083654377609491, + 0.00534894410520792, + 0.003184533677995205, + 0.003117050975561142, + 0.012066383846104145, + 0.007876772433519363, + 0.01413621287792921, + 0.018074145540595055, + 0.007705993484705687, + 0.006466687191277742, + 0.009816101752221584, + 0.007001848891377449, + 0.0041589280590415, + 0.004007265903055668, + 0.005111359059810638, + 0.004933721851557493, + 0.0051613072864711285, + 0.0036918174009770155, + 0.002768137725070119, + 0.0024396704975515604, + 0.007138918153941631, + 0.003527828026562929, + 0.006306997966021299, + 0.005500671453773975, + 0.005756875034421682, + 0.009866363368928432, + 0.004043424502015114, + 0.004862957168370485, + 0.014550603926181793, + 0.0057999854907393456, + 0.01833837479352951, + 0.0067220102064311504, + 0.007271694950759411, + 0.008367124944925308, + 0.004083922132849693, + 0.005013759713619947, + 0.05477166175842285, + 0.02642226591706276, + 0.08004176616668701, + 0.040364667773246765, + 0.026185400784015656, + 0.034126050770282745, + 0.04632659628987312, + 0.024232711642980576, + 0.014399523846805096, + 0.006039174739271402, + 0.007478022016584873, + 0.009840171784162521, + 0.008835738524794579, + 0.01375509425997734, + 0.016445694491267204, + 0.008051726967096329, + 0.019702116027474403, + 0.013153821229934692, + 0.010720866732299328, + 0.014609779231250286, + 0.02291281335055828, + 0.018502362072467804, + 0.03443080559372902, + 0.008136066608130932 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 9, + 0 + ], + "selected_reward": -0.13020484149456024 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"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 massive nurse shark visible resting on a fine, white sandy seabed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene bathed in a deep, monochromatic sapphire blue color?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does sunlight stream from above, creating shifting specks of light and shadow on the shark's back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the shot framed as a narrow, documentary-style view with the shark taking center stage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the shark's signature barbels near its mouth sway gently with the current?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011091247200965881 + ], + "opposite_logprob_tokens": [ + -4.511091232299805 + ], + "logprob_reward": -0.011091247200965881, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011091247200965881, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12242159247398376, + 0.05757589638233185, + 0.18729323148727417, + 0.06501621752977371, + 0.0549352765083313, + 0.024080846458673477, + 0.07783305644989014, + 0.3243839740753174, + 0.03254567086696625, + 0.05391426384449005 + ], + "credit_map": [ + 0.017772160470485687, + 0.013698468916118145, + 0.031070830300450325, + 0.012277270667254925, + 0.012064831331372261, + 0.01259629800915718, + 0.007859503850340843, + 0.015082227997481823, + 0.0072807129472494125, + 0.004916631616652012, + 0.009337869472801685, + 0.006423828657716513, + 0.011674635112285614, + 0.00868968479335308, + 0.004394677933305502, + 0.004857853055000305, + 0.03272111713886261, + 0.01937650889158249, + 0.03561213240027428, + 0.021680505946278572, + 0.018376220017671585, + 0.021486857905983925, + 0.015616640448570251, + 0.022423233836889267, + 0.009313289076089859, + 0.006466548424214125, + 0.011121895164251328, + 0.015369349159300327, + 0.006645190995186567, + 0.004321544896811247, + 0.006951060611754656, + 0.004827337805181742, + 0.008301804773509502, + 0.00683009484782815, + 0.005251632072031498, + 0.006505756638944149, + 0.009618773125112057, + 0.010465627536177635, + 0.003982420079410076, + 0.003979172091931105, + 0.0036625785287469625, + 0.0017002563690766692, + 0.004081801511347294, + 0.00336469360627234, + 0.00313585321418941, + 0.003505677217617631, + 0.002041277475655079, + 0.002588708186522126, + 0.026074320077896118, + 0.0035342874471098185, + 0.03141365200281143, + 0.004209450911730528, + 0.0020163448061794043, + 0.003941116854548454, + 0.0028496100567281246, + 0.0037942754570394754, + 0.06179049611091614, + 0.025687698274850845, + 0.048723254352808, + 0.040320370346307755, + 0.03598105534911156, + 0.042075593024492264, + 0.046687811613082886, + 0.023117700591683388, + 0.005253657698631287, + 0.0031022648327052593, + 0.0030471475329250097, + 0.003402500879019499, + 0.003497776575386524, + 0.006365603767335415, + 0.005190315190702677, + 0.0026864015962928534, + 0.00767896044999361, + 0.005020389333367348, + 0.0047123427502810955, + 0.005353001412004232, + 0.009269127622246742, + 0.0073099820874631405, + 0.011039015837013721, + 0.003531447611749172 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 2, + 0 + ], + "selected_reward": -0.011091247200965881 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"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 massive nurse shark visible resting on a fine, white sandy seabed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene bathed in a deep, monochromatic sapphire blue color?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does sunlight stream from above, creating shifting specks of light and shadow on the shark's back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the shot framed as a narrow, documentary-style view with the shark taking center stage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the shark's signature barbels near its mouth sway gently with the current?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018193040043115616 + ], + "opposite_logprob_tokens": [ + -4.018193244934082 + ], + "logprob_reward": -0.018193040043115616, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018193040043115616, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12190960347652435, + 0.13891786336898804, + 0.08003177493810654, + 0.19326379895210266, + 0.08883824944496155, + 0.09031728655099869, + 0.0858425423502922, + 0.12356801331043243, + 0.03591034188866615, + 0.041400566697120667 + ], + "credit_map": [ + 0.02094261720776558, + 0.013661668635904789, + 0.024288956075906754, + 0.010154619812965393, + 0.016430383548140526, + 0.014859390445053577, + 0.007313488516956568, + 0.014258467592298985, + 0.019135836511850357, + 0.012163705192506313, + 0.02881282940506935, + 0.017326626926660538, + 0.020447304472327232, + 0.018170824274420738, + 0.009198706597089767, + 0.013662014156579971, + 0.012018810957670212, + 0.008870060555636883, + 0.017258528620004654, + 0.011140401475131512, + 0.00783848762512207, + 0.010525820776820183, + 0.005304674152284861, + 0.007074977736920118, + 0.029617995023727417, + 0.018715545535087585, + 0.04537176340818405, + 0.03564801067113876, + 0.01516919955611229, + 0.016627412289381027, + 0.015500885434448719, + 0.01661297120153904, + 0.013634203933179379, + 0.01015445590019226, + 0.01151756476610899, + 0.007924629375338554, + 0.011408276855945587, + 0.01911851577460766, + 0.007130355108529329, + 0.00795024260878563, + 0.013056930154561996, + 0.008722664788365364, + 0.013850671239197254, + 0.011626540683209896, + 0.011076007969677448, + 0.011066829785704613, + 0.010603678412735462, + 0.010313952341675758, + 0.02796066179871559, + 0.0043832832016050816, + 0.03376525640487671, + 0.00517357187345624, + 0.003480750136077404, + 0.004599312786012888, + 0.0032382954377681017, + 0.0032413937151432037, + 0.017626319080591202, + 0.019269121810793877, + 0.01656617596745491, + 0.012010753154754639, + 0.013190323486924171, + 0.020052315667271614, + 0.013412414118647575, + 0.01144057884812355, + 0.007211361080408096, + 0.0022513538133352995, + 0.0032287430949509144, + 0.004732328467071056, + 0.0023757361341267824, + 0.0030031639616936445, + 0.009553137235343456, + 0.003554518101736903, + 0.004554494749754667, + 0.005430401302874088, + 0.004941256251186132, + 0.003954629879444838, + 0.009089110419154167, + 0.006321290507912636, + 0.0043807183392345905, + 0.0027286638505756855 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 1, + 7 + ], + "selected_reward": -0.018193040043115616 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a massive nurse shark visible resting on a fine, white sandy seabed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene bathed in a deep, monochromatic sapphire blue color?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does sunlight stream from above, creating shifting specks of light and shadow on the shark's back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the shot framed as a narrow, documentary-style view with the shark taking center stage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the shark's signature barbels near its mouth sway gently with the current?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"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": [ + 0.09775713086128235, + 0.0436689667403698, + 0.048385728150606155, + 0.41215139627456665, + 0.06262563914060593, + 0.02886396087706089, + 0.05672302842140198, + 0.12570668756961823, + 0.0747985690832138, + 0.04931894689798355 + ], + "credit_map": [ + 0.014254610985517502, + 0.013919558376073837, + 0.017548564821481705, + 0.007585563231259584, + 0.017352882772684097, + 0.010751619003713131, + 0.005689026787877083, + 0.01065529603511095, + 0.005514625459909439, + 0.0037103970535099506, + 0.0072944918647408485, + 0.005466571543365717, + 0.0061195604503154755, + 0.006306516006588936, + 0.003482149913907051, + 0.005774652119725943, + 0.00825358647853136, + 0.006096729543060064, + 0.008073822595179081, + 0.005787798203527927, + 0.005093454848974943, + 0.005804137792438269, + 0.0036467730533331633, + 0.005629423074424267, + 0.06058839336037636, + 0.04308458790183067, + 0.10212454199790955, + 0.05334993824362755, + 0.03344869613647461, + 0.05232538655400276, + 0.0342620350420475, + 0.03296774998307228, + 0.00790683925151825, + 0.008826222270727158, + 0.008272163569927216, + 0.008434437215328217, + 0.012863844633102417, + 0.006538654677569866, + 0.005574247799813747, + 0.004209219012409449, + 0.0045011891052126884, + 0.0018095786217600107, + 0.006457719020545483, + 0.003018759423866868, + 0.002833383623510599, + 0.004347996786236763, + 0.0023678920697420835, + 0.003527440596371889, + 0.015070602297782898, + 0.004686716012656689, + 0.018630338832736015, + 0.004623047541826963, + 0.004397368989884853, + 0.0038235385436564684, + 0.0027358136139810085, + 0.002755596535280347, + 0.022380774840712547, + 0.011107129976153374, + 0.0225103497505188, + 0.014741037040948868, + 0.013010104186832905, + 0.01658794842660427, + 0.016033325344324112, + 0.009336014278233051, + 0.02056737244129181, + 0.004018561914563179, + 0.004881284665316343, + 0.00845694076269865, + 0.0036110286600887775, + 0.0034992871806025505, + 0.025523068383336067, + 0.004241019021719694, + 0.006550217512995005, + 0.005655835382640362, + 0.008132738061249256, + 0.0061975507996976376, + 0.007204857654869556, + 0.007957017980515957, + 0.004780569113790989, + 0.0028401538729667664 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 7, + 0 + ], + "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 a massive nurse shark visible resting on a fine, white sandy seabed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene bathed in a deep, monochromatic sapphire blue color?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does sunlight stream from above, creating shifting specks of light and shadow on the shark's back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the shot framed as a narrow, documentary-style view with the shark taking center stage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the shark's signature barbels near its mouth sway gently with the current?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.054941657930612564 + ], + "opposite_logprob_tokens": [ + -2.9299416542053223 + ], + "logprob_reward": -0.054941657930612564, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.054941657930612564, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12012127041816711, + 0.07185451686382294, + 0.05507705360651016, + 0.1816204935312271, + 0.05648284777998924, + 0.033683303743600845, + 0.038265131413936615, + 0.12295257300138474, + 0.2023332715034485, + 0.11760963499546051 + ], + "credit_map": [ + 0.015871472656726837, + 0.011148317717015743, + 0.024331098422408104, + 0.008262711577117443, + 0.011845664121210575, + 0.015116211958229542, + 0.009740578942000866, + 0.023805199190974236, + 0.009403209201991558, + 0.006983144208788872, + 0.009631380438804626, + 0.0072190891951322556, + 0.014134707860648632, + 0.008296576328575611, + 0.004524038173258305, + 0.011662358418107033, + 0.008378114551305771, + 0.007923354394733906, + 0.011993779800832272, + 0.005320191383361816, + 0.005648409482091665, + 0.00634366599842906, + 0.0034092464484274387, + 0.006060282699763775, + 0.028167298063635826, + 0.017118941992521286, + 0.046523455530405045, + 0.0326271653175354, + 0.015046457760035992, + 0.014453611336648464, + 0.013406455516815186, + 0.014277072623372078, + 0.007262180093675852, + 0.0058058579452335835, + 0.0074076890014112, + 0.00851878710091114, + 0.008329090662300587, + 0.008657295256853104, + 0.005941223353147507, + 0.004560714587569237, + 0.004769324790686369, + 0.003161825705319643, + 0.004871573764830828, + 0.005137295462191105, + 0.005513917654752731, + 0.004021887667477131, + 0.002825748175382614, + 0.003381724003702402, + 0.005598403979092836, + 0.003377686021849513, + 0.00688666757196188, + 0.004435767885297537, + 0.005473801866173744, + 0.006548185367137194, + 0.0031853499822318554, + 0.0027592587284743786, + 0.023529646918177605, + 0.010561813600361347, + 0.01665680482983589, + 0.014897659420967102, + 0.01293550617992878, + 0.016252968460321426, + 0.017806993797421455, + 0.010311154648661613, + 0.04541848972439766, + 0.015053179115056992, + 0.011229468509554863, + 0.023638645187020302, + 0.016892259940505028, + 0.023946557193994522, + 0.055732306092977524, + 0.010422336868941784, + 0.015154949389398098, + 0.010041101835668087, + 0.009476040489971638, + 0.013272539712488651, + 0.019043689593672752, + 0.016706028953194618, + 0.026927195489406586, + 0.0069880676455795765 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 3, + 7 + ], + "selected_reward": -0.054941657930612564 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "narrow, documentary-style shot, suspended almost motionlessly in the water, captures a massive nurse shark, resting peacefully on the fine, white sandy seabed, taking center stage. The scene is bathed in a deep, almost monochromatic sapphire blue, creating a tranquil and mysterious atmosphere of the deep sea. The shark's gray-brown skin appears thick against the blue water. Its broad, flat head rests close to the sand, its small, deep-set eyes seemingly closed in repose, and its signature barbels near its mouth sway gently with the barely perceptible current. Sunlight streaming from above penetrates the deep water, constantly refracted and distorted by the ripples on the surface, creating a constantly shifting, dancing web of light. These shifting specks of light and shadow, like a silent narrative, slowly and rhythmically glide across the shark's broad, smooth back and body curves, vividly highlighting the texture of its skin and the contours of its muscles. The image, in its profound stillness, reveals a primal and majestic vitality.", + "seed": 31477, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000156/group_3/index_3603_seed_31477_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": -0.021613454259932042, + "vlm_logprob": -0.021613454259932042, + "vlm_yes_no_margin": 4.05, + "vlm_token_credit": -0.021613454259932042, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"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 massive nurse shark visible resting on a fine, white sandy seabed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene bathed in a deep, monochromatic sapphire blue color?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does sunlight stream from above, creating shifting specks of light and shadow on the shark's back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the shot framed as a narrow, documentary-style view with the shark taking center stage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the shark's signature barbels near its mouth sway gently with the current?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.028882987797260284 + ], + "opposite_logprob_tokens": [ + -3.6538829803466797 + ], + "logprob_reward": -0.028882987797260284, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.028882987797260284, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2318548858165741, + 0.05279902368783951, + 0.10437892377376556, + 0.021039636805653572, + 0.04603336378931999, + 0.07394716143608093, + 0.03623087331652641, + 0.08903078734874725, + 0.2728515565395355, + 0.07183381915092468 + ], + "credit_map": [ + 0.021278252825140953, + 0.029813865199685097, + 0.03169526532292366, + 0.03190033510327339, + 0.03943874314427376, + 0.02874183841049671, + 0.031081922352313995, + 0.01790465973317623, + 0.00836234912276268, + 0.010218288749456406, + 0.00825780350714922, + 0.005810956470668316, + 0.005276354029774666, + 0.005238926969468594, + 0.004719267599284649, + 0.00491507351398468, + 0.01766051910817623, + 0.01988229900598526, + 0.018908290192484856, + 0.008640014566481113, + 0.010835578665137291, + 0.010970023460686207, + 0.007791121955960989, + 0.009691075421869755, + 0.0035391286946833134, + 0.0026683590840548277, + 0.0025295400992035866, + 0.0021664216183125973, + 0.0026259806472808123, + 0.004026383627206087, + 0.001794095034711063, + 0.0016897268360480666, + 0.005679040681570768, + 0.005952638108283281, + 0.006333579309284687, + 0.004720477852970362, + 0.007203152868896723, + 0.005964280106127262, + 0.006220697425305843, + 0.003959495574235916, + 0.009890607558190823, + 0.012439773418009281, + 0.0071741510182619095, + 0.007792762015014887, + 0.00878951046615839, + 0.00979998055845499, + 0.007430254016071558, + 0.01063012145459652, + 0.006014770362526178, + 0.005759881343692541, + 0.004764341749250889, + 0.0037763812579214573, + 0.004088744521141052, + 0.004373140633106232, + 0.003850368084385991, + 0.003603244898840785, + 0.013972307555377483, + 0.017093803733587265, + 0.009467607364058495, + 0.013888667337596416, + 0.011375068686902523, + 0.007480080239474773, + 0.008184822276234627, + 0.007568421773612499, + 0.04745373874902725, + 0.0689326748251915, + 0.021566003561019897, + 0.028468240052461624, + 0.04119565337896347, + 0.030772283673286438, + 0.01697014644742012, + 0.017492832615971565, + 0.010110653936862946, + 0.007252678275108337, + 0.006737978663295507, + 0.006778386887162924, + 0.00939777959138155, + 0.016815774142742157, + 0.006545856595039368, + 0.00819470640271902 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 2 + ], + "selected_reward": -0.028882987797260284 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"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 massive nurse shark visible resting on a fine, white sandy seabed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene bathed in a deep, monochromatic sapphire blue color?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does sunlight stream from above, creating shifting specks of light and shadow on the shark's back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the shot framed as a narrow, documentary-style view with the shark taking center stage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the shark's signature barbels near its mouth sway gently with the current?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005269684828817844 + ], + "opposite_logprob_tokens": [ + -5.255269527435303 + ], + "logprob_reward": -0.005269684828817844, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005269684828817844, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07128903269767761, + 0.06814660131931305, + 0.22253064811229706, + 0.07637840509414673, + 0.05972164124250412, + 0.05686669424176216, + 0.0323033332824707, + 0.29318973422050476, + 0.0816405862569809, + 0.03793332725763321 + ], + "credit_map": [ + 0.01017563696950674, + 0.012601565569639206, + 0.01270713284611702, + 0.006517430301755667, + 0.008111453615128994, + 0.0070669748820364475, + 0.008172289468348026, + 0.005936546716839075, + 0.011034172028303146, + 0.01725936494767666, + 0.008917090483009815, + 0.0046791560016572475, + 0.008704278618097305, + 0.007453979924321175, + 0.005408142693340778, + 0.004690420348197222, + 0.06380734592676163, + 0.08856013417243958, + 0.012115629389882088, + 0.006652790121734142, + 0.012085773050785065, + 0.026557864621281624, + 0.006954920943826437, + 0.005796193145215511, + 0.013981747440993786, + 0.012586289085447788, + 0.009838176891207695, + 0.0072402083314955235, + 0.008698326535522938, + 0.01221301406621933, + 0.005837581120431423, + 0.005983063485473394, + 0.008663838729262352, + 0.008835556916892529, + 0.007964005693793297, + 0.004538011737167835, + 0.00795398186892271, + 0.008008713833987713, + 0.006512476596981287, + 0.007245056796818972, + 0.009087273851037025, + 0.012306124903261662, + 0.007885115221142769, + 0.0055176373571157455, + 0.0062157167121768, + 0.005536404438316822, + 0.0046898298896849155, + 0.005628592800348997, + 0.00633267592638731, + 0.006138869561254978, + 0.004114654380828142, + 0.0027585874777287245, + 0.0033495815005153418, + 0.004127092659473419, + 0.0034312931820750237, + 0.0020505802240222692, + 0.0482790470123291, + 0.05891811102628708, + 0.04767735302448273, + 0.042426906526088715, + 0.029415057972073555, + 0.021063383668661118, + 0.026206452399492264, + 0.019203444942831993, + 0.011507256887853146, + 0.015699174255132675, + 0.00524442782625556, + 0.007454907987266779, + 0.013220392167568207, + 0.015532162971794605, + 0.005217370577156544, + 0.0077648949809372425, + 0.005391131155192852, + 0.00310980761423707, + 0.006262646988034248, + 0.0034027190413326025, + 0.005728806834667921, + 0.007381049916148186, + 0.0032467488199472427, + 0.0034104175865650177 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 2, + 8 + ], + "selected_reward": -0.005269684828817844 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"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 massive nurse shark visible resting on a fine, white sandy seabed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene bathed in a deep, monochromatic sapphire blue color?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does sunlight stream from above, creating shifting specks of light and shadow on the shark's back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the shot framed as a narrow, documentary-style view with the shark taking center stage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the shark's signature barbels near its mouth sway gently with the current?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012545948848128319 + ], + "opposite_logprob_tokens": [ + -4.387546062469482 + ], + "logprob_reward": -0.012545948848128319, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012545948848128319, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10594728589057922, + 0.08359396457672119, + 0.13775084912776947, + 0.04839029908180237, + 0.050937771797180176, + 0.07028192281723022, + 0.05497904121875763, + 0.10091960430145264, + 0.20730815827846527, + 0.1398911476135254 + ], + "credit_map": [ + 0.014325534924864769, + 0.024539703503251076, + 0.00887328665703535, + 0.013967078179121017, + 0.013003590516746044, + 0.011678466573357582, + 0.012299164198338985, + 0.00726046459749341, + 0.014854906126856804, + 0.024303071200847626, + 0.008050225675106049, + 0.007339705713093281, + 0.006652333773672581, + 0.009608826600015163, + 0.006385833956301212, + 0.006399059202522039, + 0.03375837951898575, + 0.04777713865041733, + 0.010500017553567886, + 0.005586516112089157, + 0.008864455856382847, + 0.016962623223662376, + 0.0070160189643502235, + 0.007285698316991329, + 0.008375086821615696, + 0.007760866545140743, + 0.007510917726904154, + 0.004201201256364584, + 0.005754263140261173, + 0.005244613159447908, + 0.00419737258926034, + 0.005345979705452919, + 0.00765112554654479, + 0.010033435188233852, + 0.0074156299233436584, + 0.0037579035852104425, + 0.005469537805765867, + 0.006183065474033356, + 0.005663941614329815, + 0.004763133358210325, + 0.009668971411883831, + 0.009085492230951786, + 0.008554216474294662, + 0.008324169553816319, + 0.008486210368573666, + 0.008980278857052326, + 0.007950528524816036, + 0.009232057258486748, + 0.008248450234532356, + 0.009783630259335041, + 0.005312802270054817, + 0.005886266008019447, + 0.0077842180617153645, + 0.005912971217185259, + 0.006447248160839081, + 0.005603454075753689, + 0.015257162973284721, + 0.01848703995347023, + 0.021417075768113136, + 0.014116705395281315, + 0.008556028828024864, + 0.007152904290705919, + 0.008845062926411629, + 0.00708762276917696, + 0.032370973378419876, + 0.04776837304234505, + 0.015449626371264458, + 0.02002849243581295, + 0.032984863966703415, + 0.029258068650960922, + 0.013138934969902039, + 0.016308806836605072, + 0.020043237134814262, + 0.0102227246388793, + 0.012768213637173176, + 0.010090737603604794, + 0.02084922231733799, + 0.040660470724105835, + 0.009867045097053051, + 0.015389503911137581 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 2 + ], + "selected_reward": -0.012545948848128319 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"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 massive nurse shark visible resting on a fine, white sandy seabed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene bathed in a deep, monochromatic sapphire blue color?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does sunlight stream from above, creating shifting specks of light and shadow on the shark's back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the shot framed as a narrow, documentary-style view with the shark taking center stage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the shark's signature barbels near its mouth sway gently with the current?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03808639943599701 + ], + "opposite_logprob_tokens": [ + -3.288086414337158 + ], + "logprob_reward": -0.03808639943599701, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03808639943599701, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08396046608686447, + 0.07133202254772186, + 0.2832449972629547, + 0.029523836448788643, + 0.14322802424430847, + 0.10849302262067795, + 0.027129532769322395, + 0.07559522986412048, + 0.1345202773809433, + 0.04297268018126488 + ], + "credit_map": [ + 0.008282559923827648, + 0.010584825649857521, + 0.011706619523465633, + 0.009411361999809742, + 0.014517274685204029, + 0.011288300156593323, + 0.011742874048650265, + 0.006426640320569277, + 0.007535375189036131, + 0.009207033552229404, + 0.009708340279757977, + 0.006414124742150307, + 0.014597560279071331, + 0.008822786621749401, + 0.00947160366922617, + 0.005575194489210844, + 0.08295142650604248, + 0.11629623174667358, + 0.009171528741717339, + 0.008013316430151463, + 0.013585343956947327, + 0.03575311228632927, + 0.007776730228215456, + 0.009697278961539268, + 0.005342818796634674, + 0.0033039841800928116, + 0.0034988150000572205, + 0.0026679192669689655, + 0.004792523104697466, + 0.005029093474149704, + 0.0030519107822328806, + 0.0018367693992331624, + 0.017844339832663536, + 0.02058476023375988, + 0.019214153289794922, + 0.013191517442464828, + 0.02196117490530014, + 0.01636095903813839, + 0.020400451496243477, + 0.013670659624040127, + 0.014714997261762619, + 0.01519052404910326, + 0.012799517251551151, + 0.011083676479756832, + 0.017900707200169563, + 0.014524850063025951, + 0.013188770040869713, + 0.009089983068406582, + 0.0031591844744980335, + 0.0031941388733685017, + 0.0030486835166811943, + 0.003256414085626602, + 0.003072354942560196, + 0.0043686311691999435, + 0.004538499750196934, + 0.002491623628884554, + 0.009803099557757378, + 0.010661308653652668, + 0.010940310545265675, + 0.010959034785628319, + 0.010686427354812622, + 0.008202403783798218, + 0.00650386419147253, + 0.007838782854378223, + 0.024292120710015297, + 0.03559340536594391, + 0.011123769916594028, + 0.013573605567216873, + 0.020412296056747437, + 0.014087528921663761, + 0.007557156961411238, + 0.007880380377173424, + 0.006195405498147011, + 0.004239747300744057, + 0.004339905455708504, + 0.004567737225443125, + 0.006486209575086832, + 0.007983798161149025, + 0.0038237336557358503, + 0.005336141213774681 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 4, + 8 + ], + "selected_reward": -0.03808639943599701 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"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 massive nurse shark visible resting on a fine, white sandy seabed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene bathed in a deep, monochromatic sapphire blue color?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does sunlight stream from above, creating shifting specks of light and shadow on the shark's back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the shot framed as a narrow, documentary-style view with the shark taking center stage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the shark's signature barbels near its mouth sway gently with the current?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02328225038945675 + ], + "opposite_logprob_tokens": [ + -3.773282289505005 + ], + "logprob_reward": -0.02328225038945675, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.02328225038945675, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11591190844774246, + 0.08885838836431503, + 0.13319720327854156, + 0.034745387732982635, + 0.06670785695314407, + 0.05519234761595726, + 0.071856789290905, + 0.2086857259273529, + 0.0502265989780426, + 0.17461775243282318 + ], + "credit_map": [ + 0.0125741520896554, + 0.017181571573019028, + 0.01070308592170477, + 0.013143100775778294, + 0.020170193165540695, + 0.014156820252537727, + 0.016511278226971626, + 0.011471714824438095, + 0.01271356362849474, + 0.020022910088300705, + 0.007860427722334862, + 0.009626777842640877, + 0.010448169894516468, + 0.009879758581519127, + 0.009517780505120754, + 0.0087890001013875, + 0.040538374334573746, + 0.05692058801651001, + 0.004425234626978636, + 0.0031410257797688246, + 0.006307288073003292, + 0.016760913655161858, + 0.002538645640015602, + 0.0025651201140135527, + 0.006398663856089115, + 0.005415358114987612, + 0.003798315767198801, + 0.0028773921076208353, + 0.004507311619818211, + 0.004730912856757641, + 0.00386843946762383, + 0.003148995339870453, + 0.010153695940971375, + 0.009497955441474915, + 0.007927163504064083, + 0.005243923980742693, + 0.008601675741374493, + 0.01085443515330553, + 0.006819715723395348, + 0.007609295193105936, + 0.007128279190510511, + 0.007964901626110077, + 0.007200927007943392, + 0.006916866172105074, + 0.006852501537650824, + 0.005813330877572298, + 0.0067375218495726585, + 0.006578023079782724, + 0.01162815187126398, + 0.010335753671824932, + 0.010151362046599388, + 0.006655669771134853, + 0.007443297654390335, + 0.008982320316135883, + 0.009347246028482914, + 0.007312992122024298, + 0.032392773777246475, + 0.03352506458759308, + 0.043541070073843, + 0.031209053471684456, + 0.015017248690128326, + 0.016108838841319084, + 0.019803578034043312, + 0.01708810403943062, + 0.007155207917094231, + 0.010448407381772995, + 0.0037247631698846817, + 0.0044871666468679905, + 0.007401998154819012, + 0.0073107327334582806, + 0.004023493267595768, + 0.005674831103533506, + 0.023619968444108963, + 0.012676105834543705, + 0.014285854063928127, + 0.01341345813125372, + 0.02601677179336548, + 0.049710482358932495, + 0.013301624916493893, + 0.021593481302261353 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 9, + 2 + ], + "selected_reward": -0.02328225038945675 + } + ], + "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/0000156/group_4/index_tensor([9113, 9113])_group_4_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000156/group_4/index_tensor([9113, 9113])_group_4_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..6aa516b6abb05751cd53756b8be3589d5a011d4c --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000156/group_4/index_tensor([9113, 9113])_group_4_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "low-angle, high-speed camera shot, filmed in the style of a sports documentary, closely follows a heavily modified truck across the vast desert. Above it, the rusted metal skeleton of a massive, abandoned satellite receiving station stands silently, its mottled dish antenna casting a vast shadow in the low, setting sun. The matte black truck performs a breathtaking wheelie stunt with astonishing balance, its front end lifted high, almost perpendicular to the ground, revealing its complex, reinforced suspension system with red springs. Its massive, deeply treaded rear wheels furiously dig into the soft sand, sending billowing golden dust billowing backward and to the sides in a spectacular dust wave. Some grains of sand caught in the camera's path create a natural, gritty effect, enhancing the scene's grittiness. The soft golden light of the afternoon shines from the back of the truck, forming a clear halo, softly outlining the hard metal contours of the car body and the edges of the flying dust. The whole scene is full of primitive power and thrilling dynamic aesthetics.", + "seed": 15437, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000156/group_4/index_9113_seed_15437_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": -0.02300945243332535, + "vlm_logprob": -0.02300945243332535, + "vlm_yes_no_margin": 4.6, + "vlm_token_credit": -0.02300945243332535, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a heavily modified matte black truck?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the truck performing a wheelie stunt with its front end lifted high and almost perpendicular to the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a rusted metal skeleton of a massive abandoned satellite receiving station visible in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene take place in a vast desert with billowing golden dust kicked up by the truck's rear wheels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot filmed from a low angle with a high-speed camera following the truck?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.037351470440626144 + ], + "opposite_logprob_tokens": [ + -4.537351608276367 + ], + "logprob_reward": -0.037351470440626144, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.037351470440626144, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12508897483348846, + 0.22390732169151306, + 0.03421814739704132, + 0.12377773225307465, + 0.02844151295721531, + 0.08132340759038925, + 0.1271994262933731, + 0.12565073370933533, + 0.05547022819519043, + 0.07492246478796005 + ], + "credit_map": [ + 0.017273126170039177, + 0.011027531698346138, + 0.02930058166384697, + 0.015574386343359947, + 0.013433202169835567, + 0.01260296069085598, + 0.012621195986866951, + 0.013255991041660309, + 0.023146385326981544, + 0.011175377294421196, + 0.07740480452775955, + 0.019300511106848717, + 0.04405681788921356, + 0.016118723899126053, + 0.017939649522304535, + 0.014765048399567604, + 0.00314434920437634, + 0.004763511940836906, + 0.0058917999267578125, + 0.004913911689072847, + 0.00454008299857378, + 0.0032952071633189917, + 0.004250198137015104, + 0.0034190856385976076, + 0.020146192982792854, + 0.012315786443650723, + 0.010868419893085957, + 0.006136092822998762, + 0.017923785373568535, + 0.04135438799858093, + 0.006502652075141668, + 0.008530412800610065, + 0.003337597707286477, + 0.0034535888116806746, + 0.002967364853248, + 0.0033716929610818624, + 0.0027043416630476713, + 0.004484680015593767, + 0.0041066743433475494, + 0.004015572834759951, + 0.007535714190453291, + 0.009977176785469055, + 0.010858508758246899, + 0.013088577426970005, + 0.008782383985817432, + 0.007442370057106018, + 0.00657843379303813, + 0.017060240730643272, + 0.014036417007446289, + 0.017864584922790527, + 0.014846713282167912, + 0.01305617019534111, + 0.013104071840643883, + 0.021281585097312927, + 0.010774229653179646, + 0.022235652431845665, + 0.016684429720044136, + 0.011997451074421406, + 0.0174102745950222, + 0.010962908156216145, + 0.016084151342511177, + 0.01573367230594158, + 0.012705138884484768, + 0.02407270297408104, + 0.006574444007128477, + 0.004372107796370983, + 0.005405641160905361, + 0.008079517632722855, + 0.00653039338067174, + 0.007671317085623741, + 0.0077454699203372, + 0.009091334417462349, + 0.019242094829678535, + 0.02230902761220932, + 0.0033751458395272493, + 0.005061883945018053, + 0.005892612040042877, + 0.006255569867789745, + 0.005941128358244896, + 0.006845004390925169 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 6, + 7 + ], + "selected_reward": -0.037351470440626144 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a heavily modified matte black truck?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the truck performing a wheelie stunt with its front end lifted high and almost perpendicular to the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a rusted metal skeleton of a massive abandoned satellite receiving station visible in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene take place in a vast desert with billowing golden dust kicked up by the truck's rear wheels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot filmed from a low angle with a high-speed camera following the truck?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01821855828166008 + ], + "opposite_logprob_tokens": [ + -4.018218517303467 + ], + "logprob_reward": -0.01821855828166008, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.01821855828166008, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11197450011968613, + 0.10449978709220886, + 0.02761854976415634, + 0.033254507929086685, + 0.018451189622282982, + 0.16563017666339874, + 0.06256920844316483, + 0.05313851684331894, + 0.04396889731287956, + 0.3788946568965912 + ], + "credit_map": [ + 0.014537349343299866, + 0.00625217892229557, + 0.039865490049123764, + 0.012109577655792236, + 0.007439690176397562, + 0.006597455125302076, + 0.007359596900641918, + 0.017813151702284813, + 0.010088560171425343, + 0.00679605221375823, + 0.03383132815361023, + 0.011137397028505802, + 0.019375009462237358, + 0.007644338067620993, + 0.008598418906331062, + 0.007028676103800535, + 0.0025151122827082872, + 0.004253698047250509, + 0.004672044422477484, + 0.0034502996131777763, + 0.003335312008857727, + 0.0034483042545616627, + 0.002929672598838806, + 0.0030141035094857216, + 0.00424190191552043, + 0.0039552864618599415, + 0.004355129320174456, + 0.0032193060033023357, + 0.004625553730875254, + 0.005837602540850639, + 0.003960156813263893, + 0.003059566719457507, + 0.002131657674908638, + 0.0025686186272650957, + 0.003021234180778265, + 0.00243469700217247, + 0.0019132976885885, + 0.0025877589359879494, + 0.0018924149917438626, + 0.0019015104044228792, + 0.012521819211542606, + 0.023629965260624886, + 0.018943816423416138, + 0.04131727293133736, + 0.028686782345175743, + 0.014331795275211334, + 0.012004914693534374, + 0.01419379934668541, + 0.008132065646350384, + 0.00968105811625719, + 0.010951180942356586, + 0.008041062392294407, + 0.005676130298525095, + 0.004163934383541346, + 0.007685836870223284, + 0.00823794212192297, + 0.007539867889136076, + 0.0049959877505898476, + 0.008190200664103031, + 0.004884613212198019, + 0.006689423229545355, + 0.0060758632607758045, + 0.0072797369211912155, + 0.00748281879350543, + 0.0053444355726242065, + 0.003677415894344449, + 0.0051241228356957436, + 0.0038623083382844925, + 0.004242085851728916, + 0.0068764276802539825, + 0.004930989816784859, + 0.009911109693348408, + 0.09098616987466812, + 0.12280470877885818, + 0.01800060085952282, + 0.018201492726802826, + 0.02897653542459011, + 0.047020018100738525, + 0.031416911631822586, + 0.02148822508752346 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 5, + 0 + ], + "selected_reward": -0.01821855828166008 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a heavily modified matte black truck?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the truck performing a wheelie stunt with its front end lifted high and almost perpendicular to the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a rusted metal skeleton of a massive abandoned satellite receiving station visible in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene take place in a vast desert with billowing golden dust kicked up by the truck's rear wheels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot filmed from a low angle with a high-speed camera following the truck?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04864773899316788 + ], + "opposite_logprob_tokens": [ + -3.04864764213562 + ], + "logprob_reward": -0.04864773899316788, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04864773899316788, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.45974621176719666, + 0.1531282514333725, + 0.016791896894574165, + 0.03316210210323334, + 0.02715705893933773, + 0.10235172510147095, + 0.07753855735063553, + 0.058751508593559265, + 0.03752593696117401, + 0.03384681046009064 + ], + "credit_map": [ + 0.07191826403141022, + 0.02403210662305355, + 0.16688689589500427, + 0.05168435350060463, + 0.02480313554406166, + 0.030885418877005577, + 0.030087288469076157, + 0.05944870784878731, + 0.01636037789285183, + 0.009029473178088665, + 0.05380427837371826, + 0.012933861464262009, + 0.027418412268161774, + 0.0113609554246068, + 0.012516045942902565, + 0.009704836644232273, + 0.0017418967327103019, + 0.002340956823900342, + 0.002128664869815111, + 0.001998358638957143, + 0.0025809917133301497, + 0.001921670394949615, + 0.002131102606654167, + 0.0019482544157654047, + 0.00324436672963202, + 0.004068631678819656, + 0.005739981308579445, + 0.0036606290377676487, + 0.0053977277129888535, + 0.0037863932084292173, + 0.003606007667258382, + 0.003658364526927471, + 0.003003173042088747, + 0.004445550963282585, + 0.003965543583035469, + 0.0029885992407798767, + 0.0031673097983002663, + 0.0044077318161726, + 0.0032484401017427444, + 0.0019307100446894765, + 0.010122274048626423, + 0.012841061688959599, + 0.01482661534100771, + 0.019607840105891228, + 0.012106252834200859, + 0.01354648731648922, + 0.008360026404261589, + 0.01094115525484085, + 0.009699425660073757, + 0.012741026468575, + 0.008375120349228382, + 0.006749437190592289, + 0.006310740485787392, + 0.01746545359492302, + 0.006002070847898722, + 0.010195276699960232, + 0.0077812871895730495, + 0.007132661063224077, + 0.009351721033453941, + 0.0054738507606089115, + 0.009118936024606228, + 0.0072188181802630424, + 0.005635750945657492, + 0.007038479205220938, + 0.004499522037804127, + 0.003248447086662054, + 0.005693279206752777, + 0.00432754959911108, + 0.005290589760988951, + 0.004125659354031086, + 0.004275952465832233, + 0.006064934656023979, + 0.006027912721037865, + 0.008466154336929321, + 0.0021855931263417006, + 0.0026916342321783304, + 0.0030274109449237585, + 0.005017993971705437, + 0.003542352933436632, + 0.0028877549339085817 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 5 + ], + "selected_reward": -0.04864773899316788 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a heavily modified matte black truck?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the truck performing a wheelie stunt with its front end lifted high and almost perpendicular to the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a rusted metal skeleton of a massive abandoned satellite receiving station visible in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene take place in a vast desert with billowing golden dust kicked up by the truck's rear wheels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot filmed from a low angle with a high-speed camera following the truck?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.001042656716890633 + ], + "opposite_logprob_tokens": [ + -6.87604284286499 + ], + "logprob_reward": -0.001042656716890633, + "yes_no_margin_reward": 6.875, + "token_credit_reward": -0.001042656716890633, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09305954724550247, + 0.2384868711233139, + 0.015838341787457466, + 0.040779031813144684, + 0.01116075087338686, + 0.12016357481479645, + 0.2381957769393921, + 0.06415164470672607, + 0.05269356444478035, + 0.12547089159488678 + ], + "credit_map": [ + 0.014397176913917065, + 0.00760337058454752, + 0.018414512276649475, + 0.015030772425234318, + 0.008046495728194714, + 0.006854776293039322, + 0.009380808100104332, + 0.013331649824976921, + 0.026013396680355072, + 0.01112490613013506, + 0.08494379371404648, + 0.01953743025660515, + 0.04391675442457199, + 0.0167124904692173, + 0.02009359933435917, + 0.016144489869475365, + 0.0016259968979284167, + 0.002031044801697135, + 0.002152510918676853, + 0.0023953483905643225, + 0.0029365660157054663, + 0.0018722466193139553, + 0.0012097047874704003, + 0.0016149238217622042, + 0.007805509027093649, + 0.0033991325180977583, + 0.0033673010766506195, + 0.0021784487180411816, + 0.004430596251040697, + 0.01351950503885746, + 0.00236497656442225, + 0.0037135661114007235, + 0.0014021563110873103, + 0.0015013226075097919, + 0.001543334568850696, + 0.0011662428732961416, + 0.0012233342276886106, + 0.0016609546728432178, + 0.001243429258465767, + 0.0014199777506291866, + 0.011531240306794643, + 0.0237499438226223, + 0.015222014859318733, + 0.017967719584703445, + 0.013035204261541367, + 0.013186792843043804, + 0.009744293987751007, + 0.015726374462246895, + 0.03241907060146332, + 0.012707924470305443, + 0.03375774621963501, + 0.011633464135229588, + 0.012708940543234348, + 0.04592050984501839, + 0.019235316663980484, + 0.0698128193616867, + 0.008942081592977047, + 0.0062454561702907085, + 0.008986743167042732, + 0.006588628515601158, + 0.008860508911311626, + 0.008738317526876926, + 0.005554152186959982, + 0.01023576594889164, + 0.008982149884104729, + 0.008759596385061741, + 0.0028980367351323366, + 0.0038845334202051163, + 0.004885809496045113, + 0.01444462314248085, + 0.003759060287848115, + 0.005079763475805521, + 0.016272302716970444, + 0.028790317475795746, + 0.008524858392775059, + 0.011461990885436535, + 0.01043593231588602, + 0.020970415323972702, + 0.014048207551240921, + 0.014966875314712524 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 6, + 9 + ], + "selected_reward": -0.001042656716890633 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a heavily modified matte black truck?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the truck performing a wheelie stunt with its front end lifted high and almost perpendicular to the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a rusted metal skeleton of a massive abandoned satellite receiving station visible in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene take place in a vast desert with billowing golden dust kicked up by the truck's rear wheels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot filmed from a low angle with a high-speed camera following the truck?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009786837734282017 + ], + "opposite_logprob_tokens": [ + -4.634786605834961 + ], + "logprob_reward": -0.009786837734282017, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009786837734282017, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12581698596477509, + 0.2557780146598816, + 0.04393475875258446, + 0.055981431156396866, + 0.0329039990901947, + 0.13451723754405975, + 0.07045494019985199, + 0.07958374172449112, + 0.12049118429422379, + 0.08053769916296005 + ], + "credit_map": [ + 0.017807627096772194, + 0.009874582290649414, + 0.02357431687414646, + 0.01896658167243004, + 0.012109645642340183, + 0.011786331422626972, + 0.014887099154293537, + 0.016810797154903412, + 0.027087166905403137, + 0.01740610972046852, + 0.08442271500825882, + 0.021655919030308723, + 0.04990772902965546, + 0.018869299441576004, + 0.02010343037545681, + 0.01632561720907688, + 0.004917771089822054, + 0.005513329058885574, + 0.0071884519420564175, + 0.006723690778017044, + 0.005098275374621153, + 0.005128426477313042, + 0.004236843902617693, + 0.005127966403961182, + 0.007612777873873711, + 0.005548686720430851, + 0.008514785207808018, + 0.0040221731178462505, + 0.00680199358612299, + 0.012760619632899761, + 0.005067882128059864, + 0.005652515683323145, + 0.003926401492208242, + 0.004214593209326267, + 0.004885628819465637, + 0.0036972633097320795, + 0.004908801056444645, + 0.004936252720654011, + 0.003661162918433547, + 0.0026738967280834913, + 0.012964598834514618, + 0.02851938270032406, + 0.01832546852529049, + 0.022528674453496933, + 0.013049236498773098, + 0.009875238873064518, + 0.012630538083612919, + 0.016624094918370247, + 0.007529673166573048, + 0.009811564348638058, + 0.007390814367681742, + 0.005673707462847233, + 0.007262601051479578, + 0.00790905486792326, + 0.00904666818678379, + 0.015830861404538155, + 0.01012891624122858, + 0.009854860603809357, + 0.009531041607260704, + 0.007991456426680088, + 0.010433854535222054, + 0.010819767601788044, + 0.006813591346144676, + 0.01401025615632534, + 0.015036207623779774, + 0.008452103473246098, + 0.013017292134463787, + 0.01010496262460947, + 0.011522927321493626, + 0.020754175260663033, + 0.013172704726457596, + 0.028430810198187828, + 0.021757809445261955, + 0.02666158229112625, + 0.003270163433626294, + 0.0036964602768421173, + 0.005082021001726389, + 0.007512813434004784, + 0.0044653145596385, + 0.00809154100716114 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 5, + 0 + ], + "selected_reward": -0.009786837734282017 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle, high-speed camera shot, filmed in the style of a sports documentary, closely follows a heavily modified truck across the vast desert. Above it, the rusted metal skeleton of a massive, abandoned satellite receiving station stands silently, its mottled dish antenna casting a vast shadow in the low, setting sun. The matte black truck performs a breathtaking wheelie stunt with astonishing balance, its front end lifted high, almost perpendicular to the ground, revealing its complex, reinforced suspension system with red springs. Its massive, deeply treaded rear wheels furiously dig into the soft sand, sending billowing golden dust billowing backward and to the sides in a spectacular dust wave. Some grains of sand caught in the camera's path create a natural, gritty effect, enhancing the scene's grittiness. The soft golden light of the afternoon shines from the back of the truck, forming a clear halo, softly outlining the hard metal contours of the car body and the edges of the flying dust. The whole scene is full of primitive power and thrilling dynamic aesthetics.", + "seed": 15438, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000156/group_4/index_9113_seed_15438_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": -0.01242375749279745, + "vlm_logprob": -0.01242375749279745, + "vlm_yes_no_margin": 5.3, + "vlm_token_credit": -0.01242375749279745, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a heavily modified matte black truck?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the truck performing a wheelie stunt with its front end lifted high and almost perpendicular to the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a rusted metal skeleton of a massive abandoned satellite receiving station visible in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene take place in a vast desert with billowing golden dust kicked up by the truck's rear wheels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot filmed from a low angle with a high-speed camera following the truck?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.031571730971336365 + ], + "opposite_logprob_tokens": [ + -4.531571865081787 + ], + "logprob_reward": -0.031571730971336365, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.031571730971336365, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1070018783211708, + 0.03969544172286987, + 0.033496368676424026, + 0.11399658769369125, + 0.06614957749843597, + 0.07269390672445297, + 0.09468524903059006, + 0.2816011309623718, + 0.05302019789814949, + 0.1376597285270691 + ], + "credit_map": [ + 0.009559894911944866, + 0.012106397189199924, + 0.007751626893877983, + 0.009631920605897903, + 0.030605893582105637, + 0.01091176550835371, + 0.01342707872390747, + 0.013007299043238163, + 0.0038785291835665703, + 0.0043776375241577625, + 0.003987446893006563, + 0.004947177600115538, + 0.010729726403951645, + 0.0039479839615523815, + 0.004045356530696154, + 0.003781580366194248, + 0.003162405453622341, + 0.003497984493151307, + 0.005607178434729576, + 0.005080064758658409, + 0.004752553999423981, + 0.0031899793539196253, + 0.0030559804290533066, + 0.005150224547833204, + 0.008578947745263577, + 0.007225031964480877, + 0.026788534596562386, + 0.014463666826486588, + 0.013726816512644291, + 0.015355195850133896, + 0.012181313708424568, + 0.015677079558372498, + 0.01100656483322382, + 0.005618288181722164, + 0.011496718972921371, + 0.0070493691600859165, + 0.006147958803921938, + 0.013805033639073372, + 0.005307632498443127, + 0.005718009080737829, + 0.007403884548693895, + 0.01126696914434433, + 0.012430156581103802, + 0.00866147130727768, + 0.006209895014762878, + 0.008560670539736748, + 0.008382984437048435, + 0.009777877479791641, + 0.010362284258008003, + 0.009201325476169586, + 0.019299481064081192, + 0.011838418431580067, + 0.008647220209240913, + 0.009428084827959538, + 0.013016236945986748, + 0.012892194092273712, + 0.027438102290034294, + 0.009370563551783562, + 0.09273840487003326, + 0.010339541360735893, + 0.00892571359872818, + 0.11251586675643921, + 0.009102143347263336, + 0.011170800775289536, + 0.00608796626329422, + 0.005217812955379486, + 0.009947369806468487, + 0.005808848887681961, + 0.008257347159087658, + 0.004462442826479673, + 0.004465735983103514, + 0.00877267587929964, + 0.013701985590159893, + 0.01650647632777691, + 0.016651825979351997, + 0.022228460758924484, + 0.01657017692923546, + 0.017016658559441566, + 0.019534887745976448, + 0.015449253842234612 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 9, + 3 + ], + "selected_reward": -0.031571730971336365 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a heavily modified matte black truck?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the truck performing a wheelie stunt with its front end lifted high and almost perpendicular to the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a rusted metal skeleton of a massive abandoned satellite receiving station visible in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene take place in a vast desert with billowing golden dust kicked up by the truck's rear wheels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot filmed from a low angle with a high-speed camera following the truck?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018210014328360558 + ], + "opposite_logprob_tokens": [ + -4.018209934234619 + ], + "logprob_reward": -0.018210014328360558, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018210014328360558, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05460462346673012, + 0.05338691174983978, + 0.018039051443338394, + 0.1268259882926941, + 0.062281087040901184, + 0.04398408904671669, + 0.0602809302508831, + 0.4643332064151764, + 0.054201774299144745, + 0.062062304466962814 + ], + "credit_map": [ + 0.006398211698979139, + 0.010400143451988697, + 0.005282278638333082, + 0.005819916259497404, + 0.006515675690025091, + 0.005987434182316065, + 0.007279284298419952, + 0.006921680644154549, + 0.007040182128548622, + 0.007541566155850887, + 0.011125123128294945, + 0.006319258362054825, + 0.004741447977721691, + 0.006144600920379162, + 0.0055977459996938705, + 0.004876988008618355, + 0.002659410936757922, + 0.0019740452989935875, + 0.0024004876613616943, + 0.002699444303289056, + 0.001838531345129013, + 0.0017343483632430434, + 0.002459991490468383, + 0.002272792859002948, + 0.010459923185408115, + 0.007923859171569347, + 0.032203640788793564, + 0.01646769605576992, + 0.014122018590569496, + 0.01690317504107952, + 0.013050714507699013, + 0.015694964677095413, + 0.008914071135222912, + 0.005019985605031252, + 0.012367538176476955, + 0.0056122844107449055, + 0.0050356569699943066, + 0.013471057638525963, + 0.0059547582641243935, + 0.005905734375119209, + 0.005690884310752153, + 0.004592400509864092, + 0.008594672195613384, + 0.00556263467296958, + 0.004283652175217867, + 0.005732137244194746, + 0.005538725759834051, + 0.003988984972238541, + 0.007499608676880598, + 0.004677167162299156, + 0.014933619648218155, + 0.008773286826908588, + 0.0040227570571005344, + 0.005974095314741135, + 0.00702196080237627, + 0.007378434296697378, + 0.04227184131741524, + 0.011611136607825756, + 0.15671519935131073, + 0.01842154748737812, + 0.016759244725108147, + 0.18315951526165009, + 0.017592333257198334, + 0.017802361398935318, + 0.007343824487179518, + 0.006977458484470844, + 0.007712101098150015, + 0.006447474937886, + 0.006893571466207504, + 0.005235159769654274, + 0.005307267885655165, + 0.008284917101264, + 0.007491410244256258, + 0.005944402888417244, + 0.012879830785095692, + 0.00800152774900198, + 0.005999192129820585, + 0.00545552046969533, + 0.008644037880003452, + 0.007646379992365837 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 3, + 4 + ], + "selected_reward": -0.018210014328360558 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a heavily modified matte black truck?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the truck performing a wheelie stunt with its front end lifted high and almost perpendicular to the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a rusted metal skeleton of a massive abandoned satellite receiving station visible in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene take place in a vast desert with billowing golden dust kicked up by the truck's rear wheels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot filmed from a low angle with a high-speed camera following the truck?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0059483409859240055 + ], + "opposite_logprob_tokens": [ + -5.130948543548584 + ], + "logprob_reward": -0.0059483409859240055, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.0059483409859240055, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19781960546970367, + 0.0742991492152214, + 0.040934085845947266, + 0.04091964289546013, + 0.0572928786277771, + 0.05307868495583534, + 0.047434065490961075, + 0.3211456835269928, + 0.06832010298967361, + 0.0987560823559761 + ], + "credit_map": [ + 0.022520586848258972, + 0.040917787700891495, + 0.01859264448285103, + 0.017524097114801407, + 0.028670797124505043, + 0.018470272421836853, + 0.026375873014330864, + 0.02474755421280861, + 0.0077758594416081905, + 0.00943237729370594, + 0.013407383114099503, + 0.008225515484809875, + 0.0075178309343755245, + 0.008540930226445198, + 0.011023754253983498, + 0.008375492878258228, + 0.0032899982761591673, + 0.004660820122808218, + 0.004570668563246727, + 0.004637335892766714, + 0.005540062207728624, + 0.007576075382530689, + 0.003921672701835632, + 0.0067374506033957005, + 0.003806467168033123, + 0.003114072373136878, + 0.006250730715692043, + 0.006965653970837593, + 0.007355417590588331, + 0.003939454909414053, + 0.004518077243119478, + 0.004969767760485411, + 0.007325214333832264, + 0.006398654542863369, + 0.009024097584187984, + 0.006561586167663336, + 0.007202105596661568, + 0.006310091353952885, + 0.006649285554885864, + 0.007821842096745968, + 0.005287373438477516, + 0.007372067775577307, + 0.009756968356668949, + 0.00599607452750206, + 0.004742463119328022, + 0.006413517519831657, + 0.0061417268589138985, + 0.00736849382519722, + 0.004372574854642153, + 0.005393566098064184, + 0.00712309405207634, + 0.006478957366198301, + 0.0063146185129880905, + 0.004239785019308329, + 0.00801299698650837, + 0.005498474929481745, + 0.028987308964133263, + 0.006523017305880785, + 0.11277318000793457, + 0.014156948775053024, + 0.006767311133444309, + 0.13029304146766663, + 0.00826340727508068, + 0.01338146161288023, + 0.006568820681422949, + 0.006875894498080015, + 0.013486267067492008, + 0.008585700765252113, + 0.006661772262305021, + 0.005932803265750408, + 0.008084757253527641, + 0.012124087661504745, + 0.011845829896628857, + 0.013262084685266018, + 0.01720072142779827, + 0.01308745238929987, + 0.009064698591828346, + 0.010809972882270813, + 0.011997734196484089, + 0.011487585492432117 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 9 + ], + "selected_reward": -0.0059483409859240055 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a heavily modified matte black truck?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the truck performing a wheelie stunt with its front end lifted high and almost perpendicular to the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a rusted metal skeleton of a massive abandoned satellite receiving station visible in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene take place in a vast desert with billowing golden dust kicked up by the truck's rear wheels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot filmed from a low angle with a high-speed camera following the truck?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0004365683125797659 + ], + "opposite_logprob_tokens": [ + -7.750436782836914 + ], + "logprob_reward": -0.0004365683125797659, + "yes_no_margin_reward": 7.75, + "token_credit_reward": -0.0004365683125797659, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09509140253067017, + 0.07007554173469543, + 0.033606212586164474, + 0.03492587059736252, + 0.08852314949035645, + 0.060977786779403687, + 0.17900174856185913, + 0.21900269389152527, + 0.13166441023349762, + 0.08713124692440033 + ], + "credit_map": [ + 0.01024083886295557, + 0.013400396332144737, + 0.008953632786870003, + 0.008547344245016575, + 0.020309364423155785, + 0.009454034268856049, + 0.012691869400441647, + 0.011493915691971779, + 0.006720425095409155, + 0.00846407376229763, + 0.007388112600892782, + 0.010207271203398705, + 0.014973175711929798, + 0.006499285809695721, + 0.008411751128733158, + 0.00741144735366106, + 0.0021688418928533792, + 0.003033536486327648, + 0.004341576248407364, + 0.004681233782321215, + 0.0026367499958723783, + 0.005220927298069, + 0.0035004490055143833, + 0.008022896014153957, + 0.004180077463388443, + 0.002544811926782131, + 0.007611729204654694, + 0.004012127406895161, + 0.002952283015474677, + 0.004497880581766367, + 0.003985103685408831, + 0.005141850560903549, + 0.008636532351374626, + 0.010754445567727089, + 0.017948204651474953, + 0.00948378350585699, + 0.006848465651273727, + 0.010517166927456856, + 0.011421626433730125, + 0.012912916950881481, + 0.00697470223531127, + 0.009398823603987694, + 0.011780529282987118, + 0.006298565771430731, + 0.00604684604331851, + 0.007645501755177975, + 0.004623717628419399, + 0.00820909533649683, + 0.020653076469898224, + 0.01055415440350771, + 0.04015830159187317, + 0.021295880898833275, + 0.011168121360242367, + 0.02176351100206375, + 0.020275508984923363, + 0.03313319385051727, + 0.019338171929121017, + 0.007345074322074652, + 0.07218993455171585, + 0.010506263934075832, + 0.007551639806479216, + 0.08232045918703079, + 0.008564078249037266, + 0.011187058873474598, + 0.01623288169503212, + 0.01215776614844799, + 0.019217848777770996, + 0.019673604518175125, + 0.013187301345169544, + 0.010641640052199364, + 0.01678306795656681, + 0.023770296946167946, + 0.013466508127748966, + 0.008742596954107285, + 0.02148061990737915, + 0.008912710472941399, + 0.007625453639775515, + 0.009676017798483372, + 0.008458035066723824, + 0.008769294247031212 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 6, + 8 + ], + "selected_reward": -0.0004365683125797659 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a heavily modified matte black truck?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the truck performing a wheelie stunt with its front end lifted high and almost perpendicular to the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a rusted metal skeleton of a massive abandoned satellite receiving station visible in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene take place in a vast desert with billowing golden dust kicked up by the truck's rear wheels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot filmed from a low angle with a high-speed camera following the truck?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005952132865786552 + ], + "opposite_logprob_tokens": [ + -5.13095235824585 + ], + "logprob_reward": -0.005952132865786552, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005952132865786552, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11639149487018585, + 0.05866555869579315, + 0.032565779983997345, + 0.045426350086927414, + 0.0763758197426796, + 0.08786652982234955, + 0.09651529788970947, + 0.2453528195619583, + 0.052740175276994705, + 0.18810024857521057 + ], + "credit_map": [ + 0.010477776639163494, + 0.01938742958009243, + 0.012061500921845436, + 0.011051315814256668, + 0.018029645085334778, + 0.010310113430023193, + 0.01873794011771679, + 0.016335777938365936, + 0.007365765981376171, + 0.007175643928349018, + 0.011221738532185555, + 0.007058130111545324, + 0.006096353754401207, + 0.007380112539976835, + 0.006647281348705292, + 0.005720534827560186, + 0.005751843564212322, + 0.003770207753404975, + 0.004202669020742178, + 0.004368587397038937, + 0.003747542854398489, + 0.003016475588083267, + 0.00407551322132349, + 0.003632941050454974, + 0.00670260563492775, + 0.004690895322710276, + 0.011800073087215424, + 0.004795285407453775, + 0.005808422341942787, + 0.004241219721734524, + 0.0033216974698007107, + 0.004066146910190582, + 0.01108346227556467, + 0.00505263963714242, + 0.01284112874418497, + 0.005581878591328859, + 0.00612899474799633, + 0.01937597617506981, + 0.007815920747816563, + 0.008495819754898548, + 0.010764247737824917, + 0.007920771837234497, + 0.014388342387974262, + 0.01094128005206585, + 0.0069452086463570595, + 0.011354797519743443, + 0.016467034816741943, + 0.009084846824407578, + 0.008251070976257324, + 0.010130628943443298, + 0.014927650801837444, + 0.013563706539571285, + 0.011363530531525612, + 0.009672924876213074, + 0.016154469922184944, + 0.01245131529867649, + 0.0188346765935421, + 0.0105031318962574, + 0.07440333068370819, + 0.016358230262994766, + 0.015346942469477654, + 0.08256618678569794, + 0.0174750704318285, + 0.009865233674645424, + 0.005727482959628105, + 0.0074005709029734135, + 0.007910197600722313, + 0.006236208602786064, + 0.006060282699763775, + 0.0040275477804243565, + 0.0076102446764707565, + 0.007767637725919485, + 0.028929641470313072, + 0.01630380004644394, + 0.043842293322086334, + 0.020476961508393288, + 0.012486173771321774, + 0.020852750167250633, + 0.023218609392642975, + 0.021990004926919937 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 9, + 0 + ], + "selected_reward": -0.005952132865786552 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle, high-speed camera shot, filmed in the style of a sports documentary, closely follows a heavily modified truck across the vast desert. Above it, the rusted metal skeleton of a massive, abandoned satellite receiving station stands silently, its mottled dish antenna casting a vast shadow in the low, setting sun. The matte black truck performs a breathtaking wheelie stunt with astonishing balance, its front end lifted high, almost perpendicular to the ground, revealing its complex, reinforced suspension system with red springs. Its massive, deeply treaded rear wheels furiously dig into the soft sand, sending billowing golden dust billowing backward and to the sides in a spectacular dust wave. Some grains of sand caught in the camera's path create a natural, gritty effect, enhancing the scene's grittiness. The soft golden light of the afternoon shines from the back of the truck, forming a clear halo, softly outlining the hard metal contours of the car body and the edges of the flying dust. The whole scene is full of primitive power and thrilling dynamic aesthetics.", + "seed": 15439, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000156/group_4/index_9113_seed_15439_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": -0.00823023875709623, + "vlm_logprob": -0.00823023875709623, + "vlm_yes_no_margin": 5.575, + "vlm_token_credit": -0.00823023875709623, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a heavily modified matte black truck?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the truck performing a wheelie stunt with its front end lifted high and almost perpendicular to the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a rusted metal skeleton of a massive abandoned satellite receiving station visible in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene take place in a vast desert with billowing golden dust kicked up by the truck's rear wheels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot filmed from a low angle with a high-speed camera following the truck?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016144851222634315 + ], + "opposite_logprob_tokens": [ + -5.641144752502441 + ], + "logprob_reward": -0.016144851222634315, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.016144851222634315, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11389698088169098, + 0.0621490404009819, + 0.0630456954240799, + 0.10489478707313538, + 0.09975786507129669, + 0.07162434607744217, + 0.14842379093170166, + 0.10830021649599075, + 0.15035578608512878, + 0.07755156606435776 + ], + "credit_map": [ + 0.015714656561613083, + 0.014995734207332134, + 0.015881506726145744, + 0.009742885828018188, + 0.016623804345726967, + 0.015145361423492432, + 0.007049894426018, + 0.018743125721812248, + 0.009823789820075035, + 0.007531861774623394, + 0.015198986977338791, + 0.004915463272482157, + 0.00590388011187315, + 0.0059141782112419605, + 0.005400979891419411, + 0.00745989428833127, + 0.00703034270554781, + 0.010690003633499146, + 0.006666248664259911, + 0.008352982811629772, + 0.01036688033491373, + 0.0072297630831599236, + 0.006796781439334154, + 0.00591268902644515, + 0.012690446339547634, + 0.007157571148127317, + 0.02512362413108349, + 0.009102202020585537, + 0.016537917777895927, + 0.013413895852863789, + 0.007774104829877615, + 0.013095013797283173, + 0.02164122648537159, + 0.016409166157245636, + 0.011868619360029697, + 0.010232350789010525, + 0.008483916521072388, + 0.008730596862733364, + 0.009761281311511993, + 0.012630700133740902, + 0.0063356622122228146, + 0.005734444595873356, + 0.00836532935500145, + 0.008729970082640648, + 0.01616494543850422, + 0.007573483977466822, + 0.01314017828553915, + 0.005580327473580837, + 0.014590030536055565, + 0.020630476996302605, + 0.01670854166150093, + 0.015466244891285896, + 0.039291054010391235, + 0.014469126239418983, + 0.01568707451224327, + 0.011581222526729107, + 0.011808976531028748, + 0.010749360546469688, + 0.01822570525109768, + 0.017847906798124313, + 0.014333629049360752, + 0.00899926945567131, + 0.012473426759243011, + 0.013861932791769505, + 0.017259027808904648, + 0.007737603969871998, + 0.04306718707084656, + 0.02044045366346836, + 0.01360158622264862, + 0.012254059314727783, + 0.013049628585577011, + 0.022946229204535484, + 0.009511895477771759, + 0.007749645970761776, + 0.00899210199713707, + 0.008344746194779873, + 0.007123473100364208, + 0.015304124914109707, + 0.008649162948131561, + 0.011876407079398632 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 6, + 0 + ], + "selected_reward": -0.016144851222634315 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a heavily modified matte black truck?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the truck performing a wheelie stunt with its front end lifted high and almost perpendicular to the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a rusted metal skeleton of a massive abandoned satellite receiving station visible in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene take place in a vast desert with billowing golden dust kicked up by the truck's rear wheels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot filmed from a low angle with a high-speed camera following the truck?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011093958280980587 + ], + "opposite_logprob_tokens": [ + -4.511094093322754 + ], + "logprob_reward": -0.011093958280980587, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011093958280980587, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0917062908411026, + 0.04686139151453972, + 0.08166081458330154, + 0.11760026961565018, + 0.1239335909485817, + 0.09813501685857773, + 0.1231929287314415, + 0.1727806031703949, + 0.11849699169397354, + 0.025632165372371674 + ], + "credit_map": [ + 0.012583896517753601, + 0.010685817338526249, + 0.013069319538772106, + 0.0072233048267662525, + 0.014114150777459145, + 0.01867692545056343, + 0.005624080542474985, + 0.009728790260851383, + 0.005202400032430887, + 0.004204982426017523, + 0.007046512793749571, + 0.005294641945511103, + 0.006105567794293165, + 0.005344450008124113, + 0.007539265789091587, + 0.006123567000031471, + 0.009507752023637295, + 0.010412775911390781, + 0.0154352318495512, + 0.006850586272776127, + 0.0112897465005517, + 0.008544431999325752, + 0.007037697359919548, + 0.01258259080350399, + 0.01283368468284607, + 0.00865220557898283, + 0.02413194440305233, + 0.011666407808661461, + 0.01975402794778347, + 0.013109899125993252, + 0.013613772578537464, + 0.013838316313922405, + 0.033396366983652115, + 0.01850138045847416, + 0.017305726185441017, + 0.010338588617742062, + 0.010119464248418808, + 0.009967664256691933, + 0.011796141043305397, + 0.01250824797898531, + 0.008932963944971561, + 0.007282281760126352, + 0.010974673554301262, + 0.012751118279993534, + 0.022524014115333557, + 0.010292287915945053, + 0.01954103820025921, + 0.005836636293679476, + 0.010878893546760082, + 0.015195825137197971, + 0.013494079001247883, + 0.012822411023080349, + 0.030782869085669518, + 0.012166937813162804, + 0.018229497596621513, + 0.009622414596378803, + 0.013614541850984097, + 0.02095743454992771, + 0.018836332485079765, + 0.032892920076847076, + 0.03247848153114319, + 0.012995888479053974, + 0.02646823786199093, + 0.01453674677759409, + 0.022418156266212463, + 0.004900255706161261, + 0.04909870773553848, + 0.009784507565200329, + 0.004822473973035812, + 0.004277489148080349, + 0.005804182030260563, + 0.01739121787250042, + 0.0029490157030522823, + 0.0029582944698631763, + 0.003916314337402582, + 0.0030529333744198084, + 0.0029940844979137182, + 0.003794310847297311, + 0.002899874234572053, + 0.003067334648221731 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 4, + 6 + ], + "selected_reward": -0.011093958280980587 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a heavily modified matte black truck?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the truck performing a wheelie stunt with its front end lifted high and almost perpendicular to the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a rusted metal skeleton of a massive abandoned satellite receiving station visible in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene take place in a vast desert with billowing golden dust kicked up by the truck's rear wheels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot filmed from a low angle with a high-speed camera following the truck?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006735123693943024 + ], + "opposite_logprob_tokens": [ + -5.006735324859619 + ], + "logprob_reward": -0.006735123693943024, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006735123693943024, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.35204586386680603, + 0.04870184138417244, + 0.2890264689922333, + 0.01625494286417961, + 0.05473970249295235, + 0.03407026454806328, + 0.038067180663347244, + 0.08301587402820587, + 0.06483839452266693, + 0.01923944242298603 + ], + "credit_map": [ + 0.05242588371038437, + 0.053328827023506165, + 0.03893156722187996, + 0.022590938955545425, + 0.049392201006412506, + 0.07256431877613068, + 0.01657681353390217, + 0.0462353378534317, + 0.005922944750636816, + 0.006755545269697905, + 0.010124907828867435, + 0.0034232656471431255, + 0.008240987546741962, + 0.0041911406442523, + 0.004484445787966251, + 0.005558605305850506, + 0.026784827932715416, + 0.0544009804725647, + 0.026504989713430405, + 0.039500296115875244, + 0.045067064464092255, + 0.031910039484500885, + 0.032798901200294495, + 0.03205939382314682, + 0.0020678648725152016, + 0.0016259741969406605, + 0.0026909413281828165, + 0.0017244743648916483, + 0.0018627322278916836, + 0.0021360106766223907, + 0.001965255243703723, + 0.0021816904190927744, + 0.014407234266400337, + 0.010406781919300556, + 0.004999015014618635, + 0.00510287843644619, + 0.0037196315824985504, + 0.004248333629220724, + 0.00520592974498868, + 0.006649903021752834, + 0.006198039278388023, + 0.004349694121629, + 0.003877784591168165, + 0.002881290391087532, + 0.0056387884542346, + 0.0034602507948875427, + 0.004436534829437733, + 0.003227885579690337, + 0.0034027984365820885, + 0.0036102267913520336, + 0.003951575141400099, + 0.004616767633706331, + 0.008336060680449009, + 0.0035420707426965237, + 0.007000827696174383, + 0.0036068554036319256, + 0.012194826267659664, + 0.007331257686018944, + 0.015547201968729496, + 0.010511399246752262, + 0.007938433438539505, + 0.006810332182794809, + 0.007403193973004818, + 0.015279233455657959, + 0.007057631853967905, + 0.00376063771545887, + 0.019642489030957222, + 0.006338405888527632, + 0.004456603433936834, + 0.004426793195307255, + 0.005645874422043562, + 0.013509970158338547, + 0.0026950868777930737, + 0.002212181920185685, + 0.0021143914200365543, + 0.0019178790971636772, + 0.0020251271780580282, + 0.003355756402015686, + 0.002994321286678314, + 0.0019247003365308046 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 7 + ], + "selected_reward": -0.006735123693943024 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a heavily modified matte black truck?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the truck performing a wheelie stunt with its front end lifted high and almost perpendicular to the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a rusted metal skeleton of a massive abandoned satellite receiving station visible in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene take place in a vast desert with billowing golden dust kicked up by the truck's rear wheels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot filmed from a low angle with a high-speed camera following the truck?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0004371640970930457 + ], + "opposite_logprob_tokens": [ + -7.750437259674072 + ], + "logprob_reward": -0.0004371640970930457, + "yes_no_margin_reward": 7.75, + "token_credit_reward": -0.0004371640970930457, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.163744255900383, + 0.08248887956142426, + 0.07563336193561554, + 0.04393303021788597, + 0.10770290344953537, + 0.052838947623968124, + 0.04850851371884346, + 0.03745381534099579, + 0.355173259973526, + 0.032523006200790405 + ], + "credit_map": [ + 0.02641143836081028, + 0.021297143772244453, + 0.02588583528995514, + 0.00989669281989336, + 0.0274039376527071, + 0.024763526394963264, + 0.009160399436950684, + 0.018925277516245842, + 0.010034852661192417, + 0.00915038026869297, + 0.020685993134975433, + 0.00696589658036828, + 0.009749043732881546, + 0.006264788098633289, + 0.010373537428677082, + 0.009264390915632248, + 0.010050777345895767, + 0.007515012752264738, + 0.016129033640027046, + 0.006040959153324366, + 0.0051306746900081635, + 0.010242542251944542, + 0.006709769368171692, + 0.013814590871334076, + 0.005101347342133522, + 0.002525902120396495, + 0.01118327584117651, + 0.004693963564932346, + 0.0037441172171384096, + 0.0060853008180856705, + 0.0035973547492176294, + 0.007001771125942469, + 0.028379565104842186, + 0.02018866129219532, + 0.011352898553013802, + 0.010327312164008617, + 0.005658634472638369, + 0.008731532841920853, + 0.00978983286768198, + 0.013274460099637508, + 0.011154765263199806, + 0.007117439992725849, + 0.006642007268965244, + 0.006014397367835045, + 0.005179903004318476, + 0.004974851850420237, + 0.006865991745144129, + 0.004889584146440029, + 0.005210498813539743, + 0.005517751444131136, + 0.009069057181477547, + 0.006446012295782566, + 0.006496647838503122, + 0.00493182148784399, + 0.0049529606476426125, + 0.005883766803890467, + 0.005614448804408312, + 0.0038621602579951286, + 0.004752363543957472, + 0.005302642937749624, + 0.005476451013237238, + 0.002981659723445773, + 0.004469261970371008, + 0.004994827788323164, + 0.10026168078184128, + 0.012223534286022186, + 0.14605459570884705, + 0.035130348056554794, + 0.009408185258507729, + 0.011147584766149521, + 0.023475350812077522, + 0.01747198775410652, + 0.00456133671104908, + 0.002746569225564599, + 0.0037974449805915356, + 0.004363473504781723, + 0.0028648183215409517, + 0.007232703268527985, + 0.004249393939971924, + 0.0027072676457464695 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 4 + ], + "selected_reward": -0.0004371640970930457 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a heavily modified matte black truck?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the truck performing a wheelie stunt with its front end lifted high and almost perpendicular to the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a rusted metal skeleton of a massive abandoned satellite receiving station visible in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene take place in a vast desert with billowing golden dust kicked up by the truck's rear wheels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot filmed from a low angle with a high-speed camera following the truck?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006740096490830183 + ], + "opposite_logprob_tokens": [ + -5.006740093231201 + ], + "logprob_reward": -0.006740096490830183, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006740096490830183, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08599253743886948, + 0.03327690064907074, + 0.053767379373311996, + 0.07150866091251373, + 0.075153648853302, + 0.07023714482784271, + 0.05548475310206413, + 0.07688003033399582, + 0.4268796145915985, + 0.050819288939237595 + ], + "credit_map": [ + 0.013272857293486595, + 0.009588300250470638, + 0.011876082979142666, + 0.009629282169044018, + 0.013329241424798965, + 0.010859327390789986, + 0.005285068415105343, + 0.012152377516031265, + 0.004512342624366283, + 0.0031788498163223267, + 0.006728184875100851, + 0.003159347688779235, + 0.004580708686262369, + 0.003921265713870525, + 0.003517195349559188, + 0.003679006127640605, + 0.0064538344740867615, + 0.004568302072584629, + 0.011045628227293491, + 0.004792606923729181, + 0.005575605668127537, + 0.006924382410943508, + 0.004847370553761721, + 0.009559651836752892, + 0.009053793735802174, + 0.008310751989483833, + 0.012207395397126675, + 0.005638636648654938, + 0.011558931320905685, + 0.007870430126786232, + 0.00872220192104578, + 0.008146516047418118, + 0.008847752586007118, + 0.006102659273892641, + 0.011940663680434227, + 0.01018017902970314, + 0.008948729373514652, + 0.006730994675308466, + 0.011615694500505924, + 0.010786972008645535, + 0.008008633740246296, + 0.006126098800450563, + 0.011208158917725086, + 0.007863478735089302, + 0.012870430015027523, + 0.007565013132989407, + 0.009867355227470398, + 0.006727972533553839, + 0.00879835058003664, + 0.005855065770447254, + 0.00807859655469656, + 0.0072880880907177925, + 0.006766245700418949, + 0.005960856098681688, + 0.006719396449625492, + 0.006018153857439756, + 0.008273530751466751, + 0.0088947219774127, + 0.009447603486478329, + 0.014032337814569473, + 0.012893270701169968, + 0.005570243112742901, + 0.010701076127588749, + 0.0070672533474862576, + 0.0713902935385704, + 0.017803603783249855, + 0.1723482459783554, + 0.034036971628665924, + 0.013359401375055313, + 0.014698414131999016, + 0.020534655079245567, + 0.08270803093910217, + 0.005775160621851683, + 0.0059153735637664795, + 0.007649434730410576, + 0.005981246940791607, + 0.005420984234660864, + 0.009425931610167027, + 0.004837405867874622, + 0.005813749972730875 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 7 + ], + "selected_reward": -0.006740096490830183 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle, high-speed camera shot, filmed in the style of a sports documentary, closely follows a heavily modified truck across the vast desert. Above it, the rusted metal skeleton of a massive, abandoned satellite receiving station stands silently, its mottled dish antenna casting a vast shadow in the low, setting sun. The matte black truck performs a breathtaking wheelie stunt with astonishing balance, its front end lifted high, almost perpendicular to the ground, revealing its complex, reinforced suspension system with red springs. Its massive, deeply treaded rear wheels furiously dig into the soft sand, sending billowing golden dust billowing backward and to the sides in a spectacular dust wave. Some grains of sand caught in the camera's path create a natural, gritty effect, enhancing the scene's grittiness. The soft golden light of the afternoon shines from the back of the truck, forming a clear halo, softly outlining the hard metal contours of the car body and the edges of the flying dust. The whole scene is full of primitive power and thrilling dynamic aesthetics.", + "seed": 15440, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000156/group_4/index_9113_seed_15440_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": -0.009847506927326322, + "vlm_logprob": -0.009847506927326322, + "vlm_yes_no_margin": 5.225, + "vlm_token_credit": -0.009847506927326322, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a heavily modified matte black truck?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the truck performing a wheelie stunt with its front end lifted high and almost perpendicular to the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a rusted metal skeleton of a massive abandoned satellite receiving station visible in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene take place in a vast desert with billowing golden dust kicked up by the truck's rear wheels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot filmed from a low angle with a high-speed camera following the truck?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016507476568222046 + ], + "opposite_logprob_tokens": [ + -5.016507625579834 + ], + "logprob_reward": -0.016507476568222046, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.016507476568222046, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1325695961713791, + 0.12320158630609512, + 0.06992657482624054, + 0.0993458554148674, + 0.04823145270347595, + 0.05326513573527336, + 0.17024439573287964, + 0.08732333779335022, + 0.1248626857995987, + 0.09102938324213028 + ], + "credit_map": [ + 0.0221845842897892, + 0.022560182958841324, + 0.008299117907881737, + 0.010449882596731186, + 0.016354402527213097, + 0.018411889672279358, + 0.015132259577512741, + 0.019177274778485298, + 0.0263617392629385, + 0.013848790898919106, + 0.007735070772469044, + 0.01564009115099907, + 0.012215190567076206, + 0.01538475975394249, + 0.017900465056300163, + 0.014115477912127972, + 0.009044767357409, + 0.007673767860978842, + 0.006510098930448294, + 0.00891736801713705, + 0.016734249889850616, + 0.006011410150676966, + 0.00826224870979786, + 0.006772662978619337, + 0.00748235359787941, + 0.007502458523958921, + 0.020330606028437614, + 0.013646923005580902, + 0.01165002305060625, + 0.016850927844643593, + 0.013082727789878845, + 0.008799831382930279, + 0.00639037461951375, + 0.004653135780245066, + 0.0066934251226484776, + 0.004909173119813204, + 0.005526985973119736, + 0.005861502140760422, + 0.008441686630249023, + 0.005755169782787561, + 0.00517584104090929, + 0.005928883329033852, + 0.0059732841327786446, + 0.006778930313885212, + 0.005570185836404562, + 0.004811191000044346, + 0.01275626290589571, + 0.006270559970289469, + 0.011975151486694813, + 0.01908239722251892, + 0.022700415924191475, + 0.02332254871726036, + 0.026909498497843742, + 0.024408312514424324, + 0.013515235856175423, + 0.028330843895673752, + 0.01288302056491375, + 0.011712216772139072, + 0.00871342234313488, + 0.009059097617864609, + 0.018152665346860886, + 0.012018774636089802, + 0.005986509378999472, + 0.008797631599009037, + 0.016290616244077682, + 0.011107531376183033, + 0.030245359987020493, + 0.017094504088163376, + 0.010158540681004524, + 0.00949569046497345, + 0.01629161648452282, + 0.014178821817040443, + 0.010070634074509144, + 0.011167422868311405, + 0.012459782883524895, + 0.016601353883743286, + 0.01007696520537138, + 0.013355094939470291, + 0.011755095794796944, + 0.005543034989386797 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 8 + ], + "selected_reward": -0.016507476568222046 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a heavily modified matte black truck?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the truck performing a wheelie stunt with its front end lifted high and almost perpendicular to the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a rusted metal skeleton of a massive abandoned satellite receiving station visible in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene take place in a vast desert with billowing golden dust kicked up by the truck's rear wheels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot filmed from a low angle with a high-speed camera following the truck?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014219158329069614 + ], + "opposite_logprob_tokens": [ + -4.264219284057617 + ], + "logprob_reward": -0.014219158329069614, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014219158329069614, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1973789632320404, + 0.04417992755770683, + 0.03453269973397255, + 0.0956946611404419, + 0.09534305334091187, + 0.07114214450120926, + 0.06843987852334976, + 0.2256881296634674, + 0.12165780365467072, + 0.045942723751068115 + ], + "credit_map": [ + 0.031812723726034164, + 0.02262502908706665, + 0.009930932894349098, + 0.017019130289554596, + 0.027433834969997406, + 0.03745115175843239, + 0.03222644701600075, + 0.018879709765315056, + 0.009962721727788448, + 0.006544190924614668, + 0.0027589257806539536, + 0.004927198402583599, + 0.005124641116708517, + 0.005472433753311634, + 0.005574547685682774, + 0.003815269097685814, + 0.005173176061362028, + 0.004470741841942072, + 0.0036309286952018738, + 0.006796258967369795, + 0.0051803793758153915, + 0.0035219970159232616, + 0.0031253602355718613, + 0.0026338594034314156, + 0.007848169654607773, + 0.006061546504497528, + 0.010318404994904995, + 0.009398717433214188, + 0.016183100640773773, + 0.010102706961333752, + 0.02095934934914112, + 0.014822661876678467, + 0.031908292323350906, + 0.008624166250228882, + 0.0349455364048481, + 0.005727757234126329, + 0.0032672358211129904, + 0.0034452336840331554, + 0.0031170635484158993, + 0.004307773895561695, + 0.005737373139709234, + 0.009629367850720882, + 0.004881526809185743, + 0.026855429634451866, + 0.003428063355386257, + 0.003773102769628167, + 0.012869263999164104, + 0.003968025557696819, + 0.004573673475533724, + 0.01022361870855093, + 0.00962871965020895, + 0.009846003726124763, + 0.009539392776787281, + 0.00943987350910902, + 0.005324605852365494, + 0.009863986633718014, + 0.031822338700294495, + 0.02176641672849655, + 0.039917320013046265, + 0.023974059149622917, + 0.02700250782072544, + 0.015505407936871052, + 0.01780029386281967, + 0.04789978638291359, + 0.016768470406532288, + 0.012406613677740097, + 0.027508916333317757, + 0.015314947813749313, + 0.013114799745380878, + 0.010431688278913498, + 0.011812134645879269, + 0.014300231821835041, + 0.004617581609636545, + 0.005089010577648878, + 0.0077706631273031235, + 0.008350453339517117, + 0.0037616780027747154, + 0.007334332913160324, + 0.0048932465724647045, + 0.00412575900554657 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 8 + ], + "selected_reward": -0.014219158329069614 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a heavily modified matte black truck?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the truck performing a wheelie stunt with its front end lifted high and almost perpendicular to the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a rusted metal skeleton of a massive abandoned satellite receiving station visible in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene take place in a vast desert with billowing golden dust kicked up by the truck's rear wheels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot filmed from a low angle with a high-speed camera following the truck?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012537825852632523 + ], + "opposite_logprob_tokens": [ + -4.387537956237793 + ], + "logprob_reward": -0.012537825852632523, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012537825852632523, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16515475511550903, + 0.14645232260227203, + 0.1545897275209427, + 0.06671197712421417, + 0.13231155276298523, + 0.025998760014772415, + 0.08273845911026001, + 0.1385262906551361, + 0.030117884278297424, + 0.05739828199148178 + ], + "credit_map": [ + 0.03069062903523445, + 0.02575516328215599, + 0.012746656313538551, + 0.011831443756818771, + 0.01870514079928398, + 0.02085254341363907, + 0.01598002202808857, + 0.028593163937330246, + 0.028264470398426056, + 0.02311924658715725, + 0.010070252232253551, + 0.018235983327031136, + 0.016303518787026405, + 0.01740763522684574, + 0.01886753737926483, + 0.014183668419718742, + 0.022658908739686012, + 0.017084326595067978, + 0.014101723209023476, + 0.02668812684714794, + 0.03514223173260689, + 0.01459430530667305, + 0.010375394485890865, + 0.013944700360298157, + 0.009040682576596737, + 0.009101387113332748, + 0.009626981802284718, + 0.005601946264505386, + 0.008334430865943432, + 0.009618834592401981, + 0.009021787904202938, + 0.006365926470607519, + 0.03616452217102051, + 0.01612147130072117, + 0.038511972874403, + 0.011678098700940609, + 0.009189172647893429, + 0.0072527918964624405, + 0.006715401075780392, + 0.006678119767457247, + 0.0037036172579973936, + 0.003064317861571908, + 0.00234258477576077, + 0.0024064169265329838, + 0.005240635015070438, + 0.0032847875263541937, + 0.003282198216766119, + 0.002674202434718609, + 0.005107332020998001, + 0.008299756795167923, + 0.012051324360072613, + 0.010683203116059303, + 0.012120872735977173, + 0.014882029965519905, + 0.006841430906206369, + 0.012752506881952286, + 0.023220451548695564, + 0.01282353512942791, + 0.017937755212187767, + 0.020886873826384544, + 0.019619956612586975, + 0.01574387401342392, + 0.010682755149900913, + 0.01761108823120594, + 0.004299030173569918, + 0.0033836604561656713, + 0.008372055366635323, + 0.004235296044498682, + 0.002057854551821947, + 0.0018192187417298555, + 0.002858969382941723, + 0.003091799793764949, + 0.005368490237742662, + 0.008278971537947655, + 0.011459572240710258, + 0.013436884619295597, + 0.003671986050903797, + 0.005615217611193657, + 0.005327184684574604, + 0.004239975940436125 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 1 + ], + "selected_reward": -0.012537825852632523 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a heavily modified matte black truck?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the truck performing a wheelie stunt with its front end lifted high and almost perpendicular to the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a rusted metal skeleton of a massive abandoned satellite receiving station visible in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene take place in a vast desert with billowing golden dust kicked up by the truck's rear wheels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot filmed from a low angle with a high-speed camera following the truck?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0007172634359449148 + ], + "opposite_logprob_tokens": [ + -7.2507171630859375 + ], + "logprob_reward": -0.0007172634359449148, + "yes_no_margin_reward": 7.25, + "token_credit_reward": -0.0007172634359449148, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08949687331914902, + 0.01998303271830082, + 0.03513197973370552, + 0.02118644304573536, + 0.09951327741146088, + 0.07625721395015717, + 0.1401691436767578, + 0.09680278599262238, + 0.1935737580060959, + 0.22788552939891815 + ], + "credit_map": [ + 0.016641857102513313, + 0.012916206382215023, + 0.006783837452530861, + 0.006776436232030392, + 0.01103594433516264, + 0.012663006782531738, + 0.008488326333463192, + 0.014191255904734135, + 0.0032269316725432873, + 0.0025187262799590826, + 0.0017981190467253327, + 0.0025290248449891806, + 0.002194470725953579, + 0.0023783270735293627, + 0.0030623062048107386, + 0.002275124890729785, + 0.004444696940481663, + 0.0034893639385700226, + 0.0032559309620410204, + 0.006369677372276783, + 0.007123662158846855, + 0.00314992549829185, + 0.003706389805302024, + 0.003592332359403372, + 0.003017377108335495, + 0.002616662997752428, + 0.002526482567191124, + 0.002017420018091798, + 0.0029203372541815042, + 0.00236701313406229, + 0.0033171772956848145, + 0.0024039712734520435, + 0.0320613794028759, + 0.008371902629733086, + 0.03468847647309303, + 0.006351841613650322, + 0.005414543207734823, + 0.0041936966590583324, + 0.003772786119952798, + 0.0046586389653384686, + 0.006592043209820986, + 0.010336616076529026, + 0.004869736731052399, + 0.028161045163869858, + 0.003960882313549519, + 0.0046310387551784515, + 0.013934212736785412, + 0.0037716366350650787, + 0.011729047633707523, + 0.017277294769883156, + 0.011503881774842739, + 0.01959964819252491, + 0.026992913335561752, + 0.022267168387770653, + 0.012566562741994858, + 0.018232624977827072, + 0.010526703670620918, + 0.010146214626729488, + 0.015912476927042007, + 0.014920098707079887, + 0.016890404745936394, + 0.007421502377837896, + 0.006007055751979351, + 0.014978322200477123, + 0.030603865161538124, + 0.018151530995965004, + 0.057946138083934784, + 0.026613423600792885, + 0.011941281147301197, + 0.011231007985770702, + 0.015653522685170174, + 0.021432986482977867, + 0.026803191751241684, + 0.03189175948500633, + 0.04664251208305359, + 0.05865052342414856, + 0.011980226263403893, + 0.017601098865270615, + 0.02172533981502056, + 0.012590857222676277 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 6 + ], + "selected_reward": -0.0007172634359449148 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a heavily modified matte black truck?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the truck performing a wheelie stunt with its front end lifted high and almost perpendicular to the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a rusted metal skeleton of a massive abandoned satellite receiving station visible in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene take place in a vast desert with billowing golden dust kicked up by the truck's rear wheels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot filmed from a low angle with a high-speed camera following the truck?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00525581045076251 + ], + "opposite_logprob_tokens": [ + -5.255255699157715 + ], + "logprob_reward": -0.00525581045076251, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.00525581045076251, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10048802942037582, + 0.04695441946387291, + 0.031264595687389374, + 0.048395294696092606, + 0.10110242664813995, + 0.1484634131193161, + 0.09422548115253448, + 0.09073836356401443, + 0.1792822629213333, + 0.15908578038215637 + ], + "credit_map": [ + 0.01393685769289732, + 0.011758527718484402, + 0.006435702554881573, + 0.008686486631631851, + 0.014548360370099545, + 0.01312723197042942, + 0.019398916512727737, + 0.012595935724675655, + 0.009001954458653927, + 0.0052233426831662655, + 0.0040194750763475895, + 0.005244499538093805, + 0.005616689100861549, + 0.005306655541062355, + 0.007342720404267311, + 0.005199075676500797, + 0.003725295653566718, + 0.003228133777156472, + 0.005164762958884239, + 0.00430634617805481, + 0.00441930815577507, + 0.002713874913752079, + 0.0037427295465022326, + 0.003964138682931662, + 0.0068336352705955505, + 0.0055642761290073395, + 0.0072466651909053326, + 0.005099375732243061, + 0.005406889133155346, + 0.007218159269541502, + 0.004929859656840563, + 0.006096424534916878, + 0.015049319714307785, + 0.019730277359485626, + 0.016207221895456314, + 0.014234923757612705, + 0.011178004555404186, + 0.010023931972682476, + 0.007863902486860752, + 0.006814825348556042, + 0.011950411833822727, + 0.014721007086336613, + 0.010921658016741276, + 0.02617485821247101, + 0.01747455634176731, + 0.01344601996243, + 0.0401127003133297, + 0.013662188313901424, + 0.007411561906337738, + 0.009848416782915592, + 0.012932994402945042, + 0.010577745735645294, + 0.014196704141795635, + 0.014006607234477997, + 0.008027919568121433, + 0.01722351275384426, + 0.00646484037861228, + 0.009292161092162132, + 0.01427379809319973, + 0.012081737630069256, + 0.019183583557605743, + 0.009822696447372437, + 0.006855295971035957, + 0.01276423316448927, + 0.026561850681900978, + 0.01221313513815403, + 0.049629613757133484, + 0.02571914531290531, + 0.011909565888345242, + 0.009163321927189827, + 0.022402819246053696, + 0.02168278396129608, + 0.017724137753248215, + 0.023674679920077324, + 0.035664718598127365, + 0.03827940672636032, + 0.007121676579117775, + 0.009229164570569992, + 0.013595564290881157, + 0.013796414248645306 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 5 + ], + "selected_reward": -0.00525581045076251 + } + ], + "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/0000156/group_5/index_tensor([9911, 9911])_group_5_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000156/group_5/index_tensor([9911, 9911])_group_5_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..fbafb705b45cf2c3701d4171de1e57182380e792 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000156/group_5/index_tensor([9911, 9911])_group_5_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "This stunning long sequence, rendered in a captivating Pixar style, begins with the camera gliding smoothly in a slow, sweeping aerial shot, following the magnificent, ancient stone fortifications as they majestically wind through a lush, rolling green landscape. The walls themselves are constructed from massive, weathered gray stones, each with softened edges and unique textures, dotted with a soft, velvety moss. Sturdy cylindrical watchtowers with conical roofs of warm red tiles punctuate the length of the walls, their simple forms yet heroic presence against the horizon. The entire scene is bathed in the warm golden glow of the afternoon sun, casting long, soft shadows and highlighting the gentle contours of the hills. In the distance, the azure sky is adorned with a few perfectly fluffy, stylized white clouds. The camera's steady, graceful movement along the fortifications creates a sense of peaceful grandeur and epic adventure, inviting the viewer into a world of whimsical history and breathtaking beauty.", + "seed": 28656, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000156/group_5/index_9911_seed_28656_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": -0.03437331853783689, + "vlm_logprob": -0.03437331853783689, + "vlm_yes_no_margin": 5.125, + "vlm_token_credit": -0.03437331853783689, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a Pixar-style animation aesthetic?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are massive weathered gray stone walls with softened edges and moss visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are sturdy cylindrical watchtowers with conical red tiled roofs present along the walls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set in a lush rolling green landscape under a warm golden afternoon sun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a slow sweeping aerial shot gliding along the fortifications?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07747498154640198 + ], + "opposite_logprob_tokens": [ + -2.702475070953369 + ], + "logprob_reward": -0.07747498154640198, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07747498154640198, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08387600630521774, + 0.018583105877041817, + 0.054880525916814804, + 0.05138279125094414, + 0.13869857788085938, + 0.11747702211141586, + 0.1017860472202301, + 0.1544482260942459, + 0.05335664376616478, + 0.22551105916500092 + ], + "credit_map": [ + 0.008637483231723309, + 0.00868726521730423, + 0.006683274172246456, + 0.014667441137135029, + 0.011302681639790535, + 0.008725912310183048, + 0.01666446402668953, + 0.00850748736411333, + 0.0015511694364249706, + 0.002363222651183605, + 0.0020602899603545666, + 0.0035309898667037487, + 0.002659638412296772, + 0.001579012256115675, + 0.0027210162952542305, + 0.002117767231538892, + 0.004902568645775318, + 0.005682763643562794, + 0.011400114744901657, + 0.006961342412978411, + 0.007023491431027651, + 0.004734794143587351, + 0.00924267154186964, + 0.00493277981877327, + 0.008892390877008438, + 0.004305593203753233, + 0.00693888682872057, + 0.010420160368084908, + 0.005116336978971958, + 0.004729738458991051, + 0.007574585732072592, + 0.003405100665986538, + 0.020893914625048637, + 0.014343120157718658, + 0.028309304267168045, + 0.017795616760849953, + 0.01104295626282692, + 0.01009751670062542, + 0.023117709904909134, + 0.013098432682454586, + 0.02591431327164173, + 0.009454675950109959, + 0.01828618533909321, + 0.012694067321717739, + 0.0056088767014443874, + 0.006972332485020161, + 0.013765105046331882, + 0.02478145994246006, + 0.007458588574081659, + 0.009324703365564346, + 0.015581842511892319, + 0.01541974488645792, + 0.010381583124399185, + 0.008880781941115856, + 0.019225366413593292, + 0.015513437800109386, + 0.006576689891517162, + 0.010634100995957851, + 0.017227526754140854, + 0.012849515303969383, + 0.01901574432849884, + 0.033652614802122116, + 0.024650586768984795, + 0.02984144724905491, + 0.0038289502263069153, + 0.0032931105233728886, + 0.003701664973050356, + 0.017991814762353897, + 0.0024209031835198402, + 0.0021318327635526657, + 0.013030952773988247, + 0.006957418285310268, + 0.00815320760011673, + 0.009839940816164017, + 0.03538113459944725, + 0.08043939620256424, + 0.009074495173990726, + 0.007830760441720486, + 0.04933406785130501, + 0.025458041578531265 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 7, + 4 + ], + "selected_reward": -0.07747498154640198 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a Pixar-style animation aesthetic?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are massive weathered gray stone walls with softened edges and moss visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are sturdy cylindrical watchtowers with conical red tiled roofs present along the walls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set in a lush rolling green landscape under a warm golden afternoon sun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a slow sweeping aerial shot gliding along the fortifications?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08907752484083176 + ], + "opposite_logprob_tokens": [ + -2.4640774726867676 + ], + "logprob_reward": -0.08907752484083176, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08907752484083176, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06849654763936996, + 0.0217344518750906, + 0.07443852722644806, + 0.04074031487107277, + 0.2448476254940033, + 0.04777786135673523, + 0.16390906274318695, + 0.16635440289974213, + 0.06221002712845802, + 0.10949119180440903 + ], + "credit_map": [ + 0.005527343600988388, + 0.006511660758405924, + 0.003897295566275716, + 0.01407338585704565, + 0.007156328298151493, + 0.006427236832678318, + 0.017152726650238037, + 0.00775056192651391, + 0.0014538511168211699, + 0.002350315684452653, + 0.002102350816130638, + 0.004154629539698362, + 0.0031124865636229515, + 0.0030127926729619503, + 0.0029960472602397203, + 0.0025519777555018663, + 0.006755832117050886, + 0.010693809948861599, + 0.009145048446953297, + 0.01150758657604456, + 0.009984069503843784, + 0.007708131801337004, + 0.010556472465395927, + 0.008087568916380405, + 0.0046271090395748615, + 0.004554815124720335, + 0.004537614062428474, + 0.004118111915886402, + 0.00575510086491704, + 0.007163939066231251, + 0.007239112164825201, + 0.002744507510215044, + 0.024964340031147003, + 0.016543123871088028, + 0.08742770552635193, + 0.031409118324518204, + 0.010170594789087772, + 0.01092720776796341, + 0.03594934940338135, + 0.027456171810626984, + 0.00953080877661705, + 0.004239568021148443, + 0.005237332079559565, + 0.004447545390576124, + 0.0036794773768633604, + 0.0033151113893836737, + 0.006697382312268019, + 0.010630634613335133, + 0.025218553841114044, + 0.014788195490837097, + 0.019380657002329826, + 0.021470168605446815, + 0.009131962433457375, + 0.010438026860356331, + 0.0322786383330822, + 0.031202860176563263, + 0.011738753877580166, + 0.011657210998237133, + 0.01865905523300171, + 0.02060248889029026, + 0.012366898357868195, + 0.015537477098405361, + 0.04234074801206589, + 0.03345175459980965, + 0.007775274571031332, + 0.0059303706511855125, + 0.005538887344300747, + 0.0097551504150033, + 0.005994662642478943, + 0.004414210561662912, + 0.012808462604880333, + 0.009993002749979496, + 0.006500849965959787, + 0.007774182129651308, + 0.018328651785850525, + 0.028909830376505852, + 0.005341114476323128, + 0.004391443450003862, + 0.026080811396241188, + 0.012164299376308918 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 7, + 6 + ], + "selected_reward": -0.08907752484083176 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a Pixar-style animation aesthetic?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are massive weathered gray stone walls with softened edges and moss visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are sturdy cylindrical watchtowers with conical red tiled roofs present along the walls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set in a lush rolling green landscape under a warm golden afternoon sun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a slow sweeping aerial shot gliding along the fortifications?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0003045333724003285 + ], + "opposite_logprob_tokens": [ + -8.125304222106934 + ], + "logprob_reward": -0.0003045333724003285, + "yes_no_margin_reward": 8.125, + "token_credit_reward": -0.0003045333724003285, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0995292216539383, + 0.03354829549789429, + 0.14271509647369385, + 0.03700021654367447, + 0.15603230893611908, + 0.04327882081270218, + 0.05983167886734009, + 0.23159970343112946, + 0.10133890062570572, + 0.0951257273554802 + ], + "credit_map": [ + 0.007786629255861044, + 0.010390715673565865, + 0.004563953727483749, + 0.018507439643144608, + 0.01844092644751072, + 0.010606917552649975, + 0.019035683944821358, + 0.010196961462497711, + 0.0033327986020594835, + 0.004052113275974989, + 0.002768054837360978, + 0.004249112214893103, + 0.00604633241891861, + 0.002833063481375575, + 0.004814488347619772, + 0.005452335812151432, + 0.014779296703636646, + 0.021968185901641846, + 0.014216450043022633, + 0.01973516307771206, + 0.023086845874786377, + 0.017164230346679688, + 0.01923958957195282, + 0.012525358237326145, + 0.0033216034062206745, + 0.002723495941609144, + 0.004984189756214619, + 0.004582674708217382, + 0.005780229344964027, + 0.002936489647254348, + 0.00971267744898796, + 0.0029588602483272552, + 0.01478007435798645, + 0.0102236308157444, + 0.05568976327776909, + 0.019630039110779762, + 0.005821120459586382, + 0.007119360845535994, + 0.02418525330722332, + 0.018583083525300026, + 0.006621577311307192, + 0.0040090265683829784, + 0.005577579140663147, + 0.004379913676530123, + 0.003917979542165995, + 0.003469930961728096, + 0.008469204418361187, + 0.00683361291885376, + 0.005996036343276501, + 0.006252908147871494, + 0.008630719035863876, + 0.008308777585625648, + 0.004239640664309263, + 0.0056618028320372105, + 0.011512543074786663, + 0.009229259565472603, + 0.012523987330496311, + 0.0189836248755455, + 0.018086742609739304, + 0.019941171631217003, + 0.030152536928653717, + 0.04144096374511719, + 0.04607255756855011, + 0.04439812898635864, + 0.0076302094385027885, + 0.005775313824415207, + 0.006480567157268524, + 0.03281249478459358, + 0.004988783970475197, + 0.004975351970642805, + 0.026345958933234215, + 0.012330230325460434, + 0.005945796146988869, + 0.0052928985096514225, + 0.016009951010346413, + 0.027336593717336655, + 0.004472642671316862, + 0.003915554378181696, + 0.021985359489917755, + 0.010166940279304981 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 4, + 2 + ], + "selected_reward": -0.0003045333724003285 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a Pixar-style animation aesthetic?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are massive weathered gray stone walls with softened edges and moss visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are sturdy cylindrical watchtowers with conical red tiled roofs present along the walls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set in a lush rolling green landscape under a warm golden afternoon sun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a slow sweeping aerial shot gliding along the fortifications?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0009189196862280369 + ], + "opposite_logprob_tokens": [ + -7.000918865203857 + ], + "logprob_reward": -0.0009189196862280369, + "yes_no_margin_reward": 7.0, + "token_credit_reward": -0.0009189196862280369, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05257124453783035, + 0.028200017288327217, + 0.03058374859392643, + 0.020645080134272575, + 0.056692734360694885, + 0.20206207036972046, + 0.2296261489391327, + 0.07548879832029343, + 0.06257356703281403, + 0.2415565848350525 + ], + "credit_map": [ + 0.007848551496863365, + 0.008565704338252544, + 0.002714981324970722, + 0.004994501359760761, + 0.007650234270840883, + 0.007633007597178221, + 0.006076245103031397, + 0.007088015787303448, + 0.003058814210817218, + 0.006235507782548666, + 0.0025233954656869173, + 0.004470545332878828, + 0.0034007178619503975, + 0.0019531074212864041, + 0.004122510552406311, + 0.0024354159832000732, + 0.003967741504311562, + 0.0054567549377679825, + 0.004304397851228714, + 0.00351118971593678, + 0.002955240197479725, + 0.0025975473690778017, + 0.004447761923074722, + 0.0033431141637265682, + 0.0030981730669736862, + 0.0028243684209883213, + 0.002797584980726242, + 0.004188837483525276, + 0.0014146285830065608, + 0.0010624006390571594, + 0.003383657895028591, + 0.0018754308111965656, + 0.0050443001091480255, + 0.003336208639666438, + 0.021464986726641655, + 0.008582581765949726, + 0.0027749615255743265, + 0.0019258821848779917, + 0.008697118610143661, + 0.004866693168878555, + 0.04822544753551483, + 0.01759667508304119, + 0.02130839042365551, + 0.016373060643672943, + 0.01061362773180008, + 0.014095387421548367, + 0.024890150874853134, + 0.04895932227373123, + 0.02111656963825226, + 0.009566240012645721, + 0.01725434511899948, + 0.009216273203492165, + 0.009853191673755646, + 0.016648443415760994, + 0.07055166363716125, + 0.07541941851377487, + 0.0048567368648946285, + 0.005848736502230167, + 0.009559434838593006, + 0.008235675282776356, + 0.00849451869726181, + 0.012021726928651333, + 0.012096961960196495, + 0.014375008642673492, + 0.0042021619156003, + 0.003342523705214262, + 0.003168218769133091, + 0.017691174522042274, + 0.0037135861348360777, + 0.0029374677687883377, + 0.0176708921790123, + 0.009847540408372879, + 0.012517530471086502, + 0.020399892702698708, + 0.03751020133495331, + 0.06792862713336945, + 0.018062401562929153, + 0.01815991848707199, + 0.04177248850464821, + 0.02520553395152092 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 6, + 5 + ], + "selected_reward": -0.0009189196862280369 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a Pixar-style animation aesthetic?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are massive weathered gray stone walls with softened edges and moss visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are sturdy cylindrical watchtowers with conical red tiled roofs present along the walls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set in a lush rolling green landscape under a warm golden afternoon sun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a slow sweeping aerial shot gliding along the fortifications?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0040906332433223724 + ], + "opposite_logprob_tokens": [ + -5.504090785980225 + ], + "logprob_reward": -0.0040906332433223724, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.0040906332433223724, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07215812802314758, + 0.052342575043439865, + 0.05090976879000664, + 0.03160114958882332, + 0.1273036152124405, + 0.07484975457191467, + 0.05434400960803032, + 0.11100905388593674, + 0.13142292201519012, + 0.29405897855758667 + ], + "credit_map": [ + 0.006232762709259987, + 0.008020784705877304, + 0.007442301604896784, + 0.011724027805030346, + 0.010518884286284447, + 0.007880149409174919, + 0.01109339203685522, + 0.009245835244655609, + 0.004518670961260796, + 0.005773533135652542, + 0.00668885326012969, + 0.011462227441370487, + 0.005239115096628666, + 0.003969225101172924, + 0.009202229790389538, + 0.005488724447786808, + 0.0047297351993620396, + 0.007487324066460133, + 0.008133180439472198, + 0.005567333661019802, + 0.005550692323595285, + 0.0050710695795714855, + 0.008771196007728577, + 0.005599241703748703, + 0.00373577862046659, + 0.00368138263002038, + 0.004147191997617483, + 0.005640214309096336, + 0.003885007230564952, + 0.0023560707923024893, + 0.0055201551876962185, + 0.002635349752381444, + 0.012952521443367004, + 0.008745711296796799, + 0.04666312411427498, + 0.019393494352698326, + 0.004803669173270464, + 0.004009577911347151, + 0.020564408972859383, + 0.010171111673116684, + 0.018068909645080566, + 0.006713981740176678, + 0.008949468843638897, + 0.0075781699270009995, + 0.00390971964225173, + 0.005003991536796093, + 0.008408854715526104, + 0.016216665506362915, + 0.004790615756064653, + 0.004636676516383886, + 0.007037373725324869, + 0.005310714710503817, + 0.0031103675719350576, + 0.004424210637807846, + 0.012881437316536903, + 0.012152619659900665, + 0.007598710712045431, + 0.010656371712684631, + 0.011229380033910275, + 0.01409622747451067, + 0.012529575265944004, + 0.015503255650401115, + 0.01902693323791027, + 0.020368607714772224, + 0.015180704183876514, + 0.007815993390977383, + 0.00914558582007885, + 0.02866784669458866, + 0.008489003404974937, + 0.007986952550709248, + 0.030219659209251404, + 0.023917194455862045, + 0.010286294855177402, + 0.01269838958978653, + 0.04678087309002876, + 0.11297372728586197, + 0.00975564494729042, + 0.007259901612997055, + 0.061533961445093155, + 0.032770201563835144 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 4 + ], + "selected_reward": -0.0040906332433223724 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "This stunning long sequence, rendered in a captivating Pixar style, begins with the camera gliding smoothly in a slow, sweeping aerial shot, following the magnificent, ancient stone fortifications as they majestically wind through a lush, rolling green landscape. The walls themselves are constructed from massive, weathered gray stones, each with softened edges and unique textures, dotted with a soft, velvety moss. Sturdy cylindrical watchtowers with conical roofs of warm red tiles punctuate the length of the walls, their simple forms yet heroic presence against the horizon. The entire scene is bathed in the warm golden glow of the afternoon sun, casting long, soft shadows and highlighting the gentle contours of the hills. In the distance, the azure sky is adorned with a few perfectly fluffy, stylized white clouds. The camera's steady, graceful movement along the fortifications creates a sense of peaceful grandeur and epic adventure, inviting the viewer into a world of whimsical history and breathtaking beauty.", + "seed": 28657, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000156/group_5/index_9911_seed_28657_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": -0.017091004992835224, + "vlm_logprob": -0.017091004992835224, + "vlm_yes_no_margin": 5.475, + "vlm_token_credit": -0.017091004992835224, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a Pixar-style animation aesthetic?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are massive weathered gray stone walls with softened edges and moss visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are sturdy cylindrical watchtowers with conical red tiled roofs present along the walls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set in a lush rolling green landscape under a warm golden afternoon sun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a slow sweeping aerial shot gliding along the fortifications?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07047808170318604 + ], + "opposite_logprob_tokens": [ + -2.8204779624938965 + ], + "logprob_reward": -0.07047808170318604, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.07047808170318604, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.32901522517204285, + 0.048959407955408096, + 0.048416074365377426, + 0.04777573049068451, + 0.04863882064819336, + 0.02258336916565895, + 0.028141355141997337, + 0.13601168990135193, + 0.24315033853054047, + 0.04730789735913277 + ], + "credit_map": [ + 0.04437974467873573, + 0.027844388037919998, + 0.06113395467400551, + 0.03338078781962395, + 0.029532339423894882, + 0.028897173702716827, + 0.048051219433546066, + 0.055795639753341675, + 0.0037071481347084045, + 0.002617965918034315, + 0.006259121932089329, + 0.00390612636692822, + 0.004524220712482929, + 0.0028518391773104668, + 0.012915607541799545, + 0.012177382595837116, + 0.0031016303692013025, + 0.003207228146493435, + 0.0076739429496228695, + 0.00625378405675292, + 0.0026090736500918865, + 0.004612505901604891, + 0.0157021414488554, + 0.005255772732198238, + 0.004207899793982506, + 0.004801936447620392, + 0.006673117633908987, + 0.006539661902934313, + 0.004808348603546619, + 0.005516654346138239, + 0.007824287749826908, + 0.007403827738016844, + 0.002659672638401389, + 0.0038404082879424095, + 0.006666339468210936, + 0.010330897755920887, + 0.0037559557240456343, + 0.006360322702676058, + 0.009430699981749058, + 0.0055945273488759995, + 0.0022517736069858074, + 0.002264134120196104, + 0.0029104314744472504, + 0.0031042806804180145, + 0.0019261043053120375, + 0.0029656486585736275, + 0.004160900600254536, + 0.003000096883624792, + 0.0022333210799843073, + 0.0027282137889415026, + 0.0035815928131341934, + 0.0038978327065706253, + 0.0023881320375949144, + 0.00417191069573164, + 0.004771160427480936, + 0.004369194153696299, + 0.008967610076069832, + 0.015032847411930561, + 0.01817963644862175, + 0.011333538219332695, + 0.01936866156756878, + 0.018873518332839012, + 0.014651834964752197, + 0.029604053124785423, + 0.02755485661327839, + 0.02446586824953556, + 0.026443159207701683, + 0.020164290443062782, + 0.02735474333167076, + 0.05365736037492752, + 0.024957962334156036, + 0.03855210170149803, + 0.006096140947192907, + 0.003477697027847171, + 0.006360936909914017, + 0.005078671965748072, + 0.004162800032645464, + 0.006887323688715696, + 0.005241635255515575, + 0.010002694092690945 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 7 + ], + "selected_reward": -0.07047808170318604 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a Pixar-style animation aesthetic?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are massive weathered gray stone walls with softened edges and moss visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are sturdy cylindrical watchtowers with conical red tiled roofs present along the walls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set in a lush rolling green landscape under a warm golden afternoon sun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a slow sweeping aerial shot gliding along the fortifications?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007660536095499992 + ], + "opposite_logprob_tokens": [ + -4.882660388946533 + ], + "logprob_reward": -0.007660536095499992, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007660536095499992, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.29634836316108704, + 0.13936300575733185, + 0.12080176919698715, + 0.03647252172231674, + 0.017955634742975235, + 0.044951438903808594, + 0.031771473586559296, + 0.19947239756584167, + 0.07826191186904907, + 0.03460146486759186 + ], + "credit_map": [ + 0.03270649537444115, + 0.023382844403386116, + 0.03909926488995552, + 0.03258449211716652, + 0.048627182841300964, + 0.023043259978294373, + 0.06140702962875366, + 0.03549778461456299, + 0.010126729495823383, + 0.006987452041357756, + 0.014544899575412273, + 0.007368257734924555, + 0.012856269255280495, + 0.007404677104204893, + 0.043426334857940674, + 0.03664839640259743, + 0.008701203390955925, + 0.010464055463671684, + 0.018152492120862007, + 0.018917201086878777, + 0.008756226859986782, + 0.010968323796987534, + 0.030883558094501495, + 0.013958710245788097, + 0.002460488583892584, + 0.002034858101978898, + 0.006160947028547525, + 0.0029781449120491743, + 0.0028441976755857468, + 0.0022335359826684, + 0.01113129872828722, + 0.006629051174968481, + 0.0009852258954197168, + 0.001909859711304307, + 0.002306348644196987, + 0.003870559623464942, + 0.0013925882522016764, + 0.0022306335158646107, + 0.002739502349868417, + 0.002520916983485222, + 0.0027622065972536802, + 0.003573052119463682, + 0.004869251511991024, + 0.006201182492077351, + 0.0059050340205430984, + 0.006593456957489252, + 0.00987927708774805, + 0.0051679788157343864, + 0.002477527130395174, + 0.0038936370983719826, + 0.004709120839834213, + 0.00585583783686161, + 0.0025009606033563614, + 0.0042665558867156506, + 0.004725106991827488, + 0.0033427297603338957, + 0.014730376191437244, + 0.0248358566313982, + 0.029702330008149147, + 0.014046766795217991, + 0.02675694227218628, + 0.03176169842481613, + 0.01798309199512005, + 0.03965533524751663, + 0.006772077176719904, + 0.006670274306088686, + 0.008284665644168854, + 0.008348909206688404, + 0.008819928392767906, + 0.015001369640231133, + 0.012591168284416199, + 0.01177352387458086, + 0.002846628660336137, + 0.0033740305807441473, + 0.00425131618976593, + 0.0043940916657447815, + 0.003945923410356045, + 0.002710386412218213, + 0.006905886810272932, + 0.006173199974000454 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 1 + ], + "selected_reward": -0.007660536095499992 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a Pixar-style animation aesthetic?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are massive weathered gray stone walls with softened edges and moss visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are sturdy cylindrical watchtowers with conical red tiled roofs present along the walls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set in a lush rolling green landscape under a warm golden afternoon sun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a slow sweeping aerial shot gliding along the fortifications?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0005616756388917565 + ], + "opposite_logprob_tokens": [ + -7.500561714172363 + ], + "logprob_reward": -0.0005616756388917565, + "yes_no_margin_reward": 7.5, + "token_credit_reward": -0.0005616756388917565, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18991176784038544, + 0.08271872252225876, + 0.14098088443279266, + 0.1137220486998558, + 0.15539683401584625, + 0.10634443163871765, + 0.06039457023143768, + 0.032040566205978394, + 0.09934822469949722, + 0.01914193667471409 + ], + "credit_map": [ + 0.017213277518749237, + 0.016122570261359215, + 0.023169951513409615, + 0.020193103700876236, + 0.030674979090690613, + 0.021709444001317024, + 0.03351800888776779, + 0.027310431003570557, + 0.005379846319556236, + 0.0044038789346814156, + 0.007969414815306664, + 0.005413684528321028, + 0.011825491674244404, + 0.005431447643786669, + 0.023051658645272255, + 0.019243305549025536, + 0.008699620142579079, + 0.012539060786366463, + 0.021484483033418655, + 0.02993990108370781, + 0.009521015919744968, + 0.0114906532689929, + 0.03194821998476982, + 0.015357943251729012, + 0.009386472404003143, + 0.007610645145177841, + 0.021628957241773605, + 0.01255283784121275, + 0.009035876952111721, + 0.010799692943692207, + 0.0263691209256649, + 0.016338462010025978, + 0.012282381765544415, + 0.01100122183561325, + 0.014143136329948902, + 0.017412874847650528, + 0.01847068779170513, + 0.043768059462308884, + 0.017380241304636, + 0.020938249304890633, + 0.010638921521604061, + 0.009331419132649899, + 0.00782561581581831, + 0.008732592687010765, + 0.010952416807413101, + 0.029320096597075462, + 0.012808366678655148, + 0.0167350135743618, + 0.005168448202311993, + 0.00656422134488821, + 0.007917510345578194, + 0.009211246855556965, + 0.008069993928074837, + 0.007969455793499947, + 0.00904445443302393, + 0.006449244916439056, + 0.0030296642798930407, + 0.003369109472259879, + 0.006107195746153593, + 0.00403390359133482, + 0.003138255327939987, + 0.0030873070936650038, + 0.0043805004097521305, + 0.00489463284611702, + 0.01073854323476553, + 0.010384747758507729, + 0.010388781316578388, + 0.007029409520328045, + 0.01110062561929226, + 0.020129183307290077, + 0.013731639832258224, + 0.015845295041799545, + 0.0016847287770360708, + 0.0028295109514147043, + 0.0023320314940065145, + 0.0024207914248108864, + 0.002076655626296997, + 0.0019534286111593246, + 0.0030120157171040773, + 0.0028327754698693752 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 2 + ], + "selected_reward": -0.0005616756388917565 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a Pixar-style animation aesthetic?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are massive weathered gray stone walls with softened edges and moss visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are sturdy cylindrical watchtowers with conical red tiled roofs present along the walls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set in a lush rolling green landscape under a warm golden afternoon sun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a slow sweeping aerial shot gliding along the fortifications?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0008116050157696009 + ], + "opposite_logprob_tokens": [ + -7.125811576843262 + ], + "logprob_reward": -0.0008116050157696009, + "yes_no_margin_reward": 7.125, + "token_credit_reward": -0.0008116050157696009, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.32120954990386963, + 0.07486709207296371, + 0.13096170127391815, + 0.0625620186328888, + 0.0357663668692112, + 0.04514003545045853, + 0.10508359968662262, + 0.04098609834909439, + 0.13877303898334503, + 0.04465049132704735 + ], + "credit_map": [ + 0.036395683884620667, + 0.023765243589878082, + 0.06917133182287216, + 0.04227806627750397, + 0.02538701705634594, + 0.0257854163646698, + 0.0427299365401268, + 0.055696845054626465, + 0.006260741967707872, + 0.004986910615116358, + 0.009779062122106552, + 0.006864291150122881, + 0.008968965150415897, + 0.0055207847617566586, + 0.018446970731019974, + 0.014039366506040096, + 0.009180724620819092, + 0.011563774198293686, + 0.021214505657553673, + 0.019369781017303467, + 0.008683552965521812, + 0.01003427803516388, + 0.03484724834561348, + 0.016067828983068466, + 0.006609791424125433, + 0.004239041358232498, + 0.014687437564134598, + 0.007115120533853769, + 0.003568868851289153, + 0.005338764749467373, + 0.01372155174612999, + 0.0072814407758414745, + 0.00247375201433897, + 0.0037231568712741137, + 0.0037818399723619223, + 0.003823402803391218, + 0.003552917391061783, + 0.009151258505880833, + 0.004813903942704201, + 0.0044461325742304325, + 0.003948159981518984, + 0.004974646028131247, + 0.0045736744068562984, + 0.0057071116752922535, + 0.005832751281559467, + 0.006082578562200069, + 0.008808211423456669, + 0.005212897900491953, + 0.007757185027003288, + 0.010228033177554607, + 0.013098020106554031, + 0.019354963675141335, + 0.014754971489310265, + 0.008449259214103222, + 0.017321353778243065, + 0.014119806699454784, + 0.0035034434404224157, + 0.0030432159546762705, + 0.005807945504784584, + 0.004091558512300253, + 0.005470894742757082, + 0.0068378145806491375, + 0.005206078290939331, + 0.007025145925581455, + 0.014153577387332916, + 0.013032264076173306, + 0.015029577538371086, + 0.01084402110427618, + 0.016425689682364464, + 0.030483609065413475, + 0.01622582972049713, + 0.022578461095690727, + 0.004924871493130922, + 0.004256049171090126, + 0.0062219081446528435, + 0.004912284202873707, + 0.0038676979020237923, + 0.006502148229628801, + 0.0042398362420499325, + 0.009725692681968212 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 2 + ], + "selected_reward": -0.0008116050157696009 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a Pixar-style animation aesthetic?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are massive weathered gray stone walls with softened edges and moss visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are sturdy cylindrical watchtowers with conical red tiled roofs present along the walls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set in a lush rolling green landscape under a warm golden afternoon sun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a slow sweeping aerial shot gliding along the fortifications?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0059431265108287334 + ], + "opposite_logprob_tokens": [ + -5.130943298339844 + ], + "logprob_reward": -0.0059431265108287334, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.0059431265108287334, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2242981344461441, + 0.052719611674547195, + 0.10123448818922043, + 0.06893014907836914, + 0.06932603567838669, + 0.06543248891830444, + 0.13705511391162872, + 0.17430850863456726, + 0.075291708111763, + 0.031403783708810806 + ], + "credit_map": [ + 0.01654115505516529, + 0.017130866646766663, + 0.028129244223237038, + 0.024308940395712852, + 0.034304846078157425, + 0.023852486163377762, + 0.047194939106702805, + 0.03283565118908882, + 0.003885335521772504, + 0.004967729561030865, + 0.00630587711930275, + 0.005531749222427607, + 0.009910723194479942, + 0.0047860704362392426, + 0.011427801102399826, + 0.0059043229557573795, + 0.008397279307246208, + 0.009985986165702343, + 0.015497570857405663, + 0.01441168412566185, + 0.008179991506040096, + 0.009222233667969704, + 0.024633606895804405, + 0.010906139388680458, + 0.006181760225445032, + 0.005832735449075699, + 0.010956563986837864, + 0.008055206388235092, + 0.006573953200131655, + 0.007930884137749672, + 0.01465135533362627, + 0.008747687563300133, + 0.003798185847699642, + 0.005358316004276276, + 0.007395231630653143, + 0.015327614732086658, + 0.005371593404561281, + 0.010776534676551819, + 0.012356055900454521, + 0.008942502550780773, + 0.005070172715932131, + 0.005690217018127441, + 0.00860336422920227, + 0.008453895337879658, + 0.008973861113190651, + 0.006698212586343288, + 0.014600259251892567, + 0.007342499680817127, + 0.009707671590149403, + 0.012650055810809135, + 0.018757464364171028, + 0.02713148109614849, + 0.017473990097641945, + 0.013694263063371181, + 0.022895393893122673, + 0.014744771644473076, + 0.014554694294929504, + 0.01976238563656807, + 0.021232271566987038, + 0.014555629342794418, + 0.025274472311139107, + 0.029053164646029472, + 0.01667287014424801, + 0.03320300206542015, + 0.00648675300180912, + 0.006709320470690727, + 0.00769633986055851, + 0.008110951632261276, + 0.009428642690181732, + 0.009676317684352398, + 0.015278372913599014, + 0.01190500520169735, + 0.0032974551431834698, + 0.0032700190786272287, + 0.0037468133959919214, + 0.0036357075441628695, + 0.004084783140569925, + 0.004317793063819408, + 0.004275718238204718, + 0.0047754934057593346 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 6 + ], + "selected_reward": -0.0059431265108287334 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "This stunning long sequence, rendered in a captivating Pixar style, begins with the camera gliding smoothly in a slow, sweeping aerial shot, following the magnificent, ancient stone fortifications as they majestically wind through a lush, rolling green landscape. The walls themselves are constructed from massive, weathered gray stones, each with softened edges and unique textures, dotted with a soft, velvety moss. Sturdy cylindrical watchtowers with conical roofs of warm red tiles punctuate the length of the walls, their simple forms yet heroic presence against the horizon. The entire scene is bathed in the warm golden glow of the afternoon sun, casting long, soft shadows and highlighting the gentle contours of the hills. In the distance, the azure sky is adorned with a few perfectly fluffy, stylized white clouds. The camera's steady, graceful movement along the fortifications creates a sense of peaceful grandeur and epic adventure, inviting the viewer into a world of whimsical history and breathtaking beauty.", + "seed": 28658, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000156/group_5/index_9911_seed_28658_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": -0.007752715540118515, + "vlm_logprob": -0.007752715540118515, + "vlm_yes_no_margin": 5.6, + "vlm_token_credit": -0.007752715540118515, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a Pixar-style animation aesthetic?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are massive weathered gray stone walls with softened edges and moss visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are sturdy cylindrical watchtowers with conical red tiled roofs present along the walls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set in a lush rolling green landscape under a warm golden afternoon sun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a slow sweeping aerial shot gliding along the fortifications?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02646334283053875 + ], + "opposite_logprob_tokens": [ + -3.776463270187378 + ], + "logprob_reward": -0.02646334283053875, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.02646334283053875, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.438084214925766, + 0.06969312578439713, + 0.049105048179626465, + 0.06813627481460571, + 0.05782533064484596, + 0.09326612204313278, + 0.04096819460391998, + 0.042218685150146484, + 0.10322196036577225, + 0.03748098760843277 + ], + "credit_map": [ + 0.04412597790360451, + 0.048763569444417953, + 0.049310196191072464, + 0.08227557688951492, + 0.0484645739197731, + 0.03888365998864174, + 0.06270907074213028, + 0.06355161219835281, + 0.005615002941340208, + 0.006518108770251274, + 0.006915678735822439, + 0.011731570586562157, + 0.007238644640892744, + 0.01341934222728014, + 0.010692319832742214, + 0.007562458980828524, + 0.005467808805406094, + 0.0062095895409584045, + 0.0052948566153645515, + 0.011560935527086258, + 0.004970366135239601, + 0.004396690055727959, + 0.006832457613199949, + 0.004372347611933947, + 0.0057649570517241955, + 0.004859603941440582, + 0.00729239359498024, + 0.00922655500471592, + 0.004411336500197649, + 0.005230715498328209, + 0.01780172437429428, + 0.01354899350553751, + 0.007427085191011429, + 0.0046060397289693356, + 0.007206973619759083, + 0.009920584969222546, + 0.0058973017148673534, + 0.006069657392799854, + 0.007729797624051571, + 0.008967901580035686, + 0.012630005367100239, + 0.010362030938267708, + 0.014243566431105137, + 0.012632260099053383, + 0.012149084359407425, + 0.009188656695187092, + 0.01023142784833908, + 0.011829103343188763, + 0.007473664358258247, + 0.004955267067998648, + 0.00419518630951643, + 0.0060722180642187595, + 0.0034625083208084106, + 0.004122821614146233, + 0.004413146525621414, + 0.0062733846716582775, + 0.002879895269870758, + 0.0029048004653304815, + 0.006194948218762875, + 0.0114930160343647, + 0.0028961459174752235, + 0.002301179338246584, + 0.006938110571354628, + 0.006610596086829901, + 0.009724962525069714, + 0.008391655050218105, + 0.01860181987285614, + 0.016590459272265434, + 0.011837818659842014, + 0.008976462297141552, + 0.013938027434051037, + 0.015160766430199146, + 0.003033560933545232, + 0.002446643775328994, + 0.005334849935024977, + 0.005696564447134733, + 0.0036342439707368612, + 0.0028248345479369164, + 0.009491970762610435, + 0.005018323659896851 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 5 + ], + "selected_reward": -0.02646334283053875 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a Pixar-style animation aesthetic?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are massive weathered gray stone walls with softened edges and moss visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are sturdy cylindrical watchtowers with conical red tiled roofs present along the walls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set in a lush rolling green landscape under a warm golden afternoon sun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a slow sweeping aerial shot gliding along the fortifications?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0067548975348472595 + ], + "opposite_logprob_tokens": [ + -5.0067548751831055 + ], + "logprob_reward": -0.0067548975348472595, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.0067548975348472595, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5059424042701721, + 0.1252494901418686, + 0.03475836664438248, + 0.06666775792837143, + 0.08138375729322433, + 0.04669135808944702, + 0.05569034069776535, + 0.04727989807724953, + 0.020529767498373985, + 0.01580680161714554 + ], + "credit_map": [ + 0.056394919753074646, + 0.04938197880983353, + 0.059184443205595016, + 0.09967899322509766, + 0.04964883252978325, + 0.04256690666079521, + 0.07399433106184006, + 0.07509204000234604, + 0.012571590952575207, + 0.019102925434708595, + 0.015271797776222229, + 0.021208658814430237, + 0.013416913338005543, + 0.011779025197029114, + 0.017512718215584755, + 0.014385879971086979, + 0.0044194296933710575, + 0.0029409436974674463, + 0.004943161271512508, + 0.005924317054450512, + 0.0027339837979525328, + 0.0023315653670579195, + 0.007355465088039637, + 0.004109506960958242, + 0.005187694914638996, + 0.0034851646050810814, + 0.005072639323771, + 0.01027159858494997, + 0.003183596534654498, + 0.004114299546927214, + 0.020466646179556847, + 0.014886124059557915, + 0.003967180848121643, + 0.0029959252569824457, + 0.005812725983560085, + 0.031196564435958862, + 0.004743339028209448, + 0.004715109709650278, + 0.007343195844441652, + 0.02060973271727562, + 0.004373239353299141, + 0.005344953387975693, + 0.0077391136437654495, + 0.008829106576740742, + 0.004104658029973507, + 0.002651325659826398, + 0.006471253465861082, + 0.007177713792771101, + 0.004646108951419592, + 0.004659522324800491, + 0.006020022090524435, + 0.013953390531241894, + 0.0037156976759433746, + 0.0023564044386148453, + 0.012513557448983192, + 0.007825642824172974, + 0.004706865642219782, + 0.0033631587866693735, + 0.006891306024044752, + 0.010554593056440353, + 0.003727744799107313, + 0.002726873615756631, + 0.008910912089049816, + 0.006398450583219528, + 0.0021279766224324703, + 0.001747912960126996, + 0.0036922607105225325, + 0.0030431728810071945, + 0.001822095480747521, + 0.001761274877935648, + 0.003011585446074605, + 0.00332349119707942, + 0.0016159492079168558, + 0.0010280366986989975, + 0.0020375167950987816, + 0.0022296763490885496, + 0.001674426137469709, + 0.001128413830883801, + 0.0039918567053973675, + 0.0021009286865592003 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 4 + ], + "selected_reward": -0.0067548975348472595 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a Pixar-style animation aesthetic?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are massive weathered gray stone walls with softened edges and moss visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are sturdy cylindrical watchtowers with conical red tiled roofs present along the walls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set in a lush rolling green landscape under a warm golden afternoon sun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a slow sweeping aerial shot gliding along the fortifications?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.001178999780677259 + ], + "opposite_logprob_tokens": [ + -6.751179218292236 + ], + "logprob_reward": -0.001178999780677259, + "yes_no_margin_reward": 6.75, + "token_credit_reward": -0.001178999780677259, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.33997800946235657, + 0.12842242419719696, + 0.04840138554573059, + 0.13435962796211243, + 0.22727519273757935, + 0.04439301788806915, + 0.021831631660461426, + 0.023373868316411972, + 0.01521384809166193, + 0.01675105094909668 + ], + "credit_map": [ + 0.038214728236198425, + 0.032100796699523926, + 0.040571145713329315, + 0.06787044554948807, + 0.033265918493270874, + 0.027504542842507362, + 0.048949889838695526, + 0.051500510424375534, + 0.01785600744187832, + 0.025045696645975113, + 0.009879384189844131, + 0.018009569495916367, + 0.01565016247332096, + 0.014674135483801365, + 0.017905255779623985, + 0.009402216412127018, + 0.006451214198023081, + 0.0031969763804227114, + 0.006781920790672302, + 0.007130283862352371, + 0.0032964206766337156, + 0.0033221521880477667, + 0.012333407066762447, + 0.005889007356017828, + 0.007719739805907011, + 0.005476032849401236, + 0.011694974265992641, + 0.02774806134402752, + 0.010159309022128582, + 0.005958654452115297, + 0.04997870326042175, + 0.015624161809682846, + 0.016144178807735443, + 0.009449944831430912, + 0.016494857147336006, + 0.09003958106040955, + 0.005803752224892378, + 0.01122352946549654, + 0.02032340131700039, + 0.05779595300555229, + 0.005091503728181124, + 0.003023152006790042, + 0.009482559747993946, + 0.008121168240904808, + 0.003655967302620411, + 0.0036802212707698345, + 0.005235937424004078, + 0.0061025056056678295, + 0.0023834595922380686, + 0.001949803321622312, + 0.002143874065950513, + 0.004016944207251072, + 0.0012635826133191586, + 0.0020492179319262505, + 0.004229028709232807, + 0.003795720636844635, + 0.0018886808538809419, + 0.0015439944108948112, + 0.0033486727625131607, + 0.006783610209822655, + 0.0017047065775841475, + 0.0012797658564522862, + 0.0038303823675960302, + 0.0029940546955913305, + 0.0016709299525246024, + 0.001192699302919209, + 0.002800705609843135, + 0.0022839789744466543, + 0.0014917432563379407, + 0.0008938767714425921, + 0.002607783069834113, + 0.002272130688652396, + 0.001649729092605412, + 0.001317713176831603, + 0.002934602089226246, + 0.003347279503941536, + 0.0014816243201494217, + 0.00115690054371953, + 0.002239766763523221, + 0.0026234365068376064 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 3 + ], + "selected_reward": -0.001178999780677259 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a Pixar-style animation aesthetic?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are massive weathered gray stone walls with softened edges and moss visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are sturdy cylindrical watchtowers with conical red tiled roofs present along the walls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set in a lush rolling green landscape under a warm golden afternoon sun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a slow sweeping aerial shot gliding along the fortifications?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.001177928177639842 + ], + "opposite_logprob_tokens": [ + -6.751177787780762 + ], + "logprob_reward": -0.001177928177639842, + "yes_no_margin_reward": 6.75, + "token_credit_reward": -0.001177928177639842, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1515931338071823, + 0.06379002332687378, + 0.07159161567687988, + 0.18154887855052948, + 0.06307639926671982, + 0.14109955728054047, + 0.1313987672328949, + 0.09731942415237427, + 0.057793028652668, + 0.0407891683280468 + ], + "credit_map": [ + 0.015377683565020561, + 0.01617293246090412, + 0.01653522625565529, + 0.02407216466963291, + 0.016805103048682213, + 0.016250034794211388, + 0.019198082387447357, + 0.02718190848827362, + 0.008244069293141365, + 0.007574120536446571, + 0.010215530171990395, + 0.011245152913033962, + 0.007156653795391321, + 0.004997643176466227, + 0.009092950262129307, + 0.005263902712613344, + 0.0077668484300374985, + 0.0060218386352062225, + 0.01162257045507431, + 0.011133565567433834, + 0.008765649981796741, + 0.007195195648819208, + 0.010156212374567986, + 0.008929739706218243, + 0.007715904153883457, + 0.006116591859608889, + 0.008081152103841305, + 0.03959004953503609, + 0.008205498568713665, + 0.006882579065859318, + 0.07295145094394684, + 0.032005663961172104, + 0.01113057229667902, + 0.005171017721295357, + 0.008117866702377796, + 0.00829232856631279, + 0.01049947552382946, + 0.004401441663503647, + 0.008335260674357414, + 0.007128440774977207, + 0.012579960748553276, + 0.010676992125809193, + 0.030914682894945145, + 0.028141070157289505, + 0.009419878013432026, + 0.009395642206072807, + 0.020023636519908905, + 0.01994769088923931, + 0.02615521289408207, + 0.02251438982784748, + 0.014446660876274109, + 0.017393551766872406, + 0.008937246166169643, + 0.006462450604885817, + 0.015862835571169853, + 0.01962641440331936, + 0.008338920772075653, + 0.00799708254635334, + 0.015253129415214062, + 0.019772036001086235, + 0.007336339447647333, + 0.008190327323973179, + 0.01019678171724081, + 0.020234817638993263, + 0.005649810656905174, + 0.004292087629437447, + 0.011041262187063694, + 0.012377741746604443, + 0.004690659698098898, + 0.004997406154870987, + 0.0064279199577867985, + 0.00831613875925541, + 0.0041622500866651535, + 0.0031474626157432795, + 0.007607622537761927, + 0.00567849213257432, + 0.004009129945188761, + 0.0023569369222968817, + 0.007258229423314333, + 0.0065690441988408566 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 5 + ], + "selected_reward": -0.001177928177639842 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a Pixar-style animation aesthetic?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are massive weathered gray stone walls with softened edges and moss visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are sturdy cylindrical watchtowers with conical red tiled roofs present along the walls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set in a lush rolling green landscape under a warm golden afternoon sun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a slow sweeping aerial shot gliding along the fortifications?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0031884093768894672 + ], + "opposite_logprob_tokens": [ + -5.753188610076904 + ], + "logprob_reward": -0.0031884093768894672, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.0031884093768894672, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.37333881855010986, + 0.03973683714866638, + 0.04774056747555733, + 0.08430339395999908, + 0.19338546693325043, + 0.11117859929800034, + 0.028369557112455368, + 0.03154337778687477, + 0.07172200083732605, + 0.018681365996599197 + ], + "credit_map": [ + 0.04288005456328392, + 0.03617887943983078, + 0.04555247724056244, + 0.07498709857463837, + 0.0369010828435421, + 0.03165807947516441, + 0.05120689049363136, + 0.05397424474358559, + 0.004629292991012335, + 0.005517642945051193, + 0.004783467389643192, + 0.005510709248483181, + 0.004551870282739401, + 0.004491873551160097, + 0.006062271073460579, + 0.004189709201455116, + 0.007249674294143915, + 0.006002993322908878, + 0.007081273943185806, + 0.00799507275223732, + 0.004841383080929518, + 0.0038890540599823, + 0.0063238246366381645, + 0.004357292316854, + 0.007683485746383667, + 0.007049919571727514, + 0.01004521269351244, + 0.016900669783353806, + 0.005918528418987989, + 0.007033082190901041, + 0.019849274307489395, + 0.009823222644627094, + 0.009177747182548046, + 0.006532356142997742, + 0.014012783765792847, + 0.0789819061756134, + 0.007029977161437273, + 0.0075318035669624805, + 0.01703784242272377, + 0.05308104678988457, + 0.010383632965385914, + 0.008635146543383598, + 0.02513519860804081, + 0.02045838162302971, + 0.007799969520419836, + 0.007137374021112919, + 0.016681920737028122, + 0.014946984127163887, + 0.00549251539632678, + 0.0037645185366272926, + 0.002413750858977437, + 0.004567749332636595, + 0.001539579126983881, + 0.0011636982671916485, + 0.0034755063243210316, + 0.005952241364866495, + 0.002108102198690176, + 0.002159895608201623, + 0.004196309484541416, + 0.007566362153738737, + 0.0023454155307263136, + 0.002755420748144388, + 0.005125842057168484, + 0.005286034196615219, + 0.006705807521939278, + 0.005176776088774204, + 0.012602008879184723, + 0.015359153039753437, + 0.005119585897773504, + 0.006031710654497147, + 0.010800286196172237, + 0.009926668368279934, + 0.0016148948343470693, + 0.001510245376266539, + 0.002743859775364399, + 0.004727962426841259, + 0.0010366225615143776, + 0.0017729573883116245, + 0.0025449502281844616, + 0.002729873638600111 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 5 + ], + "selected_reward": -0.0031884093768894672 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "This stunning long sequence, rendered in a captivating Pixar style, begins with the camera gliding smoothly in a slow, sweeping aerial shot, following the magnificent, ancient stone fortifications as they majestically wind through a lush, rolling green landscape. The walls themselves are constructed from massive, weathered gray stones, each with softened edges and unique textures, dotted with a soft, velvety moss. Sturdy cylindrical watchtowers with conical roofs of warm red tiles punctuate the length of the walls, their simple forms yet heroic presence against the horizon. The entire scene is bathed in the warm golden glow of the afternoon sun, casting long, soft shadows and highlighting the gentle contours of the hills. In the distance, the azure sky is adorned with a few perfectly fluffy, stylized white clouds. The camera's steady, graceful movement along the fortifications creates a sense of peaceful grandeur and epic adventure, inviting the viewer into a world of whimsical history and breathtaking beauty.", + "seed": 28659, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000156/group_5/index_9911_seed_28659_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": -0.012272734369616956, + "vlm_logprob": -0.012272734369616956, + "vlm_yes_no_margin": 5.475, + "vlm_token_credit": -0.012272734369616956, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a Pixar-style animation aesthetic?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are massive weathered gray stone walls with softened edges and moss visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are sturdy cylindrical watchtowers with conical red tiled roofs present along the walls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set in a lush rolling green landscape under a warm golden afternoon sun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a slow sweeping aerial shot gliding along the fortifications?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04660825431346893 + ], + "opposite_logprob_tokens": [ + -3.2966082096099854 + ], + "logprob_reward": -0.04660825431346893, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.04660825431346893, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20585258305072784, + 0.09487257152795792, + 0.041061218827962875, + 0.05689631775021553, + 0.08104376494884491, + 0.07531988620758057, + 0.057219166308641434, + 0.1158188208937645, + 0.07257401943206787, + 0.1993415653705597 + ], + "credit_map": [ + 0.011857825331389904, + 0.009414475411176682, + 0.06539639830589294, + 0.02367449551820755, + 0.007740975823253393, + 0.007793971803039312, + 0.06257735192775726, + 0.017397096380591393, + 0.007703830022364855, + 0.008742067031562328, + 0.01223490945994854, + 0.017263691872358322, + 0.010899272747337818, + 0.005768368951976299, + 0.0201361533254385, + 0.012124280445277691, + 0.002995620248839259, + 0.00462165754288435, + 0.004222316667437553, + 0.008934205397963524, + 0.0037565273232758045, + 0.0033439986873418093, + 0.007370561361312866, + 0.005816332064568996, + 0.0037884004414081573, + 0.005230523180216551, + 0.009333346970379353, + 0.00996420532464981, + 0.003313429420813918, + 0.0034835159312933683, + 0.010701377876102924, + 0.011081518605351448, + 0.007947665639221668, + 0.011086415499448776, + 0.012686479836702347, + 0.014059807173907757, + 0.005437567364424467, + 0.01003627572208643, + 0.010514244437217712, + 0.009275305084884167, + 0.008956383913755417, + 0.008733384311199188, + 0.01061543170362711, + 0.007926758378744125, + 0.006943269167095423, + 0.005661060567945242, + 0.012689796276390553, + 0.01379379816353321, + 0.006701803766191006, + 0.00496324896812439, + 0.006504842545837164, + 0.005852558184415102, + 0.0066300444304943085, + 0.0036229670513421297, + 0.008758538402616978, + 0.014185160398483276, + 0.012131785042583942, + 0.01060580462217331, + 0.018177291378378868, + 0.01794065721333027, + 0.0176573246717453, + 0.0077324556186795235, + 0.015157349407672882, + 0.016416147351264954, + 0.003989651333540678, + 0.00423017144203186, + 0.0237728338688612, + 0.009973657317459583, + 0.005955482367426157, + 0.003937703091651201, + 0.005433930084109306, + 0.015280586667358875, + 0.0393335334956646, + 0.025422777980566025, + 0.027728082612156868, + 0.027074651792645454, + 0.02381233498454094, + 0.014906851574778557, + 0.019612375646829605, + 0.021450968459248543 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 7 + ], + "selected_reward": -0.04660825431346893 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a Pixar-style animation aesthetic?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are massive weathered gray stone walls with softened edges and moss visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are sturdy cylindrical watchtowers with conical red tiled roofs present along the walls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set in a lush rolling green landscape under a warm golden afternoon sun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a slow sweeping aerial shot gliding along the fortifications?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009808086790144444 + ], + "opposite_logprob_tokens": [ + -4.63480806350708 + ], + "logprob_reward": -0.009808086790144444, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009808086790144444, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12670063972473145, + 0.21760281920433044, + 0.07130298763513565, + 0.07149888575077057, + 0.2058730125427246, + 0.026159750297665596, + 0.05072023347020149, + 0.08892188966274261, + 0.03097562864422798, + 0.11024414002895355 + ], + "credit_map": [ + 0.0069017913192510605, + 0.007282945327460766, + 0.04079190641641617, + 0.0184087622910738, + 0.006364160217344761, + 0.004469551146030426, + 0.02930140122771263, + 0.013180122710764408, + 0.025278493762016296, + 0.018152572214603424, + 0.026555990800261497, + 0.03429104760289192, + 0.020504456013441086, + 0.01632046326994896, + 0.05037762224674225, + 0.02612217329442501, + 0.00505530321970582, + 0.006946968846023083, + 0.010404652915894985, + 0.012084838934242725, + 0.007526475936174393, + 0.005648876540362835, + 0.011026568710803986, + 0.01260930486023426, + 0.004807457793504, + 0.005150134209543467, + 0.00835462473332882, + 0.015640677884221077, + 0.0037249408196657896, + 0.003536378964781761, + 0.0190025232732296, + 0.011282149702310562, + 0.016678733751177788, + 0.023152103647589684, + 0.02472434565424919, + 0.039092954248189926, + 0.013376489281654358, + 0.028831610456109047, + 0.031411346048116684, + 0.02860543131828308, + 0.002148086205124855, + 0.002427298342809081, + 0.0039025929290801287, + 0.0038713831454515457, + 0.0028241141699254513, + 0.0023728732485324144, + 0.005157942418009043, + 0.0034554596059024334, + 0.005203807260841131, + 0.004466194659471512, + 0.008856394328176975, + 0.0062176818028092384, + 0.004948804154992104, + 0.0025991727598011494, + 0.008500651456415653, + 0.009927527979016304, + 0.00682316068559885, + 0.01312448550015688, + 0.013037999160587788, + 0.016083791851997375, + 0.011142462491989136, + 0.006369961891323328, + 0.010317780077457428, + 0.012022250331938267, + 0.003080658381804824, + 0.004375384654849768, + 0.005483624991029501, + 0.005369818769395351, + 0.004444444086402655, + 0.002364585641771555, + 0.003095578635111451, + 0.002761530689895153, + 0.011093080043792725, + 0.011217543855309486, + 0.01828875206410885, + 0.0164883304387331, + 0.01759224943816662, + 0.007780192419886589, + 0.011539754457771778, + 0.01624423824250698 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 4, + 0 + ], + "selected_reward": -0.009808086790144444 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a Pixar-style animation aesthetic?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are massive weathered gray stone walls with softened edges and moss visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are sturdy cylindrical watchtowers with conical red tiled roofs present along the walls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set in a lush rolling green landscape under a warm golden afternoon sun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a slow sweeping aerial shot gliding along the fortifications?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0008131535141728818 + ], + "opposite_logprob_tokens": [ + -7.125813007354736 + ], + "logprob_reward": -0.0008131535141728818, + "yes_no_margin_reward": 7.125, + "token_credit_reward": -0.0008131535141728818, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09222206473350525, + 0.18341663479804993, + 0.2040645182132721, + 0.06363644450902939, + 0.04649357125163078, + 0.0359475240111351, + 0.06504324078559875, + 0.12984363734722137, + 0.03918638452887535, + 0.14014597237110138 + ], + "credit_map": [ + 0.007441488094627857, + 0.00910697691142559, + 0.02163449488580227, + 0.01743430271744728, + 0.006052726414054632, + 0.006293328013271093, + 0.016335399821400642, + 0.007923347875475883, + 0.023436445742845535, + 0.01626616157591343, + 0.02059471793472767, + 0.02639716863632202, + 0.01822228729724884, + 0.015058266930282116, + 0.04087189957499504, + 0.022569701075553894, + 0.014841534197330475, + 0.024782666936516762, + 0.02101009525358677, + 0.0353979617357254, + 0.018563775345683098, + 0.017641378566622734, + 0.02999475784599781, + 0.041832342743873596, + 0.005103549920022488, + 0.00855251681059599, + 0.00667978823184967, + 0.010039567947387695, + 0.006769553292542696, + 0.005975463427603245, + 0.010331107303500175, + 0.010184891521930695, + 0.003922051750123501, + 0.004622724372893572, + 0.006485460791736841, + 0.00795020256191492, + 0.0032667822670191526, + 0.006715299095958471, + 0.0064809732139110565, + 0.0070500802248716354, + 0.003013395704329014, + 0.0037775724194943905, + 0.004659994505345821, + 0.0066487789154052734, + 0.004182137083262205, + 0.002454893197864294, + 0.006301527842879295, + 0.004909224808216095, + 0.006581256631761789, + 0.006589293945580721, + 0.010626199655234814, + 0.007926234975457191, + 0.009362253360450268, + 0.0037494231946766376, + 0.011276504956185818, + 0.008932081051170826, + 0.008407442830502987, + 0.01171069499105215, + 0.021835338324308395, + 0.02239951118826866, + 0.019040508195757866, + 0.008218297734856606, + 0.02169666439294815, + 0.016535179689526558, + 0.005420085974037647, + 0.005175899248570204, + 0.0065169804729521275, + 0.006496830377727747, + 0.003739653853699565, + 0.00248670089058578, + 0.0032545828726142645, + 0.006095650605857372, + 0.012954769656062126, + 0.01470787450671196, + 0.023720094934105873, + 0.023124687373638153, + 0.019570542499423027, + 0.01053374633193016, + 0.014739382080733776, + 0.020794866606593132 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 1, + 9 + ], + "selected_reward": -0.0008131535141728818 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a Pixar-style animation aesthetic?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are massive weathered gray stone walls with softened edges and moss visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are sturdy cylindrical watchtowers with conical red tiled roofs present along the walls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set in a lush rolling green landscape under a warm golden afternoon sun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a slow sweeping aerial shot gliding along the fortifications?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0019382515456527472 + ], + "opposite_logprob_tokens": [ + -6.251938343048096 + ], + "logprob_reward": -0.0019382515456527472, + "yes_no_margin_reward": 6.25, + "token_credit_reward": -0.0019382515456527472, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.25009864568710327, + 0.029531201347708702, + 0.02948812022805214, + 0.08885029703378677, + 0.1670152246952057, + 0.04777895659208298, + 0.04361603409051895, + 0.09594788402318954, + 0.10350766777992249, + 0.14416596293449402 + ], + "credit_map": [ + 0.01479299459606409, + 0.010779649950563908, + 0.07774911820888519, + 0.04533923417329788, + 0.00814818311482668, + 0.008415808901190758, + 0.05904976651072502, + 0.02582387812435627, + 0.00266189593821764, + 0.0033318656496703625, + 0.003384815761819482, + 0.0046129957772791386, + 0.0037481384351849556, + 0.0018722196109592915, + 0.006659678183495998, + 0.0032595936208963394, + 0.0017803091323003173, + 0.0020801867358386517, + 0.0065861912444233894, + 0.004539821296930313, + 0.003585495287552476, + 0.0025607976131141186, + 0.004253178369253874, + 0.004102142993360758, + 0.007005670107901096, + 0.007004277314990759, + 0.015355391427874565, + 0.015599343925714493, + 0.007384748663753271, + 0.0054159765131771564, + 0.01778983511030674, + 0.013295053504407406, + 0.015948541462421417, + 0.015428757295012474, + 0.01853318139910698, + 0.0339038111269474, + 0.008958270773291588, + 0.017970984801650047, + 0.026359431445598602, + 0.02991223707795143, + 0.0036699583288282156, + 0.0059328521601855755, + 0.008403729647397995, + 0.008141865022480488, + 0.0036814259365200996, + 0.0037349355407059193, + 0.008175615221261978, + 0.0060385712422430515, + 0.004457133822143078, + 0.0052171507850289345, + 0.006814531050622463, + 0.004811829887330532, + 0.0066957795061171055, + 0.002798250177875161, + 0.00563357537612319, + 0.007187786046415567, + 0.005870400462299585, + 0.0076024290174245834, + 0.016655411571264267, + 0.014807766303420067, + 0.013895572163164616, + 0.005957826506346464, + 0.01944371871650219, + 0.01171476487070322, + 0.008313586004078388, + 0.012388917617499828, + 0.020914766937494278, + 0.0141457449644804, + 0.012207718566060066, + 0.007156977429986, + 0.008681435137987137, + 0.01969851925969124, + 0.013677039183676243, + 0.017104260623455048, + 0.019518466666340828, + 0.022743264213204384, + 0.02585338056087494, + 0.013213005848228931, + 0.014147398993372917, + 0.01790914684534073 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 9 + ], + "selected_reward": -0.0019382515456527472 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a Pixar-style animation aesthetic?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are massive weathered gray stone walls with softened edges and moss visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are sturdy cylindrical watchtowers with conical red tiled roofs present along the walls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set in a lush rolling green landscape under a warm golden afternoon sun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a slow sweeping aerial shot gliding along the fortifications?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002195925684645772 + ], + "opposite_logprob_tokens": [ + -6.127195835113525 + ], + "logprob_reward": -0.002195925684645772, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.002195925684645772, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08624517172574997, + 0.13195742666721344, + 0.038343511521816254, + 0.06476330757141113, + 0.21475185453891754, + 0.30396854877471924, + 0.024193478748202324, + 0.046404946595430374, + 0.05947006121277809, + 0.029901700094342232 + ], + "credit_map": [ + 0.007067064754664898, + 0.0074203405529260635, + 0.011024010367691517, + 0.02042226679623127, + 0.008467060513794422, + 0.009963610209524632, + 0.012422545813024044, + 0.009458273649215698, + 0.008036477491259575, + 0.009187934920191765, + 0.015501297079026699, + 0.02004420757293701, + 0.011226840317249298, + 0.005643017124384642, + 0.043320752680301666, + 0.018996883183717728, + 0.0028083622455596924, + 0.002987191081047058, + 0.006655439734458923, + 0.008297104388475418, + 0.0035690313670784235, + 0.0028060884214937687, + 0.0052919513545930386, + 0.005928341299295425, + 0.005677311215549707, + 0.005660476628690958, + 0.008551687933504581, + 0.009187047369778156, + 0.004220444709062576, + 0.0034571841824799776, + 0.012920301407575607, + 0.015088852494955063, + 0.017796583473682404, + 0.02048349939286709, + 0.041604265570640564, + 0.0444888211786747, + 0.009378264658153057, + 0.01981017366051674, + 0.030538050457835197, + 0.030652204528450966, + 0.016141654923558235, + 0.019422998651862144, + 0.036984335631132126, + 0.051239386200904846, + 0.02093327045440674, + 0.01691335439682007, + 0.09158304333686829, + 0.050750501453876495, + 0.002320332219824195, + 0.002164585283026099, + 0.004150220658630133, + 0.0029999695252627134, + 0.002077751327306032, + 0.0018954025581479073, + 0.004001797642558813, + 0.004583419766277075, + 0.00525784632191062, + 0.00526827946305275, + 0.00481807254254818, + 0.006813961546868086, + 0.005970193073153496, + 0.003589920001104474, + 0.006530932150781155, + 0.008155737072229385, + 0.005323381628841162, + 0.009267413057386875, + 0.008393551222980022, + 0.009398640133440495, + 0.010030191391706467, + 0.004925830289721489, + 0.004714467562735081, + 0.007416586857289076, + 0.004141367040574551, + 0.0038305551279336214, + 0.005043797194957733, + 0.004429277498275042, + 0.0024235460441559553, + 0.0028872692491859198, + 0.003231566632166505, + 0.00391432223841548 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 4, + 1 + ], + "selected_reward": -0.002195925684645772 + } + ], + "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/0000156/group_6/index_tensor([41, 41])_group_6_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000156/group_6/index_tensor([41, 41])_group_6_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..4043c1876414581b878f778f47b64d98177aa65f --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000156/group_6/index_tensor([41, 41])_group_6_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "Captured from a slightly low angle with a static camera, this ultra-detailed macro shot focuses intensely on a pale, slender hand suspended against a stark, absolute black background, a non-reflective blank. The hand is suspended in a state of predatory tension, its wrist gracefully arched back, poised to strike with the deadly grace of a snake's head. Its skin is an almost translucent alabaster, so fine that delicate, faint blue veins can be seen beneath the surface. The long, graceful fingers curl inward, but not in a fist; instead, the index and middle fingers are slightly apart and pointed forward, mimicking fangs poised to thrust. A sharp, key-like light from the side carves the hand out of the darkness, creating a dramatic black-and-white contrast. Bright highlights gleam from the spotless, natural nails and depict the tense tendons on the back of the palm, while deep, velvety shadows pool in the crevices between the fingers. Every muscle is visibly tensed, a subtle, controlled effort exerted, creating a palpable aura of imminent, precise, and threatening action.", + "seed": 8556, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000156/group_6/index_41_seed_8556_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": -0.01085654590278864, + "vlm_logprob": -0.01085654590278864, + "vlm_yes_no_margin": 4.925, + "vlm_token_credit": -0.01085654590278864, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a pale, slender human hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a stark, absolute black non-reflective surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the hand positioned with the wrist arched back and fingers curled inward with the index and middle fingers slightly apart?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lighting a sharp, key-like light from the side creating high contrast with bright highlights on the nails and deep shadows between the fingers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera angle slightly low and is the camera static?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005565860774368048 + ], + "opposite_logprob_tokens": [ + -6.255565643310547 + ], + "logprob_reward": -0.005565860774368048, + "yes_no_margin_reward": 6.25, + "token_credit_reward": -0.005565860774368048, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1453646421432495, + 0.09909062832593918, + 0.08302060514688492, + 0.08922228217124939, + 0.0916454941034317, + 0.11361536383628845, + 0.11380251497030258, + 0.05647548288106918, + 0.0995878130197525, + 0.10817522555589676 + ], + "credit_map": [ + 0.021145077422261238, + 0.013010737486183643, + 0.02038794942200184, + 0.014051205478608608, + 0.018120253458619118, + 0.020394863560795784, + 0.016279373317956924, + 0.021975189447402954, + 0.011293666437268257, + 0.009534315206110477, + 0.013618272729218006, + 0.007845351472496986, + 0.010459723882377148, + 0.01682082749903202, + 0.013954829424619675, + 0.015563640743494034, + 0.014009732753038406, + 0.006227308884263039, + 0.006349774543195963, + 0.006092824507504702, + 0.00826118141412735, + 0.016075650230050087, + 0.006663074716925621, + 0.019341059029102325, + 0.005775049328804016, + 0.0058285086415708065, + 0.006793498992919922, + 0.004785384051501751, + 0.0080203702673316, + 0.0259495060890913, + 0.006010161247104406, + 0.026059797033667564, + 0.00885215774178505, + 0.008544460870325565, + 0.01606864482164383, + 0.009055890142917633, + 0.012146523222327232, + 0.013663562946021557, + 0.00925549678504467, + 0.014058742672204971, + 0.011515263468027115, + 0.007418130990117788, + 0.012793499045073986, + 0.010121893137693405, + 0.005861131474375725, + 0.03007790632545948, + 0.008371052332222462, + 0.027456486597657204, + 0.012471737340092659, + 0.01258657407015562, + 0.021748116239905357, + 0.01102360151708126, + 0.009148265235126019, + 0.019558683037757874, + 0.011494533158838749, + 0.015770992264151573, + 0.009415879845619202, + 0.0055790850892663, + 0.006495128385722637, + 0.005040696822106838, + 0.0040306695736944675, + 0.010167578235268593, + 0.006489701569080353, + 0.009256738238036633, + 0.007562597282230854, + 0.00926485937088728, + 0.010159138590097427, + 0.01614229753613472, + 0.013251503929495811, + 0.016473937779664993, + 0.012507637031376362, + 0.014225835911929607, + 0.015229441225528717, + 0.007857704535126686, + 0.010947179049253464, + 0.01605120301246643, + 0.011141935363411903, + 0.01772484742105007, + 0.01480442751199007, + 0.014418468810617924 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 5 + ], + "selected_reward": -0.005565860774368048 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a pale, slender human hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a stark, absolute black non-reflective surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the hand positioned with the wrist arched back and fingers curled inward with the index and middle fingers slightly apart?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lighting a sharp, key-like light from the side creating high contrast with bright highlights on the nails and deep shadows between the fingers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera angle slightly low and is the camera static?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"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": [ + 0.1595144420862198, + 0.15622863173484802, + 0.037539251148700714, + 0.22070206701755524, + 0.10030759125947952, + 0.04913126677274704, + 0.07534858584403992, + 0.0666806623339653, + 0.06101587787270546, + 0.07353170216083527 + ], + "credit_map": [ + 0.022845463827252388, + 0.012136086821556091, + 0.015141957439482212, + 0.010430281050503254, + 0.012791897170245647, + 0.03076871484518051, + 0.012885956093668938, + 0.042514074593782425, + 0.024719327688217163, + 0.01252591423690319, + 0.020437417551875114, + 0.010806737467646599, + 0.015073182992637157, + 0.02722872793674469, + 0.016728807240724564, + 0.02870849333703518, + 0.004881007131189108, + 0.0028914546128362417, + 0.0034932855051010847, + 0.0032058393117040396, + 0.0037524024955928326, + 0.008523035794496536, + 0.0031798917334526777, + 0.007612329907715321, + 0.026705021038651466, + 0.015913421288132668, + 0.016322312876582146, + 0.014602434821426868, + 0.01717015728354454, + 0.05342860519886017, + 0.019321756437420845, + 0.05723833665251732, + 0.010213415138423443, + 0.009648304432630539, + 0.014456139877438545, + 0.010079358704388142, + 0.013788093812763691, + 0.014886175282299519, + 0.01084903720766306, + 0.016387058421969414, + 0.006864555180072784, + 0.003350552637130022, + 0.005334768444299698, + 0.0035943593829870224, + 0.0030221326742321253, + 0.01124568097293377, + 0.003738682484254241, + 0.01198052428662777, + 0.00855063833296299, + 0.005646351259201765, + 0.007661215960979462, + 0.00862807221710682, + 0.005165254697203636, + 0.014261199161410332, + 0.009586784988641739, + 0.01584906317293644, + 0.005606719758361578, + 0.004404359497129917, + 0.008766883052885532, + 0.0065764463506639, + 0.0050806282088160515, + 0.012893006205558777, + 0.011629992164671421, + 0.011722614988684654, + 0.00920182652771473, + 0.0035990793257951736, + 0.006005194038152695, + 0.010451123118400574, + 0.004496237728744745, + 0.010151049122214317, + 0.009615745395421982, + 0.007495610974729061, + 0.007018707692623138, + 0.004210811108350754, + 0.006260800175368786, + 0.007283850107342005, + 0.004894403275102377, + 0.017563987523317337, + 0.01183781586587429, + 0.014461321756243706 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 1 + ], + "selected_reward": -0.014212576672434807 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a pale, slender human hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a stark, absolute black non-reflective surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the hand positioned with the wrist arched back and fingers curled inward with the index and middle fingers slightly apart?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lighting a sharp, key-like light from the side creating high contrast with bright highlights on the nails and deep shadows between the fingers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera angle slightly low and is the camera static?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005952606908977032 + ], + "opposite_logprob_tokens": [ + -5.130952835083008 + ], + "logprob_reward": -0.005952606908977032, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005952606908977032, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.119207002222538, + 0.18398578464984894, + 0.08940526843070984, + 0.08066760003566742, + 0.07494113594293594, + 0.08812496066093445, + 0.05918304994702339, + 0.06787191331386566, + 0.1143745705485344, + 0.12223873287439346 + ], + "credit_map": [ + 0.01580469310283661, + 0.007283331360667944, + 0.011862347833812237, + 0.008502613753080368, + 0.01200239546597004, + 0.022898143157362938, + 0.009415091015398502, + 0.03143838047981262, + 0.015170510858297348, + 0.01500261202454567, + 0.01042712852358818, + 0.011913921684026718, + 0.028909627348184586, + 0.044969748705625534, + 0.017318112775683403, + 0.04027411714196205, + 0.005744134541600943, + 0.009748500771820545, + 0.007420307025313377, + 0.007837353274226189, + 0.01317827869206667, + 0.020377393811941147, + 0.008109831251204014, + 0.016989465802907944, + 0.012884172610938549, + 0.006074303761124611, + 0.007695245090872049, + 0.004899770487099886, + 0.007691169623285532, + 0.014553643763065338, + 0.007632510270923376, + 0.019236788153648376, + 0.009863289073109627, + 0.009374184533953667, + 0.008707036264240742, + 0.00695539778098464, + 0.011879736557602882, + 0.011130866594612598, + 0.0082201287150383, + 0.008810499683022499, + 0.007683677598834038, + 0.004733344539999962, + 0.006637170910835266, + 0.005994938313961029, + 0.004428260959684849, + 0.02855580486357212, + 0.005777658894658089, + 0.02431410737335682, + 0.005296074785292149, + 0.0056411996483802795, + 0.007942669093608856, + 0.006054497789591551, + 0.003749456722289324, + 0.012224574573338032, + 0.006108785048127174, + 0.012165792286396027, + 0.009113098494708538, + 0.007320379838347435, + 0.008907891809940338, + 0.006994260940700769, + 0.0058936928398907185, + 0.011059303767979145, + 0.00921708345413208, + 0.009366204962134361, + 0.0068327332846820354, + 0.009720939211547375, + 0.009897910058498383, + 0.025121217593550682, + 0.017960557714104652, + 0.01613817736506462, + 0.016301849856972694, + 0.012401187792420387, + 0.00786580704152584, + 0.011470316909253597, + 0.01212346088141203, + 0.02274129167199135, + 0.017183400690555573, + 0.017433827742934227, + 0.020499559119343758, + 0.012921071611344814 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 9, + 0 + ], + "selected_reward": -0.005952606908977032 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a pale, slender human hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a stark, absolute black non-reflective surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the hand positioned with the wrist arched back and fingers curled inward with the index and middle fingers slightly apart?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lighting a sharp, key-like light from the side creating high contrast with bright highlights on the nails and deep shadows between the fingers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera angle slightly low and is the camera static?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005255692172795534 + ], + "opposite_logprob_tokens": [ + -5.255255699157715 + ], + "logprob_reward": -0.005255692172795534, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005255692172795534, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14119747281074524, + 0.1220296248793602, + 0.13746735453605652, + 0.1954486072063446, + 0.10336168855428696, + 0.07557084411382675, + 0.0406547486782074, + 0.03683527186512947, + 0.0348268561065197, + 0.11260757595300674 + ], + "credit_map": [ + 0.018104558810591698, + 0.012407789938151836, + 0.0171045009046793, + 0.012559919618070126, + 0.016033900901675224, + 0.022383932024240494, + 0.014842192642390728, + 0.027760673314332962, + 0.014803610742092133, + 0.012030902318656445, + 0.016183409839868546, + 0.011498129926621914, + 0.00958411954343319, + 0.022585909813642502, + 0.015061858110129833, + 0.020281676203012466, + 0.018809586763381958, + 0.008836541324853897, + 0.010529414750635624, + 0.01144406944513321, + 0.009134123101830482, + 0.03325207903981209, + 0.01025607530027628, + 0.03520546853542328, + 0.020418904721736908, + 0.007946965284645557, + 0.014630836434662342, + 0.009140567854046822, + 0.009159812703728676, + 0.056550443172454834, + 0.014888787642121315, + 0.06271227449178696, + 0.008806449361145496, + 0.009212300181388855, + 0.014693228527903557, + 0.010154694318771362, + 0.012873455882072449, + 0.01912294141948223, + 0.010864149779081345, + 0.017634468153119087, + 0.004830691963434219, + 0.004847817122936249, + 0.005916442722082138, + 0.005032661836594343, + 0.0050682262517511845, + 0.02084617130458355, + 0.007034321781247854, + 0.02199450507760048, + 0.007437421474605799, + 0.0028644343838095665, + 0.004719270393252373, + 0.004023009445518255, + 0.0026101197581738234, + 0.0066315592266619205, + 0.005622584372758865, + 0.006746345665305853, + 0.0039673144929111, + 0.0029523842968046665, + 0.0037785908207297325, + 0.006404910236597061, + 0.0032668672502040863, + 0.00649198517203331, + 0.004384866915643215, + 0.005588351283222437, + 0.003033977234736085, + 0.0027465971652418375, + 0.003978151828050613, + 0.006053545046597719, + 0.0038099195808172226, + 0.005541463382542133, + 0.00529721612110734, + 0.004365986678749323, + 0.014670301228761673, + 0.008566741831600666, + 0.013867231085896492, + 0.01796598918735981, + 0.010438833385705948, + 0.017036521807312965, + 0.015175956301391125, + 0.014885998331010342 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 2 + ], + "selected_reward": -0.005255692172795534 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a pale, slender human hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a stark, absolute black non-reflective surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the hand positioned with the wrist arched back and fingers curled inward with the index and middle fingers slightly apart?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lighting a sharp, key-like light from the side creating high contrast with bright highlights on the nails and deep shadows between the fingers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera angle slightly low and is the camera static?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023295992985367775 + ], + "opposite_logprob_tokens": [ + -3.7732958793640137 + ], + "logprob_reward": -0.023295992985367775, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023295992985367775, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07222194969654083, + 0.04485037550330162, + 0.28669050335884094, + 0.14419059455394745, + 0.06150978058576584, + 0.06656018644571304, + 0.045769669115543365, + 0.09166673570871353, + 0.09443509578704834, + 0.09210509061813354 + ], + "credit_map": [ + 0.013224505819380283, + 0.006933408789336681, + 0.008559255860745907, + 0.005477115046232939, + 0.006208605598658323, + 0.013792361132800579, + 0.005974090192466974, + 0.012052611447870731, + 0.0067752753384411335, + 0.002904255408793688, + 0.0036712498404085636, + 0.003595892107114196, + 0.004746485035866499, + 0.008749382570385933, + 0.003733135759830475, + 0.010674705728888512, + 0.03997599333524704, + 0.01099440548568964, + 0.013561137020587921, + 0.013178265653550625, + 0.01001417264342308, + 0.08139419555664062, + 0.014633561484515667, + 0.10293877869844437, + 0.019726837053894997, + 0.008316739462316036, + 0.010948819108307362, + 0.00794090237468481, + 0.010959344916045666, + 0.03468066453933716, + 0.009219313971698284, + 0.042397983372211456, + 0.0048411390744149685, + 0.009254730306565762, + 0.008284549228847027, + 0.005496550351381302, + 0.010303385555744171, + 0.008406499400734901, + 0.007014011032879353, + 0.007908918894827366, + 0.0046201725490391254, + 0.0031178928911685944, + 0.005095086060464382, + 0.004105342552065849, + 0.0031000724993646145, + 0.022604523226618767, + 0.004522136878222227, + 0.01939496584236622, + 0.005661383271217346, + 0.004162497818470001, + 0.00430423766374588, + 0.004685126710683107, + 0.00449489988386631, + 0.008708405308425426, + 0.005194398574531078, + 0.008558725006878376, + 0.014995248056948185, + 0.010337242856621742, + 0.01369291078299284, + 0.00941198505461216, + 0.0073663475923240185, + 0.014203626662492752, + 0.010191703215241432, + 0.011467684991657734, + 0.010274538770318031, + 0.0076273903250694275, + 0.008985522203147411, + 0.017802515998482704, + 0.013102101162075996, + 0.013843485154211521, + 0.010665369220077991, + 0.012134176678955555, + 0.007843107916414738, + 0.007515958976000547, + 0.012827401049435139, + 0.014806948602199554, + 0.0073149842210114, + 0.015905018895864487, + 0.01313786394894123, + 0.012753820046782494 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 3, + 8 + ], + "selected_reward": -0.023295992985367775 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "Captured from a slightly low angle with a static camera, this ultra-detailed macro shot focuses intensely on a pale, slender hand suspended against a stark, absolute black background, a non-reflective blank. The hand is suspended in a state of predatory tension, its wrist gracefully arched back, poised to strike with the deadly grace of a snake's head. Its skin is an almost translucent alabaster, so fine that delicate, faint blue veins can be seen beneath the surface. The long, graceful fingers curl inward, but not in a fist; instead, the index and middle fingers are slightly apart and pointed forward, mimicking fangs poised to thrust. A sharp, key-like light from the side carves the hand out of the darkness, creating a dramatic black-and-white contrast. Bright highlights gleam from the spotless, natural nails and depict the tense tendons on the back of the palm, while deep, velvety shadows pool in the crevices between the fingers. Every muscle is visibly tensed, a subtle, controlled effort exerted, creating a palpable aura of imminent, precise, and threatening action.", + "seed": 8557, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000156/group_6/index_41_seed_8557_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": -0.03342230841517448, + "vlm_logprob": -0.03342230841517448, + "vlm_yes_no_margin": 3.975, + "vlm_token_credit": -0.03342230841517448, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a pale, slender human hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a stark, absolute black non-reflective surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the hand positioned with the wrist arched back and fingers curled inward with the index and middle fingers slightly apart?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lighting a sharp, key-like light from the side creating high contrast with bright highlights on the nails and deep shadows between the fingers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera angle slightly low and is the camera static?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008459093980491161 + ], + "opposite_logprob_tokens": [ + -5.258459091186523 + ], + "logprob_reward": -0.008459093980491161, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.008459093980491161, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.29318222403526306, + 0.10623268783092499, + 0.103387251496315, + 0.06661701202392578, + 0.08216726034879684, + 0.08646291494369507, + 0.08232386410236359, + 0.05137847736477852, + 0.04264308884739876, + 0.08560522645711899 + ], + "credit_map": [ + 0.029117073863744736, + 0.02432999573647976, + 0.031043346971273422, + 0.020579416304826736, + 0.06791865825653076, + 0.059109799563884735, + 0.03464760258793831, + 0.026436317712068558, + 0.010977597907185555, + 0.020308103412389755, + 0.014530474320054054, + 0.013308475725352764, + 0.00922544114291668, + 0.012239037081599236, + 0.017825298011302948, + 0.007818254642188549, + 0.011201980523765087, + 0.015707194805145264, + 0.013515153899788857, + 0.007313137874007225, + 0.00972838420420885, + 0.013118412345647812, + 0.016904529184103012, + 0.015898453071713448, + 0.012382796965539455, + 0.010825015604496002, + 0.009222197346389294, + 0.008239595219492912, + 0.00591533025726676, + 0.004847236908972263, + 0.0065618338994681835, + 0.008622998371720314, + 0.01475928071886301, + 0.009822867810726166, + 0.013993681408464909, + 0.00679083401337266, + 0.007250221911817789, + 0.007952053099870682, + 0.009527681395411491, + 0.012070638127624989, + 0.018235696479678154, + 0.007881160825490952, + 0.01679265685379505, + 0.007251534145325422, + 0.005410307552665472, + 0.004248768091201782, + 0.013470323756337166, + 0.013172462582588196, + 0.01238427683711052, + 0.004720235243439674, + 0.012978151440620422, + 0.008656824938952923, + 0.006214614026248455, + 0.009597575291991234, + 0.01158843282610178, + 0.016183745115995407, + 0.006366860121488571, + 0.003792700357735157, + 0.00725028570741415, + 0.006175250746309757, + 0.005385912023484707, + 0.003763267071917653, + 0.008589932695031166, + 0.010054266080260277, + 0.0068662771955132484, + 0.004930888768285513, + 0.0066100070253014565, + 0.005160401575267315, + 0.005426323506981134, + 0.003249096916988492, + 0.004981074016541243, + 0.005419011693447828, + 0.010229963809251785, + 0.012826623395085335, + 0.011283395811915398, + 0.010165182873606682, + 0.009637129493057728, + 0.006584441289305687, + 0.011501156724989414, + 0.013377328403294086 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.008459093980491161 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a pale, slender human hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a stark, absolute black non-reflective surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the hand positioned with the wrist arched back and fingers curled inward with the index and middle fingers slightly apart?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lighting a sharp, key-like light from the side creating high contrast with bright highlights on the nails and deep shadows between the fingers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera angle slightly low and is the camera static?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04866999387741089 + ], + "opposite_logprob_tokens": [ + -3.0486700534820557 + ], + "logprob_reward": -0.04866999387741089, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04866999387741089, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1636543720960617, + 0.06077081710100174, + 0.1488519161939621, + 0.15892605483531952, + 0.08074797689914703, + 0.1863422393798828, + 0.07567594945430756, + 0.043722834438085556, + 0.0432964451611042, + 0.038011334836483 + ], + "credit_map": [ + 0.015378464013338089, + 0.01419073436409235, + 0.02470199204981327, + 0.014659663662314415, + 0.023749954998493195, + 0.02661050111055374, + 0.027276048436760902, + 0.017087003216147423, + 0.008918928913772106, + 0.008173925802111626, + 0.01047973521053791, + 0.006318800617009401, + 0.00704595260322094, + 0.007877192460000515, + 0.007377280853688717, + 0.004579000640660524, + 0.019738471135497093, + 0.025586873292922974, + 0.019393157213926315, + 0.011697327718138695, + 0.012481195852160454, + 0.022754477337002754, + 0.023299340158700943, + 0.013901074416935444, + 0.029041064903140068, + 0.024591324850916862, + 0.023111766204237938, + 0.02149386517703533, + 0.016959037631750107, + 0.01084830891340971, + 0.013203677721321583, + 0.019677013158798218, + 0.019930899143218994, + 0.006216029170900583, + 0.01726808212697506, + 0.008613030426204205, + 0.0043317233212292194, + 0.007831299677491188, + 0.007740675006061792, + 0.008816235698759556, + 0.042371828109025955, + 0.014913873746991158, + 0.036969058215618134, + 0.018103305250406265, + 0.010992487892508507, + 0.01046667993068695, + 0.022488543763756752, + 0.030036456882953644, + 0.02187458984553814, + 0.005959805566817522, + 0.01578964665532112, + 0.0055669816210865974, + 0.004387149587273598, + 0.005826311651617289, + 0.008467106148600578, + 0.007804356049746275, + 0.010289456695318222, + 0.0027508458588272333, + 0.007567108608782291, + 0.004973056260496378, + 0.003296258859336376, + 0.00258423900231719, + 0.0065920851193368435, + 0.00566978519782424, + 0.005418836604803801, + 0.003701667068526149, + 0.005680696573108435, + 0.006756692659109831, + 0.004350637551397085, + 0.004152373876422644, + 0.0065199630334973335, + 0.006715577561408281, + 0.005194965749979019, + 0.0035865488462150097, + 0.005409045610576868, + 0.0034489270765334368, + 0.00524200638756156, + 0.004355276003479958, + 0.004496205132454634, + 0.006278358865529299 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 3 + ], + "selected_reward": -0.04866999387741089 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a pale, slender human hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a stark, absolute black non-reflective surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the hand positioned with the wrist arched back and fingers curled inward with the index and middle fingers slightly apart?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lighting a sharp, key-like light from the side creating high contrast with bright highlights on the nails and deep shadows between the fingers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera angle slightly low and is the camera static?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"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": [ + 0.22314311563968658, + 0.06516954302787781, + 0.12740643322467804, + 0.08462947607040405, + 0.1342029720544815, + 0.06570225954055786, + 0.1077146977186203, + 0.07515206933021545, + 0.07029601186513901, + 0.04658345878124237 + ], + "credit_map": [ + 0.0134676368907094, + 0.014468137174844742, + 0.017399754375219345, + 0.013673363253474236, + 0.0714363381266594, + 0.05509793758392334, + 0.019147200509905815, + 0.018452730029821396, + 0.007181820459663868, + 0.009538976475596428, + 0.01034708134829998, + 0.007848859764635563, + 0.008992986753582954, + 0.008237343281507492, + 0.00871730875223875, + 0.0043051657266914845, + 0.012954790145158768, + 0.015078887343406677, + 0.01704435609281063, + 0.01365391444414854, + 0.013417945243418217, + 0.012310998514294624, + 0.02348381280899048, + 0.019461728632450104, + 0.012268410064280033, + 0.01199729461222887, + 0.012409861199557781, + 0.008235222660005093, + 0.006822776049375534, + 0.008459712378680706, + 0.01010182686150074, + 0.014334374107420444, + 0.014273202046751976, + 0.012263067997992039, + 0.02108047716319561, + 0.012635708786547184, + 0.009562202729284763, + 0.012602935545146465, + 0.02021096646785736, + 0.03157441318035126, + 0.008139211684465408, + 0.005935778841376305, + 0.008157254196703434, + 0.009830762632191181, + 0.005392484832555056, + 0.00451112724840641, + 0.011314276605844498, + 0.01242135651409626, + 0.0118505097925663, + 0.008109807036817074, + 0.013943199999630451, + 0.01154583040624857, + 0.010166377760469913, + 0.018767911940813065, + 0.015946950763463974, + 0.017384106293320656, + 0.009249343536794186, + 0.0061788856983184814, + 0.01304507628083229, + 0.013744879513978958, + 0.008526669815182686, + 0.005646996200084686, + 0.009628940373659134, + 0.009131277911365032, + 0.008308183401823044, + 0.006134623661637306, + 0.011342352256178856, + 0.009588400833308697, + 0.0053147501312196255, + 0.007758304011076689, + 0.010432617738842964, + 0.011416779831051826, + 0.005814060103148222, + 0.0039476873353123665, + 0.00691570620983839, + 0.0038780986797064543, + 0.0040225801058113575, + 0.006206430960446596, + 0.006797445006668568, + 0.009001454338431358 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 2 + ], + "selected_reward": -0.014200471341609955 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a pale, slender human hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a stark, absolute black non-reflective surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the hand positioned with the wrist arched back and fingers curled inward with the index and middle fingers slightly apart?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lighting a sharp, key-like light from the side creating high contrast with bright highlights on the nails and deep shadows between the fingers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera angle slightly low and is the camera static?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006740451790392399 + ], + "opposite_logprob_tokens": [ + -5.006740570068359 + ], + "logprob_reward": -0.006740451790392399, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006740451790392399, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2375103235244751, + 0.12756024301052094, + 0.16157132387161255, + 0.07263357937335968, + 0.1041613295674324, + 0.0743805319070816, + 0.053112369030714035, + 0.03741602599620819, + 0.06384218484163284, + 0.06781206279993057 + ], + "credit_map": [ + 0.020224357023835182, + 0.023531876504421234, + 0.03419400379061699, + 0.017648613080382347, + 0.03327091410756111, + 0.041777294129133224, + 0.03911067917943001, + 0.027752574533224106, + 0.016304131597280502, + 0.017576446756720543, + 0.02158971317112446, + 0.011055505834519863, + 0.01476715411990881, + 0.01610243134200573, + 0.01935449242591858, + 0.010810378938913345, + 0.0203403327614069, + 0.030964307487010956, + 0.019790610298514366, + 0.014849635772407055, + 0.014939520508050919, + 0.018995676189661026, + 0.024912940338253975, + 0.01677830144762993, + 0.01407794002443552, + 0.010690344497561455, + 0.011040696874260902, + 0.008554470725357533, + 0.008336469531059265, + 0.005526450928300619, + 0.006889647338539362, + 0.007517565507441759, + 0.020715901628136635, + 0.012492364272475243, + 0.01803511567413807, + 0.010045777074992657, + 0.005062924697995186, + 0.008735385723412037, + 0.01334883738309145, + 0.015725020319223404, + 0.01731790043413639, + 0.007218654733151197, + 0.013270989060401917, + 0.009010488167405128, + 0.005399554036557674, + 0.004796496592462063, + 0.009598791599273682, + 0.007767656352370977, + 0.011509409174323082, + 0.006672773044556379, + 0.008469115011394024, + 0.005300972610712051, + 0.004607247654348612, + 0.003862265031784773, + 0.006570893805474043, + 0.00611968943849206, + 0.006878249812871218, + 0.0027842053677886724, + 0.0053098020143806934, + 0.005763381253927946, + 0.003540925681591034, + 0.0029616551473736763, + 0.004368724767118692, + 0.005809082183986902, + 0.009458495303988457, + 0.005522280465811491, + 0.008039523847401142, + 0.007582962047308683, + 0.006971664261072874, + 0.005810785572975874, + 0.005979937966912985, + 0.01447653491050005, + 0.009239552542567253, + 0.00678444467484951, + 0.009357409551739693, + 0.005645548924803734, + 0.010117834433913231, + 0.007641638163477182, + 0.00871504656970501, + 0.010310585610568523 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 1 + ], + "selected_reward": -0.006740451790392399 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a pale, slender human hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a stark, absolute black non-reflective surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the hand positioned with the wrist arched back and fingers curled inward with the index and middle fingers slightly apart?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lighting a sharp, key-like light from the side creating high contrast with bright highlights on the nails and deep shadows between the fingers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera angle slightly low and is the camera static?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08904153108596802 + ], + "opposite_logprob_tokens": [ + -2.4640414714813232 + ], + "logprob_reward": -0.08904153108596802, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08904153108596802, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22147655487060547, + 0.11774051189422607, + 0.08788441121578217, + 0.0840597078204155, + 0.2325037717819214, + 0.050743501633405685, + 0.026117607951164246, + 0.04107725992798805, + 0.04962238296866417, + 0.08877427130937576 + ], + "credit_map": [ + 0.019615136086940765, + 0.01597864180803299, + 0.021395549178123474, + 0.014918224886059761, + 0.04926755279302597, + 0.04756703972816467, + 0.03443429246544838, + 0.018300127238035202, + 0.015745382755994797, + 0.015452025458216667, + 0.017809052020311356, + 0.012428884394466877, + 0.016808493062853813, + 0.01478582713752985, + 0.01757103018462658, + 0.007139828056097031, + 0.011969641782343388, + 0.010510662570595741, + 0.01233926136046648, + 0.009721367619931698, + 0.00942085962742567, + 0.011451583355665207, + 0.011392333544790745, + 0.011078698560595512, + 0.014574402943253517, + 0.01137001346796751, + 0.011900302022695541, + 0.007171190343797207, + 0.00975865963846445, + 0.009739778004586697, + 0.007784231565892696, + 0.011761126108467579, + 0.028807401657104492, + 0.02347254939377308, + 0.035464808344841, + 0.018276430666446686, + 0.01955263502895832, + 0.025825683027505875, + 0.02991185523569584, + 0.05119240656495094, + 0.008003472350537777, + 0.00439782440662384, + 0.007051773369312286, + 0.003550242865458131, + 0.003637998830527067, + 0.004603616893291473, + 0.007827444933354855, + 0.011671128682792187, + 0.0046274615451693535, + 0.0014630843652412295, + 0.005057381931692362, + 0.002148705068975687, + 0.0021083212923258543, + 0.0015886763576418161, + 0.0031438437290489674, + 0.005980133544653654, + 0.004949907306581736, + 0.00340058165602386, + 0.006277179811149836, + 0.008256394416093826, + 0.0028038653545081615, + 0.00375500600785017, + 0.006213491316884756, + 0.005420837085694075, + 0.0067258612252771854, + 0.004015443846583366, + 0.006920260842889547, + 0.005967673379927874, + 0.004571221768856049, + 0.00592384347692132, + 0.005839596036821604, + 0.009658481925725937, + 0.011183615773916245, + 0.00855494849383831, + 0.012743175029754639, + 0.008387615904211998, + 0.012311353348195553, + 0.010008827783167362, + 0.010124373249709606, + 0.015460360795259476 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 1 + ], + "selected_reward": -0.08904153108596802 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "Captured from a slightly low angle with a static camera, this ultra-detailed macro shot focuses intensely on a pale, slender hand suspended against a stark, absolute black background, a non-reflective blank. The hand is suspended in a state of predatory tension, its wrist gracefully arched back, poised to strike with the deadly grace of a snake's head. Its skin is an almost translucent alabaster, so fine that delicate, faint blue veins can be seen beneath the surface. The long, graceful fingers curl inward, but not in a fist; instead, the index and middle fingers are slightly apart and pointed forward, mimicking fangs poised to thrust. A sharp, key-like light from the side carves the hand out of the darkness, creating a dramatic black-and-white contrast. Bright highlights gleam from the spotless, natural nails and depict the tense tendons on the back of the palm, while deep, velvety shadows pool in the crevices between the fingers. Every muscle is visibly tensed, a subtle, controlled effort exerted, creating a palpable aura of imminent, precise, and threatening action.", + "seed": 8558, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000156/group_6/index_41_seed_8558_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": -0.0572314677760005, + "vlm_logprob": -0.0572314677760005, + "vlm_yes_no_margin": 3.55, + "vlm_token_credit": -0.0572314677760005, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a pale, slender human hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a stark, absolute black non-reflective surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the hand positioned with the wrist arched back and fingers curled inward with the index and middle fingers slightly apart?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lighting a sharp, key-like light from the side creating high contrast with bright highlights on the nails and deep shadows between the fingers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera angle slightly low and is the camera static?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012746528722345829 + ], + "opposite_logprob_tokens": [ + -4.762746334075928 + ], + "logprob_reward": -0.012746528722345829, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.012746528722345829, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06510011106729507, + 0.04646151885390282, + 0.05342414975166321, + 0.033234383910894394, + 0.00909013208001852, + 0.01749052107334137, + 0.25811854004859924, + 0.03406408429145813, + 0.11896374076604843, + 0.36405283212661743 + ], + "credit_map": [ + 0.00932727474719286, + 0.0065920716151595116, + 0.010334746912121773, + 0.0073230373673141, + 0.007483021356165409, + 0.0054224831983447075, + 0.007938846945762634, + 0.010678625665605068, + 0.003358070272952318, + 0.007616906892508268, + 0.009498740546405315, + 0.006216070149093866, + 0.005524578504264355, + 0.004612440709024668, + 0.005609924905002117, + 0.004024788271635771, + 0.004338355734944344, + 0.008949772454798222, + 0.013920461758971214, + 0.008818047121167183, + 0.007171098608523607, + 0.004110424313694239, + 0.002890628296881914, + 0.0032253635581582785, + 0.003386765718460083, + 0.004468907136470079, + 0.009316373616456985, + 0.005541671067476273, + 0.003546329215168953, + 0.002331711817532778, + 0.0018767507281154394, + 0.0027658732142299414, + 0.0007221686537377536, + 0.0012692961608991027, + 0.0020525073632597923, + 0.0016118095954880118, + 0.0008574813255108893, + 0.0007346029742620885, + 0.0007190852193161845, + 0.0011231808457523584, + 0.001693683909252286, + 0.002372976392507553, + 0.002877837512642145, + 0.002696085488423705, + 0.002074274467304349, + 0.0017822989029809833, + 0.0016836319118738174, + 0.002309731673449278, + 0.012254721485078335, + 0.01476301159709692, + 0.08549412339925766, + 0.032257188111543655, + 0.005794962868094444, + 0.012102591805160046, + 0.007339990697801113, + 0.08811192214488983, + 0.005637308117002249, + 0.004465978592634201, + 0.006777812261134386, + 0.003564781043678522, + 0.0024596811272203922, + 0.002572052413597703, + 0.003075195709243417, + 0.00551127502694726, + 0.020321818068623543, + 0.014198926277458668, + 0.025089126080274582, + 0.010614973492920399, + 0.010208995081484318, + 0.01048124860972166, + 0.015844494104385376, + 0.012204165570437908, + 0.03092183731496334, + 0.04986792430281639, + 0.08768371492624283, + 0.03016597591340542, + 0.025285881012678146, + 0.02233361452817917, + 0.04851463437080383, + 0.0692792609333992 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 6, + 8 + ], + "selected_reward": -0.012746528722345829 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a pale, slender human hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a stark, absolute black non-reflective surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the hand positioned with the wrist arched back and fingers curled inward with the index and middle fingers slightly apart?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lighting a sharp, key-like light from the side creating high contrast with bright highlights on the nails and deep shadows between the fingers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera angle slightly low and is the camera static?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.16037379205226898 + ], + "opposite_logprob_tokens": [ + -1.9103738069534302 + ], + "logprob_reward": -0.16037379205226898, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.16037379205226898, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08818509429693222, + 0.18277589976787567, + 0.1405632644891739, + 0.04221760109066963, + 0.04458998888731003, + 0.03704805299639702, + 0.3441484570503235, + 0.026308787986636162, + 0.0578848272562027, + 0.03627806156873703 + ], + "credit_map": [ + 0.014031988568603992, + 0.010876879096031189, + 0.013892372138798237, + 0.013501576147973537, + 0.008253056555986404, + 0.005956314038485289, + 0.008023595437407494, + 0.013649305328726768, + 0.013633627444505692, + 0.026130761951208115, + 0.022821802645921707, + 0.03322572261095047, + 0.02383352257311344, + 0.0207343939691782, + 0.021492060273885727, + 0.020904002711176872, + 0.011114224791526794, + 0.021488074213266373, + 0.0361572690308094, + 0.023136401548981667, + 0.017112217843532562, + 0.011994881555438042, + 0.00895289983600378, + 0.01060729380697012, + 0.004696156829595566, + 0.007196372840553522, + 0.010537047870457172, + 0.006674271076917648, + 0.0055341715924441814, + 0.002581208012998104, + 0.002050075214356184, + 0.002948288805782795, + 0.004919085651636124, + 0.004653905052691698, + 0.012701114639639854, + 0.008305596187710762, + 0.003698376240208745, + 0.0028734812512993813, + 0.0026361255440860987, + 0.004802298732101917, + 0.0046844640746712685, + 0.0050544836558401585, + 0.007697717752307653, + 0.006011372897773981, + 0.003153699915856123, + 0.003228485817089677, + 0.003050068626180291, + 0.00416775606572628, + 0.016124771907925606, + 0.02070511318743229, + 0.11674551665782928, + 0.04200043901801109, + 0.007529471535235643, + 0.01609637774527073, + 0.009580530226230621, + 0.11536619812250137, + 0.0032889109570533037, + 0.006560392677783966, + 0.004166232887655497, + 0.00312602031044662, + 0.0028315214440226555, + 0.0014053178019821644, + 0.0019179433584213257, + 0.0030124459881335497, + 0.010991014540195465, + 0.007796307560056448, + 0.011023552156984806, + 0.005323618650436401, + 0.004391822498291731, + 0.005526402499526739, + 0.006899354513734579, + 0.005932749714702368, + 0.005194723140448332, + 0.004373061936348677, + 0.005794273689389229, + 0.0038213240914046764, + 0.0029981695115566254, + 0.002463960787281394, + 0.006530008744448423, + 0.005102534778416157 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 1, + 2 + ], + "selected_reward": -0.16037379205226898 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a pale, slender human hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a stark, absolute black non-reflective surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the hand positioned with the wrist arched back and fingers curled inward with the index and middle fingers slightly apart?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lighting a sharp, key-like light from the side creating high contrast with bright highlights on the nails and deep shadows between the fingers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera angle slightly low and is the camera static?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009788372553884983 + ], + "opposite_logprob_tokens": [ + -4.634788513183594 + ], + "logprob_reward": -0.009788372553884983, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009788372553884983, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11420629918575287, + 0.14167386293411255, + 0.15965251624584198, + 0.03635747358202934, + 0.060376156121492386, + 0.05154945328831673, + 0.10672523081302643, + 0.15747474133968353, + 0.0876634493470192, + 0.08432082831859589 + ], + "credit_map": [ + 0.018958423286676407, + 0.014518966898322105, + 0.016218580305576324, + 0.011474703438580036, + 0.015282981097698212, + 0.00894726812839508, + 0.011543684639036655, + 0.0172616858035326, + 0.008863294497132301, + 0.01657186634838581, + 0.022878259420394897, + 0.0194233451038599, + 0.016906660050153732, + 0.014211280271410942, + 0.025785163044929504, + 0.017033999785780907, + 0.013832730241119862, + 0.017729241400957108, + 0.04774109274148941, + 0.031084353104233742, + 0.015386836603283882, + 0.014546230435371399, + 0.008743891492486, + 0.010588143020868301, + 0.00344888586550951, + 0.005182759370654821, + 0.006799258291721344, + 0.0053603085689246655, + 0.005309506319463253, + 0.0028407287318259478, + 0.0031767438631504774, + 0.004239282105118036, + 0.005244385451078415, + 0.007976917549967766, + 0.008560078218579292, + 0.009168851189315319, + 0.007614227943122387, + 0.006698388606309891, + 0.006541724316775799, + 0.008571585640311241, + 0.005104898940771818, + 0.008523548021912575, + 0.00777570391073823, + 0.007205219939351082, + 0.005786596331745386, + 0.004175481386482716, + 0.006093279924243689, + 0.0068847257643938065, + 0.01026066392660141, + 0.021268488839268684, + 0.029611237347126007, + 0.014079822227358818, + 0.00887453742325306, + 0.00721572432667017, + 0.00794199388474226, + 0.007472760044038296, + 0.021744849160313606, + 0.02618616260588169, + 0.029560377821326256, + 0.020233890041708946, + 0.013649633154273033, + 0.011800024658441544, + 0.017859866842627525, + 0.016439929604530334, + 0.014810462482273579, + 0.01129293441772461, + 0.01615002378821373, + 0.00813443586230278, + 0.006899432744830847, + 0.008263981901109219, + 0.013086183927953243, + 0.00902599561959505, + 0.010065498761832714, + 0.0091589679941535, + 0.01876109093427658, + 0.007018445059657097, + 0.006407284643501043, + 0.004807162564247847, + 0.009819978848099709, + 0.018282398581504822 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 7, + 1 + ], + "selected_reward": -0.009788372553884983 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a pale, slender human hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a stark, absolute black non-reflective surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the hand positioned with the wrist arched back and fingers curled inward with the index and middle fingers slightly apart?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lighting a sharp, key-like light from the side creating high contrast with bright highlights on the nails and deep shadows between the fingers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera angle slightly low and is the camera static?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014203292317688465 + ], + "opposite_logprob_tokens": [ + -4.264203071594238 + ], + "logprob_reward": -0.014203292317688465, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014203292317688465, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10171127319335938, + 0.1369764655828476, + 0.04598453268408775, + 0.03347838670015335, + 0.08045049756765366, + 0.03860574960708618, + 0.34000280499458313, + 0.056585852056741714, + 0.07130707800388336, + 0.09489736706018448 + ], + "credit_map": [ + 0.01658329740166664, + 0.014636090956628323, + 0.012919926084578037, + 0.01539600919932127, + 0.011043420992791653, + 0.00874543096870184, + 0.00802953727543354, + 0.014357559382915497, + 0.010055113583803177, + 0.023163866251707077, + 0.019722601398825645, + 0.02121836133301258, + 0.01675356924533844, + 0.015282392501831055, + 0.015691839158535004, + 0.015088725835084915, + 0.00427820673212409, + 0.008970487862825394, + 0.0063996696844697, + 0.006925590801984072, + 0.006946152076125145, + 0.004039111081510782, + 0.004364147782325745, + 0.004061168059706688, + 0.0029940863605588675, + 0.00420207716524601, + 0.004791222978383303, + 0.00505693256855011, + 0.0038575376383960247, + 0.00438073743134737, + 0.0027494975365698338, + 0.005446294788271189, + 0.00870569609105587, + 0.00861594919115305, + 0.025635670870542526, + 0.01601032167673111, + 0.0057718208990991116, + 0.0037147218827158213, + 0.0039032769855111837, + 0.0080930320546031, + 0.005043945275247097, + 0.004610980860888958, + 0.007160936947911978, + 0.00646757148206234, + 0.0041525098495185375, + 0.0035874524619430304, + 0.0031057174783200026, + 0.004476635251194239, + 0.016977591440081596, + 0.023117849603295326, + 0.1183394119143486, + 0.03927736356854439, + 0.009524492546916008, + 0.015632372349500656, + 0.011534236371517181, + 0.10559947043657303, + 0.007188647985458374, + 0.00877484679222107, + 0.008750082924962044, + 0.0071670436300337315, + 0.005157842300832272, + 0.004533824976533651, + 0.006010358687490225, + 0.009003207087516785, + 0.007204659748822451, + 0.005726453382521868, + 0.025121979415416718, + 0.008876170963048935, + 0.004182541277259588, + 0.0042100222781300545, + 0.005081370007246733, + 0.01090388372540474, + 0.011773592792451382, + 0.009168010205030441, + 0.020760178565979004, + 0.008072178810834885, + 0.0085296630859375, + 0.005399120971560478, + 0.011576367542147636, + 0.01961825042963028 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 1, + 0 + ], + "selected_reward": -0.014203292317688465 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a pale, slender human hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a stark, absolute black non-reflective surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the hand positioned with the wrist arched back and fingers curled inward with the index and middle fingers slightly apart?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lighting a sharp, key-like light from the side creating high contrast with bright highlights on the nails and deep shadows between the fingers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera angle slightly low and is the camera static?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08904535323381424 + ], + "opposite_logprob_tokens": [ + -2.464045286178589 + ], + "logprob_reward": -0.08904535323381424, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08904535323381424, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12454823404550552, + 0.10324154794216156, + 0.09687504172325134, + 0.048768218606710434, + 0.03259946405887604, + 0.03527885675430298, + 0.27236270904541016, + 0.08825290203094482, + 0.13551181554794312, + 0.06256125122308731 + ], + "credit_map": [ + 0.02062792144715786, + 0.012975172139704227, + 0.022142138332128525, + 0.01759372092783451, + 0.011104884557425976, + 0.008307067677378654, + 0.013450486585497856, + 0.01834682933986187, + 0.00615497212857008, + 0.014714010991156101, + 0.011386321857571602, + 0.012378759682178497, + 0.014964759349822998, + 0.012894661165773869, + 0.016515886411070824, + 0.014232169836759567, + 0.007459982763975859, + 0.026786739006638527, + 0.01781630888581276, + 0.011334562674164772, + 0.015471852384507656, + 0.007590032648295164, + 0.004950063768774271, + 0.005465491209179163, + 0.004505612887442112, + 0.008349400945007801, + 0.01020374707877636, + 0.006997244898229837, + 0.0059273820370435715, + 0.004236999433487654, + 0.0029041622765362263, + 0.005643664393573999, + 0.0030615918803960085, + 0.004492125473916531, + 0.006198478862643242, + 0.004502397030591965, + 0.004759268369525671, + 0.0026128743775188923, + 0.0033259615302085876, + 0.003646765137091279, + 0.003923967480659485, + 0.004192874301224947, + 0.0065016294829547405, + 0.006301881279796362, + 0.0038725105114281178, + 0.0026540469843894243, + 0.0028039137832820415, + 0.005028031766414642, + 0.016404731199145317, + 0.016308438032865524, + 0.08092297613620758, + 0.03282096982002258, + 0.009448894299566746, + 0.014547680504620075, + 0.013599696569144726, + 0.08830930292606354, + 0.00973475445061922, + 0.010637126863002777, + 0.02127552591264248, + 0.011552631855010986, + 0.007734636310487986, + 0.0070069157518446445, + 0.010456139221787453, + 0.009855164214968681, + 0.020036891102790833, + 0.01531219482421875, + 0.03149799630045891, + 0.011978032998740673, + 0.010746908374130726, + 0.012499985285103321, + 0.019045572727918625, + 0.014394224621355534, + 0.006928944494575262, + 0.0063254088163375854, + 0.015200705267488956, + 0.005781253799796104, + 0.004886066075414419, + 0.0037962920032441616, + 0.007364867255091667, + 0.012277710251510143 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 8, + 0 + ], + "selected_reward": -0.08904535323381424 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "Captured from a slightly low angle with a static camera, this ultra-detailed macro shot focuses intensely on a pale, slender hand suspended against a stark, absolute black background, a non-reflective blank. The hand is suspended in a state of predatory tension, its wrist gracefully arched back, poised to strike with the deadly grace of a snake's head. Its skin is an almost translucent alabaster, so fine that delicate, faint blue veins can be seen beneath the surface. The long, graceful fingers curl inward, but not in a fist; instead, the index and middle fingers are slightly apart and pointed forward, mimicking fangs poised to thrust. A sharp, key-like light from the side carves the hand out of the darkness, creating a dramatic black-and-white contrast. Bright highlights gleam from the spotless, natural nails and depict the tense tendons on the back of the palm, while deep, velvety shadows pool in the crevices between the fingers. Every muscle is visibly tensed, a subtle, controlled effort exerted, creating a palpable aura of imminent, precise, and threatening action.", + "seed": 8559, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000156/group_6/index_41_seed_8559_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": -0.016536964662373067, + "vlm_logprob": -0.016536964662373067, + "vlm_yes_no_margin": 4.225, + "vlm_token_credit": -0.016536964662373067, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a pale, slender human hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a stark, absolute black non-reflective surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the hand positioned with the wrist arched back and fingers curled inward with the index and middle fingers slightly apart?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lighting a sharp, key-like light from the side creating high contrast with bright highlights on the nails and deep shadows between the fingers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera angle slightly low and is the camera static?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.013913542963564396 + ], + "opposite_logprob_tokens": [ + -4.513913631439209 + ], + "logprob_reward": -0.013913542963564396, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.013913542963564396, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14520958065986633, + 0.07750606536865234, + 0.0758805125951767, + 0.16400247812271118, + 0.1740323305130005, + 0.03406853973865509, + 0.0653904601931572, + 0.17131908237934113, + 0.039573315531015396, + 0.053017646074295044 + ], + "credit_map": [ + 0.021109245717525482, + 0.019118908792734146, + 0.020219983533024788, + 0.01933952607214451, + 0.014404406771063805, + 0.013872088864445686, + 0.021391114220023155, + 0.01575431227684021, + 0.009300136007368565, + 0.010276020504534245, + 0.013405893929302692, + 0.015312152914702892, + 0.004537047818303108, + 0.005207644775509834, + 0.006760541815310717, + 0.01270662434399128, + 0.01762363873422146, + 0.009053058922290802, + 0.014832701534032822, + 0.008993435651063919, + 0.0033353748731315136, + 0.005194380413740873, + 0.006139806471765041, + 0.010708116926252842, + 0.022172441706061363, + 0.025496285408735275, + 0.031476788222789764, + 0.029401930049061775, + 0.009470773860812187, + 0.009691178798675537, + 0.011773965321481228, + 0.02451910451054573, + 0.037356194108724594, + 0.021722273901104927, + 0.03164096549153328, + 0.0222890954464674, + 0.007626571226865053, + 0.01438394095748663, + 0.01043232623487711, + 0.028580954298377037, + 0.006529102101922035, + 0.003666351782158017, + 0.007769891060888767, + 0.003242528298869729, + 0.001605037716217339, + 0.0032277340069413185, + 0.0025017522275447845, + 0.0055261459201574326, + 0.009205386973917484, + 0.006310975644737482, + 0.013887523673474789, + 0.011699418537318707, + 0.003306294558569789, + 0.005189602263271809, + 0.005497716832906008, + 0.010293545201420784, + 0.023595105856657028, + 0.022889545187354088, + 0.027106251567602158, + 0.03250407427549362, + 0.01266621332615614, + 0.013359126634895802, + 0.015546254813671112, + 0.02365252375602722, + 0.003966012503951788, + 0.004974655341356993, + 0.009350528009235859, + 0.005572045221924782, + 0.0035158998798578978, + 0.0031402751337736845, + 0.0038897907361388206, + 0.005164110101759434, + 0.006349023897200823, + 0.004952582530677319, + 0.015939287841320038, + 0.005108669865876436, + 0.003396298037841916, + 0.005385247059166431, + 0.002967726206406951, + 0.00891881249845028 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 7, + 3 + ], + "selected_reward": -0.013913542963564396 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a pale, slender human hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a stark, absolute black non-reflective surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the hand positioned with the wrist arched back and fingers curled inward with the index and middle fingers slightly apart?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lighting a sharp, key-like light from the side creating high contrast with bright highlights on the nails and deep shadows between the fingers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera angle slightly low and is the camera static?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014211636036634445 + ], + "opposite_logprob_tokens": [ + -4.264211654663086 + ], + "logprob_reward": -0.014211636036634445, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014211636036634445, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2661172151565552, + 0.2592403292655945, + 0.02641785889863968, + 0.10888823866844177, + 0.1278039515018463, + 0.07430103421211243, + 0.046203237026929855, + 0.02121626026928425, + 0.023217864334583282, + 0.04659394919872284 + ], + "credit_map": [ + 0.02880457602441311, + 0.027058279141783714, + 0.04831739142537117, + 0.06313418596982956, + 0.013476505875587463, + 0.017470473423600197, + 0.02714669704437256, + 0.040709156543016434, + 0.0349709689617157, + 0.03039109706878662, + 0.045772116631269455, + 0.05311228334903717, + 0.014941496774554253, + 0.016080986708402634, + 0.02369506284594536, + 0.04027634859085083, + 0.004317932296544313, + 0.0036188452504575253, + 0.005668549798429012, + 0.002909178612753749, + 0.002647233661264181, + 0.0018267116975039244, + 0.0025469628162682056, + 0.0028824482578784227, + 0.01635742001235485, + 0.013809115625917912, + 0.020352309569716454, + 0.020551880821585655, + 0.006510763429105282, + 0.007054791320115328, + 0.007727365475147963, + 0.016524609178304672, + 0.014036830514669418, + 0.023545466363430023, + 0.022149786353111267, + 0.01995820924639702, + 0.009975627064704895, + 0.008904610760509968, + 0.010523301549255848, + 0.018710127100348473, + 0.013032727874815464, + 0.008579015731811523, + 0.013517437502741814, + 0.01200161874294281, + 0.003631570376455784, + 0.005358211230486631, + 0.005410067271441221, + 0.012770400382578373, + 0.006718744058161974, + 0.0049480656161904335, + 0.007579224184155464, + 0.009080405347049236, + 0.0024624397046864033, + 0.005030816420912743, + 0.004247791599482298, + 0.006135758012533188, + 0.0020001917146146297, + 0.003193745855242014, + 0.0033871973864734173, + 0.0034394124522805214, + 0.001644299947656691, + 0.0022673828061670065, + 0.0015597448218613863, + 0.003724287496879697, + 0.003170065116137266, + 0.002937727840617299, + 0.005013888236135244, + 0.0030216658487915993, + 0.001272087567485869, + 0.0022179100196808577, + 0.001978711225092411, + 0.003605811856687069, + 0.004315195605158806, + 0.0072519839741289616, + 0.012577870860695839, + 0.005255850963294506, + 0.0030776956118643284, + 0.0035990893375128508, + 0.0026972664054483175, + 0.00781900156289339 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 4 + ], + "selected_reward": -0.014211636036634445 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a pale, slender human hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a stark, absolute black non-reflective surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the hand positioned with the wrist arched back and fingers curled inward with the index and middle fingers slightly apart?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lighting a sharp, key-like light from the side creating high contrast with bright highlights on the nails and deep shadows between the fingers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera angle slightly low and is the camera static?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02057822048664093 + ], + "opposite_logprob_tokens": [ + -3.895578145980835 + ], + "logprob_reward": -0.02057822048664093, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.02057822048664093, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05260293185710907, + 0.2174738496541977, + 0.06399578601121902, + 0.15313513576984406, + 0.09114612638950348, + 0.11217601597309113, + 0.08542875945568085, + 0.05619027465581894, + 0.11237405985593796, + 0.05547698587179184 + ], + "credit_map": [ + 0.007303662598133087, + 0.0076885949820280075, + 0.007041524164378643, + 0.005441425833851099, + 0.004928491078317165, + 0.006146319676190615, + 0.00681511964648962, + 0.0072378008626401424, + 0.022810664027929306, + 0.019227659329771996, + 0.03541063144803047, + 0.05307590588927269, + 0.012777355499565601, + 0.013402448035776615, + 0.02685408852994442, + 0.03391510620713234, + 0.005927698221057653, + 0.01040660124272108, + 0.013858658261597157, + 0.008491341955959797, + 0.006056108977645636, + 0.006878469604998827, + 0.005675520747900009, + 0.0067013963125646114, + 0.02328803576529026, + 0.024501517415046692, + 0.027249189093708992, + 0.023239003494381905, + 0.008745762519538403, + 0.009812021628022194, + 0.013637462630867958, + 0.022662153467535973, + 0.016352351754903793, + 0.012489541433751583, + 0.014745496213436127, + 0.010244039818644524, + 0.008829698897898197, + 0.006868856493383646, + 0.006505914032459259, + 0.015110238455235958, + 0.018728893250226974, + 0.013400258496403694, + 0.026781069114804268, + 0.011984569951891899, + 0.005511556286364794, + 0.00826601404696703, + 0.006242827512323856, + 0.021260835230350494, + 0.009364506229758263, + 0.0106274439021945, + 0.01500626653432846, + 0.013175534084439278, + 0.00671136099845171, + 0.009021257050335407, + 0.00913467537611723, + 0.012387719936668873, + 0.004318772815167904, + 0.0065773725509643555, + 0.011426894925534725, + 0.011325071565806866, + 0.003463038941845298, + 0.0040749614126980305, + 0.005398688837885857, + 0.009605476632714272, + 0.012771173380315304, + 0.010707263834774494, + 0.019758136942982674, + 0.011015789583325386, + 0.016830524429678917, + 0.012095923535525799, + 0.010725234635174274, + 0.018470028415322304, + 0.005719801876693964, + 0.007759853266179562, + 0.01663082279264927, + 0.005021884571760893, + 0.0036364372354000807, + 0.0040070428512990475, + 0.003244198625907302, + 0.00945695023983717 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 3, + 8 + ], + "selected_reward": -0.02057822048664093 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a pale, slender human hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a stark, absolute black non-reflective surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the hand positioned with the wrist arched back and fingers curled inward with the index and middle fingers slightly apart?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lighting a sharp, key-like light from the side creating high contrast with bright highlights on the nails and deep shadows between the fingers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera angle slightly low and is the camera static?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0076294224709272385 + ], + "opposite_logprob_tokens": [ + -4.88262939453125 + ], + "logprob_reward": -0.0076294224709272385, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.0076294224709272385, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11850319057703018, + 0.190143883228302, + 0.055722903460264206, + 0.17715291678905487, + 0.20552855730056763, + 0.07405399531126022, + 0.030097596347332, + 0.029470067471265793, + 0.015127897262573242, + 0.10419904440641403 + ], + "credit_map": [ + 0.012962976470589638, + 0.014691747725009918, + 0.02503846026957035, + 0.023723118007183075, + 0.006846092641353607, + 0.009919807314872742, + 0.010156969539821148, + 0.015164022333920002, + 0.022401366382837296, + 0.021768145263195038, + 0.036990076303482056, + 0.03999736160039902, + 0.010118667036294937, + 0.01021963357925415, + 0.019119037315249443, + 0.029529599472880363, + 0.0068902974016964436, + 0.0076060062274336815, + 0.011582083068788052, + 0.011123865842819214, + 0.00269560469314456, + 0.0038232794031500816, + 0.004878478590399027, + 0.007123288232833147, + 0.027211250737309456, + 0.022386115044355392, + 0.03250807523727417, + 0.031230948865413666, + 0.01082900632172823, + 0.011548422276973724, + 0.013226312585175037, + 0.028212780132889748, + 0.029854092746973038, + 0.026380790397524834, + 0.047269683331251144, + 0.030338415876030922, + 0.01151698362082243, + 0.010832088068127632, + 0.016873300075531006, + 0.032463189214468, + 0.013372095301747322, + 0.009085244499146938, + 0.013559844344854355, + 0.011287480592727661, + 0.0037175421603024006, + 0.005476607009768486, + 0.005113156978040934, + 0.012442028149962425, + 0.004229975398629904, + 0.0026633720844984055, + 0.007255408447235823, + 0.004442804493010044, + 0.0011368400882929564, + 0.002558249281719327, + 0.0027038080152124166, + 0.005107139237225056, + 0.002548884367570281, + 0.004385571926832199, + 0.005941544193774462, + 0.004835478495806456, + 0.001940950402058661, + 0.002479011658579111, + 0.002460471587255597, + 0.004878156818449497, + 0.0015289963921532035, + 0.0017183683812618256, + 0.0025652798358350992, + 0.001743120839819312, + 0.0015935678966343403, + 0.0018237276235595345, + 0.0018965683411806822, + 0.0022582681849598885, + 0.011213081888854504, + 0.012484868988394737, + 0.03400961682200432, + 0.008301018737256527, + 0.008525364100933075, + 0.006330681499093771, + 0.00504370266571641, + 0.01829070784151554 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 1, + 3 + ], + "selected_reward": -0.0076294224709272385 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a pale, slender human hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a stark, absolute black non-reflective surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the hand positioned with the wrist arched back and fingers curled inward with the index and middle fingers slightly apart?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lighting a sharp, key-like light from the side creating high contrast with bright highlights on the nails and deep shadows between the fingers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera angle slightly low and is the camera static?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02635200135409832 + ], + "opposite_logprob_tokens": [ + -3.6513519287109375 + ], + "logprob_reward": -0.02635200135409832, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.02635200135409832, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08364754915237427, + 0.10829632729291916, + 0.062472596764564514, + 0.09764713048934937, + 0.26676806807518005, + 0.10614100843667984, + 0.17212358117103577, + 0.03150292858481407, + 0.02681739255785942, + 0.0445835143327713 + ], + "credit_map": [ + 0.010322795249521732, + 0.00908899400383234, + 0.016722749918699265, + 0.016166530549526215, + 0.0053086658008396626, + 0.007703131530433893, + 0.00804193690419197, + 0.01029274147003889, + 0.011484099552035332, + 0.010612326674163342, + 0.020719552412629128, + 0.02554011158645153, + 0.006031235679984093, + 0.006719167344272137, + 0.010932383127510548, + 0.016257448121905327, + 0.013601611368358135, + 0.004567818716168404, + 0.01199391856789589, + 0.008732019923627377, + 0.004218903370201588, + 0.005114083178341389, + 0.005347719416022301, + 0.008896516636013985, + 0.017722435295581818, + 0.009478636085987091, + 0.019569234922528267, + 0.016402432695031166, + 0.00535869924351573, + 0.0077811372466385365, + 0.0062803360633552074, + 0.015054216608405113, + 0.045989736914634705, + 0.043445296585559845, + 0.048545755445957184, + 0.03135019913315773, + 0.02113662101328373, + 0.015936195850372314, + 0.017315339297056198, + 0.04304890334606171, + 0.019290607422590256, + 0.015247967094182968, + 0.016293199732899666, + 0.013219538144767284, + 0.007709560450166464, + 0.008757540956139565, + 0.008101989515125751, + 0.017520615831017494, + 0.02701386623084545, + 0.018031835556030273, + 0.03299726918339729, + 0.029270775616168976, + 0.009732192382216454, + 0.018178412690758705, + 0.013807637616991997, + 0.023091597482562065, + 0.004170081112533808, + 0.00466818967834115, + 0.004109293222427368, + 0.0063889953307807446, + 0.002088315784931183, + 0.0029951375909149647, + 0.0023948880843818188, + 0.004688027314841747, + 0.0025422086473554373, + 0.002582496963441372, + 0.004718927666544914, + 0.003162143984809518, + 0.003132964251562953, + 0.003607541788369417, + 0.0024508724454790354, + 0.004620236344635487, + 0.005187283270061016, + 0.006146078929305077, + 0.011131037957966328, + 0.0043904734775424, + 0.0029526755679398775, + 0.004431720823049545, + 0.0029239216819405556, + 0.007420323323458433 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 6, + 1 + ], + "selected_reward": -0.02635200135409832 + } + ], + "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/0000156/group_7/index_tensor([7835, 7835])_group_7_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000156/group_7/index_tensor([7835, 7835])_group_7_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..a174e70a10525402bfbd627b69d9406f7d2a95b3 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000156/group_7/index_tensor([7835, 7835])_group_7_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "static, low-angle, wide-angle shot establishes a perfect central composition, dominated by a massive, brutalist concrete sculpture rendered in a bold, graphic novel art style. An impenetrable monolith composed of interlocking, sharply angled geometric panels, the sculpture stands tall against a desolate, empty sky. The entire scene is defined by a palette of rich black ink and stark, high-contrast colors; deep, solid black shadows, cast by a harsh, unseen light source, carve dramatic forms into the sculpture's rough, pitted concrete texture. The ground begins to collapse beneath its immense weight; initially, tiny, hairline cracks appear at the base of the sculpture, like precisely drawn ink lines. These cracks then quickly and decisively spiderweb outward, fracturing the barren earth in radial patterns. The action captures a sense of immense pressure as large chunks of the jagged ground crack and are visibly thrust upward, revealing a growing abyss of pure, inky blackness as the cracks widen and the earth itself crumbles.", + "seed": 17672, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000156/group_7/index_7835_seed_17672_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": -0.01676245592534542, + "vlm_logprob": -0.01676245592534542, + "vlm_yes_no_margin": 4.375, + "vlm_token_credit": -0.01676245592534542, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return 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 brutalist concrete sculpture composed of interlocking geometric panels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style a bold graphic novel art style with rich black ink and stark high-contrast colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene show the ground cracking at the base of the sculpture and fracturing outward in radial patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are large chunks of the jagged ground visibly thrust upward to reveal a growing abyss of pure blackness?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a static, low-angle, wide-angle shot establishing a perfect central composition?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03447234258055687 + ], + "opposite_logprob_tokens": [ + -3.5344722270965576 + ], + "logprob_reward": -0.03447234258055687, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.03447234258055687, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09060666710138321, + 0.029219811782240868, + 0.06275445967912674, + 0.11671782284975052, + 0.04144219309091568, + 0.04627125337719917, + 0.061938460916280746, + 0.10147683322429657, + 0.2501886487007141, + 0.19938385486602783 + ], + "credit_map": [ + 0.014455307275056839, + 0.012050481513142586, + 0.01365287508815527, + 0.010612444952130318, + 0.011900696903467178, + 0.006683055777102709, + 0.01063546072691679, + 0.010616345331072807, + 0.004007191862910986, + 0.004095757380127907, + 0.0041744341142475605, + 0.004260945599526167, + 0.0036387061700224876, + 0.0023385295644402504, + 0.0033733577001839876, + 0.0033308877609670162, + 0.005315191578119993, + 0.004498539492487907, + 0.009794731624424458, + 0.011545604094862938, + 0.0043447888456285, + 0.0032054234761744738, + 0.008375437930226326, + 0.015674741938710213, + 0.010114955715835094, + 0.004108488094061613, + 0.008844331838190556, + 0.02497175894677639, + 0.007855978794395924, + 0.006002436857670546, + 0.037688057869672775, + 0.01713181473314762, + 0.003006097860634327, + 0.003905761055648327, + 0.004470264073461294, + 0.008511921390891075, + 0.003114855382591486, + 0.005134184844791889, + 0.009047433733940125, + 0.004251671954989433, + 0.0053727589547634125, + 0.004107700195163488, + 0.005343702156096697, + 0.007426731754094362, + 0.0031702821142971516, + 0.005726971197873354, + 0.008869143202900887, + 0.006253961473703384, + 0.0050033265724778175, + 0.005127640441060066, + 0.006175947841256857, + 0.004662158899009228, + 0.004483554046601057, + 0.018223091959953308, + 0.013079886324703693, + 0.005182850640267134, + 0.008263329043984413, + 0.017302092164754868, + 0.010465936735272408, + 0.01045601349323988, + 0.023542620241642, + 0.013351299799978733, + 0.008693162351846695, + 0.009402381256222725, + 0.007415106985718012, + 0.10172605514526367, + 0.006806483957916498, + 0.0066892909817397594, + 0.10218381881713867, + 0.012048275209963322, + 0.007785423193126917, + 0.005534181371331215, + 0.02821279875934124, + 0.04954929277300835, + 0.016223590821027756, + 0.008000357076525688, + 0.050896693021059036, + 0.02139575034379959, + 0.006007845979183912, + 0.01909751072525978 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 3 + ], + "selected_reward": -0.03447234258055687 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return 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 brutalist concrete sculpture composed of interlocking geometric panels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style a bold graphic novel art style with rich black ink and stark high-contrast colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene show the ground cracking at the base of the sculpture and fracturing outward in radial patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are large chunks of the jagged ground visibly thrust upward to reveal a growing abyss of pure blackness?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a static, low-angle, wide-angle shot establishing a perfect central composition?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011090421117842197 + ], + "opposite_logprob_tokens": [ + -4.511090278625488 + ], + "logprob_reward": -0.011090421117842197, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011090421117842197, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.150248184800148, + 0.10121005028486252, + 0.06730253994464874, + 0.02148243971168995, + 0.040799155831336975, + 0.025661468505859375, + 0.07632111757993698, + 0.11428295075893402, + 0.20934611558914185, + 0.19334597885608673 + ], + "credit_map": [ + 0.02100714109838009, + 0.013849022798240185, + 0.02242719568312168, + 0.022698933258652687, + 0.01843775436282158, + 0.011437623761594296, + 0.020854024216532707, + 0.01953648030757904, + 0.014672700315713882, + 0.008710554800927639, + 0.010767794214189053, + 0.01747644878923893, + 0.01167154312133789, + 0.006309895776212215, + 0.017398307099938393, + 0.014202800579369068, + 0.004212906118482351, + 0.0042440458200871944, + 0.008529151789844036, + 0.014377878978848457, + 0.005042660515755415, + 0.004037454258650541, + 0.00830045435577631, + 0.01855798438191414, + 0.0017059891251847148, + 0.0010851932456716895, + 0.003465350717306137, + 0.0038883164525032043, + 0.0019081932259723544, + 0.002188318409025669, + 0.003390104044228792, + 0.0038509757723659277, + 0.0035154290962964296, + 0.003551215399056673, + 0.007141382433474064, + 0.008315074257552624, + 0.0037306391168385744, + 0.003603754099458456, + 0.007977880537509918, + 0.0029637820553034544, + 0.0027444087900221348, + 0.0020065566059201956, + 0.0029967036098241806, + 0.004583153873682022, + 0.0017580530839040875, + 0.003974812105298042, + 0.0038345088250935078, + 0.0037632694002240896, + 0.009684860706329346, + 0.00750252278521657, + 0.007494935765862465, + 0.006224732846021652, + 0.007527444511651993, + 0.018004195764660835, + 0.013883139938116074, + 0.0059992847964167595, + 0.008937939070165157, + 0.02193225920200348, + 0.00997613649815321, + 0.011215449310839176, + 0.02927693910896778, + 0.014098681509494781, + 0.009208029136061668, + 0.009637516923248768, + 0.01138475351035595, + 0.0634213238954544, + 0.008812693879008293, + 0.0065676067024469376, + 0.08007818460464478, + 0.01715431548655033, + 0.012211497873067856, + 0.009715721942484379, + 0.02277565933763981, + 0.0398513525724411, + 0.030841000378131866, + 0.008878407068550587, + 0.040274687111377716, + 0.021000640466809273, + 0.008340492844581604, + 0.021383754909038544 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 0 + ], + "selected_reward": -0.011090421117842197 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return 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 brutalist concrete sculpture composed of interlocking geometric panels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style a bold graphic novel art style with rich black ink and stark high-contrast colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene show the ground cracking at the base of the sculpture and fracturing outward in radial patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are large chunks of the jagged ground visibly thrust upward to reveal a growing abyss of pure blackness?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a static, low-angle, wide-angle shot establishing a perfect central composition?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005956161767244339 + ], + "opposite_logprob_tokens": [ + -5.130956172943115 + ], + "logprob_reward": -0.005956161767244339, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005956161767244339, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.02659669518470764, + 0.009799400344491005, + 0.01405931543558836, + 0.019178882241249084, + 0.031185347586870193, + 0.01419917680323124, + 0.019064176827669144, + 0.03128708899021149, + 0.1954410821199417, + 0.639188826084137 + ], + "credit_map": [ + 0.003908294718712568, + 0.003179645398631692, + 0.0038918517529964447, + 0.0033718179911375046, + 0.003485822817310691, + 0.0017095899675041437, + 0.0037665863055735826, + 0.003283088095486164, + 0.0016187827568501234, + 0.0010442690690979362, + 0.0016248428728431463, + 0.0014406649861484766, + 0.0012697818456217647, + 0.0005938218091614544, + 0.0010029603727161884, + 0.0012042769230902195, + 0.0011159727582708001, + 0.001029760460369289, + 0.0026354556903243065, + 0.0023510903120040894, + 0.001230988884344697, + 0.0013337862910702825, + 0.002347604837268591, + 0.0020146567840129137, + 0.0027083461172878742, + 0.0010751554509624839, + 0.0023970752954483032, + 0.004205101169645786, + 0.001269445987418294, + 0.001154825440607965, + 0.003085806267336011, + 0.003283129306510091, + 0.001488691894337535, + 0.0014102464774623513, + 0.003782037878409028, + 0.009047041647136211, + 0.0014655321137979627, + 0.0018731129821389914, + 0.010047540068626404, + 0.002071147318929434, + 0.001117417006753385, + 0.0013266524765640497, + 0.0011041199322789907, + 0.0028855944983661175, + 0.0007211347692646086, + 0.002218756126239896, + 0.003373536514118314, + 0.0014519658870995045, + 0.002148303436115384, + 0.002236188855022192, + 0.002459545386955142, + 0.0020194384269416332, + 0.001630220329388976, + 0.0034707903396338224, + 0.0027063179295510054, + 0.0023933728225529194, + 0.0032544720452278852, + 0.005363973323255777, + 0.003573458408936858, + 0.003486906411126256, + 0.005837768781930208, + 0.004194808658212423, + 0.002482103416696191, + 0.0030935986433178186, + 0.004988613538444042, + 0.08021029084920883, + 0.005141234491020441, + 0.005936256609857082, + 0.07930213212966919, + 0.008261663839221, + 0.005225963890552521, + 0.006374940276145935, + 0.0949699804186821, + 0.13000454008579254, + 0.08095661550760269, + 0.028042199090123177, + 0.13867893815040588, + 0.058625444769859314, + 0.025653444230556488, + 0.0822577252984047 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 7 + ], + "selected_reward": -0.005956161767244339 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return 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 brutalist concrete sculpture composed of interlocking geometric panels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style a bold graphic novel art style with rich black ink and stark high-contrast colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene show the ground cracking at the base of the sculpture and fracturing outward in radial patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are large chunks of the jagged ground visibly thrust upward to reveal a growing abyss of pure blackness?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a static, low-angle, wide-angle shot establishing a perfect central composition?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"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": [ + 0.09129398316144943, + 0.028671257197856903, + 0.04274452477693558, + 0.06903824955224991, + 0.09587813913822174, + 0.08201898634433746, + 0.0666293278336525, + 0.15582159161567688, + 0.11721287667751312, + 0.2506910562515259 + ], + "credit_map": [ + 0.013571061193943024, + 0.011763133108615875, + 0.014395360834896564, + 0.009909109212458134, + 0.01239551231265068, + 0.006151572801172733, + 0.011924339458346367, + 0.01118389517068863, + 0.0035621162969619036, + 0.0025636949576437473, + 0.00410956796258688, + 0.004238836467266083, + 0.0029340730980038643, + 0.0018864382291212678, + 0.00540316104888916, + 0.003973369486629963, + 0.003555529285222292, + 0.003030747640877962, + 0.007155091501772404, + 0.009188813157379627, + 0.0030139596201479435, + 0.002980221062898636, + 0.006463756784796715, + 0.007356409449130297, + 0.002476699184626341, + 0.001833838876336813, + 0.0042081247083842754, + 0.02174757421016693, + 0.002279429929330945, + 0.0026952389162033796, + 0.026591259986162186, + 0.007206080015748739, + 0.00596236577257514, + 0.005951109807938337, + 0.011329050175845623, + 0.027470886707305908, + 0.004408715292811394, + 0.0055134231224656105, + 0.02891283668577671, + 0.006329752039164305, + 0.00370403123088181, + 0.0034408196806907654, + 0.004723662044852972, + 0.02369876578450203, + 0.0035015931352972984, + 0.010331372730433941, + 0.027298755943775177, + 0.005319984629750252, + 0.007511953357607126, + 0.007614089176058769, + 0.006862767972052097, + 0.007010300178080797, + 0.005784474778920412, + 0.015679067000746727, + 0.011077937670052052, + 0.005088741425424814, + 0.01117731723934412, + 0.029771653935313225, + 0.011697364039719105, + 0.014229705557227135, + 0.04720936715602875, + 0.01955881342291832, + 0.011477298103272915, + 0.010700065642595291, + 0.008716476149857044, + 0.029537316411733627, + 0.005656568333506584, + 0.004411259666085243, + 0.04082426056265831, + 0.01166069321334362, + 0.009126625023782253, + 0.007279683835804462, + 0.03414809703826904, + 0.05352230370044708, + 0.0294223390519619, + 0.013291919603943825, + 0.05625202879309654, + 0.02552851289510727, + 0.011264367029070854, + 0.027261486276984215 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 7, + 8 + ], + "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 main subject a massive brutalist concrete sculpture composed of interlocking geometric panels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style a bold graphic novel art style with rich black ink and stark high-contrast colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene show the ground cracking at the base of the sculpture and fracturing outward in radial patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are large chunks of the jagged ground visibly thrust upward to reveal a growing abyss of pure blackness?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a static, low-angle, wide-angle shot establishing a perfect central composition?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026346080005168915 + ], + "opposite_logprob_tokens": [ + -3.65134596824646 + ], + "logprob_reward": -0.026346080005168915, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026346080005168915, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09202534705400467, + 0.06886136531829834, + 0.05595829337835312, + 0.21818195283412933, + 0.12256895750761032, + 0.04292678460478783, + 0.040237609297037125, + 0.08431318402290344, + 0.1117234006524086, + 0.16320307552814484 + ], + "credit_map": [ + 0.013394415378570557, + 0.00962630845606327, + 0.015362652018666267, + 0.010918708518147469, + 0.010518078692257404, + 0.007158063352108002, + 0.012229816988110542, + 0.012817303650081158, + 0.008208319544792175, + 0.007599320728331804, + 0.008108523674309254, + 0.012829728424549103, + 0.00823643896728754, + 0.004641886334866285, + 0.00984368659555912, + 0.009393456391990185, + 0.002876324113458395, + 0.002823655027896166, + 0.004673844203352928, + 0.01223872322589159, + 0.006470740307122469, + 0.003198903053998947, + 0.006282174494117498, + 0.0173939298838377, + 0.005750070791691542, + 0.0036604669876396656, + 0.00526463845744729, + 0.07942384481430054, + 0.006006953772157431, + 0.005154700949788094, + 0.09412848204374313, + 0.018792787566781044, + 0.006611375138163567, + 0.005592165514826775, + 0.01914541982114315, + 0.03486942499876022, + 0.005005755461752415, + 0.005588564556092024, + 0.03694181516766548, + 0.008814435452222824, + 0.002605590270832181, + 0.0026350326370447874, + 0.0033870802726596594, + 0.010496154427528381, + 0.002569672418758273, + 0.006355548743158579, + 0.01106616947799921, + 0.0038115379866212606, + 0.004485142882913351, + 0.004535313695669174, + 0.004439620766788721, + 0.004617687314748764, + 0.0035603067371994257, + 0.00824830960482359, + 0.005695983301848173, + 0.004655246622860432, + 0.00680032093077898, + 0.015055662952363491, + 0.007788392249494791, + 0.00926433689892292, + 0.020282531157135963, + 0.01105019636452198, + 0.007147128693759441, + 0.006924613378942013, + 0.003862137673422694, + 0.04355631023645401, + 0.0032782070338726044, + 0.0026762112975120544, + 0.045050401240587234, + 0.005477573722600937, + 0.0044050561264157295, + 0.0034175014588981867, + 0.019271399825811386, + 0.03979300707578659, + 0.016800547018647194, + 0.0062965345568954945, + 0.04255813732743263, + 0.015802307054400444, + 0.005473561584949493, + 0.017207598313689232 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 9, + 4 + ], + "selected_reward": -0.026346080005168915 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, low-angle, wide-angle shot establishes a perfect central composition, dominated by a massive, brutalist concrete sculpture rendered in a bold, graphic novel art style. An impenetrable monolith composed of interlocking, sharply angled geometric panels, the sculpture stands tall against a desolate, empty sky. The entire scene is defined by a palette of rich black ink and stark, high-contrast colors; deep, solid black shadows, cast by a harsh, unseen light source, carve dramatic forms into the sculpture's rough, pitted concrete texture. The ground begins to collapse beneath its immense weight; initially, tiny, hairline cracks appear at the base of the sculpture, like precisely drawn ink lines. These cracks then quickly and decisively spiderweb outward, fracturing the barren earth in radial patterns. The action captures a sense of immense pressure as large chunks of the jagged ground crack and are visibly thrust upward, revealing a growing abyss of pure, inky blackness as the cracks widen and the earth itself crumbles.", + "seed": 17673, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000156/group_7/index_7835_seed_17673_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": -0.00814493540674448, + "vlm_logprob": -0.00814493540674448, + "vlm_yes_no_margin": 5.125, + "vlm_token_credit": -0.00814493540674448, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return 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 brutalist concrete sculpture composed of interlocking geometric panels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style a bold graphic novel art style with rich black ink and stark high-contrast colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene show the ground cracking at the base of the sculpture and fracturing outward in radial patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are large chunks of the jagged ground visibly thrust upward to reveal a growing abyss of pure blackness?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a static, low-angle, wide-angle shot establishing a perfect central composition?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007724415510892868 + ], + "opposite_logprob_tokens": [ + -5.507724285125732 + ], + "logprob_reward": -0.007724415510892868, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.007724415510892868, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06336449831724167, + 0.0184953473508358, + 0.016504252329468727, + 0.014731889590620995, + 0.11719629168510437, + 0.056908007711172104, + 0.04402033984661102, + 0.049289051443338394, + 0.5553910732269287, + 0.06409917771816254 + ], + "credit_map": [ + 0.007947742938995361, + 0.007235625293105841, + 0.004953401163220406, + 0.0039475224912166595, + 0.009490812197327614, + 0.008957396261394024, + 0.008861439302563667, + 0.011970562860369682, + 0.0023964070715010166, + 0.002350939903408289, + 0.0021083359606564045, + 0.0013799407752230763, + 0.002701352583244443, + 0.0022769998759031296, + 0.0030620689503848553, + 0.0022193032782524824, + 0.002351514296606183, + 0.001629610895179212, + 0.001392718288116157, + 0.0010235434165224433, + 0.002714607398957014, + 0.002418006071820855, + 0.002080645877867937, + 0.002893606899306178, + 0.0018972752150148153, + 0.0013528433628380299, + 0.0019099715864285827, + 0.001076582120731473, + 0.0022780008148401976, + 0.0021472417283803225, + 0.002309568226337433, + 0.0017604072345420718, + 0.02056439034640789, + 0.009721493348479271, + 0.016544988378882408, + 0.00760747492313385, + 0.01701994799077511, + 0.014405470341444016, + 0.017413891851902008, + 0.013918635435402393, + 0.007601873017847538, + 0.003695569932460785, + 0.006334478966891766, + 0.006531374529004097, + 0.006989894900470972, + 0.007037999574095011, + 0.008478631265461445, + 0.01023818552494049, + 0.004116448108106852, + 0.004341767635196447, + 0.0039012660272419453, + 0.006322129629552364, + 0.006374425254762173, + 0.003514479845762253, + 0.008149039931595325, + 0.007300779223442078, + 0.003808761015534401, + 0.00677948584780097, + 0.002945716492831707, + 0.006949885748326778, + 0.007601573131978512, + 0.0076076495461165905, + 0.008202715776860714, + 0.005393263418227434, + 0.022512253373861313, + 0.22155188024044037, + 0.010150535963475704, + 0.017371540889143944, + 0.18187154829502106, + 0.04649890214204788, + 0.01647726260125637, + 0.03895717114210129, + 0.007445263676345348, + 0.010674345307052135, + 0.0035558410454541445, + 0.004153310786932707, + 0.01278957724571228, + 0.015586989000439644, + 0.003980236127972603, + 0.005913611967116594 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 4, + 9 + ], + "selected_reward": -0.007724415510892868 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return 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 brutalist concrete sculpture composed of interlocking geometric panels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style a bold graphic novel art style with rich black ink and stark high-contrast colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene show the ground cracking at the base of the sculpture and fracturing outward in radial patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are large chunks of the jagged ground visibly thrust upward to reveal a growing abyss of pure blackness?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a static, low-angle, wide-angle shot establishing a perfect central composition?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0067449514754116535 + ], + "opposite_logprob_tokens": [ + -5.006744861602783 + ], + "logprob_reward": -0.0067449514754116535, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.0067449514754116535, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09461965411901474, + 0.03125179931521416, + 0.02008354291319847, + 0.019734805449843407, + 0.028700578957796097, + 0.10509410500526428, + 0.04093578830361366, + 0.0696677416563034, + 0.43946829438209534, + 0.1504436731338501 + ], + "credit_map": [ + 0.013147642835974693, + 0.006953812204301357, + 0.007767934817820787, + 0.004734922666102648, + 0.014147444628179073, + 0.019308658316731453, + 0.01333220861852169, + 0.015227023512125015, + 0.004283493384718895, + 0.0015345740830525756, + 0.0036752349697053432, + 0.001599736511707306, + 0.004599249921739101, + 0.0066584995947778225, + 0.0026101055555045605, + 0.006290904246270657, + 0.0030185929499566555, + 0.0009411107166670263, + 0.0029976547230035067, + 0.0008988952031359076, + 0.0029419083148241043, + 0.004142851568758488, + 0.0013585211709141731, + 0.0037840078584849834, + 0.00322004989720881, + 0.0010639753891155124, + 0.002710724947974086, + 0.0009902823949232697, + 0.00305120344273746, + 0.0034615539480000734, + 0.002258226741105318, + 0.0029787863604724407, + 0.004253874532878399, + 0.0029138592071831226, + 0.003443238092586398, + 0.0031223881524056196, + 0.003932801540941, + 0.0020980520639568567, + 0.006282654590904713, + 0.0026537105441093445, + 0.021674124523997307, + 0.006047171540558338, + 0.009170548059046268, + 0.006745255086570978, + 0.010873429477214813, + 0.010327441617846489, + 0.011781606823205948, + 0.02847452089190483, + 0.005390106700360775, + 0.0033073527738451958, + 0.0032795104198157787, + 0.0069498601369559765, + 0.004247089847922325, + 0.004881627857685089, + 0.00728420726954937, + 0.005596030503511429, + 0.005299810320138931, + 0.011864437721669674, + 0.004637469537556171, + 0.010903041809797287, + 0.0085934828966856, + 0.009733334183692932, + 0.010084242559969425, + 0.00855192355811596, + 0.023279665037989616, + 0.16083243489265442, + 0.009648007340729237, + 0.014083421789109707, + 0.14194266498088837, + 0.032718196511268616, + 0.015598412603139877, + 0.04136548563838005, + 0.0166170634329319, + 0.025532297790050507, + 0.009702314622700214, + 0.010798005387187004, + 0.02876284345984459, + 0.024160565808415413, + 0.010636703111231327, + 0.0242338627576828 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 5 + ], + "selected_reward": -0.0067449514754116535 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive brutalist concrete sculpture composed of interlocking geometric panels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style a bold graphic novel art style with rich black ink and stark high-contrast colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene show the ground cracking at the base of the sculpture and fracturing outward in radial patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are large chunks of the jagged ground visibly thrust upward to reveal a growing abyss of pure blackness?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a static, low-angle, wide-angle shot establishing a perfect central composition?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0028224652633070946 + ], + "opposite_logprob_tokens": [ + -5.877822399139404 + ], + "logprob_reward": -0.0028224652633070946, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.0028224652633070946, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06262822449207306, + 0.024835223332047462, + 0.028031963855028152, + 0.03451881557703018, + 0.16033753752708435, + 0.045029930770397186, + 0.07416132837533951, + 0.06625878065824509, + 0.20385386049747467, + 0.3003443479537964 + ], + "credit_map": [ + 0.009253205731511116, + 0.004232826177030802, + 0.007713590748608112, + 0.002662955317646265, + 0.009255096316337585, + 0.011820470914244652, + 0.008195862174034119, + 0.009494219906628132, + 0.0043503339402377605, + 0.001584275858476758, + 0.003775469958782196, + 0.0014062612317502499, + 0.0037102142814546824, + 0.003018222516402602, + 0.0028004113119095564, + 0.00419003376737237, + 0.0043447245843708515, + 0.002114604925736785, + 0.0034855513367801905, + 0.0016827116487547755, + 0.0038900310173630714, + 0.004561575595289469, + 0.0037848378997296095, + 0.004167926497757435, + 0.005491096060723066, + 0.0024255376774817705, + 0.005067048594355583, + 0.002381348516792059, + 0.00496194651350379, + 0.0052390750497579575, + 0.0048901597037911415, + 0.004062602296471596, + 0.027908429503440857, + 0.012520520016551018, + 0.0239389818161726, + 0.011620956473052502, + 0.023588603362441063, + 0.020444493740797043, + 0.020313354209065437, + 0.020002201199531555, + 0.008988986723124981, + 0.002458184724673629, + 0.0037984547670930624, + 0.003274552756920457, + 0.004896144382655621, + 0.0052365511655807495, + 0.0055755190551280975, + 0.010801538825035095, + 0.009571075439453125, + 0.006236457731574774, + 0.008038615807890892, + 0.014021866023540497, + 0.009258816950023174, + 0.006412931717932224, + 0.010164515115320683, + 0.010457044467329979, + 0.005860034376382828, + 0.008324774913489819, + 0.0034457664005458355, + 0.00946963019669056, + 0.008558153174817562, + 0.00962830800563097, + 0.011672461405396461, + 0.00929965078830719, + 0.013491801917552948, + 0.056602202355861664, + 0.007387246936559677, + 0.007635666523128748, + 0.06061769276857376, + 0.022616151720285416, + 0.009546656161546707, + 0.025956444442272186, + 0.03349479287862778, + 0.047753967344760895, + 0.015158142894506454, + 0.01944124884903431, + 0.04570977762341499, + 0.06761663407087326, + 0.023406649008393288, + 0.04776313528418541 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 4 + ], + "selected_reward": -0.0028224652633070946 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return 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 brutalist concrete sculpture composed of interlocking geometric panels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style a bold graphic novel art style with rich black ink and stark high-contrast colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene show the ground cracking at the base of the sculpture and fracturing outward in radial patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are large chunks of the jagged ground visibly thrust upward to reveal a growing abyss of pure blackness?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a static, low-angle, wide-angle shot establishing a perfect central composition?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005249170120805502 + ], + "opposite_logprob_tokens": [ + -5.2552490234375 + ], + "logprob_reward": -0.005249170120805502, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005249170120805502, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11622180044651031, + 0.050317127257585526, + 0.024742072448134422, + 0.043879617005586624, + 0.11171426624059677, + 0.12178479135036469, + 0.15025410056114197, + 0.05379161238670349, + 0.16995620727539062, + 0.15733838081359863 + ], + "credit_map": [ + 0.018959056586027145, + 0.010863937437534332, + 0.011370076797902584, + 0.005506465677171946, + 0.017217962071299553, + 0.020912444218993187, + 0.012543576769530773, + 0.01884828507900238, + 0.006528228987008333, + 0.004046114161610603, + 0.005505310837179422, + 0.0030362692195922136, + 0.007307450752705336, + 0.007705821190029383, + 0.0056690932251513, + 0.01051884051412344, + 0.003010255517438054, + 0.001462407293729484, + 0.0034363670274615288, + 0.001694731879979372, + 0.003451941767707467, + 0.004738803021609783, + 0.0036354223266243935, + 0.0033121448941528797, + 0.004881122149527073, + 0.0032858599442988634, + 0.007034119684249163, + 0.0045469533652067184, + 0.006124899722635746, + 0.0050555383786559105, + 0.007909487932920456, + 0.005041635129600763, + 0.017870761454105377, + 0.006474137771874666, + 0.014644099399447441, + 0.007431692909449339, + 0.014467881061136723, + 0.011982309632003307, + 0.020585203543305397, + 0.01825817860662937, + 0.028231441974639893, + 0.005113829858601093, + 0.0067229606211185455, + 0.008249517530202866, + 0.008568924851715565, + 0.011386260390281677, + 0.012934586964547634, + 0.040577273815870285, + 0.015451991930603981, + 0.01183291058987379, + 0.015563515946269035, + 0.030647829174995422, + 0.018387921154499054, + 0.013280592858791351, + 0.022915178909897804, + 0.022174160927534103, + 0.005364757496863604, + 0.006868964526802301, + 0.0032401648350059986, + 0.006332137621939182, + 0.007345648016780615, + 0.009540913626551628, + 0.008592262864112854, + 0.006506762932986021, + 0.01864154264330864, + 0.026887673884630203, + 0.01137449499219656, + 0.01284752506762743, + 0.03919638693332672, + 0.029379848390817642, + 0.013377537950873375, + 0.018251191824674606, + 0.014992937445640564, + 0.0428008995950222, + 0.007547204382717609, + 0.007831635884940624, + 0.037593238055706024, + 0.028049809858202934, + 0.009189029224216938, + 0.009333615191280842 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 6 + ], + "selected_reward": -0.005249170120805502 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return 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 brutalist concrete sculpture composed of interlocking geometric panels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style a bold graphic novel art style with rich black ink and stark high-contrast colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene show the ground cracking at the base of the sculpture and fracturing outward in radial patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are large chunks of the jagged ground visibly thrust upward to reveal a growing abyss of pure blackness?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a static, low-angle, wide-angle shot establishing a perfect central composition?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018183674663305283 + ], + "opposite_logprob_tokens": [ + -4.018183708190918 + ], + "logprob_reward": -0.018183674663305283, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018183674663305283, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10110767930746078, + 0.030211277306079865, + 0.014554880559444427, + 0.029224468395113945, + 0.17333629727363586, + 0.08607854694128036, + 0.04650260508060455, + 0.11607016623020172, + 0.34508201479911804, + 0.0578320175409317 + ], + "credit_map": [ + 0.013015901669859886, + 0.005842071492224932, + 0.008169371634721756, + 0.004241478629410267, + 0.01564403437077999, + 0.021900111809372902, + 0.01451210305094719, + 0.017782608047127724, + 0.0048988377675414085, + 0.0018495580879971385, + 0.002900509163737297, + 0.0015367093728855252, + 0.004243417177349329, + 0.004633735399693251, + 0.004571123979985714, + 0.005577386822551489, + 0.00236878520809114, + 0.0010963132372125983, + 0.001612934865988791, + 0.0009317391086369753, + 0.002245940500870347, + 0.0025048989336937666, + 0.001741198473609984, + 0.0020530701149255037, + 0.004667903296649456, + 0.0024597891606390476, + 0.003961001988500357, + 0.0016422733897343278, + 0.004292832221835852, + 0.004263077396899462, + 0.003909710794687271, + 0.004027880262583494, + 0.030887199565768242, + 0.013424164615571499, + 0.025826459750533104, + 0.011255400255322456, + 0.024995962157845497, + 0.020690515637397766, + 0.023549648001790047, + 0.022706951946020126, + 0.012298469431698322, + 0.007138131186366081, + 0.011488335207104683, + 0.010951254516839981, + 0.011031843721866608, + 0.006321486551314592, + 0.011677122674882412, + 0.015171905048191547, + 0.0035947905853390694, + 0.004261143505573273, + 0.004154874477535486, + 0.00785006582736969, + 0.00637673307210207, + 0.0061811297200620174, + 0.007980024442076683, + 0.00610384251922369, + 0.011225352063775063, + 0.016562003642320633, + 0.007257186342030764, + 0.016212571412324905, + 0.013961556367576122, + 0.019052622839808464, + 0.01952977478504181, + 0.012269090861082077, + 0.015583187341690063, + 0.13083046674728394, + 0.007122430484741926, + 0.0108102448284626, + 0.11395730823278427, + 0.025908133015036583, + 0.012125971727073193, + 0.028744284063577652, + 0.005797910038381815, + 0.012467662803828716, + 0.003979502245783806, + 0.004527983255684376, + 0.009352966211736202, + 0.010765484534204006, + 0.004086262546479702, + 0.006854248698800802 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 4, + 7 + ], + "selected_reward": -0.018183674663305283 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, low-angle, wide-angle shot establishes a perfect central composition, dominated by a massive, brutalist concrete sculpture rendered in a bold, graphic novel art style. An impenetrable monolith composed of interlocking, sharply angled geometric panels, the sculpture stands tall against a desolate, empty sky. The entire scene is defined by a palette of rich black ink and stark, high-contrast colors; deep, solid black shadows, cast by a harsh, unseen light source, carve dramatic forms into the sculpture's rough, pitted concrete texture. The ground begins to collapse beneath its immense weight; initially, tiny, hairline cracks appear at the base of the sculpture, like precisely drawn ink lines. These cracks then quickly and decisively spiderweb outward, fracturing the barren earth in radial patterns. The action captures a sense of immense pressure as large chunks of the jagged ground crack and are visibly thrust upward, revealing a growing abyss of pure, inky blackness as the cracks widen and the earth itself crumbles.", + "seed": 17674, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000156/group_7/index_7835_seed_17674_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": -0.02414236175827682, + "vlm_logprob": -0.02414236175827682, + "vlm_yes_no_margin": 4.325, + "vlm_token_credit": -0.02414236175827682, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return 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 brutalist concrete sculpture composed of interlocking geometric panels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style a bold graphic novel art style with rich black ink and stark high-contrast colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene show the ground cracking at the base of the sculpture and fracturing outward in radial patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are large chunks of the jagged ground visibly thrust upward to reveal a growing abyss of pure blackness?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a static, low-angle, wide-angle shot establishing a perfect central composition?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011218564584851265 + ], + "opposite_logprob_tokens": [ + -5.136218547821045 + ], + "logprob_reward": -0.011218564584851265, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.011218564584851265, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.32715556025505066, + 0.02638525515794754, + 0.04619026184082031, + 0.03349941968917847, + 0.2810111939907074, + 0.06999237835407257, + 0.04637724533677101, + 0.08403515070676804, + 0.03701011463999748, + 0.04834340885281563 + ], + "credit_map": [ + 0.0196053609251976, + 0.019879620522260666, + 0.011142250150442123, + 0.014342515729367733, + 0.13281181454658508, + 0.01842535473406315, + 0.0875856801867485, + 0.02336299605667591, + 0.0032942600082606077, + 0.0030750397127121687, + 0.0021759672090411186, + 0.003759683109819889, + 0.002779495669528842, + 0.00350102293305099, + 0.00468431506305933, + 0.0031154705211520195, + 0.0068163881078362465, + 0.004694964271038771, + 0.00439628679305315, + 0.005553629714995623, + 0.006182770244777203, + 0.004984698258340359, + 0.00765632838010788, + 0.005905198864638805, + 0.003681098809465766, + 0.0030444865114986897, + 0.003581774653866887, + 0.004916953854262829, + 0.004472138360142708, + 0.003117405576631427, + 0.007859372533857822, + 0.002826191019266844, + 0.02936146967113018, + 0.03695269301533699, + 0.023868855088949203, + 0.04029297083616257, + 0.046179577708244324, + 0.02503219246864319, + 0.056963928043842316, + 0.022359507158398628, + 0.005194532219320536, + 0.003958161920309067, + 0.0100757647305727, + 0.007532303687185049, + 0.006760690361261368, + 0.0030688666738569736, + 0.020846975967288017, + 0.012555081397294998, + 0.005260334350168705, + 0.005641934461891651, + 0.006924451794475317, + 0.005416337866336107, + 0.0066198501735925674, + 0.005674644839018583, + 0.005473498720675707, + 0.005366191733628511, + 0.0072328788228333, + 0.01070820540189743, + 0.010078244842588902, + 0.009619815275073051, + 0.015438363887369633, + 0.009859318844974041, + 0.00999152846634388, + 0.01110678818076849, + 0.002960589714348316, + 0.00589659484103322, + 0.0049019320867955685, + 0.004572364967316389, + 0.005705568473786116, + 0.004259415902197361, + 0.0047699869610369205, + 0.003943662159144878, + 0.005475839599967003, + 0.006161805707961321, + 0.005922084208577871, + 0.003504397813230753, + 0.007799274288117886, + 0.008150555193424225, + 0.0038792507257312536, + 0.0074502043426036835 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 7 + ], + "selected_reward": -0.011218564584851265 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return 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 brutalist concrete sculpture composed of interlocking geometric panels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style a bold graphic novel art style with rich black ink and stark high-contrast colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene show the ground cracking at the base of the sculpture and fracturing outward in radial patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are large chunks of the jagged ground visibly thrust upward to reveal a growing abyss of pure blackness?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a static, low-angle, wide-angle shot establishing a perfect central composition?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02059772238135338 + ], + "opposite_logprob_tokens": [ + -3.8955976963043213 + ], + "logprob_reward": -0.02059772238135338, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.02059772238135338, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05492228642106056, + 0.008655829355120659, + 0.05077480524778366, + 0.025696799159049988, + 0.7329463362693787, + 0.007541806437075138, + 0.013731456361711025, + 0.0513586662709713, + 0.039220090955495834, + 0.015151909552514553 + ], + "credit_map": [ + 0.003330124542117119, + 0.003736294573172927, + 0.0019297266844660044, + 0.0025692402850836515, + 0.0213179849088192, + 0.002974059199914336, + 0.014953380450606346, + 0.004111472051590681, + 0.001227945671416819, + 0.0009245727560482919, + 0.0007937109330669045, + 0.001431903918273747, + 0.000793253886513412, + 0.0010011137928813696, + 0.0015163249336183071, + 0.0009670039871707559, + 0.00987478531897068, + 0.006574796047061682, + 0.0033935874234884977, + 0.005937501788139343, + 0.006204111967235804, + 0.006399971898645163, + 0.006744735408574343, + 0.005645317491143942, + 0.0030059912241995335, + 0.003772648749873042, + 0.0020549006294459105, + 0.003931999672204256, + 0.003202924272045493, + 0.002160081174224615, + 0.005537575110793114, + 0.002030677627772093, + 0.07757195830345154, + 0.09510305523872375, + 0.06126556918025017, + 0.10602829605340958, + 0.11897407472133636, + 0.0663745105266571, + 0.15141187608242035, + 0.056217025965452194, + 0.0011740201152861118, + 0.000853432749863714, + 0.0009152733837254345, + 0.0009048643405549228, + 0.001286147627979517, + 0.0008057635859586298, + 0.0007662465213797987, + 0.0008360585779882967, + 0.0011572433868423104, + 0.0017466803546994925, + 0.0023335881996899843, + 0.0015544024063274264, + 0.0020999102853238583, + 0.0012590390397235751, + 0.001948863035067916, + 0.0016317303525283933, + 0.005559245124459267, + 0.00766998715698719, + 0.005994499661028385, + 0.0053652189671993256, + 0.010108891874551773, + 0.00529763987287879, + 0.005077290814369917, + 0.006285892333835363, + 0.0022387178614735603, + 0.0035829658154398203, + 0.005501516163349152, + 0.0077415467239916325, + 0.006061956286430359, + 0.0029856024775654078, + 0.006997857708483934, + 0.004109928850084543, + 0.00169044837821275, + 0.0015745172277092934, + 0.0018113494152203202, + 0.001039218157529831, + 0.002263553673401475, + 0.0023835445754230022, + 0.0013577162753790617, + 0.0030315620824694633 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 7 + ], + "selected_reward": -0.02059772238135338 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return 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 brutalist concrete sculpture composed of interlocking geometric panels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style a bold graphic novel art style with rich black ink and stark high-contrast colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene show the ground cracking at the base of the sculpture and fracturing outward in radial patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are large chunks of the jagged ground visibly thrust upward to reveal a growing abyss of pure blackness?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a static, low-angle, wide-angle shot establishing a perfect central composition?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0028255560901015997 + ], + "opposite_logprob_tokens": [ + -5.877825736999512 + ], + "logprob_reward": -0.0028255560901015997, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.0028255560901015997, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13772441446781158, + 0.021574197337031364, + 0.03369801118969917, + 0.05586257576942444, + 0.4015665054321289, + 0.05591532588005066, + 0.04245195910334587, + 0.09380589425563812, + 0.06796614825725555, + 0.08943498134613037 + ], + "credit_map": [ + 0.008412903174757957, + 0.009523209184408188, + 0.006166031118482351, + 0.00897540058940649, + 0.051685232669115067, + 0.010748098604381084, + 0.030015382915735245, + 0.012198138050734997, + 0.002912973752245307, + 0.0023053691256791353, + 0.0022827256470918655, + 0.002414791379123926, + 0.0024899619165807962, + 0.00227154279127717, + 0.003343678545206785, + 0.0035531551111489534, + 0.00665443018078804, + 0.0030095346737653017, + 0.002866918221116066, + 0.0035467848647385836, + 0.0034466988872736692, + 0.005097604356706142, + 0.004931270144879818, + 0.004144768230617046, + 0.004720799159258604, + 0.004550394136458635, + 0.004060510080307722, + 0.008321304805576801, + 0.004822822753340006, + 0.0035990215837955475, + 0.018227683380246162, + 0.007560038007795811, + 0.04168963432312012, + 0.054356176406145096, + 0.0330904982984066, + 0.057554155588150024, + 0.06585271656513214, + 0.036387186497449875, + 0.08281399309635162, + 0.029822109267115593, + 0.0043845949694514275, + 0.005330126732587814, + 0.009601952508091927, + 0.004061136394739151, + 0.019586903974413872, + 0.0036226355005055666, + 0.005864458158612251, + 0.003463519737124443, + 0.003902291879057884, + 0.005149718839675188, + 0.006104464177042246, + 0.00463892100378871, + 0.007159499451518059, + 0.0045743416994810104, + 0.005710093304514885, + 0.005212625954300165, + 0.007206262089312077, + 0.013385539874434471, + 0.013930938206613064, + 0.013364859856665134, + 0.013276866637170315, + 0.008560753427445889, + 0.012232166714966297, + 0.011848509311676025, + 0.0056394292041659355, + 0.01125156506896019, + 0.008003149181604385, + 0.008859819732606411, + 0.011857841163873672, + 0.005578035023063421, + 0.010870198719203472, + 0.005906113889068365, + 0.008335969410836697, + 0.011892778798937798, + 0.01064808014780283, + 0.012189607135951519, + 0.013310380280017853, + 0.011464227922260761, + 0.009052877314388752, + 0.012541059404611588 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 7 + ], + "selected_reward": -0.0028255560901015997 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return 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 brutalist concrete sculpture composed of interlocking geometric panels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style a bold graphic novel art style with rich black ink and stark high-contrast colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene show the ground cracking at the base of the sculpture and fracturing outward in radial patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are large chunks of the jagged ground visibly thrust upward to reveal a growing abyss of pure blackness?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a static, low-angle, wide-angle shot establishing a perfect central composition?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"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": [ + 0.16120122373104095, + 0.03449905291199684, + 0.11202690005302429, + 0.06944828480482101, + 0.25462606549263, + 0.041083548218011856, + 0.05960152670741081, + 0.10196048021316528, + 0.038225241005420685, + 0.12732771039009094 + ], + "credit_map": [ + 0.010831254534423351, + 0.009113444946706295, + 0.007174542639404535, + 0.007871766574680805, + 0.06377005577087402, + 0.012386219576001167, + 0.037243057042360306, + 0.01281089335680008, + 0.004242402501404285, + 0.003166954731568694, + 0.004352246876806021, + 0.00357109890319407, + 0.00537826307117939, + 0.004067542962729931, + 0.005292573943734169, + 0.004427972715348005, + 0.01770588383078575, + 0.01741630770266056, + 0.008156844414770603, + 0.01367967575788498, + 0.017712559551000595, + 0.012434013187885284, + 0.015085060149431229, + 0.009836559183895588, + 0.0061665610410273075, + 0.00692377844825387, + 0.004803255666047335, + 0.012183154933154583, + 0.005845093633979559, + 0.003958933055400848, + 0.023769136518239975, + 0.005798365920782089, + 0.02571272663772106, + 0.03414693474769592, + 0.021225126460194588, + 0.03718385100364685, + 0.04122386500239372, + 0.022156698629260063, + 0.05358632281422615, + 0.019390521571040154, + 0.0037021504249423742, + 0.003197234123945236, + 0.004616500344127417, + 0.00414083618670702, + 0.00785776600241661, + 0.002728064777329564, + 0.004158821888267994, + 0.010682170279324055, + 0.006645037326961756, + 0.0071528139524161816, + 0.008094885386526585, + 0.009137983433902264, + 0.010024771094322205, + 0.005975620821118355, + 0.007619910407811403, + 0.004950504284352064, + 0.009360075928270817, + 0.01316047739237547, + 0.01333212573081255, + 0.016902495175600052, + 0.015220316126942635, + 0.01278958935290575, + 0.011180785484611988, + 0.010014613159000874, + 0.0034356964752078056, + 0.004353560972958803, + 0.006522682961076498, + 0.004360653925687075, + 0.006011601537466049, + 0.0038040627259761095, + 0.0053051733411848545, + 0.004431812558323145, + 0.013110392726957798, + 0.02225467376410961, + 0.014436203986406326, + 0.007999375462532043, + 0.020462315529584885, + 0.0249947477132082, + 0.009567765519022942, + 0.014502243138849735 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 9 + ], + "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 main subject a massive brutalist concrete sculpture composed of interlocking geometric panels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style a bold graphic novel art style with rich black ink and stark high-contrast colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene show the ground cracking at the base of the sculpture and fracturing outward in radial patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are large chunks of the jagged ground visibly thrust upward to reveal a growing abyss of pure blackness?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a static, low-angle, wide-angle shot establishing a perfect central composition?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0700078159570694 + ], + "opposite_logprob_tokens": [ + -2.695007801055908 + ], + "logprob_reward": -0.0700078159570694, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.0700078159570694, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19318614900112152, + 0.06596583127975464, + 0.17786762118339539, + 0.11427813023328781, + 0.07282664626836777, + 0.06476061046123505, + 0.034887563437223434, + 0.08196382969617844, + 0.10641518980264664, + 0.08784844726324081 + ], + "credit_map": [ + 0.014100749045610428, + 0.016022933647036552, + 0.007880199700593948, + 0.012944478541612625, + 0.06465284526348114, + 0.013443946838378906, + 0.04944895952939987, + 0.014692014083266258, + 0.007320770528167486, + 0.006217268295586109, + 0.005718388594686985, + 0.01030109915882349, + 0.011684347875416279, + 0.0057796575129032135, + 0.012000077404081821, + 0.006944215856492519, + 0.02687760628759861, + 0.019226836040616035, + 0.013905505649745464, + 0.020319722592830658, + 0.021898025646805763, + 0.028979234397411346, + 0.026348302140831947, + 0.020312368869781494, + 0.010093304328620434, + 0.013408026657998562, + 0.012160689570009708, + 0.013264328241348267, + 0.015663141384720802, + 0.01093219593167305, + 0.02434605173766613, + 0.01441038679331541, + 0.008683095686137676, + 0.008582168258726597, + 0.006037914659827948, + 0.009246728383004665, + 0.013496055267751217, + 0.009836780838668346, + 0.010738939046859741, + 0.006204958539456129, + 0.0043411171063780785, + 0.003862948389723897, + 0.016887400299310684, + 0.005628317128866911, + 0.00823911465704441, + 0.0030644310172647238, + 0.016621438786387444, + 0.006115840747952461, + 0.0034375975374132395, + 0.00493730278685689, + 0.0046609919518232346, + 0.004328120034188032, + 0.0042352755554020405, + 0.0042823259718716145, + 0.004110772628337145, + 0.004895172547549009, + 0.007986265234649181, + 0.011078841984272003, + 0.012494515627622604, + 0.014239427633583546, + 0.011621061712503433, + 0.009428534656763077, + 0.008122275583446026, + 0.006992897018790245, + 0.006323042791336775, + 0.01082849595695734, + 0.017218412831425667, + 0.021630750969052315, + 0.01544963289052248, + 0.009367651306092739, + 0.015275019221007824, + 0.010322172194719315, + 0.008057335391640663, + 0.011779396794736385, + 0.009766648523509502, + 0.010305874980986118, + 0.014032223261892796, + 0.011362641118466854, + 0.00997810810804367, + 0.012566209770739079 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 3 + ], + "selected_reward": -0.0700078159570694 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, low-angle, wide-angle shot establishes a perfect central composition, dominated by a massive, brutalist concrete sculpture rendered in a bold, graphic novel art style. An impenetrable monolith composed of interlocking, sharply angled geometric panels, the sculpture stands tall against a desolate, empty sky. The entire scene is defined by a palette of rich black ink and stark, high-contrast colors; deep, solid black shadows, cast by a harsh, unseen light source, carve dramatic forms into the sculpture's rough, pitted concrete texture. The ground begins to collapse beneath its immense weight; initially, tiny, hairline cracks appear at the base of the sculpture, like precisely drawn ink lines. These cracks then quickly and decisively spiderweb outward, fracturing the barren earth in radial patterns. The action captures a sense of immense pressure as large chunks of the jagged ground crack and are visibly thrust upward, revealing a growing abyss of pure, inky blackness as the cracks widen and the earth itself crumbles.", + "seed": 17675, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000156/group_7/index_7835_seed_17675_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": -0.016778803989291192, + "vlm_logprob": -0.016778803989291192, + "vlm_yes_no_margin": 4.425, + "vlm_token_credit": -0.016778803989291192, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return 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 brutalist concrete sculpture composed of interlocking geometric panels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style a bold graphic novel art style with rich black ink and stark high-contrast colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene show the ground cracking at the base of the sculpture and fracturing outward in radial patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are large chunks of the jagged ground visibly thrust upward to reveal a growing abyss of pure blackness?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a static, low-angle, wide-angle shot establishing a perfect central composition?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.024148032069206238 + ], + "opposite_logprob_tokens": [ + -4.024147987365723 + ], + "logprob_reward": -0.024148032069206238, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.024148032069206238, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05651998519897461, + 0.1365327090024948, + 0.0408257320523262, + 0.15738993883132935, + 0.1672690510749817, + 0.11967486143112183, + 0.1910010725259781, + 0.04753398895263672, + 0.037334028631448746, + 0.04591857269406319 + ], + "credit_map": [ + 0.009302245453000069, + 0.0039056583773344755, + 0.013805770315229893, + 0.005480694118887186, + 0.004655716009438038, + 0.005343624856323004, + 0.007839739322662354, + 0.006186538841575384, + 0.019228721037507057, + 0.017228877171874046, + 0.018963908776640892, + 0.02060992643237114, + 0.018238656222820282, + 0.01150339562445879, + 0.016272831708192825, + 0.014486393891274929, + 0.005395111627876759, + 0.0031331703066825867, + 0.007614701054990292, + 0.003034933703020215, + 0.004232161212712526, + 0.004791593644768, + 0.005905431229621172, + 0.006718628108501434, + 0.02740248665213585, + 0.008371974341571331, + 0.02206125110387802, + 0.008688646368682384, + 0.012075585313141346, + 0.040908120572566986, + 0.013688508421182632, + 0.02419336326420307, + 0.01859172247350216, + 0.017264751717448235, + 0.03783551976084709, + 0.02092949114739895, + 0.020397624000906944, + 0.019250497221946716, + 0.01513119600713253, + 0.017868243157863617, + 0.01459458190947771, + 0.012478995136916637, + 0.01539573259651661, + 0.022775938734412193, + 0.01178030576556921, + 0.011317338794469833, + 0.015272507444024086, + 0.016059471294283867, + 0.019858092069625854, + 0.025571724399924278, + 0.02437470480799675, + 0.033935025334358215, + 0.019035184755921364, + 0.02509346976876259, + 0.027962012216448784, + 0.015170849859714508, + 0.005313835106790066, + 0.0049869646318256855, + 0.007616488263010979, + 0.008073913864791393, + 0.004322605207562447, + 0.005848490167409182, + 0.005466161761432886, + 0.005905528552830219, + 0.0047647543251514435, + 0.005085421726107597, + 0.003347779391333461, + 0.002368837594985962, + 0.004842543974518776, + 0.006596069782972336, + 0.0032200098503381014, + 0.007108614780008793, + 0.003469133051112294, + 0.011039583943784237, + 0.003848036052659154, + 0.004291342105716467, + 0.012466655112802982, + 0.003784686792641878, + 0.004434448201209307, + 0.0025846848729997873 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 4, + 3 + ], + "selected_reward": -0.024148032069206238 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return 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 brutalist concrete sculpture composed of interlocking geometric panels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style a bold graphic novel art style with rich black ink and stark high-contrast colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene show the ground cracking at the base of the sculpture and fracturing outward in radial patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are large chunks of the jagged ground visibly thrust upward to reveal a growing abyss of pure blackness?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a static, low-angle, wide-angle shot establishing a perfect central composition?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014210578985512257 + ], + "opposite_logprob_tokens": [ + -4.2642107009887695 + ], + "logprob_reward": -0.014210578985512257, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014210578985512257, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03224458172917366, + 0.04869607463479042, + 0.06241000443696976, + 0.5612788796424866, + 0.016059303656220436, + 0.031244302168488503, + 0.10217052698135376, + 0.03569541126489639, + 0.027418866753578186, + 0.082782082259655 + ], + "credit_map": [ + 0.007669165730476379, + 0.0024449657648801804, + 0.005187202710658312, + 0.002175609115511179, + 0.002791855949908495, + 0.005264576058834791, + 0.0035490598529577255, + 0.003162145847454667, + 0.0076357233338057995, + 0.005794256925582886, + 0.009588619694113731, + 0.004499022848904133, + 0.004776248708367348, + 0.005182511638849974, + 0.0065772742964327335, + 0.00464241299778223, + 0.011855432763695717, + 0.003279131604358554, + 0.008439143188297749, + 0.0028889428358525038, + 0.002575326943770051, + 0.006725550163537264, + 0.004196772817522287, + 0.02244970016181469, + 0.10173849761486053, + 0.026583051308989525, + 0.1035565584897995, + 0.02609460987150669, + 0.03796354681253433, + 0.14532768726348877, + 0.04101651906967163, + 0.07899840176105499, + 0.002195625798776746, + 0.001330355298705399, + 0.002713909838348627, + 0.002140953205525875, + 0.0027349938172847033, + 0.0013925557723268867, + 0.001857264549471438, + 0.0016936451429501176, + 0.0037638891953974962, + 0.004170222207903862, + 0.0038414408918470144, + 0.005272827576845884, + 0.002503541996702552, + 0.003448224626481533, + 0.0049803960137069225, + 0.003263757098466158, + 0.013719452545046806, + 0.015315480530261993, + 0.0088708670809865, + 0.016707058995962143, + 0.011850079521536827, + 0.010566259734332561, + 0.01667974330484867, + 0.008461571298539639, + 0.005075155291706324, + 0.004318240564316511, + 0.004032476805150509, + 0.005292136687785387, + 0.0036911761853843927, + 0.004929094575345516, + 0.004503933247178793, + 0.0038531958125531673, + 0.003093759063631296, + 0.004017367493361235, + 0.0027193804271519184, + 0.0020405075047165155, + 0.004874648991972208, + 0.005592117086052895, + 0.0027342636603862047, + 0.0023468194995075464, + 0.005032432731240988, + 0.022300614044070244, + 0.007508100941777229, + 0.004798367153853178, + 0.02784254588186741, + 0.005626968573778868, + 0.004218066111207008, + 0.00545498076826334 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 6, + 9 + ], + "selected_reward": -0.014210578985512257 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return 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 brutalist concrete sculpture composed of interlocking geometric panels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style a bold graphic novel art style with rich black ink and stark high-contrast colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene show the ground cracking at the base of the sculpture and fracturing outward in radial patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are large chunks of the jagged ground visibly thrust upward to reveal a growing abyss of pure blackness?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a static, low-angle, wide-angle shot establishing a perfect central composition?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008642885833978653 + ], + "opposite_logprob_tokens": [ + -4.758642673492432 + ], + "logprob_reward": -0.008642885833978653, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008642885833978653, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03854318708181381, + 0.019908037036657333, + 0.05966946482658386, + 0.4078523814678192, + 0.03051379881799221, + 0.02439415268599987, + 0.09117124229669571, + 0.07677198201417923, + 0.13415507972240448, + 0.11702066659927368 + ], + "credit_map": [ + 0.007491022348403931, + 0.0032728658989071846, + 0.006190272513777018, + 0.0031541408970952034, + 0.002792201703414321, + 0.0057653505355119705, + 0.007043079938739538, + 0.0028342504519969225, + 0.00319404317997396, + 0.001687288167886436, + 0.004740269389003515, + 0.0019721975550055504, + 0.0016121909720823169, + 0.0022021338809281588, + 0.002296997932717204, + 0.0022029143292456865, + 0.013052213937044144, + 0.0024691999424248934, + 0.003907278645783663, + 0.002027614740654826, + 0.0026327152736485004, + 0.005652043502777815, + 0.004024917725473642, + 0.025903481990098953, + 0.0723450556397438, + 0.019087037071585655, + 0.07779087871313095, + 0.019404903054237366, + 0.027646435424685478, + 0.1055525690317154, + 0.02929059974849224, + 0.05673488602042198, + 0.0025984502863138914, + 0.0026145000010728836, + 0.008341345004737377, + 0.0035965233109891415, + 0.0035634676460176706, + 0.0030717288609594107, + 0.0030515226535499096, + 0.0036762605886906385, + 0.002610035939142108, + 0.002332726726308465, + 0.003591749584302306, + 0.003168871160596609, + 0.002336914651095867, + 0.002934100804850459, + 0.0035169401671737432, + 0.003902814816683531, + 0.012667997740209103, + 0.01277904398739338, + 0.007876062765717506, + 0.01404930092394352, + 0.009648507460951805, + 0.010216430760920048, + 0.014515013433992863, + 0.009418885223567486, + 0.007075028959661722, + 0.006608064752072096, + 0.012555008754134178, + 0.012390273623168468, + 0.005307502578943968, + 0.009289679117500782, + 0.018069107085466385, + 0.00547731714323163, + 0.02186334691941738, + 0.014996961690485477, + 0.012080392800271511, + 0.008761887438595295, + 0.016737639904022217, + 0.020044289529323578, + 0.011492308229207993, + 0.028178248554468155, + 0.010900216177105904, + 0.027318544685840607, + 0.01262608915567398, + 0.007283185143023729, + 0.03121199458837509, + 0.01386831235140562, + 0.007312887813895941, + 0.006499434355646372 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 8, + 9 + ], + "selected_reward": -0.008642885833978653 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return 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 brutalist concrete sculpture composed of interlocking geometric panels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style a bold graphic novel art style with rich black ink and stark high-contrast colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene show the ground cracking at the base of the sculpture and fracturing outward in radial patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are large chunks of the jagged ground visibly thrust upward to reveal a growing abyss of pure blackness?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a static, low-angle, wide-angle shot establishing a perfect central composition?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0031929248943924904 + ], + "opposite_logprob_tokens": [ + -5.753192901611328 + ], + "logprob_reward": -0.0031929248943924904, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.0031929248943924904, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04257192835211754, + 0.056279245764017105, + 0.028012346476316452, + 0.4707176685333252, + 0.06398838758468628, + 0.045255620032548904, + 0.04635652154684067, + 0.07001307606697083, + 0.05640965700149536, + 0.12039555609226227 + ], + "credit_map": [ + 0.007188404444605112, + 0.004101448226720095, + 0.00804846826940775, + 0.0039473632350564, + 0.004427150823175907, + 0.004184326622635126, + 0.005929654464125633, + 0.004745111335068941, + 0.008236301131546497, + 0.0061968048103153706, + 0.008265183307230473, + 0.007412034552544355, + 0.007582671009004116, + 0.004597607534378767, + 0.008290394209325314, + 0.005698251072317362, + 0.005326941143721342, + 0.0021679645869880915, + 0.002345280721783638, + 0.0016349494690075517, + 0.0018486345652490854, + 0.002589876065030694, + 0.002542939968407154, + 0.009555760771036148, + 0.08348812907934189, + 0.022168420255184174, + 0.08837484568357468, + 0.022156327962875366, + 0.03198743239045143, + 0.12108990550041199, + 0.035041600465774536, + 0.06641100347042084, + 0.005802846048027277, + 0.006145160645246506, + 0.017152680084109306, + 0.008020089939236641, + 0.009359314106404781, + 0.004759062081575394, + 0.00591134512796998, + 0.006837890017777681, + 0.00597135815769434, + 0.004684969317167997, + 0.005708429496735334, + 0.009320426732301712, + 0.004135610535740852, + 0.0053873746655881405, + 0.005615238565951586, + 0.004432214889675379, + 0.005449633114039898, + 0.0064217825420200825, + 0.003968563862144947, + 0.008781400509178638, + 0.004779440816491842, + 0.00559078436344862, + 0.006963101681321859, + 0.00440181465819478, + 0.007739041931927204, + 0.008018857799470425, + 0.012349395081400871, + 0.011989166028797626, + 0.006339270155876875, + 0.006110129412263632, + 0.012415707111358643, + 0.005051502492278814, + 0.008386198431253433, + 0.008671112358570099, + 0.006013832520693541, + 0.006581307388842106, + 0.00820718239992857, + 0.007509111426770687, + 0.006986845750361681, + 0.004054071847349405, + 0.00871811993420124, + 0.026676759123802185, + 0.012723787687718868, + 0.012610753066837788, + 0.03282753750681877, + 0.010874415747821331, + 0.008559402078390121, + 0.007404768839478493 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 9, + 7 + ], + "selected_reward": -0.0031929248943924904 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive brutalist concrete sculpture composed of interlocking geometric panels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style a bold graphic novel art style with rich black ink and stark high-contrast colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene show the ground cracking at the base of the sculpture and fracturing outward in radial patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are large chunks of the jagged ground visibly thrust upward to reveal a growing abyss of pure blackness?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a static, low-angle, wide-angle shot establishing a perfect central composition?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03369959816336632 + ], + "opposite_logprob_tokens": [ + -3.4086995124816895 + ], + "logprob_reward": -0.03369959816336632, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03369959816336632, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08581819385290146, + 0.13894905149936676, + 0.11673404276371002, + 0.35128530859947205, + 0.06978020071983337, + 0.047690946608781815, + 0.08200683444738388, + 0.039032917469739914, + 0.029867375269532204, + 0.038835156708955765 + ], + "credit_map": [ + 0.018742062151432037, + 0.007913709618151188, + 0.011253457516431808, + 0.0068422467447817326, + 0.006955153774470091, + 0.013266217894852161, + 0.011932470835745335, + 0.00891287811100483, + 0.020308149978518486, + 0.014658056199550629, + 0.021879468113183975, + 0.02110636793076992, + 0.017697805538773537, + 0.012507357634603977, + 0.016824640333652496, + 0.013967206701636314, + 0.023064659908413887, + 0.006393247749656439, + 0.012644301168620586, + 0.005728086922317743, + 0.005992948543280363, + 0.0091351717710495, + 0.008923772722482681, + 0.044851843267679214, + 0.06236250698566437, + 0.016642164438962936, + 0.06698408722877502, + 0.016544807702302933, + 0.02302711084485054, + 0.0904988944530487, + 0.025637906044721603, + 0.04958781972527504, + 0.0065163117833435535, + 0.006520690396428108, + 0.017714930698275566, + 0.008445150218904018, + 0.011326917447149754, + 0.005262717604637146, + 0.006010175682604313, + 0.007983304560184479, + 0.004570689518004656, + 0.004412312991917133, + 0.007368858903646469, + 0.00640438636764884, + 0.0068372758105397224, + 0.00568646052852273, + 0.006532041355967522, + 0.005878923460841179, + 0.007735305931419134, + 0.01310544554144144, + 0.008748780004680157, + 0.014796088449656963, + 0.009953505359590054, + 0.007860817946493626, + 0.012639682739973068, + 0.007167208474129438, + 0.005564607679843903, + 0.003858240321278572, + 0.00532363448292017, + 0.005841811653226614, + 0.0026705723721534014, + 0.0037197775673121214, + 0.006859517190605402, + 0.005194755271077156, + 0.004218265879899263, + 0.003291183151304722, + 0.0026810443960130215, + 0.0032362621277570724, + 0.004288641270250082, + 0.0042481799609959126, + 0.00428736861795187, + 0.003616430563852191, + 0.004243805073201656, + 0.007024243474006653, + 0.005159018095582724, + 0.0039035980589687824, + 0.006385377608239651, + 0.005378407891839743, + 0.003321745665743947, + 0.003418962936848402 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 1, + 2 + ], + "selected_reward": -0.03369959816336632 + } + ], + "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/0000156/group_8/index_tensor([8685, 8685])_group_8_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000156/group_8/index_tensor([8685, 8685])_group_8_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..1931a898d3a1b3c545ec1fba55f8ee95e2c90f3a --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000156/group_8/index_tensor([8685, 8685])_group_8_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "Shot from a slightly overhead angle, the camera pans incredibly slowly to the right, delicately tracking the movement of a hand writing on an elegant, dark wood desk. The hand's slender fingers firmly and firmly grasp a delicate black fountain pen with a gold-rimmed barrel. Its gleaming metallic iridium tip glides effortlessly and steadily across a sheet of thick, off-white stationery, leaving behind a trail of evenly inked, dark blue cursive. At this moment, a soft, warm, slightly iridescent lens flare, as if carefully choreographed, slowly sweeps across the left side of the frame, grazing the moving metal tip and causing it to flare with a dazzling, fleeting brilliance. The light comes from a warm, off-screen desk lamp, casting a soft, focused beam that casts a subtle, shifting shadow on the paper. A stack of heavy books is faintly visible in the background, but completely blurred. The main colors of the entire picture are warm brown and beige, with dark blue ink and golden light on the tip of the pen as embellishments, creating an atmosphere of extreme tranquility, concentration and full of classical bookishness.", + "seed": 27925, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000156/group_8/index_8685_seed_27925_dp_16_rank_32_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 8, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.0750897977501154, + "vlm_logprob": -0.0750897977501154, + "vlm_yes_no_margin": 3.1, + "vlm_token_credit": -0.0750897977501154, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"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 hand holding a black fountain pen with a gold-rimmed barrel visible writing on a sheet of paper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the pen tip metallic and iridescent, leaving dark blue cursive ink on off-white stationery?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene set on an elegant dark wood desk with a stack of heavy books faintly visible and blurred in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera pan slowly to the right while tracking the movement of the hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a soft, warm, slightly iridescent lens flare sweeping across the left side of the frame that grazes the pen tip?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.045812610536813736 + ], + "opposite_logprob_tokens": [ + -3.2958126068115234 + ], + "logprob_reward": -0.045812610536813736, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.045812610536813736, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2748090326786041, + 0.012096047401428223, + 0.01519789919257164, + 0.009367898106575012, + 0.025943810120224953, + 0.03334007039666176, + 0.024117089807987213, + 0.024659480899572372, + 0.5515461564064026, + 0.02892250381410122 + ], + "credit_map": [ + 0.019907156005501747, + 0.016703292727470398, + 0.06678247451782227, + 0.07254400849342346, + 0.022064533084630966, + 0.038894783705472946, + 0.013994204811751842, + 0.023918606340885162, + 0.0016038393368944526, + 0.0011294834548607469, + 0.0018398902611806989, + 0.0011751723941415548, + 0.0022039299365133047, + 0.001717903302051127, + 0.0010920539498329163, + 0.0013337760465219617, + 0.0015486518386751413, + 0.0014260929310694337, + 0.001800001715309918, + 0.0019597404170781374, + 0.0019567592535167933, + 0.001723824068903923, + 0.0025954630691558123, + 0.002187367295846343, + 0.000820548040792346, + 0.0007682466530241072, + 0.0010053220903500915, + 0.0009784518042579293, + 0.0011953338980674744, + 0.0011475620558485389, + 0.0020918294321745634, + 0.0013606053544208407, + 0.002187731210142374, + 0.0032831132411956787, + 0.004858320578932762, + 0.005756868980824947, + 0.0016748320776969194, + 0.0023931548930704594, + 0.0031675980426371098, + 0.00262219225987792, + 0.0014917491935193539, + 0.0011425843695178628, + 0.013473378494381905, + 0.013150516897439957, + 0.0008144831517711282, + 0.0012776459334418178, + 0.0012523930054157972, + 0.0007373221451416612, + 0.0026364449877291918, + 0.0028343284502625465, + 0.0060414946638047695, + 0.005646660923957825, + 0.0020207681227475405, + 0.002361586084589362, + 0.0012981316540390253, + 0.0012776777148246765, + 0.004841621499508619, + 0.00351392338052392, + 0.002375348238274455, + 0.0032450954895466566, + 0.002642326056957245, + 0.0029881885275244713, + 0.00246327999047935, + 0.0025897002778947353, + 0.02365943230688572, + 0.01682530902326107, + 0.23567074537277222, + 0.21740573644638062, + 0.011494479142129421, + 0.01407825667411089, + 0.020456865429878235, + 0.011955350637435913, + 0.002649752888828516, + 0.002738163573667407, + 0.004430443048477173, + 0.004675703123211861, + 0.003030293621122837, + 0.002913728589192033, + 0.005379183683544397, + 0.0031052392441779375 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 5 + ], + "selected_reward": -0.045812610536813736 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"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 hand holding a black fountain pen with a gold-rimmed barrel visible writing on a sheet of paper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the pen tip metallic and iridescent, leaving dark blue cursive ink on off-white stationery?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene set on an elegant dark wood desk with a stack of heavy books faintly visible and blurred in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera pan slowly to the right while tracking the movement of the hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a soft, warm, slightly iridescent lens flare sweeping across the left side of the frame that grazes the pen tip?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020611852407455444 + ], + "opposite_logprob_tokens": [ + -3.8956117630004883 + ], + "logprob_reward": -0.020611852407455444, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020611852407455444, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23995934426784515, + 0.02127149887382984, + 0.02885962836444378, + 0.01527454610913992, + 0.023737965151667595, + 0.03608192503452301, + 0.03035716898739338, + 0.03240079805254936, + 0.5308170318603516, + 0.04124012216925621 + ], + "credit_map": [ + 0.026539241895079613, + 0.022575821727514267, + 0.028216831386089325, + 0.03746955469250679, + 0.028731834143400192, + 0.05179939791560173, + 0.01671913079917431, + 0.027907544746994972, + 0.002321118488907814, + 0.002080643782392144, + 0.003680296242237091, + 0.0025409942027181387, + 0.00344357592985034, + 0.0027314212638884783, + 0.0017859870567917824, + 0.0026874595787376165, + 0.0036673115100711584, + 0.0022367883939296007, + 0.003321092575788498, + 0.003101024078205228, + 0.003290269523859024, + 0.0029910204466432333, + 0.006173711735755205, + 0.004078411962836981, + 0.0028747711330652237, + 0.0016159095102921128, + 0.001970401033759117, + 0.0016839804593473673, + 0.0012570839608088136, + 0.0015510942321270704, + 0.0026053166948258877, + 0.0017159883864223957, + 0.003263619262725115, + 0.0034357979893684387, + 0.003687128657475114, + 0.004138532094657421, + 0.001867125160060823, + 0.0022384803742170334, + 0.0028906476218253374, + 0.00221663317643106, + 0.0018695782637223601, + 0.002179639646783471, + 0.013180256821215153, + 0.01308115292340517, + 0.0011492581106722355, + 0.001622489420697093, + 0.00169094349257648, + 0.0013086062390357256, + 0.0034581958316266537, + 0.005204346962273121, + 0.003455613274127245, + 0.004394113551825285, + 0.0034187682904303074, + 0.004044961649924517, + 0.0029333194252103567, + 0.0034478495363146067, + 0.004660657607018948, + 0.004330975003540516, + 0.003086484968662262, + 0.003270986257120967, + 0.0038784248754382133, + 0.00287715345621109, + 0.005355730652809143, + 0.004940386395901442, + 0.023452313616871834, + 0.01735769584774971, + 0.22617076337337494, + 0.2081509828567505, + 0.012484078295528889, + 0.013527421280741692, + 0.018444068729877472, + 0.011229694820940495, + 0.0036260313354432583, + 0.003050760831683874, + 0.006785284727811813, + 0.007742110174149275, + 0.0035212985239923, + 0.004371023736894131, + 0.006979798898100853, + 0.005163812544196844 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 9 + ], + "selected_reward": -0.020611852407455444 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"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 hand holding a black fountain pen with a gold-rimmed barrel visible writing on a sheet of paper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the pen tip metallic and iridescent, leaving dark blue cursive ink on off-white stationery?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene set on an elegant dark wood desk with a stack of heavy books faintly visible and blurred in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera pan slowly to the right while tracking the movement of the hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a soft, warm, slightly iridescent lens flare sweeping across the left side of the frame that grazes the pen tip?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023284928873181343 + ], + "opposite_logprob_tokens": [ + -3.773284912109375 + ], + "logprob_reward": -0.023284928873181343, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023284928873181343, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09434846043586731, + 0.019180353730916977, + 0.020974984392523766, + 0.02927122637629509, + 0.16903094947338104, + 0.0219742301851511, + 0.02505112998187542, + 0.032507169991731644, + 0.574090301990509, + 0.01357116736471653 + ], + "credit_map": [ + 0.009154105558991432, + 0.00992725882679224, + 0.0153088653460145, + 0.015249043703079224, + 0.009496101178228855, + 0.01639728993177414, + 0.007065512239933014, + 0.01175028458237648, + 0.003227022709324956, + 0.002503485418856144, + 0.003009896259754896, + 0.002151428023353219, + 0.0018068946665152907, + 0.00241107982583344, + 0.0026383795775473118, + 0.0014321665512397885, + 0.0036684616934508085, + 0.0031847055070102215, + 0.003014632733538747, + 0.002050882438197732, + 0.0017177558038383722, + 0.0028334553353488445, + 0.0023787820246070623, + 0.0021263083908706903, + 0.005171452183276415, + 0.0047075217589735985, + 0.004178663715720177, + 0.0032300588209182024, + 0.003096704836934805, + 0.003772685071453452, + 0.002713739639148116, + 0.0024004001170396805, + 0.008827927522361279, + 0.013124926015734673, + 0.061933472752571106, + 0.06459572166204453, + 0.00291122286580503, + 0.006902392953634262, + 0.0059115164913237095, + 0.004823783878237009, + 0.0013987993588671088, + 0.0012271229643374681, + 0.007871500216424465, + 0.007855407893657684, + 0.0006805668817833066, + 0.00116213399451226, + 0.0009703648393042386, + 0.0008083335123956203, + 0.0016704123700037599, + 0.0021218969486653805, + 0.00827476754784584, + 0.007694825530052185, + 0.0012410276103764772, + 0.0014115066733211279, + 0.0013467913959175348, + 0.0012899007415398955, + 0.004037154372781515, + 0.004226474091410637, + 0.0039748442359268665, + 0.004697558470070362, + 0.002907381160184741, + 0.004256130196154118, + 0.004451605025678873, + 0.003956022206693888, + 0.03242674097418785, + 0.022601814940571785, + 0.23921971023082733, + 0.21876531839370728, + 0.0137048764154315, + 0.015635818243026733, + 0.020239507779479027, + 0.011496500112116337, + 0.0018743800465017557, + 0.0015528412768617272, + 0.0028817756101489067, + 0.002264078240841627, + 0.0008048325544223189, + 0.0008549921913072467, + 0.0014041541144251823, + 0.001934114028699696 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 4, + 0 + ], + "selected_reward": -0.023284928873181343 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"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 hand holding a black fountain pen with a gold-rimmed barrel visible writing on a sheet of paper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the pen tip metallic and iridescent, leaving dark blue cursive ink on off-white stationery?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene set on an elegant dark wood desk with a stack of heavy books faintly visible and blurred in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera pan slowly to the right while tracking the movement of the hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a soft, warm, slightly iridescent lens flare sweeping across the left side of the frame that grazes the pen tip?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.25205302238464355 + ], + "opposite_logprob_tokens": [ + -1.5020530223846436 + ], + "logprob_reward": -0.25205302238464355, + "yes_no_margin_reward": 1.25, + "token_credit_reward": -0.25205302238464355, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12103927880525589, + 0.018929803743958473, + 0.01352305244654417, + 0.015291321091353893, + 0.028617724776268005, + 0.04634125158190727, + 0.12973715364933014, + 0.03638029471039772, + 0.4431982636451721, + 0.14694184064865112 + ], + "credit_map": [ + 0.007050970569252968, + 0.008724117651581764, + 0.04206838086247444, + 0.040910352021455765, + 0.005214398261159658, + 0.008610674180090427, + 0.004402366001158953, + 0.0040580197237432, + 0.002286543371155858, + 0.0019524386152625084, + 0.0028552531730383635, + 0.0028313484508544207, + 0.0028863130137324333, + 0.0022438126616179943, + 0.0019767123740166426, + 0.0018973816186189651, + 0.0013305832399055362, + 0.0014740686165168881, + 0.0018044273601844907, + 0.0019066096283495426, + 0.0016846442595124245, + 0.0021361233666539192, + 0.0014258868759498, + 0.0017607087502256036, + 0.0025252888444811106, + 0.0016121241496875882, + 0.0025499160401523113, + 0.0021223481744527817, + 0.001546325976960361, + 0.00233707157894969, + 0.0012821124400943518, + 0.0013161341194063425, + 0.0034153813030570745, + 0.0030254547018557787, + 0.006930701434612274, + 0.0054644327610731125, + 0.0016331629594787955, + 0.002531474456191063, + 0.0029502329416573048, + 0.002666885731741786, + 0.002803020877763629, + 0.002416108502075076, + 0.014476524665951729, + 0.014800039120018482, + 0.002875786041840911, + 0.002717606956139207, + 0.0039241197519004345, + 0.0023280431050807238, + 0.007971223443746567, + 0.011703773401677608, + 0.0413312129676342, + 0.0397978350520134, + 0.005000328179448843, + 0.008379311300814152, + 0.007793832570314407, + 0.007759632542729378, + 0.003087845863774419, + 0.0035078844521194696, + 0.006471275817602873, + 0.007367774378508329, + 0.0025964269880205393, + 0.002858820604160428, + 0.006174118723720312, + 0.004316146019846201, + 0.030453504994511604, + 0.03192206844687462, + 0.15818139910697937, + 0.14687030017375946, + 0.01850374974310398, + 0.01910318434238434, + 0.022504981607198715, + 0.015659090131521225, + 0.019768426194787025, + 0.016554711386561394, + 0.02476964145898819, + 0.02039310894906521, + 0.014246739447116852, + 0.013644933700561523, + 0.022250501438975334, + 0.015313772484660149 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 6 + ], + "selected_reward": -0.25205302238464355 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"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 hand holding a black fountain pen with a gold-rimmed barrel visible writing on a sheet of paper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the pen tip metallic and iridescent, leaving dark blue cursive ink on off-white stationery?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene set on an elegant dark wood desk with a stack of heavy books faintly visible and blurred in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera pan slowly to the right while tracking the movement of the hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a soft, warm, slightly iridescent lens flare sweeping across the left side of the frame that grazes the pen tip?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.033686574548482895 + ], + "opposite_logprob_tokens": [ + -3.408686637878418 + ], + "logprob_reward": -0.033686574548482895, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.033686574548482895, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17369626462459564, + 0.022279713302850723, + 0.013837493024766445, + 0.025661727413535118, + 0.062392253428697586, + 0.025046350434422493, + 0.024059610441327095, + 0.037321459501981735, + 0.5636729598045349, + 0.05203212797641754 + ], + "credit_map": [ + 0.013010519556701183, + 0.00837875809520483, + 0.053432647138834, + 0.058152638375759125, + 0.009651370346546173, + 0.014325456693768501, + 0.008961393497884274, + 0.007783483248203993, + 0.0022828737273812294, + 0.0023234146647155285, + 0.003214664524421096, + 0.00308202370069921, + 0.0034742075949907303, + 0.003604394383728504, + 0.0021089701913297176, + 0.0021891642827540636, + 0.0028150747530162334, + 0.002106043975800276, + 0.0025225081481039524, + 0.0016491567948833108, + 0.0008596089901402593, + 0.0015694390749558806, + 0.0011657709255814552, + 0.0011498902458697557, + 0.002358502708375454, + 0.0031661198008805513, + 0.002442779019474983, + 0.0026820695493370295, + 0.0021120363380759954, + 0.0027387808077037334, + 0.004146014340221882, + 0.006015423219650984, + 0.006859351880848408, + 0.009163429029285908, + 0.012833126820623875, + 0.013347025960683823, + 0.0035604725126177073, + 0.005701912567019463, + 0.005519714672118425, + 0.005407219752669334, + 0.0018651032587513328, + 0.002232371363788843, + 0.0074025425128638744, + 0.007867010310292244, + 0.0010776809649541974, + 0.001231012400239706, + 0.0019997272174805403, + 0.0013709025224670768, + 0.0023981514386832714, + 0.003011524910107255, + 0.004991051275283098, + 0.00527287507429719, + 0.00188517349306494, + 0.002314600395038724, + 0.001876840484328568, + 0.002309394534677267, + 0.004393119364976883, + 0.004338491708040237, + 0.004139950033277273, + 0.005643810145556927, + 0.003336808178573847, + 0.004404823761433363, + 0.00539424829185009, + 0.005670205224305391, + 0.031023191288113594, + 0.02135768160223961, + 0.2366100400686264, + 0.22374218702316284, + 0.009150650352239609, + 0.011497207917273045, + 0.019910458475351334, + 0.010381529107689857, + 0.00802726112306118, + 0.0038187094032764435, + 0.01057655643671751, + 0.012530338019132614, + 0.003690962214022875, + 0.004562145099043846, + 0.004948557820171118, + 0.0038775999564677477 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 4 + ], + "selected_reward": -0.033686574548482895 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "Shot from a slightly overhead angle, the camera pans incredibly slowly to the right, delicately tracking the movement of a hand writing on an elegant, dark wood desk. The hand's slender fingers firmly and firmly grasp a delicate black fountain pen with a gold-rimmed barrel. Its gleaming metallic iridium tip glides effortlessly and steadily across a sheet of thick, off-white stationery, leaving behind a trail of evenly inked, dark blue cursive. At this moment, a soft, warm, slightly iridescent lens flare, as if carefully choreographed, slowly sweeps across the left side of the frame, grazing the moving metal tip and causing it to flare with a dazzling, fleeting brilliance. The light comes from a warm, off-screen desk lamp, casting a soft, focused beam that casts a subtle, shifting shadow on the paper. A stack of heavy books is faintly visible in the background, but completely blurred. The main colors of the entire picture are warm brown and beige, with dark blue ink and golden light on the tip of the pen as embellishments, creating an atmosphere of extreme tranquility, concentration and full of classical bookishness.", + "seed": 27926, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000156/group_8/index_8685_seed_27926_dp_16_rank_32_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 8, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.13902242314070462, + "vlm_logprob": -0.13902242314070462, + "vlm_yes_no_margin": 3.525, + "vlm_token_credit": -0.13902242314070462, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"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 hand holding a black fountain pen with a gold-rimmed barrel visible writing on a sheet of paper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the pen tip metallic and iridescent, leaving dark blue cursive ink on off-white stationery?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene set on an elegant dark wood desk with a stack of heavy books faintly visible and blurred in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera pan slowly to the right while tracking the movement of the hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a soft, warm, slightly iridescent lens flare sweeping across the left side of the frame that grazes the pen tip?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01186162605881691 + ], + "opposite_logprob_tokens": [ + -4.761861801147461 + ], + "logprob_reward": -0.01186162605881691, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.01186162605881691, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.25119853019714355, + 0.079686738550663, + 0.08257189393043518, + 0.11667568981647491, + 0.10705024003982544, + 0.03974512219429016, + 0.0328628309071064, + 0.08735213428735733, + 0.10276298969984055, + 0.10009385645389557 + ], + "credit_map": [ + 0.02802920714020729, + 0.030691474676132202, + 0.0299556702375412, + 0.022073715925216675, + 0.020490704104304314, + 0.030964789912104607, + 0.06274183094501495, + 0.026251109316945076, + 0.008549203164875507, + 0.008910881355404854, + 0.01064762007445097, + 0.006857826840132475, + 0.009051740169525146, + 0.012812739238142967, + 0.011612053029239178, + 0.011244677007198334, + 0.01318839006125927, + 0.011486132629215717, + 0.009485559538006783, + 0.007549686823040247, + 0.007721301168203354, + 0.011438482441008091, + 0.013882714323699474, + 0.007819633930921555, + 0.01856434904038906, + 0.01737428829073906, + 0.013722783885896206, + 0.01243541482836008, + 0.01176235917955637, + 0.012322990223765373, + 0.017849497497081757, + 0.012644006870687008, + 0.015010625123977661, + 0.015995539724826813, + 0.016176674515008926, + 0.011084325611591339, + 0.008984001353383064, + 0.012063560076057911, + 0.017865533009171486, + 0.009869980625808239, + 0.0044417851604521275, + 0.004517759662121534, + 0.005379242356866598, + 0.004934851545840502, + 0.004244204610586166, + 0.0062181963585317135, + 0.006129614543169737, + 0.0038794686552137136, + 0.004089977126568556, + 0.004610957577824593, + 0.004133895970880985, + 0.003710166784003377, + 0.004700842779129744, + 0.00450561847537756, + 0.004224488511681557, + 0.00288688438013196, + 0.00843950267881155, + 0.009101025760173798, + 0.011976446025073528, + 0.010629242286086082, + 0.009212677367031574, + 0.013651853427290916, + 0.014018449932336807, + 0.010322941467165947, + 0.009154308587312698, + 0.009770050644874573, + 0.009210171177983284, + 0.025397799909114838, + 0.010513427667319775, + 0.01580357924103737, + 0.01477015856653452, + 0.008143490180373192, + 0.009770473465323448, + 0.010843301191926003, + 0.013845601119101048, + 0.01280899066478014, + 0.014995207078754902, + 0.012098115868866444, + 0.012889068573713303, + 0.012843105010688305 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 4 + ], + "selected_reward": -0.01186162605881691 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"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 hand holding a black fountain pen with a gold-rimmed barrel visible writing on a sheet of paper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the pen tip metallic and iridescent, leaving dark blue cursive ink on off-white stationery?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene set on an elegant dark wood desk with a stack of heavy books faintly visible and blurred in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera pan slowly to the right while tracking the movement of the hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a soft, warm, slightly iridescent lens flare sweeping across the left side of the frame that grazes the pen tip?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018200181424617767 + ], + "opposite_logprob_tokens": [ + -4.018200397491455 + ], + "logprob_reward": -0.018200181424617767, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018200181424617767, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19886867702007294, + 0.07045746594667435, + 0.05274525284767151, + 0.036373432725667953, + 0.04297727346420288, + 0.04140613600611687, + 0.09831425547599792, + 0.10961813479661942, + 0.2962951064109802, + 0.05294423922896385 + ], + "credit_map": [ + 0.03316212072968483, + 0.030691303312778473, + 0.034546419978141785, + 0.019410988315939903, + 0.019021933898329735, + 0.021702438592910767, + 0.023206165060400963, + 0.017127325758337975, + 0.006797862704843283, + 0.006223015487194061, + 0.009772233664989471, + 0.007885969243943691, + 0.007976571097970009, + 0.00929179135710001, + 0.011157273314893246, + 0.011352749541401863, + 0.006389130372554064, + 0.005534531082957983, + 0.005454500671476126, + 0.00541881937533617, + 0.006533398758620024, + 0.007608647923916578, + 0.009696736931800842, + 0.00610949145630002, + 0.004272309597581625, + 0.004589488729834557, + 0.004997794982045889, + 0.00455127190798521, + 0.003653231542557478, + 0.004091184586286545, + 0.006336113438010216, + 0.0038820405025035143, + 0.005086896475404501, + 0.004866132512688637, + 0.00691927969455719, + 0.004820464178919792, + 0.005581049248576164, + 0.005575881339609623, + 0.006273205392062664, + 0.0038543695118278265, + 0.0044766138307750225, + 0.0038558831438422203, + 0.005166993476450443, + 0.0036832832265645266, + 0.005872635170817375, + 0.005311178043484688, + 0.007958026602864265, + 0.0050815255381166935, + 0.0123222591355443, + 0.009593160822987556, + 0.009317232295870781, + 0.010729165747761726, + 0.009313095360994339, + 0.012589734978973866, + 0.02107023447751999, + 0.013379387557506561, + 0.010530372150242329, + 0.015416922979056835, + 0.01021287776529789, + 0.011834978125989437, + 0.012971284799277782, + 0.01722807064652443, + 0.0201079323887825, + 0.011315700598061085, + 0.06468755751848221, + 0.028496671468019485, + 0.01690632477402687, + 0.049030546098947525, + 0.010923098772764206, + 0.08640768378973007, + 0.008648768067359924, + 0.031194472685456276, + 0.007890598848462105, + 0.005904568824917078, + 0.008769719861447811, + 0.005984017625451088, + 0.007377957459539175, + 0.0067109959200024605, + 0.00431457906961441, + 0.005991808604449034 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 7 + ], + "selected_reward": -0.018200181424617767 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"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 hand holding a black fountain pen with a gold-rimmed barrel visible writing on a sheet of paper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the pen tip metallic and iridescent, leaving dark blue cursive ink on off-white stationery?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene set on an elegant dark wood desk with a stack of heavy books faintly visible and blurred in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera pan slowly to the right while tracking the movement of the hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a soft, warm, slightly iridescent lens flare sweeping across the left side of the frame that grazes the pen tip?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005946326069533825 + ], + "opposite_logprob_tokens": [ + -5.130946159362793 + ], + "logprob_reward": -0.005946326069533825, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005946326069533825, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11142978817224503, + 0.2399236410856247, + 0.1405690312385559, + 0.16933076083660126, + 0.01815829984843731, + 0.024662084877490997, + 0.016243968158960342, + 0.01592913083732128, + 0.24088320136070251, + 0.022870101034641266 + ], + "credit_map": [ + 0.01005795318633318, + 0.011043380945920944, + 0.014063109643757343, + 0.009197473526000977, + 0.011789422482252121, + 0.016987748444080353, + 0.021474167704582214, + 0.01681654527783394, + 0.048115674406290054, + 0.04033399373292923, + 0.04237423837184906, + 0.02183029241859913, + 0.018084270879626274, + 0.019345957785844803, + 0.02571491338312626, + 0.024124329909682274, + 0.026478061452507973, + 0.021573685109615326, + 0.0223335362970829, + 0.013384095393121243, + 0.011531036347150803, + 0.014325467869639397, + 0.018261615186929703, + 0.012681540101766586, + 0.02996082417666912, + 0.029270391911268234, + 0.027247503399848938, + 0.01550230011343956, + 0.01773434691131115, + 0.017206115648150444, + 0.02067084051668644, + 0.011738455854356289, + 0.00332225882448256, + 0.003510999958962202, + 0.0022821337915956974, + 0.0021432095672935247, + 0.001501777907833457, + 0.002127314917743206, + 0.0015677768969908357, + 0.0017028289148584008, + 0.0035453601740300655, + 0.00307233608327806, + 0.003760966006666422, + 0.0028118975460529327, + 0.0027706415858119726, + 0.003975235391408205, + 0.0025904898066073656, + 0.0021351606119424105, + 0.0016922897193580866, + 0.002196050249040127, + 0.0026478306390345097, + 0.0021486077457666397, + 0.0015109650557860732, + 0.0020467534195631742, + 0.002183853881433606, + 0.001817617448978126, + 0.001731516676954925, + 0.0023240011651068926, + 0.002553511643782258, + 0.0018882392905652523, + 0.001650255755521357, + 0.002598764607682824, + 0.0017456532223150134, + 0.0014371909201145172, + 0.05006835609674454, + 0.02281639538705349, + 0.0107387974858284, + 0.025603707879781723, + 0.008528416976332664, + 0.08510731160640717, + 0.00767457066103816, + 0.030345669016242027, + 0.0025056558661162853, + 0.00277752080000937, + 0.0029286497738212347, + 0.003435688791796565, + 0.003322051605209708, + 0.0029216993134468794, + 0.0022288826294243336, + 0.0027499550487846136 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 1, + 3 + ], + "selected_reward": -0.005946326069533825 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"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 hand holding a black fountain pen with a gold-rimmed barrel visible writing on a sheet of paper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the pen tip metallic and iridescent, leaving dark blue cursive ink on off-white stationery?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene set on an elegant dark wood desk with a stack of heavy books faintly visible and blurred in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera pan slowly to the right while tracking the movement of the hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a soft, warm, slightly iridescent lens flare sweeping across the left side of the frame that grazes the pen tip?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.6327646374702454 + ], + "opposite_logprob_tokens": [ + -0.7577646374702454 + ], + "logprob_reward": -0.6327646374702454, + "yes_no_margin_reward": 0.125, + "token_credit_reward": -0.6327646374702454, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1064157709479332, + 0.0454409085214138, + 0.0685574933886528, + 0.06448836624622345, + 0.049409255385398865, + 0.05179604887962341, + 0.05620763450860977, + 0.06956374645233154, + 0.27399444580078125, + 0.21412630379199982 + ], + "credit_map": [ + 0.01888115145266056, + 0.011547873727977276, + 0.012109389528632164, + 0.010351179167628288, + 0.009647071361541748, + 0.013820108026266098, + 0.020295215770602226, + 0.009763798676431179, + 0.0057565635070204735, + 0.005013295914977789, + 0.004952890332788229, + 0.003967869095504284, + 0.00702094379812479, + 0.006678323727101088, + 0.004604942165315151, + 0.0074460809119045734, + 0.01212759967893362, + 0.009137023240327835, + 0.00959047582000494, + 0.006584692746400833, + 0.005266277585178614, + 0.0072187501937150955, + 0.011049038730561733, + 0.007583640981465578, + 0.00876615010201931, + 0.008901543915271759, + 0.007775102276355028, + 0.00604131119325757, + 0.0066442969255149364, + 0.006870916113257408, + 0.011181343346834183, + 0.00830770842730999, + 0.007836972363293171, + 0.005936515983194113, + 0.006735756061971188, + 0.005493737291544676, + 0.004143758676946163, + 0.005394994746893644, + 0.007028400432318449, + 0.006839122157543898, + 0.005585125181823969, + 0.005425652023404837, + 0.006326751783490181, + 0.005887272302061319, + 0.006280341185629368, + 0.008521314710378647, + 0.009605123661458492, + 0.004164470359683037, + 0.0067566451616585255, + 0.00560368224978447, + 0.007169290911406279, + 0.007991663180291653, + 0.006884932518005371, + 0.007988848723471165, + 0.008564340882003307, + 0.005248233210295439, + 0.011251913383603096, + 0.00801045075058937, + 0.009961234405636787, + 0.006963989697396755, + 0.007666375022381544, + 0.01101294718682766, + 0.006357674952596426, + 0.008339165709912777, + 0.03448027744889259, + 0.05093684792518616, + 0.04156466946005821, + 0.05599076673388481, + 0.022018982097506523, + 0.03283792734146118, + 0.012334151193499565, + 0.023830819875001907, + 0.026520727202296257, + 0.02266126498579979, + 0.043171390891075134, + 0.02289874479174614, + 0.02613167092204094, + 0.03429007902741432, + 0.017359288409352303, + 0.021093139424920082 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 0 + ], + "selected_reward": -0.6327646374702454 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"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 hand holding a black fountain pen with a gold-rimmed barrel visible writing on a sheet of paper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the pen tip metallic and iridescent, leaving dark blue cursive ink on off-white stationery?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene set on an elegant dark wood desk with a stack of heavy books faintly visible and blurred in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera pan slowly to the right while tracking the movement of the hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a soft, warm, slightly iridescent lens flare sweeping across the left side of the frame that grazes the pen tip?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026339344680309296 + ], + "opposite_logprob_tokens": [ + -3.651339292526245 + ], + "logprob_reward": -0.026339344680309296, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026339344680309296, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21993707120418549, + 0.04614071175456047, + 0.06215281039476395, + 0.12195344269275665, + 0.03700072690844536, + 0.021232955157756805, + 0.0319829136133194, + 0.04589790105819702, + 0.37316152453422546, + 0.0405399352312088 + ], + "credit_map": [ + 0.026436785236001015, + 0.01773274503648281, + 0.04401988163590431, + 0.022779714316129684, + 0.021054008975625038, + 0.02145882323384285, + 0.04112418368458748, + 0.02533094212412834, + 0.006963431369513273, + 0.00595358619466424, + 0.007427913136780262, + 0.004289044067263603, + 0.00412718066945672, + 0.005308929830789566, + 0.005875897593796253, + 0.006194732151925564, + 0.01233995147049427, + 0.010753796435892582, + 0.008300207555294037, + 0.005564763210713863, + 0.004479581955820322, + 0.007067587226629257, + 0.006296340376138687, + 0.0073505849577486515, + 0.012331259436905384, + 0.01408771239221096, + 0.013975430279970169, + 0.011953072622418404, + 0.01956172101199627, + 0.014760593883693218, + 0.023760946467518806, + 0.01152271032333374, + 0.0057922410778701305, + 0.006081879138946533, + 0.00420727813616395, + 0.002850784221664071, + 0.004698156379163265, + 0.005375522188842297, + 0.004591990262269974, + 0.0034028782974928617, + 0.003180120838806033, + 0.002765311161056161, + 0.002689557848498225, + 0.002032998949289322, + 0.002476718509569764, + 0.0026257282588630915, + 0.00337168388068676, + 0.0020908366423100233, + 0.004692062735557556, + 0.0028575623873621225, + 0.004236704669892788, + 0.0032239363063126802, + 0.002579800318926573, + 0.003252227557823062, + 0.006615995895117521, + 0.004524626303464174, + 0.004052660893648863, + 0.008821775205433369, + 0.0051162890158593655, + 0.004117628559470177, + 0.0073430417105555534, + 0.005212053656578064, + 0.0062731970101594925, + 0.004961258266121149, + 0.08844684809446335, + 0.036249712109565735, + 0.014707681722939014, + 0.03022758476436138, + 0.01562123280018568, + 0.13008911907672882, + 0.019122270867228508, + 0.03869706764817238, + 0.005599416792392731, + 0.004002964589744806, + 0.00386880268342793, + 0.004167977254837751, + 0.004949992522597313, + 0.009920288808643818, + 0.0031395694240927696, + 0.004890924319624901 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 3 + ], + "selected_reward": -0.026339344680309296 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "Shot from a slightly overhead angle, the camera pans incredibly slowly to the right, delicately tracking the movement of a hand writing on an elegant, dark wood desk. The hand's slender fingers firmly and firmly grasp a delicate black fountain pen with a gold-rimmed barrel. Its gleaming metallic iridium tip glides effortlessly and steadily across a sheet of thick, off-white stationery, leaving behind a trail of evenly inked, dark blue cursive. At this moment, a soft, warm, slightly iridescent lens flare, as if carefully choreographed, slowly sweeps across the left side of the frame, grazing the moving metal tip and causing it to flare with a dazzling, fleeting brilliance. The light comes from a warm, off-screen desk lamp, casting a soft, focused beam that casts a subtle, shifting shadow on the paper. A stack of heavy books is faintly visible in the background, but completely blurred. The main colors of the entire picture are warm brown and beige, with dark blue ink and golden light on the tip of the pen as embellishments, creating an atmosphere of extreme tranquility, concentration and full of classical bookishness.", + "seed": 27927, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000156/group_8/index_8685_seed_27927_dp_17_rank_34_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 8, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.08270714823156596, + "vlm_logprob": -0.08270714823156596, + "vlm_yes_no_margin": 3.55, + "vlm_token_credit": -0.08270714823156596, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"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 hand holding a black fountain pen with a gold-rimmed barrel visible writing on a sheet of paper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the pen tip metallic and iridescent, leaving dark blue cursive ink on off-white stationery?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene set on an elegant dark wood desk with a stack of heavy books faintly visible and blurred in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera pan slowly to the right while tracking the movement of the hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a soft, warm, slightly iridescent lens flare sweeping across the left side of the frame that grazes the pen tip?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020751507952809334 + ], + "opposite_logprob_tokens": [ + -4.145751476287842 + ], + "logprob_reward": -0.020751507952809334, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.020751507952809334, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11565590649843216, + 0.025307126343250275, + 0.2605360448360443, + 0.04810840263962746, + 0.12130679935216904, + 0.10051123052835464, + 0.16769210994243622, + 0.022241387516260147, + 0.05079134181141853, + 0.08784960955381393 + ], + "credit_map": [ + 0.021176595240831375, + 0.02124069258570671, + 0.010699511505663395, + 0.00935809500515461, + 0.016383502632379532, + 0.01699749007821083, + 0.010294563136994839, + 0.009505447000265121, + 0.002348966198042035, + 0.001932105515152216, + 0.002533555496484041, + 0.0030734443571418524, + 0.0028409999795258045, + 0.0035841544158756733, + 0.0051742615178227425, + 0.0038196404930204153, + 0.02800581231713295, + 0.045419879257678986, + 0.022763162851333618, + 0.024595409631729126, + 0.04553075134754181, + 0.04931555688381195, + 0.02246350795030594, + 0.02244199626147747, + 0.006500131916254759, + 0.007279660552740097, + 0.005240672267973423, + 0.005403615068644285, + 0.007037646137177944, + 0.00662371376529336, + 0.005281099118292332, + 0.004741864278912544, + 0.011798176914453506, + 0.012823916971683502, + 0.013020503334701061, + 0.012368413619697094, + 0.024071145802736282, + 0.016770416870713234, + 0.013942969962954521, + 0.016511254012584686, + 0.008863266557455063, + 0.012557598762214184, + 0.00633276766166091, + 0.0186393354088068, + 0.01766739971935749, + 0.015211007557809353, + 0.011018951423466206, + 0.01022090669721365, + 0.012503649108111858, + 0.031160756945610046, + 0.012669031508266926, + 0.012530507519841194, + 0.04458510875701904, + 0.02379738911986351, + 0.015722731128335, + 0.014722943305969238, + 0.0020549502223730087, + 0.0018029925413429737, + 0.002046047942712903, + 0.002706695580855012, + 0.003294846974313259, + 0.0031150109134614468, + 0.0035399310290813446, + 0.003680913243442774, + 0.006192781031131744, + 0.007601552177220583, + 0.006177928764373064, + 0.00523742800578475, + 0.0066365208476781845, + 0.005756196565926075, + 0.007642289623618126, + 0.005546645727008581, + 0.010083263739943504, + 0.012708589434623718, + 0.009919365867972374, + 0.008403975516557693, + 0.017910685390233994, + 0.010291646234691143, + 0.0097747091203928, + 0.008757376112043858 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 6, + 4 + ], + "selected_reward": -0.020751507952809334 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"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 hand holding a black fountain pen with a gold-rimmed barrel visible writing on a sheet of paper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the pen tip metallic and iridescent, leaving dark blue cursive ink on off-white stationery?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene set on an elegant dark wood desk with a stack of heavy books faintly visible and blurred in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera pan slowly to the right while tracking the movement of the hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a soft, warm, slightly iridescent lens flare sweeping across the left side of the frame that grazes the pen tip?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012553129345178604 + ], + "opposite_logprob_tokens": [ + -4.3875532150268555 + ], + "logprob_reward": -0.012553129345178604, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012553129345178604, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09642259031534195, + 0.04639403149485588, + 0.12410236150026321, + 0.06176729500293732, + 0.06908956170082092, + 0.32904696464538574, + 0.09794040024280548, + 0.0696386769413948, + 0.0649057999253273, + 0.04069232568144798 + ], + "credit_map": [ + 0.01582668535411358, + 0.009345549158751965, + 0.010908925905823708, + 0.008048267103731632, + 0.016760099679231644, + 0.015092083252966404, + 0.01093273051083088, + 0.009508242830634117, + 0.006893306504935026, + 0.005745865870267153, + 0.006732266861945391, + 0.004299765452742577, + 0.005705260206013918, + 0.006445967126637697, + 0.00470914738252759, + 0.005862451624125242, + 0.011276192031800747, + 0.016667533665895462, + 0.008472083136439323, + 0.027303846552968025, + 0.014044360257685184, + 0.0263059064745903, + 0.011588672176003456, + 0.00844376627355814, + 0.006872567813843489, + 0.008323371410369873, + 0.006306018214672804, + 0.007723088841885328, + 0.009695813059806824, + 0.009345085360109806, + 0.006861511617898941, + 0.0066398391500115395, + 0.010206530801951885, + 0.0072180298157036304, + 0.007278815843164921, + 0.0062601701356470585, + 0.008341304957866669, + 0.010105443187057972, + 0.009870093315839767, + 0.009809167124330997, + 0.031082989647984505, + 0.015747178345918655, + 0.01626882143318653, + 0.12756869196891785, + 0.02294783852994442, + 0.07128575444221497, + 0.02251720055937767, + 0.021628491580486298, + 0.008897789753973484, + 0.006040126550942659, + 0.005215534474700689, + 0.03406969830393791, + 0.010548303835093975, + 0.015625208616256714, + 0.009803912602365017, + 0.007739829830825329, + 0.011568689718842506, + 0.00801458302885294, + 0.01112485583871603, + 0.0075555117800831795, + 0.005988480057567358, + 0.008741292171180248, + 0.006952338386327028, + 0.00969292689114809, + 0.006349915638566017, + 0.009288246743381023, + 0.006194860674440861, + 0.008808249607682228, + 0.00710867065936327, + 0.008249599486589432, + 0.006057599559426308, + 0.012848655693233013, + 0.004058064427226782, + 0.005402783397585154, + 0.004417711403220892, + 0.004080327693372965, + 0.005428823176771402, + 0.005079480819404125, + 0.005577880889177322, + 0.006647254340350628 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 2, + 6 + ], + "selected_reward": -0.012553129345178604 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"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 hand holding a black fountain pen with a gold-rimmed barrel visible writing on a sheet of paper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the pen tip metallic and iridescent, leaving dark blue cursive ink on off-white stationery?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene set on an elegant dark wood desk with a stack of heavy books faintly visible and blurred in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera pan slowly to the right while tracking the movement of the hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a soft, warm, slightly iridescent lens flare sweeping across the left side of the frame that grazes the pen tip?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011073562316596508 + ], + "opposite_logprob_tokens": [ + -4.511073589324951 + ], + "logprob_reward": -0.011073562316596508, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011073562316596508, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06285064667463303, + 0.03259405121207237, + 0.5781381726264954, + 0.026516437530517578, + 0.04495676979422569, + 0.058512765914201736, + 0.12718689441680908, + 0.025135910138487816, + 0.021429594606161118, + 0.022678766399621964 + ], + "credit_map": [ + 0.007369754835963249, + 0.007798434235155582, + 0.005239181686192751, + 0.006512602791190147, + 0.011001822538673878, + 0.009139765985310078, + 0.007186395116150379, + 0.008602687157690525, + 0.0042099314741790295, + 0.004106629174202681, + 0.0034680371172726154, + 0.002777338260784745, + 0.004481486976146698, + 0.006142148282378912, + 0.00393840204924345, + 0.0034700771793723106, + 0.057174574583768845, + 0.08803490549325943, + 0.05042228102684021, + 0.07925025373697281, + 0.08668365329504013, + 0.1355414092540741, + 0.03881172090768814, + 0.0422193817794323, + 0.003957333508878946, + 0.003648186568170786, + 0.002526551950722933, + 0.004467679187655449, + 0.0031557625625282526, + 0.0038615961093455553, + 0.0025252855848520994, + 0.0023740422911942005, + 0.006651751231402159, + 0.006298101972788572, + 0.005562254227697849, + 0.004730812273919582, + 0.006211608648300171, + 0.006799523718655109, + 0.004247358534485102, + 0.004455358721315861, + 0.005515066906809807, + 0.0032159604597836733, + 0.0036586427595466375, + 0.0230745580047369, + 0.002706458093598485, + 0.012466400861740112, + 0.004181095398962498, + 0.003694580402225256, + 0.01116592064499855, + 0.00831411499530077, + 0.006518276408314705, + 0.04719751328229904, + 0.012142756022512913, + 0.022300582379102707, + 0.011182979680597782, + 0.008364749141037464, + 0.004523555748164654, + 0.0022909133695065975, + 0.003138081869110465, + 0.004269198514521122, + 0.0028098865877836943, + 0.0027566589415073395, + 0.0020039603114128113, + 0.0033436582889407873, + 0.0032108123414218426, + 0.0027765657287091017, + 0.0033494390081614256, + 0.002517364453524351, + 0.0026451838202774525, + 0.0024796691723167896, + 0.002343756379559636, + 0.0021068037021905184, + 0.0027098273858428, + 0.0034734054934233427, + 0.002356282901018858, + 0.002123435726389289, + 0.003957951441407204, + 0.002479991875588894, + 0.002615726087242365, + 0.0029621473513543606 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 6, + 0 + ], + "selected_reward": -0.011073562316596508 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"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 hand holding a black fountain pen with a gold-rimmed barrel visible writing on a sheet of paper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the pen tip metallic and iridescent, leaving dark blue cursive ink on off-white stationery?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene set on an elegant dark wood desk with a stack of heavy books faintly visible and blurred in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera pan slowly to the right while tracking the movement of the hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a soft, warm, slightly iridescent lens flare sweeping across the left side of the frame that grazes the pen tip?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.34858667850494385 + ], + "opposite_logprob_tokens": [ + -1.2235866785049438 + ], + "logprob_reward": -0.34858667850494385, + "yes_no_margin_reward": 0.875, + "token_credit_reward": -0.34858667850494385, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11686842888593674, + 0.0822981745004654, + 0.17247265577316284, + 0.14751197397708893, + 0.18260806798934937, + 0.055247947573661804, + 0.06190599873661995, + 0.05445391312241554, + 0.03687284514307976, + 0.08976005017757416 + ], + "credit_map": [ + 0.017560871317982674, + 0.01810595579445362, + 0.014334945939481258, + 0.010608603246510029, + 0.012775715440511703, + 0.019252954050898552, + 0.0077021773904562, + 0.016527196392416954, + 0.009669501334428787, + 0.0107975322753191, + 0.007447164971381426, + 0.00811326876282692, + 0.012834125198423862, + 0.011921948753297329, + 0.012108853086829185, + 0.009405773133039474, + 0.017044123262166977, + 0.02713531255722046, + 0.011878623627126217, + 0.029059134423732758, + 0.02413778193295002, + 0.04062659665942192, + 0.011687053367495537, + 0.010904001072049141, + 0.01940028928220272, + 0.020203154534101486, + 0.014591686427593231, + 0.02084336243569851, + 0.02397741563618183, + 0.022773485630750656, + 0.013966365717351437, + 0.011756202206015587, + 0.02584589459002018, + 0.019553780555725098, + 0.020146053284406662, + 0.01973406970500946, + 0.022665180265903473, + 0.029162198305130005, + 0.019985126331448555, + 0.02551574632525444, + 0.005292150657624006, + 0.003823998151347041, + 0.002925825770944357, + 0.01919141411781311, + 0.004474184941500425, + 0.011641609482467175, + 0.0038438907358795404, + 0.004054867662489414, + 0.004702577833086252, + 0.00782208051532507, + 0.006679124664515257, + 0.008689737878739834, + 0.012303212657570839, + 0.007428220473229885, + 0.007061442825943232, + 0.007219596765935421, + 0.005318342708051205, + 0.005741795990616083, + 0.006377560552209616, + 0.00961616076529026, + 0.009750359691679478, + 0.00623995391651988, + 0.005030412692576647, + 0.006379321683198214, + 0.00422290712594986, + 0.0053005525842309, + 0.004782635252922773, + 0.0032083597034215927, + 0.005629844963550568, + 0.0044631133787333965, + 0.00475807162001729, + 0.004507357254624367, + 0.011622468009591103, + 0.015202246606349945, + 0.010984554886817932, + 0.009484630078077316, + 0.012607526034116745, + 0.009492573328316212, + 0.008781049400568008, + 0.01158499252051115 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 2, + 3 + ], + "selected_reward": -0.34858667850494385 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"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 hand holding a black fountain pen with a gold-rimmed barrel visible writing on a sheet of paper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the pen tip metallic and iridescent, leaving dark blue cursive ink on off-white stationery?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene set on an elegant dark wood desk with a stack of heavy books faintly visible and blurred in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera pan slowly to the right while tracking the movement of the hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a soft, warm, slightly iridescent lens flare sweeping across the left side of the frame that grazes the pen tip?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020570863038301468 + ], + "opposite_logprob_tokens": [ + -3.895570755004883 + ], + "logprob_reward": -0.020570863038301468, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020570863038301468, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06315739452838898, + 0.02230810560286045, + 0.10122701525688171, + 0.01676054671406746, + 0.03957768902182579, + 0.2794581949710846, + 0.3017050623893738, + 0.12516562640666962, + 0.013933788053691387, + 0.03670649975538254 + ], + "credit_map": [ + 0.01064269244670868, + 0.010290771722793579, + 0.008358758874237537, + 0.008415121585130692, + 0.005914182402193546, + 0.009620201773941517, + 0.0044746375642716885, + 0.005441031884402037, + 0.004272237420082092, + 0.002588162198662758, + 0.002976221963763237, + 0.0023181228898465633, + 0.002523668110370636, + 0.0031285949517041445, + 0.0019179590744897723, + 0.002583141205832362, + 0.009158884175121784, + 0.01340731792151928, + 0.008001262322068214, + 0.021243063732981682, + 0.012185564264655113, + 0.02411208488047123, + 0.006775316782295704, + 0.006343524903059006, + 0.002482357667759061, + 0.0024339579977095127, + 0.001992720179259777, + 0.00220316625200212, + 0.0023248784709721804, + 0.0024658164475113153, + 0.0014480091631412506, + 0.0014096423983573914, + 0.004916155710816383, + 0.005780579522252083, + 0.00489378347992897, + 0.0029685955960303545, + 0.006588061340153217, + 0.006340389139950275, + 0.0036207986995577812, + 0.004469325765967369, + 0.02716725878417492, + 0.011824041604995728, + 0.01348404586315155, + 0.11387556791305542, + 0.014442000538110733, + 0.061282332986593246, + 0.019313178956508636, + 0.018069764599204063, + 0.024014687165617943, + 0.030264336615800858, + 0.01715107262134552, + 0.09805933386087418, + 0.037085019052028656, + 0.04905497282743454, + 0.025519641116261482, + 0.020556019619107246, + 0.023057235404849052, + 0.010854791849851608, + 0.018199775367975235, + 0.013385564088821411, + 0.014318571425974369, + 0.013748452998697758, + 0.011702788062393665, + 0.019898463040590286, + 0.0018383028218522668, + 0.0016391269164159894, + 0.001732647535391152, + 0.0017239786684513092, + 0.00196163821965456, + 0.001718616229481995, + 0.0016933807637542486, + 0.0016260973643511534, + 0.0053502884693443775, + 0.004819411318749189, + 0.006406791508197784, + 0.004364903550595045, + 0.0043027582578361034, + 0.0036352965980768204, + 0.0035973500926047564, + 0.004229702055454254 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 5, + 7 + ], + "selected_reward": -0.020570863038301468 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "Shot from a slightly overhead angle, the camera pans incredibly slowly to the right, delicately tracking the movement of a hand writing on an elegant, dark wood desk. The hand's slender fingers firmly and firmly grasp a delicate black fountain pen with a gold-rimmed barrel. Its gleaming metallic iridium tip glides effortlessly and steadily across a sheet of thick, off-white stationery, leaving behind a trail of evenly inked, dark blue cursive. At this moment, a soft, warm, slightly iridescent lens flare, as if carefully choreographed, slowly sweeps across the left side of the frame, grazing the moving metal tip and causing it to flare with a dazzling, fleeting brilliance. The light comes from a warm, off-screen desk lamp, casting a soft, focused beam that casts a subtle, shifting shadow on the paper. A stack of heavy books is faintly visible in the background, but completely blurred. The main colors of the entire picture are warm brown and beige, with dark blue ink and golden light on the tip of the pen as embellishments, creating an atmosphere of extreme tranquility, concentration and full of classical bookishness.", + "seed": 27928, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000156/group_8/index_8685_seed_27928_dp_17_rank_34_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 8, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.07087984010577202, + "vlm_logprob": -0.07087984010577202, + "vlm_yes_no_margin": 3.3749999523162844, + "vlm_token_credit": -0.07087984010577202, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"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 hand holding a black fountain pen with a gold-rimmed barrel visible writing on a sheet of paper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the pen tip metallic and iridescent, leaving dark blue cursive ink on off-white stationery?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene set on an elegant dark wood desk with a stack of heavy books faintly visible and blurred in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera pan slowly to the right while tracking the movement of the hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a soft, warm, slightly iridescent lens flare sweeping across the left side of the frame that grazes the pen tip?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01666623167693615 + ], + "opposite_logprob_tokens": [ + -4.391666412353516 + ], + "logprob_reward": -0.01666623167693615, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.01666623167693615, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.138457790017128, + 0.12054146081209183, + 0.06568082422018051, + 0.0941937044262886, + 0.10503596812486649, + 0.12735241651535034, + 0.04236023873090744, + 0.053665418177843094, + 0.0876101404428482, + 0.1651020646095276 + ], + "credit_map": [ + 0.010150657966732979, + 0.012278184294700623, + 0.02648126520216465, + 0.011933217756450176, + 0.0076103294268250465, + 0.010786792263388634, + 0.04896077513694763, + 0.010256562381982803, + 0.0109174232929945, + 0.011891827918589115, + 0.014212428592145443, + 0.014828654937446117, + 0.014231172390282154, + 0.02643338032066822, + 0.015827510505914688, + 0.012199065648019314, + 0.005582121200859547, + 0.010143275372684002, + 0.009620166383683681, + 0.007308707572519779, + 0.00744359428063035, + 0.011332753114402294, + 0.009609123691916466, + 0.0046410877257585526, + 0.013561690226197243, + 0.012628147378563881, + 0.01045999489724636, + 0.012685076333582401, + 0.011754334904253483, + 0.011290897615253925, + 0.013340030796825886, + 0.008473536930978298, + 0.014310476370155811, + 0.012447983026504517, + 0.009371442720293999, + 0.015107621438801289, + 0.011179392226040363, + 0.016058433800935745, + 0.013338206335902214, + 0.013222417794167995, + 0.008933319710195065, + 0.04085444286465645, + 0.005774251185357571, + 0.008316818624734879, + 0.00697742635384202, + 0.01158900372684002, + 0.005330701358616352, + 0.0395764522254467, + 0.0036408943124115467, + 0.0030165656935423613, + 0.002712556393817067, + 0.0028563279192894697, + 0.0046917470172047615, + 0.008108961395919323, + 0.008214098401367664, + 0.009119088761508465, + 0.0040639606304466724, + 0.0033563938923180103, + 0.006072950083762407, + 0.017870623618364334, + 0.004412445239722729, + 0.01047180499881506, + 0.004317576996982098, + 0.0030996669083833694, + 0.005125855561345816, + 0.004728451371192932, + 0.007598142605274916, + 0.03400725871324539, + 0.004624778404831886, + 0.022534627467393875, + 0.004946273751556873, + 0.004044756758958101, + 0.01356159895658493, + 0.009452165104448795, + 0.012833703309297562, + 0.06381621211767197, + 0.008220690302550793, + 0.019165489822626114, + 0.014276450499892235, + 0.023775773122906685 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 5 + ], + "selected_reward": -0.01666623167693615 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"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 hand holding a black fountain pen with a gold-rimmed barrel visible writing on a sheet of paper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the pen tip metallic and iridescent, leaving dark blue cursive ink on off-white stationery?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene set on an elegant dark wood desk with a stack of heavy books faintly visible and blurred in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera pan slowly to the right while tracking the movement of the hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a soft, warm, slightly iridescent lens flare sweeping across the left side of the frame that grazes the pen tip?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018200064077973366 + ], + "opposite_logprob_tokens": [ + -4.018199920654297 + ], + "logprob_reward": -0.018200064077973366, + "yes_no_margin_reward": 3.999999761581421, + "token_credit_reward": -0.018200064077973366, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1644437313079834, + 0.08643944561481476, + 0.055771321058273315, + 0.19106806814670563, + 0.10407958179712296, + 0.12663057446479797, + 0.014852828346192837, + 0.06522566080093384, + 0.11766788363456726, + 0.07382085174322128 + ], + "credit_map": [ + 0.017009831964969635, + 0.024118168279528618, + 0.034243304282426834, + 0.027013996616005898, + 0.012531637214124203, + 0.013377031311392784, + 0.026673849672079086, + 0.009475919418036938, + 0.0069702742621302605, + 0.010012359358370304, + 0.011367514729499817, + 0.011144071817398071, + 0.010377478785812855, + 0.014880959875881672, + 0.008383991196751595, + 0.013302801176905632, + 0.005074887070804834, + 0.005440364591777325, + 0.005891561042517424, + 0.007502717897295952, + 0.0061414809897542, + 0.014211973175406456, + 0.006346653215587139, + 0.005161686334758997, + 0.01704961434006691, + 0.0410572811961174, + 0.012868654914200306, + 0.04035511240363121, + 0.01330476626753807, + 0.019642360508441925, + 0.013719983398914337, + 0.03307030349969864, + 0.015130268409848213, + 0.013257541693747044, + 0.009826659224927425, + 0.01494155265390873, + 0.010471165180206299, + 0.014698958024382591, + 0.011139703914523125, + 0.014613740146160126, + 0.008296790532767773, + 0.03355173021554947, + 0.004734899383038282, + 0.008885477669537067, + 0.008093535900115967, + 0.01208010595291853, + 0.010591062717139721, + 0.040396977216005325, + 0.0014968279283493757, + 0.0013418877497315407, + 0.0015174454310908914, + 0.00133758585434407, + 0.0019534090533852577, + 0.001983514055609703, + 0.0034214595798403025, + 0.0018007009057328105, + 0.003540672594681382, + 0.0029915240593254566, + 0.005597286392003298, + 0.02017366513609886, + 0.004322102293372154, + 0.016750771552324295, + 0.004803922958672047, + 0.007045721635222435, + 0.00612292904406786, + 0.005272824317216873, + 0.009241540916264057, + 0.04208473861217499, + 0.0052036563865840435, + 0.03390614315867424, + 0.006878027692437172, + 0.008958030492067337, + 0.0061086444184184074, + 0.004658038727939129, + 0.005213453434407711, + 0.027569016441702843, + 0.004679910838603973, + 0.007125462871044874, + 0.006746417842805386, + 0.011719915084540844 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 5 + ], + "selected_reward": -0.018200064077973366 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"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 hand holding a black fountain pen with a gold-rimmed barrel visible writing on a sheet of paper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the pen tip metallic and iridescent, leaving dark blue cursive ink on off-white stationery?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene set on an elegant dark wood desk with a stack of heavy books faintly visible and blurred in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera pan slowly to the right while tracking the movement of the hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a soft, warm, slightly iridescent lens flare sweeping across the left side of the frame that grazes the pen tip?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012537237256765366 + ], + "opposite_logprob_tokens": [ + -4.387537002563477 + ], + "logprob_reward": -0.012537237256765366, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012537237256765366, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3990427553653717, + 0.044012051075696945, + 0.03976590931415558, + 0.0943434089422226, + 0.056742843240499496, + 0.03647866100072861, + 0.03484494239091873, + 0.06240200251340866, + 0.0717037171125412, + 0.1606636792421341 + ], + "credit_map": [ + 0.036497507244348526, + 0.04684486612677574, + 0.09930548816919327, + 0.03200094401836395, + 0.021289879456162453, + 0.032693564891815186, + 0.10690565407276154, + 0.023504851385951042, + 0.004669083748012781, + 0.003751373616978526, + 0.005547491833567619, + 0.006958637852221727, + 0.00422867713496089, + 0.007589795161038637, + 0.0069124470464885235, + 0.004354540724307299, + 0.003324270946905017, + 0.008210740983486176, + 0.006172917317599058, + 0.004272723104804754, + 0.005027439445257187, + 0.005142671521753073, + 0.004610073287039995, + 0.003005071310326457, + 0.008746704086661339, + 0.026940884068608284, + 0.0075707389041781425, + 0.028619449585676193, + 0.005418930668383837, + 0.007307893130928278, + 0.005080645438283682, + 0.004658164922147989, + 0.007436858955770731, + 0.007600049953907728, + 0.003615040797740221, + 0.009931537322700024, + 0.004616140853613615, + 0.007819048129022121, + 0.004430153872817755, + 0.011294013820588589, + 0.0032260941807180643, + 0.00959030818194151, + 0.003831390291452408, + 0.0076682488434016705, + 0.0017942046979442239, + 0.003491794690489769, + 0.002183401258662343, + 0.00469322269782424, + 0.0027226393576711416, + 0.0024984313640743494, + 0.0033818217925727367, + 0.002532867481932044, + 0.0021410826593637466, + 0.007612648885697126, + 0.0030597716104239225, + 0.010895678773522377, + 0.004284664522856474, + 0.0033947485499083996, + 0.005611611530184746, + 0.022958872839808464, + 0.004481971263885498, + 0.013943905010819435, + 0.004410048481076956, + 0.0033161796163767576, + 0.005017403047531843, + 0.005045686382800341, + 0.006171088665723801, + 0.02435530349612236, + 0.005829649977385998, + 0.014159142039716244, + 0.003892937209457159, + 0.007232503965497017, + 0.017550328746438026, + 0.013134735636413097, + 0.012404200620949268, + 0.050005871802568436, + 0.010345051996409893, + 0.014112765900790691, + 0.012047141790390015, + 0.031063590198755264 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 3 + ], + "selected_reward": -0.012537237256765366 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a hand holding a black fountain pen with a gold-rimmed barrel visible writing on a sheet of paper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the pen tip metallic and iridescent, leaving dark blue cursive ink on off-white stationery?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene set on an elegant dark wood desk with a stack of heavy books faintly visible and blurred in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera pan slowly to the right while tracking the movement of the hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a soft, warm, slightly iridescent lens flare sweeping across the left side of the frame that grazes the pen tip?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2520598769187927 + ], + "opposite_logprob_tokens": [ + -1.5020599365234375 + ], + "logprob_reward": -0.2520598769187927, + "yes_no_margin_reward": 1.25, + "token_credit_reward": -0.2520598769187927, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20094358921051025, + 0.035554371774196625, + 0.0431893989443779, + 0.09098765254020691, + 0.026464536786079407, + 0.039575040340423584, + 0.027538593858480453, + 0.1436009258031845, + 0.0985792726278305, + 0.29356658458709717 + ], + "credit_map": [ + 0.012981819920241833, + 0.011810779571533203, + 0.04281717166304588, + 0.01615026593208313, + 0.01548963412642479, + 0.02343038097023964, + 0.0660332441329956, + 0.012230291962623596, + 0.004171012435108423, + 0.00431158347055316, + 0.0035571891348809004, + 0.005733156111091375, + 0.0031722949352115393, + 0.00437586847692728, + 0.005050950683653355, + 0.005182319320738316, + 0.004366735462099314, + 0.009727885946631432, + 0.005150820594280958, + 0.004994505550712347, + 0.004937862046062946, + 0.006372272036969662, + 0.004385594744235277, + 0.003253720700740814, + 0.007676220033317804, + 0.02401421032845974, + 0.011292822659015656, + 0.028233567252755165, + 0.005140320397913456, + 0.006262236274778843, + 0.004664276726543903, + 0.0037040056195110083, + 0.0026408531703054905, + 0.0032903982792049646, + 0.002770174527540803, + 0.004517088178545237, + 0.003231485141441226, + 0.003820545971393585, + 0.002148758852854371, + 0.004045234061777592, + 0.005530670750886202, + 0.00611845450475812, + 0.004641154780983925, + 0.007444640155881643, + 0.003784332424402237, + 0.00580047769472003, + 0.003228113055229187, + 0.0030271976720541716, + 0.0020585504826158285, + 0.002730417298153043, + 0.0032487711869180202, + 0.0035351396072655916, + 0.002705938648432493, + 0.0046186670660972595, + 0.0028362583834677935, + 0.0058048502542078495, + 0.009581336751580238, + 0.007101569790393114, + 0.012458037585020065, + 0.02950078435242176, + 0.012695757672190666, + 0.03466874361038208, + 0.013168659061193466, + 0.024426039308309555, + 0.006416986230760813, + 0.007713895756751299, + 0.007318978663533926, + 0.030729802325367928, + 0.0079291220754385, + 0.010308122262358665, + 0.010834095068275928, + 0.01732826419174671, + 0.02013508789241314, + 0.0176004059612751, + 0.020655721426010132, + 0.10963945835828781, + 0.021078085526823997, + 0.05956033244729042, + 0.029534660279750824, + 0.015362839214503765 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 7 + ], + "selected_reward": -0.2520598769187927 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"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 hand holding a black fountain pen with a gold-rimmed barrel visible writing on a sheet of paper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the pen tip metallic and iridescent, leaving dark blue cursive ink on off-white stationery?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene set on an elegant dark wood desk with a stack of heavy books faintly visible and blurred in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera pan slowly to the right while tracking the movement of the hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a soft, warm, slightly iridescent lens flare sweeping across the left side of the frame that grazes the pen tip?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05493579059839249 + ], + "opposite_logprob_tokens": [ + -2.9299356937408447 + ], + "logprob_reward": -0.05493579059839249, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05493579059839249, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.30834266543388367, + 0.09855953603982925, + 0.03543904423713684, + 0.05163620784878731, + 0.04203716292977333, + 0.04848991334438324, + 0.032850828021764755, + 0.056978899985551834, + 0.1811961829662323, + 0.14446954429149628 + ], + "credit_map": [ + 0.023033786565065384, + 0.01960335113108158, + 0.06805556267499924, + 0.03257565572857857, + 0.02013380080461502, + 0.024978388100862503, + 0.0998804122209549, + 0.02008170820772648, + 0.008364522829651833, + 0.010604895651340485, + 0.012011298909783363, + 0.009241994470357895, + 0.012215163558721542, + 0.027943432331085205, + 0.007772760931402445, + 0.0104054631665349, + 0.004293997306376696, + 0.005707291420549154, + 0.00332478410564363, + 0.004461575765162706, + 0.004882437642663717, + 0.00620284816250205, + 0.003964619245380163, + 0.002601487562060356, + 0.005283997859805822, + 0.008704615756869316, + 0.005099549423903227, + 0.004570285324007273, + 0.00561339920386672, + 0.008128820918500423, + 0.007251964416354895, + 0.006983574479818344, + 0.005067281424999237, + 0.006844509858638048, + 0.003765630070120096, + 0.0071513717994093895, + 0.003773740492761135, + 0.004961633123457432, + 0.004414490889757872, + 0.0060585071332752705, + 0.0037154671736061573, + 0.011670291423797607, + 0.004105681087821722, + 0.003342381678521633, + 0.003359068650752306, + 0.005492680706083775, + 0.00430759321898222, + 0.01249675266444683, + 0.0025224245619028807, + 0.0028582268860191107, + 0.0031769853085279465, + 0.00298863323405385, + 0.004158286843448877, + 0.004575171507894993, + 0.005880351644009352, + 0.0066907452419400215, + 0.004908973351120949, + 0.004686197731643915, + 0.00756474444642663, + 0.015355490148067474, + 0.0051828790456056595, + 0.010243915021419525, + 0.005065257661044598, + 0.003971436060965061, + 0.008995459415018559, + 0.009748903103172779, + 0.014782479032874107, + 0.060498010367155075, + 0.009080745279788971, + 0.05054108425974846, + 0.011253992095589638, + 0.01629551872611046, + 0.009448041208088398, + 0.010561434552073479, + 0.009692724794149399, + 0.05702971667051315, + 0.008860370144248009, + 0.025500865653157234, + 0.013095918111503124, + 0.010280472226440907 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 9 + ], + "selected_reward": -0.05493579059839249 + } + ], + "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/0000156/group_9/index_tensor([768, 768])_group_9_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000156/group_9/index_tensor([768, 768])_group_9_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..7d24d6e50efc69edd9aa6ad29ab1c40eb4d45918 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000156/group_9/index_tensor([768, 768])_group_9_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "In stunning, captivatingly detailed Pixar-style execution, the camera begins a slow, graceful 360-degree rotation, arcing around a massive hot air balloon. The balloon's shell is a vibrant tapestry of vertical panels in cherry red, bright yellow, and sky blue, all with a smooth, slightly glossy texture that captures the bright morning light. A playful team logo, featuring a stylized cartoon badger wearing aviator goggles, adorns the side, rendered in bold, clean lines. As the camera continues its calming motion, it reveals a wicker basket beneath, where a team of three contestants stands. One, a woman with red hair tied in a practical ponytail and large, expressive green eyes, looks at her with friendly determination as she operates the burner. Beside her, a man with a warm smile and a neatly trimmed beard peers through a pair of oversized binoculars, while a third teammate excitedly points at the horizon. They all wear matching blue and yellow trench coats. The entire scene, set against a vast, soft blue sky dotted with a few fluffy, cotton-like clouds, creates a cheerful, adventurous atmosphere filled with optimistic energy and a spirit of friendly competition.", + "seed": 23922, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000156/group_9/index_768_seed_23922_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": -0.05437328172847629, + "vlm_logprob": -0.05437328172847629, + "vlm_yes_no_margin": 4.45, + "vlm_token_credit": -0.05437328172847629, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" 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 perform a slow, graceful 360-degree rotation around the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive hot air balloon with a shell featuring vertical panels in cherry red, bright yellow, and sky blue visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a wicker basket visible beneath the balloon containing three contestants?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the three contestants wear matching blue and yellow trench coats?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a vast, soft blue sky with a few fluffy, cotton-like clouds?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.13378497958183289 + ], + "opposite_logprob_tokens": [ + -2.1337850093841553 + ], + "logprob_reward": -0.13378497958183289, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.13378497958183289, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0934712365269661, + 0.09512782841920853, + 0.058719828724861145, + 0.052828893065452576, + 0.12739385664463043, + 0.2154698520898819, + 0.2287290096282959, + 0.05905851721763611, + 0.039611998945474625, + 0.029589012265205383 + ], + "credit_map": [ + 0.019597772508859634, + 0.012852326966822147, + 0.009594222530722618, + 0.01971161924302578, + 0.007835036143660545, + 0.006190588232129812, + 0.008177990093827248, + 0.009511680342257023, + 0.02113720215857029, + 0.01346662174910307, + 0.00988329853862524, + 0.01345051359385252, + 0.00984367448836565, + 0.007511613424867392, + 0.010545325465500355, + 0.00928957387804985, + 0.006474437657743692, + 0.008931090123951435, + 0.004332875367254019, + 0.0056144162081182, + 0.015811840072274208, + 0.010519861243665218, + 0.003721062559634447, + 0.0033142452593892813, + 0.0093906931579113, + 0.006924796383827925, + 0.005741576198488474, + 0.009211345575749874, + 0.008386172354221344, + 0.004775138571858406, + 0.004840476904064417, + 0.003558696713298559, + 0.014536537230014801, + 0.025659725069999695, + 0.019968559965491295, + 0.010678362101316452, + 0.021022582426667213, + 0.008000130765140057, + 0.01191597618162632, + 0.01561198104172945, + 0.018138958141207695, + 0.044121887534856796, + 0.02826494723558426, + 0.031016411259770393, + 0.03758782893419266, + 0.01015276089310646, + 0.019371140748262405, + 0.02681591548025608, + 0.023937545716762543, + 0.015034492127597332, + 0.0384969487786293, + 0.020243370905518532, + 0.019011151045560837, + 0.020293720066547394, + 0.027395086362957954, + 0.06431669741868973, + 0.007146860007196665, + 0.005758058279752731, + 0.007579368073493242, + 0.006898429710417986, + 0.005871675908565521, + 0.005953141488134861, + 0.006377547048032284, + 0.01347343996167183, + 0.0044380901381373405, + 0.006823268253356218, + 0.004858258180320263, + 0.004918184131383896, + 0.006086632143706083, + 0.003932945430278778, + 0.004100965801626444, + 0.004453654866665602, + 0.004070529714226723, + 0.00497282762080431, + 0.004254783969372511, + 0.0034383658785372972, + 0.0028193171601742506, + 0.0027929511852562428, + 0.003623368451371789, + 0.003616867819800973 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 5, + 4 + ], + "selected_reward": -0.13378497958183289 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" 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 perform a slow, graceful 360-degree rotation around the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive hot air balloon with a shell featuring vertical panels in cherry red, bright yellow, and sky blue visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a wicker basket visible beneath the balloon containing three contestants?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the three contestants wear matching blue and yellow trench coats?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a vast, soft blue sky with a few fluffy, cotton-like clouds?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0013506823452189565 + ], + "opposite_logprob_tokens": [ + -6.6263508796691895 + ], + "logprob_reward": -0.0013506823452189565, + "yes_no_margin_reward": 6.625, + "token_credit_reward": -0.0013506823452189565, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07361835986375809, + 0.050499334931373596, + 0.06966596096754074, + 0.07610185444355011, + 0.13513711094856262, + 0.23093870282173157, + 0.14618250727653503, + 0.06521481275558472, + 0.07649575173854828, + 0.07614561170339584 + ], + "credit_map": [ + 0.013051284477114677, + 0.009210494346916676, + 0.006037987302988768, + 0.009770981967449188, + 0.014804953709244728, + 0.007034874986857176, + 0.005750262178480625, + 0.007957527413964272, + 0.007645399309694767, + 0.006210552994161844, + 0.005061028525233269, + 0.00825576949864626, + 0.007209297269582748, + 0.005805687513202429, + 0.004690151195973158, + 0.0056214542128145695, + 0.009414871223270893, + 0.010114308446645737, + 0.0034180188085883856, + 0.007768961600959301, + 0.02007715217769146, + 0.011645760387182236, + 0.0033701835200190544, + 0.003856718773022294, + 0.005693433340638876, + 0.017637135460972786, + 0.008010843768715858, + 0.016419799998402596, + 0.013218516483902931, + 0.005173244513571262, + 0.005966643337160349, + 0.003982242662459612, + 0.017545724287629128, + 0.022847840562462807, + 0.025355596095323563, + 0.012582411989569664, + 0.01969590224325657, + 0.007668917533010244, + 0.010398617014288902, + 0.01904209889471531, + 0.024153513833880424, + 0.05570332333445549, + 0.023377669975161552, + 0.0256256815046072, + 0.045654505491256714, + 0.012428469024598598, + 0.022399429231882095, + 0.021596139296889305, + 0.014338391833007336, + 0.012524083256721497, + 0.020789604634046555, + 0.017275238409638405, + 0.010815453715622425, + 0.011587316170334816, + 0.0272001251578331, + 0.0316523052752018, + 0.007833395153284073, + 0.008390090428292751, + 0.007554106414318085, + 0.006170119624584913, + 0.005482198670506477, + 0.008762665092945099, + 0.0076405382715165615, + 0.013381710276007652, + 0.008907265961170197, + 0.01284049917012453, + 0.008401723578572273, + 0.009032937698066235, + 0.011409520171582699, + 0.009556248784065247, + 0.008264926262199879, + 0.008082645013928413, + 0.008843320421874523, + 0.01270582340657711, + 0.010133679956197739, + 0.009089089930057526, + 0.008996674790978432, + 0.011334385722875595, + 0.008008607663214207, + 0.007034039590507746 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 6, + 4 + ], + "selected_reward": -0.0013506823452189565 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" 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 perform a slow, graceful 360-degree rotation around the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive hot air balloon with a shell featuring vertical panels in cherry red, bright yellow, and sky blue visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a wicker basket visible beneath the balloon containing three contestants?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the three contestants wear matching blue and yellow trench coats?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a vast, soft blue sky with a few fluffy, cotton-like clouds?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"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": [ + 0.05825044959783554, + 0.05702388286590576, + 0.049088723957538605, + 0.04031744971871376, + 0.03377288952469826, + 0.0746527910232544, + 0.12903335690498352, + 0.10838799178600311, + 0.11727281659841537, + 0.3321996331214905 + ], + "credit_map": [ + 0.008144384250044823, + 0.00656794011592865, + 0.005551143083721399, + 0.009892293252050877, + 0.008224335499107838, + 0.0071356529369950294, + 0.005436294246464968, + 0.0072984034195542336, + 0.009356987662613392, + 0.007728658616542816, + 0.004819144494831562, + 0.00925567839294672, + 0.010704060085117817, + 0.0055472454987466335, + 0.004725540056824684, + 0.004886566661298275, + 0.006703645922243595, + 0.006041586399078369, + 0.002932156203314662, + 0.00447889743372798, + 0.0122026726603508, + 0.009420493617653847, + 0.0036371713504195213, + 0.003672099905088544, + 0.00560021074488759, + 0.00555074866861105, + 0.0036046579480171204, + 0.007045000325888395, + 0.0066207703202962875, + 0.003976996056735516, + 0.004831771366298199, + 0.0030872945208102465, + 0.0037674277555197477, + 0.008413979783654213, + 0.003032818902283907, + 0.0036469215992838144, + 0.006053987890481949, + 0.002091937232762575, + 0.004088591318577528, + 0.002677223877981305, + 0.0077985781244933605, + 0.008944587782025337, + 0.012642421759665012, + 0.0075826458632946014, + 0.008864004164934158, + 0.005858571268618107, + 0.009994324296712875, + 0.012967662885785103, + 0.012828604318201542, + 0.011061565019190311, + 0.019478578120470047, + 0.014718437567353249, + 0.009378709830343723, + 0.01187989767640829, + 0.024299846962094307, + 0.025387713685631752, + 0.014249471947550774, + 0.015683183446526527, + 0.01175754051655531, + 0.011410737410187721, + 0.013151329010725021, + 0.012838432565331459, + 0.010377796366810799, + 0.018919508904218674, + 0.014924095012247562, + 0.011396631598472595, + 0.013838615268468857, + 0.02094906195998192, + 0.017149019986391068, + 0.009996133856475353, + 0.016011014580726624, + 0.013008244335651398, + 0.030916711315512657, + 0.035102419555187225, + 0.042316410690546036, + 0.047242771834135056, + 0.0412418358027935, + 0.05741088092327118, + 0.042842797935009, + 0.03512579947710037 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 6, + 8 + ], + "selected_reward": -0.008640995249152184 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" 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 perform a slow, graceful 360-degree rotation around the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive hot air balloon with a shell featuring vertical panels in cherry red, bright yellow, and sky blue visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a wicker basket visible beneath the balloon containing three contestants?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the three contestants wear matching blue and yellow trench coats?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a vast, soft blue sky with a few fluffy, cotton-like clouds?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.12704995274543762 + ], + "opposite_logprob_tokens": [ + -2.1270499229431152 + ], + "logprob_reward": -0.12704995274543762, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.12704995274543762, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09754503518342972, + 0.1332762986421585, + 0.10173193365335464, + 0.03215987607836723, + 0.032629914581775665, + 0.06844404339790344, + 0.1417718231678009, + 0.16333366930484772, + 0.07763000577688217, + 0.15147744119167328 + ], + "credit_map": [ + 0.014320079237222672, + 0.013425069861114025, + 0.006959293503314257, + 0.009911480359733105, + 0.02324943244457245, + 0.010229022242128849, + 0.007951049134135246, + 0.011499608866870403, + 0.022871222347021103, + 0.018506180495023727, + 0.011654339730739594, + 0.02191239781677723, + 0.021481703966856003, + 0.014606399461627007, + 0.011018834076821804, + 0.011225204914808273, + 0.012143249623477459, + 0.01070449873805046, + 0.0071942005306482315, + 0.00874390546232462, + 0.01663774438202381, + 0.030839111655950546, + 0.007289667148143053, + 0.008179549127817154, + 0.005214529577642679, + 0.0062126778066158295, + 0.0032615005038678646, + 0.0050562783144414425, + 0.004484259057790041, + 0.0025473886635154486, + 0.003243082668632269, + 0.0021401571575552225, + 0.003718118416145444, + 0.005474948324263096, + 0.004035359714180231, + 0.004213017411530018, + 0.0032925282139331102, + 0.0030589019879698753, + 0.003470213385298848, + 0.005366825964301825, + 0.009426102042198181, + 0.010312472470104694, + 0.010210414417088032, + 0.005646326579153538, + 0.008389851078391075, + 0.005300537217408419, + 0.006636866368353367, + 0.012521471828222275, + 0.013276132754981518, + 0.013209628872573376, + 0.02052338235080242, + 0.01593240536749363, + 0.012663749977946281, + 0.01157226599752903, + 0.02952387183904648, + 0.025070371106266975, + 0.0178926270455122, + 0.018715260550379753, + 0.01891658827662468, + 0.015591111965477467, + 0.01945316605269909, + 0.01791844330728054, + 0.023039013147354126, + 0.03180745616555214, + 0.007942634634673595, + 0.009010864421725273, + 0.008231092244386673, + 0.007860482670366764, + 0.015527741052210331, + 0.008465136401355267, + 0.012471373192965984, + 0.008120669983327389, + 0.019350698217749596, + 0.029016049578785896, + 0.021711332723498344, + 0.01733892410993576, + 0.01792372763156891, + 0.013064616359770298, + 0.015069670043885708, + 0.01800241321325302 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 9, + 6 + ], + "selected_reward": -0.12704995274543762 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" 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 perform a slow, graceful 360-degree rotation around the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive hot air balloon with a shell featuring vertical panels in cherry red, bright yellow, and sky blue visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a wicker basket visible beneath the balloon containing three contestants?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the three contestants wear matching blue and yellow trench coats?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a vast, soft blue sky with a few fluffy, cotton-like clouds?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0010397987207397819 + ], + "opposite_logprob_tokens": [ + -6.876039981842041 + ], + "logprob_reward": -0.0010397987207397819, + "yes_no_margin_reward": 6.875, + "token_credit_reward": -0.0010397987207397819, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.047540973871946335, + 0.05886752903461456, + 0.02456645481288433, + 0.02981015294790268, + 0.08366041630506516, + 0.15701954066753387, + 0.12930090725421906, + 0.1062222346663475, + 0.11518964916467667, + 0.24782219529151917 + ], + "credit_map": [ + 0.0069237262941896915, + 0.00716381287202239, + 0.002750291721895337, + 0.008220719173550606, + 0.00999043881893158, + 0.004317429848015308, + 0.0040345573797822, + 0.004139994271099567, + 0.012737831100821495, + 0.008673569187521935, + 0.0055855256505310535, + 0.009564165957272053, + 0.005902406759560108, + 0.004797793459147215, + 0.005955796223133802, + 0.0056504360400140285, + 0.002186940750107169, + 0.00283594848588109, + 0.0020039742812514305, + 0.004337533842772245, + 0.003949770703911781, + 0.004968435503542423, + 0.0021973189432173967, + 0.0020865306723862886, + 0.0039337510243058205, + 0.00623770197853446, + 0.0029913275502622128, + 0.0028938932809978724, + 0.005112677346915007, + 0.003358529880642891, + 0.0034083896316587925, + 0.0018738813232630491, + 0.01281808316707611, + 0.012115461751818657, + 0.016552669927477837, + 0.008144207298755646, + 0.011401427909731865, + 0.004914524033665657, + 0.005909413564950228, + 0.011804625391960144, + 0.018583208322525024, + 0.032721228897571564, + 0.01834814064204693, + 0.016195988282561302, + 0.028292842209339142, + 0.008371332660317421, + 0.01567276380956173, + 0.01883402280509472, + 0.01331289391964674, + 0.012475322932004929, + 0.01914348639547825, + 0.014202829450368881, + 0.009971958585083485, + 0.01357324793934822, + 0.018236838281154633, + 0.028384316712617874, + 0.014054892584681511, + 0.008763783611357212, + 0.014111722819507122, + 0.012219409458339214, + 0.012525810860097408, + 0.012527570128440857, + 0.010868466459214687, + 0.02115057222545147, + 0.014666177332401276, + 0.02285819500684738, + 0.015693632885813713, + 0.012312100268900394, + 0.015905221924185753, + 0.010558059439063072, + 0.007496374659240246, + 0.015699878334999084, + 0.03827783465385437, + 0.05262657627463341, + 0.03386605530977249, + 0.03263966366648674, + 0.026393605396151543, + 0.018957559019327164, + 0.02143481932580471, + 0.023626074194908142 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 5, + 6 + ], + "selected_reward": -0.0010397987207397819 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In stunning, captivatingly detailed Pixar-style execution, the camera begins a slow, graceful 360-degree rotation, arcing around a massive hot air balloon. The balloon's shell is a vibrant tapestry of vertical panels in cherry red, bright yellow, and sky blue, all with a smooth, slightly glossy texture that captures the bright morning light. A playful team logo, featuring a stylized cartoon badger wearing aviator goggles, adorns the side, rendered in bold, clean lines. As the camera continues its calming motion, it reveals a wicker basket beneath, where a team of three contestants stands. One, a woman with red hair tied in a practical ponytail and large, expressive green eyes, looks at her with friendly determination as she operates the burner. Beside her, a man with a warm smile and a neatly trimmed beard peers through a pair of oversized binoculars, while a third teammate excitedly points at the horizon. They all wear matching blue and yellow trench coats. The entire scene, set against a vast, soft blue sky dotted with a few fluffy, cotton-like clouds, creates a cheerful, adventurous atmosphere filled with optimistic energy and a spirit of friendly competition.", + "seed": 23923, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000156/group_9/index_768_seed_23923_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": -0.1300702715292573, + "vlm_logprob": -0.1300702715292573, + "vlm_yes_no_margin": 3.8, + "vlm_token_credit": -0.1300702715292573, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" 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 perform a slow, graceful 360-degree rotation around the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive hot air balloon with a shell featuring vertical panels in cherry red, bright yellow, and sky blue visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a wicker basket visible beneath the balloon containing three contestants?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the three contestants wear matching blue and yellow trench coats?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a vast, soft blue sky with a few fluffy, cotton-like clouds?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.5821945667266846 + ], + "opposite_logprob_tokens": [ + -0.8321945667266846 + ], + "logprob_reward": -0.5821945667266846, + "yes_no_margin_reward": 0.25, + "token_credit_reward": -0.5821945667266846, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10706540942192078, + 0.0843692496418953, + 0.05396872013807297, + 0.07711640745401382, + 0.34010887145996094, + 0.08223628997802734, + 0.07622095942497253, + 0.04688600078225136, + 0.08723927289247513, + 0.04478882998228073 + ], + "credit_map": [ + 0.011497595347464085, + 0.0092413779348135, + 0.011624440550804138, + 0.009666391648352146, + 0.011766026727855206, + 0.027912039309740067, + 0.011379965580999851, + 0.013977568596601486, + 0.010488919913768768, + 0.00788329541683197, + 0.015467467717826366, + 0.011979805305600166, + 0.0075191110372543335, + 0.010304194875061512, + 0.010754233226180077, + 0.009972224943339825, + 0.006139295641332865, + 0.007264453452080488, + 0.007149152923375368, + 0.006651127245277166, + 0.007600388024002314, + 0.004841610323637724, + 0.008072693832218647, + 0.00624999962747097, + 0.006987832952290773, + 0.007005530409514904, + 0.007583489175885916, + 0.00726245017722249, + 0.009347639977931976, + 0.013355913572013378, + 0.016983697190880775, + 0.008589851669967175, + 0.04496647045016289, + 0.038685720413923264, + 0.03880520910024643, + 0.012129520997405052, + 0.13981974124908447, + 0.03911546990275383, + 0.01420301478356123, + 0.012383725494146347, + 0.006920319516211748, + 0.014736911281943321, + 0.011808252893388271, + 0.008074209094047546, + 0.01235022209584713, + 0.013133424334228039, + 0.00837213546037674, + 0.006840815301984549, + 0.012684036046266556, + 0.011157139204442501, + 0.008679348975419998, + 0.008069331757724285, + 0.008571557700634003, + 0.014704123139381409, + 0.007502219174057245, + 0.004853203892707825, + 0.00479782372713089, + 0.006978266406804323, + 0.005569811910390854, + 0.006635763682425022, + 0.005953348241746426, + 0.008288437500596046, + 0.004588714335113764, + 0.00407383544370532, + 0.010008142329752445, + 0.0089796157553792, + 0.008926229551434517, + 0.010732055641710758, + 0.00842966791242361, + 0.009631224907934666, + 0.020894497632980347, + 0.009637841954827309, + 0.004414543975144625, + 0.00671797851100564, + 0.0036931417416781187, + 0.004627452231943607, + 0.008124331943690777, + 0.005109670106321573, + 0.006487491074949503, + 0.005614220630377531 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 8 + ], + "selected_reward": -0.5821945667266846 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" 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 perform a slow, graceful 360-degree rotation around the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive hot air balloon with a shell featuring vertical panels in cherry red, bright yellow, and sky blue visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a wicker basket visible beneath the balloon containing three contestants?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the three contestants wear matching blue and yellow trench coats?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a vast, soft blue sky with a few fluffy, cotton-like clouds?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.001526977401226759 + ], + "opposite_logprob_tokens": [ + -6.501526832580566 + ], + "logprob_reward": -0.001526977401226759, + "yes_no_margin_reward": 6.5, + "token_credit_reward": -0.001526977401226759, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08635061979293823, + 0.1739390343427658, + 0.03489653393626213, + 0.14449036121368408, + 0.26647698879241943, + 0.08069606125354767, + 0.057210344821214676, + 0.036284781992435455, + 0.07360295951366425, + 0.046052251011133194 + ], + "credit_map": [ + 0.007999798282980919, + 0.007177533116191626, + 0.016948716714978218, + 0.012462311424314976, + 0.008968557231128216, + 0.01465911976993084, + 0.007749367039650679, + 0.01038521807640791, + 0.0240643173456192, + 0.01505184918642044, + 0.03651619330048561, + 0.024857593700289726, + 0.0164216049015522, + 0.018186837434768677, + 0.017032170668244362, + 0.021808471530675888, + 0.003552868962287903, + 0.003049273509532213, + 0.005164839792996645, + 0.004789392929524183, + 0.00377280474640429, + 0.003809180809184909, + 0.006479865405708551, + 0.004278306849300861, + 0.015204082243144512, + 0.01610477641224861, + 0.0155491279438138, + 0.014879941008985043, + 0.016459383070468903, + 0.027691323310136795, + 0.022762859240174294, + 0.01583886705338955, + 0.034780193120241165, + 0.028937600553035736, + 0.03488339111208916, + 0.009785372763872147, + 0.10871528834104538, + 0.03016016259789467, + 0.011432835832238197, + 0.007782156579196453, + 0.0072852312587201595, + 0.013281852938234806, + 0.01474427804350853, + 0.007583575788885355, + 0.011379224248230457, + 0.011475066654384136, + 0.008566227741539478, + 0.00638060225173831, + 0.010172897949814796, + 0.008149124681949615, + 0.005814400501549244, + 0.007612311746925116, + 0.006031954661011696, + 0.008880187757313251, + 0.006970468442887068, + 0.0035790002439171076, + 0.0038308685179799795, + 0.005851395428180695, + 0.003419985296204686, + 0.004455590154975653, + 0.004020501859486103, + 0.005431541241705418, + 0.0037582335062325, + 0.005516665056347847, + 0.008521000854671001, + 0.006790865678340197, + 0.00640875706449151, + 0.008194958791136742, + 0.008783767931163311, + 0.008529647253453732, + 0.01778596267104149, + 0.008588001132011414, + 0.005368244368582964, + 0.005376359913498163, + 0.004799987655133009, + 0.005675417836755514, + 0.007063486613333225, + 0.005226165056228638, + 0.007858917117118835, + 0.004683671519160271 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 1, + 3 + ], + "selected_reward": -0.001526977401226759 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" 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 perform a slow, graceful 360-degree rotation around the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive hot air balloon with a shell featuring vertical panels in cherry red, bright yellow, and sky blue visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a wicker basket visible beneath the balloon containing three contestants?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the three contestants wear matching blue and yellow trench coats?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a vast, soft blue sky with a few fluffy, cotton-like clouds?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023287026211619377 + ], + "opposite_logprob_tokens": [ + -3.773287057876587 + ], + "logprob_reward": -0.023287026211619377, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023287026211619377, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09611143171787262, + 0.05357523635029793, + 0.04546649008989334, + 0.18080264329910278, + 0.1451290100812912, + 0.12613274157047272, + 0.13582003116607666, + 0.06097640097141266, + 0.09893698245286942, + 0.05704893171787262 + ], + "credit_map": [ + 0.00986140314489603, + 0.0078076208010315895, + 0.017050521448254585, + 0.01327937375754118, + 0.011020969599485397, + 0.013895743526518345, + 0.012992331758141518, + 0.010203473269939423, + 0.005319161340594292, + 0.005484927445650101, + 0.00850530993193388, + 0.01008064579218626, + 0.00441651651635766, + 0.004101813770830631, + 0.009075534529983997, + 0.006591325160115957, + 0.004156504292041063, + 0.00522652780637145, + 0.007424057461321354, + 0.005094259511679411, + 0.0035363151691854, + 0.003329577622935176, + 0.011790448799729347, + 0.004908798262476921, + 0.017897149547934532, + 0.022084791213274002, + 0.019509466364979744, + 0.019878752529621124, + 0.0201854407787323, + 0.028523657470941544, + 0.03447572886943817, + 0.01824764721095562, + 0.017141198739409447, + 0.014005319215357304, + 0.027153443545103073, + 0.007230902556329966, + 0.049865227192640305, + 0.014154624193906784, + 0.006623375229537487, + 0.008954926393926144, + 0.009742639027535915, + 0.013415412046015263, + 0.01999945193529129, + 0.013015840202569962, + 0.015473504550755024, + 0.01613554358482361, + 0.023016102612018585, + 0.015334253199398518, + 0.00993729941546917, + 0.011775276623666286, + 0.01225387305021286, + 0.028034765273332596, + 0.01689763180911541, + 0.011237000115215778, + 0.029207687824964523, + 0.016476498916745186, + 0.010170753113925457, + 0.0077614979818463326, + 0.008351508527994156, + 0.006900194566696882, + 0.005967919714748859, + 0.00960205215960741, + 0.006805215962231159, + 0.005417262203991413, + 0.012353071942925453, + 0.00960120651870966, + 0.007798408158123493, + 0.011561908759176731, + 0.012412658892571926, + 0.012403558939695358, + 0.01904190704226494, + 0.01376426126807928, + 0.006091740448027849, + 0.00855115707963705, + 0.0049804626032710075, + 0.006996102165430784, + 0.008105557411909103, + 0.009025663137435913, + 0.006119729019701481, + 0.007178525906056166 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 4, + 6 + ], + "selected_reward": -0.023287026211619377 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" 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 perform a slow, graceful 360-degree rotation around the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive hot air balloon with a shell featuring vertical panels in cherry red, bright yellow, and sky blue visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a wicker basket visible beneath the balloon containing three contestants?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the three contestants wear matching blue and yellow trench coats?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a vast, soft blue sky with a few fluffy, cotton-like clouds?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03809397295117378 + ], + "opposite_logprob_tokens": [ + -3.2880940437316895 + ], + "logprob_reward": -0.03809397295117378, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03809397295117378, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11066260188817978, + 0.06607817113399506, + 0.04715557023882866, + 0.10319457948207855, + 0.06550131738185883, + 0.1298963725566864, + 0.10285649448633194, + 0.07557737827301025, + 0.14186237752437592, + 0.1572151482105255 + ], + "credit_map": [ + 0.010640769265592098, + 0.011535507626831532, + 0.012243402190506458, + 0.015747684985399246, + 0.00924675539135933, + 0.01894701085984707, + 0.018817761912941933, + 0.01348370686173439, + 0.008279483765363693, + 0.010299989953637123, + 0.005491768009960651, + 0.008405802771449089, + 0.008954143151640892, + 0.0069375536404550076, + 0.010509779676795006, + 0.007199648767709732, + 0.0031484158243983984, + 0.004687916953116655, + 0.007388985250145197, + 0.00752677908167243, + 0.004371150396764278, + 0.0035497841890901327, + 0.011530297808349133, + 0.004952243994921446, + 0.014403292909264565, + 0.01090318150818348, + 0.011150101199746132, + 0.010598567314445972, + 0.013282191008329391, + 0.012736971490085125, + 0.017620233818888664, + 0.0125000374391675, + 0.007595872040838003, + 0.011318639852106571, + 0.005053223576396704, + 0.005514056421816349, + 0.014669913798570633, + 0.00898807868361473, + 0.005118330474942923, + 0.007243198808282614, + 0.01103686448186636, + 0.011150684207677841, + 0.022395122796297073, + 0.01652316004037857, + 0.016534904018044472, + 0.02031755819916725, + 0.019129669293761253, + 0.012808411382138729, + 0.01088703516870737, + 0.010114983655512333, + 0.018772142007946968, + 0.017956266179680824, + 0.010157846845686436, + 0.009325261227786541, + 0.014964869245886803, + 0.010678081773221493, + 0.009440253488719463, + 0.009449474513530731, + 0.00985965970903635, + 0.01004011370241642, + 0.007471679709851742, + 0.010978150181472301, + 0.010103772394359112, + 0.008234274573624134, + 0.018458526581525803, + 0.015371783636510372, + 0.01186273992061615, + 0.01307913102209568, + 0.01635138690471649, + 0.019309936091303825, + 0.02619827724993229, + 0.021230604499578476, + 0.017917053773999214, + 0.019959833472967148, + 0.018451262265443802, + 0.022949205711483955, + 0.02426687814295292, + 0.018521713092923164, + 0.017421573400497437, + 0.017727622762322426 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 5 + ], + "selected_reward": -0.03809397295117378 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" 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 perform a slow, graceful 360-degree rotation around the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive hot air balloon with a shell featuring vertical panels in cherry red, bright yellow, and sky blue visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a wicker basket visible beneath the balloon containing three contestants?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the three contestants wear matching blue and yellow trench coats?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a vast, soft blue sky with a few fluffy, cotton-like clouds?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005248814355581999 + ], + "opposite_logprob_tokens": [ + -5.2552490234375 + ], + "logprob_reward": -0.005248814355581999, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005248814355581999, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0903606042265892, + 0.15478835999965668, + 0.34148892760276794, + 0.06334933638572693, + 0.13455842435359955, + 0.07807236909866333, + 0.04320426285266876, + 0.02340450882911682, + 0.034810688346624374, + 0.03596252575516701 + ], + "credit_map": [ + 0.008109070360660553, + 0.0067640687339007854, + 0.011793309822678566, + 0.009085359051823616, + 0.010397175326943398, + 0.02216898463666439, + 0.009001331403851509, + 0.013041302561759949, + 0.01655295491218567, + 0.024638155475258827, + 0.014767509885132313, + 0.017365802079439163, + 0.025662658736109734, + 0.025124816223978996, + 0.0137315159663558, + 0.016944944858551025, + 0.0339590422809124, + 0.050186481326818466, + 0.0332692414522171, + 0.048642244189977646, + 0.07314429432153702, + 0.022703643888235092, + 0.044651638716459274, + 0.03493234142661095, + 0.006867237854748964, + 0.005725997034460306, + 0.005756737664341927, + 0.009295834228396416, + 0.007997759617865086, + 0.007621623110026121, + 0.01402160432189703, + 0.006062543485313654, + 0.015588202513754368, + 0.015363250859081745, + 0.0071199266240000725, + 0.005632178857922554, + 0.05226818472146988, + 0.01578664779663086, + 0.007533543277531862, + 0.015266488306224346, + 0.006705596577376127, + 0.007253766991198063, + 0.013774272054433823, + 0.01084107905626297, + 0.009481571614742279, + 0.009015502408146858, + 0.01383957825601101, + 0.0071609970182180405, + 0.0059644915163517, + 0.006861129309982061, + 0.005151746794581413, + 0.005676984786987305, + 0.0047963205724954605, + 0.005427356343716383, + 0.004922665189951658, + 0.004403570666909218, + 0.002739311894401908, + 0.002468399703502655, + 0.0024387857411056757, + 0.0032850028946995735, + 0.0026964740827679634, + 0.003633486106991768, + 0.0020540114492177963, + 0.004089036490768194, + 0.003935738001018763, + 0.003833391470834613, + 0.004491024184972048, + 0.003924383781850338, + 0.004417266696691513, + 0.003926916513592005, + 0.007167598232626915, + 0.0031143685337156057, + 0.003390080062672496, + 0.0035603572614490986, + 0.0027231858111917973, + 0.0030334293842315674, + 0.004719136748462915, + 0.007524833548814058, + 0.004160046111792326, + 0.006851453799754381 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 1, + 4 + ], + "selected_reward": -0.005248814355581999 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In stunning, captivatingly detailed Pixar-style execution, the camera begins a slow, graceful 360-degree rotation, arcing around a massive hot air balloon. The balloon's shell is a vibrant tapestry of vertical panels in cherry red, bright yellow, and sky blue, all with a smooth, slightly glossy texture that captures the bright morning light. A playful team logo, featuring a stylized cartoon badger wearing aviator goggles, adorns the side, rendered in bold, clean lines. As the camera continues its calming motion, it reveals a wicker basket beneath, where a team of three contestants stands. One, a woman with red hair tied in a practical ponytail and large, expressive green eyes, looks at her with friendly determination as she operates the burner. Beside her, a man with a warm smile and a neatly trimmed beard peers through a pair of oversized binoculars, while a third teammate excitedly points at the horizon. They all wear matching blue and yellow trench coats. The entire scene, set against a vast, soft blue sky dotted with a few fluffy, cotton-like clouds, creates a cheerful, adventurous atmosphere filled with optimistic energy and a spirit of friendly competition.", + "seed": 23924, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000156/group_9/index_768_seed_23924_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": -0.3498902067542076, + "vlm_logprob": -0.3498902067542076, + "vlm_yes_no_margin": 2.225, + "vlm_token_credit": -0.3498902067542076, + "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 camera perform a slow, graceful 360-degree rotation around the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive hot air balloon with a shell featuring vertical panels in cherry red, bright yellow, and sky blue visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a wicker basket visible beneath the balloon containing three contestants?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the three contestants wear matching blue and yellow trench coats?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a vast, soft blue sky with a few fluffy, cotton-like clouds?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"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": [ + -1.5082895755767822 + ], + "opposite_logprob_tokens": [ + -0.25828951597213745 + ], + "logprob_reward": -1.5082895755767822, + "yes_no_margin_reward": -1.25, + "token_credit_reward": -1.5082895755767822, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.0779741033911705, + 0.04684539884328842, + 0.03140668943524361, + 0.07332947105169296, + 0.05777667090296745, + 0.1832713782787323, + 0.0343116894364357, + 0.024233857169747353, + 0.4304417669773102, + 0.040408987551927567 + ], + "credit_map": [ + 0.006315737031400204, + 0.013081301935017109, + 0.010200914926826954, + 0.013665157370269299, + 0.007268072571605444, + 0.009401835501194, + 0.008592961356043816, + 0.009448118507862091, + 0.004277064464986324, + 0.0059349751099944115, + 0.004375957418233156, + 0.012577280402183533, + 0.0040788655169308186, + 0.0059805442579090595, + 0.0050004408694803715, + 0.004620271734893322, + 0.0037517636083066463, + 0.00409362930804491, + 0.0033531775698065758, + 0.0076179723255336285, + 0.0025425530038774014, + 0.003189641749486327, + 0.0038095375057309866, + 0.003048413898795843, + 0.009997736662626266, + 0.008534462191164494, + 0.00780502101406455, + 0.014920089393854141, + 0.00780765013769269, + 0.006691369228065014, + 0.010346685536205769, + 0.007226457819342613, + 0.0036128738429397345, + 0.007948314771056175, + 0.005118294153362513, + 0.01261033397167921, + 0.008418614976108074, + 0.003786560380831361, + 0.007648283150047064, + 0.00863339938223362, + 0.019047334790229797, + 0.022480390965938568, + 0.025363368913531303, + 0.028754116967320442, + 0.026657674461603165, + 0.026439042761921883, + 0.020661188289523125, + 0.013868249952793121, + 0.004080784972757101, + 0.0049426197074353695, + 0.005422499496489763, + 0.0054954467341303825, + 0.0032580511178821325, + 0.004485215526074171, + 0.0036800692323595285, + 0.0029470028821378946, + 0.0029409260023385286, + 0.003345739794895053, + 0.0040608798153698444, + 0.003899703733623028, + 0.00239836354739964, + 0.0032185313757508993, + 0.002177372109144926, + 0.002192341024056077, + 0.03973468765616417, + 0.030368689447641373, + 0.11835585534572601, + 0.0429246760904789, + 0.03121786192059517, + 0.12139691412448883, + 0.022825773805379868, + 0.023617306724190712, + 0.008674859069287777, + 0.0041781337931752205, + 0.007480409927666187, + 0.004114388953894377, + 0.0059645287692546844, + 0.0034356731921434402, + 0.0035662681329995394, + 0.002994723618030548 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 5, + 0 + ], + "selected_reward": -1.5082895755767822 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" 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 perform a slow, graceful 360-degree rotation around the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive hot air balloon with a shell featuring vertical panels in cherry red, bright yellow, and sky blue visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a wicker basket visible beneath the balloon containing three contestants?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the three contestants wear matching blue and yellow trench coats?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a vast, soft blue sky with a few fluffy, cotton-like clouds?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011117536574602127 + ], + "opposite_logprob_tokens": [ + -4.511117458343506 + ], + "logprob_reward": -0.011117536574602127, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011117536574602127, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13723227381706238, + 0.09921454638242722, + 0.06830893456935883, + 0.10028783977031708, + 0.05567014217376709, + 0.07223246991634369, + 0.04934443160891533, + 0.030838586390018463, + 0.23106250166893005, + 0.15580835938453674 + ], + "credit_map": [ + 0.018642017617821693, + 0.02027261257171631, + 0.014970636926591396, + 0.018471693620085716, + 0.012802978977560997, + 0.01912202127277851, + 0.01474850531667471, + 0.018201805651187897, + 0.0070139942690730095, + 0.01258797850459814, + 0.008951978757977486, + 0.025860391557216644, + 0.010370074771344662, + 0.016032986342906952, + 0.008680417202413082, + 0.009716729633510113, + 0.008945264853537083, + 0.008671198971569538, + 0.005691061727702618, + 0.01678449846804142, + 0.008545562624931335, + 0.006705182138830423, + 0.006498625967651606, + 0.00646754028275609, + 0.013992490246891975, + 0.010741842910647392, + 0.01324156578630209, + 0.017352942377328873, + 0.01242686714977026, + 0.009310807101428509, + 0.01379419770091772, + 0.009427127428352833, + 0.003779033198952675, + 0.00836096704006195, + 0.005140552297234535, + 0.012902189046144485, + 0.007522181142121553, + 0.003281319746747613, + 0.007616045419126749, + 0.007067852653563023, + 0.007401782087981701, + 0.009351910091936588, + 0.012094964273273945, + 0.012067211791872978, + 0.008031695149838924, + 0.010434274561703205, + 0.006874910090118647, + 0.005975720938295126, + 0.005094248801469803, + 0.006755146197974682, + 0.005551139358431101, + 0.006632642354816198, + 0.004855435341596603, + 0.007514503318816423, + 0.004856200888752937, + 0.008085113018751144, + 0.0034140856005251408, + 0.0038692972157150507, + 0.0054088616743683815, + 0.0031845695339143276, + 0.0020514309871941805, + 0.006967541761696339, + 0.00268304580822587, + 0.0032597538083791733, + 0.020932387560606003, + 0.017383737489581108, + 0.062051236629486084, + 0.021061837673187256, + 0.015740778297185898, + 0.06565701216459274, + 0.013654960319399834, + 0.014580532908439636, + 0.02766496129333973, + 0.012022128328680992, + 0.03657091408967972, + 0.012675794772803783, + 0.013633507303893566, + 0.03864116594195366, + 0.005716601852327585, + 0.008883277885615826 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 0 + ], + "selected_reward": -0.011117536574602127 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" 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 perform a slow, graceful 360-degree rotation around the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive hot air balloon with a shell featuring vertical panels in cherry red, bright yellow, and sky blue visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a wicker basket visible beneath the balloon containing three contestants?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the three contestants wear matching blue and yellow trench coats?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a vast, soft blue sky with a few fluffy, cotton-like clouds?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"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": [ + 0.14170227944850922, + 0.04663802310824394, + 0.06963413953781128, + 0.08045240491628647, + 0.09571103751659393, + 0.09649533033370972, + 0.08154355734586716, + 0.020202618092298508, + 0.25465503334999084, + 0.11296561360359192 + ], + "credit_map": [ + 0.02331203781068325, + 0.016002701595425606, + 0.022339774295687675, + 0.016763025894761086, + 0.01234651729464531, + 0.018697356805205345, + 0.013722703792154789, + 0.018518155440688133, + 0.005231389310210943, + 0.00847382191568613, + 0.0047634197399020195, + 0.00814560055732727, + 0.003928752150386572, + 0.007338228635489941, + 0.004398716147989035, + 0.004358089994639158, + 0.008385946042835712, + 0.0077559151686728, + 0.005141156725585461, + 0.01803959161043167, + 0.0057454947382211685, + 0.00655401311814785, + 0.009918516501784325, + 0.008093497715890408, + 0.008118849247694016, + 0.007196635473519564, + 0.010374286212027073, + 0.014937829226255417, + 0.011072876863181591, + 0.010490825399756432, + 0.012008861638605595, + 0.006252233404666185, + 0.006178024224936962, + 0.02062315121293068, + 0.007802330423146486, + 0.011723029427230358, + 0.013781212270259857, + 0.0067930263467133045, + 0.01395034696906805, + 0.014859901741147041, + 0.009037183597683907, + 0.013966845348477364, + 0.01133517175912857, + 0.012767045758664608, + 0.01734127476811409, + 0.012918459251523018, + 0.01106199249625206, + 0.008067339658737183, + 0.009818588383495808, + 0.00772905582562089, + 0.012902661226689816, + 0.011226743459701538, + 0.009174110367894173, + 0.01557015161961317, + 0.005804180167615414, + 0.009318058378994465, + 0.002416132250800729, + 0.002354101510718465, + 0.003308298997581005, + 0.0023036841303110123, + 0.0020504773128777742, + 0.003369891317561269, + 0.0019569797441363335, + 0.0024430514313280582, + 0.02242458425462246, + 0.020489387214183807, + 0.06487646698951721, + 0.025310512632131577, + 0.020550575107336044, + 0.07166571915149689, + 0.01451941393315792, + 0.014818369410932064, + 0.01785355620086193, + 0.005814399570226669, + 0.031060699373483658, + 0.00826209131628275, + 0.007943903096020222, + 0.026111116632819176, + 0.006374355871230364, + 0.009545479901134968 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 9 + ], + "selected_reward": -0.03810831531882286 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" 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 perform a slow, graceful 360-degree rotation around the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive hot air balloon with a shell featuring vertical panels in cherry red, bright yellow, and sky blue visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a wicker basket visible beneath the balloon containing three contestants?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the three contestants wear matching blue and yellow trench coats?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a vast, soft blue sky with a few fluffy, cotton-like clouds?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11295023560523987 + ], + "opposite_logprob_tokens": [ + -2.237950325012207 + ], + "logprob_reward": -0.11295023560523987, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11295023560523987, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08375659584999084, + 0.07629913091659546, + 0.06430914998054504, + 0.1020900160074234, + 0.0810372456908226, + 0.1111699789762497, + 0.11711432784795761, + 0.026487944647669792, + 0.18317881226539612, + 0.1545567661523819 + ], + "credit_map": [ + 0.009937399066984653, + 0.012516397051513195, + 0.015798771753907204, + 0.013052917085587978, + 0.006730076391249895, + 0.006768492981791496, + 0.010373732075095177, + 0.008578804321587086, + 0.006545805837959051, + 0.01433454267680645, + 0.00699574314057827, + 0.01620923914015293, + 0.007283956743776798, + 0.01139149907976389, + 0.0059782094322144985, + 0.007560135331004858, + 0.007179121021181345, + 0.008926838636398315, + 0.004970225039869547, + 0.01361179631203413, + 0.008694151416420937, + 0.007272941526025534, + 0.007215656340122223, + 0.006438419222831726, + 0.011371512897312641, + 0.010405858978629112, + 0.014233894646167755, + 0.017635153606534004, + 0.014015238732099533, + 0.013576695695519447, + 0.013025935739278793, + 0.00782572478055954, + 0.005978838074952364, + 0.014027194119989872, + 0.00937190093100071, + 0.018336256965994835, + 0.007236238569021225, + 0.004206566605716944, + 0.013710261322557926, + 0.008169988170266151, + 0.010558649897575378, + 0.015145853161811829, + 0.013934919610619545, + 0.018467126414179802, + 0.016764312982559204, + 0.016087980940937996, + 0.011690857820212841, + 0.008520284667611122, + 0.009345181286334991, + 0.014088683761656284, + 0.015074916183948517, + 0.02397427149116993, + 0.009103164076805115, + 0.019478024914860725, + 0.010445349849760532, + 0.015604729764163494, + 0.0029420333448797464, + 0.0032592180650681257, + 0.004467880353331566, + 0.004262711852788925, + 0.002296396531164646, + 0.004780019633471966, + 0.0025397154968231916, + 0.0019399704178795218, + 0.0161045640707016, + 0.012585188262164593, + 0.04174241051077843, + 0.017089754343032837, + 0.014632699079811573, + 0.04949483647942543, + 0.014683360233902931, + 0.016845999285578728, + 0.024083558470010757, + 0.016279570758342743, + 0.027151575312018394, + 0.014485597610473633, + 0.015424060635268688, + 0.03375392407178879, + 0.011115365661680698, + 0.012263107113540173 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 6 + ], + "selected_reward": -0.11295023560523987 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" 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 perform a slow, graceful 360-degree rotation around the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive hot air balloon with a shell featuring vertical panels in cherry red, bright yellow, and sky blue visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a wicker basket visible beneath the balloon containing three contestants?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the three contestants wear matching blue and yellow trench coats?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a vast, soft blue sky with a few fluffy, cotton-like clouds?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07898537069559097 + ], + "opposite_logprob_tokens": [ + -2.5789854526519775 + ], + "logprob_reward": -0.07898537069559097, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07898537069559097, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17260092496871948, + 0.1057232990860939, + 0.07617991417646408, + 0.032331932336091995, + 0.07768909633159637, + 0.11657315492630005, + 0.019845757633447647, + 0.01529522892087698, + 0.17794349789619446, + 0.20581716299057007 + ], + "credit_map": [ + 0.018116727471351624, + 0.02477107383310795, + 0.01613871194422245, + 0.017184754833579063, + 0.016662800684571266, + 0.03449806943535805, + 0.022271059453487396, + 0.02295774407684803, + 0.01268947497010231, + 0.009987200610339642, + 0.01103710662573576, + 0.014647508971393108, + 0.01145898550748825, + 0.02138012833893299, + 0.01215659361332655, + 0.01236630603671074, + 0.010536362417042255, + 0.008119813166558743, + 0.006817582529038191, + 0.01605011709034443, + 0.011677379719913006, + 0.008146834559738636, + 0.007721899542957544, + 0.007109928876161575, + 0.002849617972970009, + 0.0032553987111896276, + 0.004399076569825411, + 0.0054764277301728725, + 0.004641324281692505, + 0.004817030392587185, + 0.003924951888620853, + 0.0029681064188480377, + 0.0067046028561890125, + 0.0061374856159091, + 0.010954665951430798, + 0.010297702625393867, + 0.009795775637030602, + 0.00828410405665636, + 0.010126760229468346, + 0.015388003550469875, + 0.012655469588935375, + 0.01463398803025484, + 0.01547566894441843, + 0.017319483682513237, + 0.01812784932553768, + 0.015575121156871319, + 0.01396110188215971, + 0.008824479766190052, + 0.00285500125028193, + 0.0014787446707487106, + 0.0024343125987797976, + 0.001985015580430627, + 0.0021337864454835653, + 0.003082204842939973, + 0.0020573665387928486, + 0.003819325240328908, + 0.002079528523609042, + 0.001982135698199272, + 0.0022816548589617014, + 0.0017120866104960442, + 0.0011341726640239358, + 0.0032192817889153957, + 0.0013558659702539444, + 0.0015305046690627933, + 0.016601234674453735, + 0.015173164196312428, + 0.042857468128204346, + 0.014388512820005417, + 0.012543467804789543, + 0.048812318593263626, + 0.013015050441026688, + 0.0145522840321064, + 0.034296631813049316, + 0.00775897828862071, + 0.06249111890792847, + 0.015313421376049519, + 0.008745312690734863, + 0.056617751717567444, + 0.006488780491054058, + 0.014105193316936493 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 0 + ], + "selected_reward": -0.07898537069559097 + } + ], + "vlm_decoded_outputs": [ + "No", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In stunning, captivatingly detailed Pixar-style execution, the camera begins a slow, graceful 360-degree rotation, arcing around a massive hot air balloon. The balloon's shell is a vibrant tapestry of vertical panels in cherry red, bright yellow, and sky blue, all with a smooth, slightly glossy texture that captures the bright morning light. A playful team logo, featuring a stylized cartoon badger wearing aviator goggles, adorns the side, rendered in bold, clean lines. As the camera continues its calming motion, it reveals a wicker basket beneath, where a team of three contestants stands. One, a woman with red hair tied in a practical ponytail and large, expressive green eyes, looks at her with friendly determination as she operates the burner. Beside her, a man with a warm smile and a neatly trimmed beard peers through a pair of oversized binoculars, while a third teammate excitedly points at the horizon. They all wear matching blue and yellow trench coats. The entire scene, set against a vast, soft blue sky dotted with a few fluffy, cotton-like clouds, creates a cheerful, adventurous atmosphere filled with optimistic energy and a spirit of friendly competition.", + "seed": 23925, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000156/group_9/index_768_seed_23925_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": -0.1068976883776486, + "vlm_logprob": -0.1068976883776486, + "vlm_yes_no_margin": 3.775, + "vlm_token_credit": -0.1068976883776486, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" 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 perform a slow, graceful 360-degree rotation around the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive hot air balloon with a shell featuring vertical panels in cherry red, bright yellow, and sky blue visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a wicker basket visible beneath the balloon containing three contestants?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the three contestants wear matching blue and yellow trench coats?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a vast, soft blue sky with a few fluffy, cotton-like clouds?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.4375190734863281 + ], + "opposite_logprob_tokens": [ + -1.0625190734863281 + ], + "logprob_reward": -0.4375190734863281, + "yes_no_margin_reward": 0.625, + "token_credit_reward": -0.4375190734863281, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14466407895088196, + 0.05648515745997429, + 0.05990089103579521, + 0.18266281485557556, + 0.12151274085044861, + 0.13159611821174622, + 0.08485530316829681, + 0.08116697520017624, + 0.07943210750818253, + 0.0577237606048584 + ], + "credit_map": [ + 0.00984052661806345, + 0.012085450813174248, + 0.029105737805366516, + 0.019028853625059128, + 0.013609778136014938, + 0.022169265896081924, + 0.01864783465862274, + 0.02017664536833763, + 0.006044303998351097, + 0.008340982720255852, + 0.007178432308137417, + 0.00825696624815464, + 0.008500316180288792, + 0.004773873370140791, + 0.006835676729679108, + 0.006554615218192339, + 0.007643148768693209, + 0.006645064800977707, + 0.010013425722718239, + 0.004481762181967497, + 0.007843567989766598, + 0.00721238786354661, + 0.0050345528870821, + 0.011026988737285137, + 0.025754110887646675, + 0.016279013827443123, + 0.026600774377584457, + 0.018779553472995758, + 0.031022075563669205, + 0.01692337542772293, + 0.025499586015939713, + 0.02180434577167034, + 0.015746623277664185, + 0.014382103458046913, + 0.021652135998010635, + 0.01768294721841812, + 0.011594854295253754, + 0.01579594984650612, + 0.00918208435177803, + 0.015476050786674023, + 0.018888574093580246, + 0.017354119569063187, + 0.01308534387499094, + 0.012947900220751762, + 0.012950314208865166, + 0.028262320905923843, + 0.00975194200873375, + 0.01835562288761139, + 0.01113197673112154, + 0.019954463467001915, + 0.007433328311890364, + 0.008905255235731602, + 0.020129496231675148, + 0.00677006971091032, + 0.005622600205242634, + 0.004908117465674877, + 0.009553580544888973, + 0.015374665148556232, + 0.007711498998105526, + 0.012888669967651367, + 0.012692959047853947, + 0.009287332184612751, + 0.006426800973713398, + 0.007231480907648802, + 0.01601080782711506, + 0.013367570005357265, + 0.008267674595117569, + 0.007710249163210392, + 0.011109726503491402, + 0.010769717395305634, + 0.006570883560925722, + 0.005625492427498102, + 0.013139557093381882, + 0.005555613432079554, + 0.006932321470230818, + 0.003133058315142989, + 0.01590871810913086, + 0.004501327872276306, + 0.003353745909407735, + 0.005199429113417864 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 5 + ], + "selected_reward": -0.4375190734863281 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" 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 perform a slow, graceful 360-degree rotation around the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive hot air balloon with a shell featuring vertical panels in cherry red, bright yellow, and sky blue visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a wicker basket visible beneath the balloon containing three contestants?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the three contestants wear matching blue and yellow trench coats?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a vast, soft blue sky with a few fluffy, cotton-like clouds?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0017288275994360447 + ], + "opposite_logprob_tokens": [ + -6.3767290115356445 + ], + "logprob_reward": -0.0017288275994360447, + "yes_no_margin_reward": 6.375, + "token_credit_reward": -0.0017288275994360447, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15382082760334015, + 0.05776528641581535, + 0.06470342725515366, + 0.10958988964557648, + 0.09663073718547821, + 0.04105949401855469, + 0.07311203330755234, + 0.10392042994499207, + 0.07562592625617981, + 0.22377189993858337 + ], + "credit_map": [ + 0.017393166199326515, + 0.020991122350096703, + 0.027012068778276443, + 0.02053157053887844, + 0.013814746402204037, + 0.0154141029343009, + 0.01852349378168583, + 0.020140575245022774, + 0.0076523274183273315, + 0.010546049103140831, + 0.007220085244625807, + 0.007788596209138632, + 0.0056109679862856865, + 0.0051737031899392605, + 0.005960077978670597, + 0.007813489995896816, + 0.004944740794599056, + 0.006961148232221603, + 0.007699114736169577, + 0.006449651438742876, + 0.012493234127759933, + 0.009799877181649208, + 0.007319412659853697, + 0.009036253206431866, + 0.010388609021902084, + 0.00875711627304554, + 0.015206759795546532, + 0.010509474202990532, + 0.019735068082809448, + 0.016438178718090057, + 0.015370219014585018, + 0.013184480369091034, + 0.011462780646979809, + 0.013964403420686722, + 0.016757367178797722, + 0.010200833901762962, + 0.013774605467915535, + 0.011825736612081528, + 0.007097749970853329, + 0.011547274887561798, + 0.006638167425990105, + 0.006313554476946592, + 0.0052263205870985985, + 0.0036769057624042034, + 0.00493536377325654, + 0.006583032198250294, + 0.003591161221265793, + 0.00409499229863286, + 0.009217989630997181, + 0.010131490416824818, + 0.006516939029097557, + 0.009164967574179173, + 0.010294875130057335, + 0.01523678284138441, + 0.0058979131281375885, + 0.006651083938777447, + 0.010489964857697487, + 0.02160418964922428, + 0.011955613270401955, + 0.018328964710235596, + 0.0164237879216671, + 0.0074408892542123795, + 0.010778898373246193, + 0.006898134481161833, + 0.013924566097557545, + 0.008900903165340424, + 0.00935660395771265, + 0.011853932403028011, + 0.008306419476866722, + 0.007084682118147612, + 0.007689579855650663, + 0.008509255014359951, + 0.01951039582490921, + 0.008869620971381664, + 0.08458494395017624, + 0.00964690838009119, + 0.07093203812837601, + 0.00977130327373743, + 0.009192260913550854, + 0.011264446191489697 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 3 + ], + "selected_reward": -0.0017288275994360447 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" 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 perform a slow, graceful 360-degree rotation around the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive hot air balloon with a shell featuring vertical panels in cherry red, bright yellow, and sky blue visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a wicker basket visible beneath the balloon containing three contestants?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the three contestants wear matching blue and yellow trench coats?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a vast, soft blue sky with a few fluffy, cotton-like clouds?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02058546058833599 + ], + "opposite_logprob_tokens": [ + -3.895585536956787 + ], + "logprob_reward": -0.02058546058833599, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.02058546058833599, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09001871198415756, + 0.01433630008250475, + 0.02308129519224167, + 0.06465871632099152, + 0.02555469237267971, + 0.03644021600484848, + 0.0659380778670311, + 0.2841137945652008, + 0.13324810564517975, + 0.2626100778579712 + ], + "credit_map": [ + 0.010632367804646492, + 0.01652303710579872, + 0.01454124879091978, + 0.010101576335728168, + 0.007331639528274536, + 0.011040984652936459, + 0.007189503405243158, + 0.012658352963626385, + 0.0018595875008031726, + 0.0019980284851044416, + 0.0020049528684467077, + 0.00227277516387403, + 0.001868368242867291, + 0.0009056265116669238, + 0.0015494386898353696, + 0.0018775227945297956, + 0.0023472043685615063, + 0.0016975520411506295, + 0.00461849058046937, + 0.00232151267118752, + 0.0032225013710558414, + 0.0021833411883562803, + 0.002262500347569585, + 0.004428192041814327, + 0.008749089203774929, + 0.005377582274377346, + 0.009966067969799042, + 0.006755900103598833, + 0.010574454441666603, + 0.005706433672457933, + 0.008630028925836086, + 0.008899154141545296, + 0.0033208809327334166, + 0.002326984191313386, + 0.006680293008685112, + 0.00273950700648129, + 0.0017570839263498783, + 0.003591073676943779, + 0.0017356046009808779, + 0.003403265727683902, + 0.004759873263537884, + 0.004414446651935577, + 0.004833129700273275, + 0.002252479549497366, + 0.004156965762376785, + 0.007768813520669937, + 0.0036198883317410946, + 0.004634617827832699, + 0.00824718177318573, + 0.007937118411064148, + 0.0076719531789422035, + 0.00783221609890461, + 0.012864276766777039, + 0.009381018579006195, + 0.005396873690187931, + 0.006607441231608391, + 0.026930131018161774, + 0.06648634374141693, + 0.0330033153295517, + 0.05547567456960678, + 0.04526863992214203, + 0.01655755750834942, + 0.026758769527077675, + 0.013633355498313904, + 0.02999522164463997, + 0.01527293212711811, + 0.014599932357668877, + 0.019626135006546974, + 0.015621567144989967, + 0.013989605009555817, + 0.010395226068794727, + 0.01374747883528471, + 0.02384018525481224, + 0.010979704558849335, + 0.08905398100614548, + 0.010788622312247753, + 0.0923054963350296, + 0.012421825900673866, + 0.010761566460132599, + 0.012458693236112595 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 9, + 8 + ], + "selected_reward": -0.02058546058833599 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" 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 perform a slow, graceful 360-degree rotation around the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive hot air balloon with a shell featuring vertical panels in cherry red, bright yellow, and sky blue visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a wicker basket visible beneath the balloon containing three contestants?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the three contestants wear matching blue and yellow trench coats?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a vast, soft blue sky with a few fluffy, cotton-like clouds?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07002037018537521 + ], + "opposite_logprob_tokens": [ + -2.6950204372406006 + ], + "logprob_reward": -0.07002037018537521, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07002037018537521, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.047905076295137405, + 0.014225324615836143, + 0.019510764628648758, + 0.012494552880525589, + 0.03120829351246357, + 0.024620410054922104, + 0.049802396446466446, + 0.04028308391571045, + 0.1051694005727768, + 0.6547806262969971 + ], + "credit_map": [ + 0.00554212462157011, + 0.0061554922722280025, + 0.009928043931722641, + 0.006610774900764227, + 0.003952595870941877, + 0.003302930388599634, + 0.005396494176238775, + 0.007016628980636597, + 0.0013129721628502011, + 0.002047850051894784, + 0.0025051748380064964, + 0.0014926447765901685, + 0.0015553792472928762, + 0.0016354838153347373, + 0.0010498298797756433, + 0.0026259920559823513, + 0.0010918974876403809, + 0.0016948560951277614, + 0.0025873694103211164, + 0.001977801788598299, + 0.0046249013394117355, + 0.0017148072365671396, + 0.0028514848090708256, + 0.0029676465783268213, + 0.00133419968187809, + 0.0016450153198093176, + 0.0017910738242790103, + 0.0011564723681658506, + 0.002053332980722189, + 0.001571936416439712, + 0.0014273948036134243, + 0.0015151287661865354, + 0.004290509503334761, + 0.002456776099279523, + 0.006383818574249744, + 0.00350708095356822, + 0.002069760812446475, + 0.006559463683515787, + 0.00198306143283844, + 0.003957828041166067, + 0.002863930072635412, + 0.0034153289161622524, + 0.0024722719099372625, + 0.0021157264709472656, + 0.0036183216143399477, + 0.00536089763045311, + 0.002293569268658757, + 0.0024803681299090385, + 0.006386567838490009, + 0.005726816598325968, + 0.0045756069011986256, + 0.00696322089061141, + 0.008537367917597294, + 0.00925265159457922, + 0.0036966113839298487, + 0.004663557279855013, + 0.004485298879444599, + 0.007712953258305788, + 0.00465351203456521, + 0.0075109596364200115, + 0.005218219012022018, + 0.0036097674164921045, + 0.0047881826758384705, + 0.0023041979875415564, + 0.0251263789832592, + 0.012585025280714035, + 0.010541990399360657, + 0.012786244042217731, + 0.013108131475746632, + 0.01207755971699953, + 0.007713169325143099, + 0.011230915784835815, + 0.0930858924984932, + 0.02739126607775688, + 0.1690976619720459, + 0.02117217145860195, + 0.25546300411224365, + 0.03261641040444374, + 0.026028556749224663, + 0.02992570959031582 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 6 + ], + "selected_reward": -0.07002037018537521 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" 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 perform a slow, graceful 360-degree rotation around the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive hot air balloon with a shell featuring vertical panels in cherry red, bright yellow, and sky blue visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a wicker basket visible beneath the balloon containing three contestants?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the three contestants wear matching blue and yellow trench coats?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a vast, soft blue sky with a few fluffy, cotton-like clouds?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004634710028767586 + ], + "opposite_logprob_tokens": [ + -5.379634857177734 + ], + "logprob_reward": -0.004634710028767586, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004634710028767586, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06528191268444061, + 0.16626544296741486, + 0.12983177602291107, + 0.19575262069702148, + 0.10396816581487656, + 0.03577878326177597, + 0.058922503143548965, + 0.127339705824852, + 0.04977135360240936, + 0.06708776205778122 + ], + "credit_map": [ + 0.00789586827158928, + 0.010225088335573673, + 0.011137627996504307, + 0.008206211030483246, + 0.005587053019553423, + 0.007574291434139013, + 0.006597941275686026, + 0.008057834580540657, + 0.012268279679119587, + 0.013989553786814213, + 0.029591143131256104, + 0.010846060700714588, + 0.022031649947166443, + 0.038606222718954086, + 0.014843309298157692, + 0.02408922277390957, + 0.012301729060709476, + 0.013667095452547073, + 0.016415828838944435, + 0.013576923869550228, + 0.024744302034378052, + 0.014281083829700947, + 0.014941425994038582, + 0.019903380423784256, + 0.017710356041789055, + 0.014179983176290989, + 0.031201155856251717, + 0.0167213324457407, + 0.029531417414546013, + 0.03921781852841377, + 0.021789466962218285, + 0.025401096791028976, + 0.01431411411613226, + 0.01162898913025856, + 0.025629926472902298, + 0.009324777871370316, + 0.009924671612679958, + 0.01093266624957323, + 0.00648691738024354, + 0.01572609692811966, + 0.004328237846493721, + 0.004940701182931662, + 0.00404008012264967, + 0.0031356606632471085, + 0.004259001929312944, + 0.008713942021131516, + 0.0021676342003047466, + 0.004193525295704603, + 0.00687045743688941, + 0.009243284352123737, + 0.005950361490249634, + 0.00707704434171319, + 0.007393731269985437, + 0.00856437161564827, + 0.006667459849268198, + 0.0071557918563485146, + 0.015166159719228745, + 0.013760022819042206, + 0.011232328601181507, + 0.00880264863371849, + 0.018778765574097633, + 0.0167307760566473, + 0.011412451043725014, + 0.031456559896469116, + 0.0106067955493927, + 0.006407051347196102, + 0.0061607384122908115, + 0.00616737175732851, + 0.0071150860749185085, + 0.005605746526271105, + 0.003977364860475063, + 0.003731201868504286, + 0.015184689313173294, + 0.003493203781545162, + 0.009083487093448639, + 0.0018109290394932032, + 0.026927268132567406, + 0.004169100895524025, + 0.002980367746204138, + 0.003438713727518916 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 1, + 2 + ], + "selected_reward": -0.004634710028767586 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file